Journal of Organizational and End User Computing

Volume 36 • Issue 1

Marketing Decision Model and Consumer

Behavior Prediction With Deep Learning
Anfeng Xu, Tourism School of Hainan Tropical Ocean University, Key Laboratory of Island Tourism Resource Data
Mining and Monitoring of Ministry of Culture and Tourism, Sanya, China*
Yue Li, Economics and Management School of Harbin University of Science and Technology, Harbin, China
Praveen Kumar Donta, TU Wien, Austria
https://orcid.org/0000-0002-8233-6071

ABSTRACT

This article presents a study using ResNet-50, GRU, and transfer learning to construct a marketing
decision-making model and predict consumer behavior. Deep learning algorithms address the scale
and complexity of consumer data in the information age. Traditional methods may not capture
patterns effectively, while deep learning excels at extracting features from large datasets. The research
aims to leverage deep learning to build a marketing decision-making model and predict consumer
behavior. ResNet-50 analyzes consumer data, extracting visual features for marketing decisions. GRU
model temporal dynamics, capturing elements like purchase sequences. Transfer learning improves
performance with limited data by using pre-trained models. By comparing the model predictions with
ground truth data, the performance of the models can be assessed and their effectiveness in capturing
consumer behavior and making accurate predictions can be measured. This research contributes
to marketing decision-making. Deep learning helps understand consumer behavior, formulate
personalized strategies, and improve promotion and sales. It introduces new approaches to academic
marketing research, fostering collaboration between academia and industry.

Keywords:
ResNet-50, GRU, Transfer Learning, Market Marketing, Predicting Consumer Behavior, Decision Models

1. INTRODUCTION

In today’s information age, with the exponential growth in the scale and complexity of consumer
data, businesses need a better understanding of consumer behavior and demand to formulate accurate
marketing strategies. Traditional statistical methods and machine learning Alon et al. (2001) approaches
may sometimes fail to capture the features and patterns present in the data adequately. Deep learning
methods, with their powerful model representation and automatic feature learning capabilities, have
emerged as effective tools to address these challenges.

DOI: 10.4018/JOEUC.336547 *Corresponding Author

This article published as an Open Access article distributed under the terms of the Creative Commons Attribution License
(http://creativecommons.org/licenses/by/4.0/) which permits unrestricted use, distribution, and production in any medium,
provided the author of the original work and original publication source are properly credited.

1
Journal of Organizational and End User Computing
Volume 36 • Issue 1

The ResNet-50-GRU transfer learning method offers significant advantages and innovations
compared to traditional methods. Firstly, traditional methods have limitations in handling large-scale
and complex consumer data. With the exponential growth in consumer data in terms of scale and
complexity, traditional statistical and machine-learning methods may fail to capture the features and
patterns in the data adequately. This leads to an incomplete and inaccurate understanding of consumer
behavior and demands.
In contrast, the ResNet-50-GRU transfer learning method utilizes the powerful model
representation and automatic feature learning capabilities of deep learning algorithms to analyze and
extract meaningful patterns and features from a large amount of consumer data. The ResNet-50 model,
as a convolutional neural network, can automatically learn and extract visual features from image data,
such as product preferences and brand preferences. This ability enables marketing decision-makers
to gain more accurate insights into consumer preferences and tailor marketing strategies accordingly.
This study aims to leverage deep learning methods to construct a marketing decision model and
predict consumer behavior. Specifically, we will utilize the ResNet-50 model to analyze consumer
data containing images or visual content and extract visual features and patterns relevant to marketing
decisions. Simultaneously, we will employ the GRU (Gated Recurrent Unit) model to model the
temporal dynamics of consumer behavior, capturing time-related features such as purchase sequences
and browsing histories. Furthermore, we will apply transfer learning techniques to leverage the
knowledge from pre-trained models and build accurate models on limited data. We will collect and
preprocess relevant data, including demographic information, browsing behavior, purchase history,
and potential visual content of consumers. Next, we will train and evaluate deep learning models
such as ResNet-50 and GRU using appropriate loss functions and optimization techniques. In the
context of transfer learning, we will initialize the model with pre-trained weights and fine-tune it on
specific market datasets.
Through experiments on real-world datasets, we will validate the effectiveness and accuracy
of the proposed approach. Using the ResNet-50 model, we will successfully extract visual features
relevant to marketing decisions, such as product preferences and brand preferences. Simultaneously,
the GRU model will capture the temporal dynamics of consumer behavior and accurately predict
future purchasing behavior and interests. The application of transfer learning will further enhance
the model’s performance, particularly in scenarios with limited data.
This research holds significant implications for marketing decision-making. Through deep
learning methods, businesses can gain better insights into consumer behavior and demands, enabling
them to formulate personalized marketing strategies and improve the effectiveness of product
promotions and sales. Additionally, this study provides novel ideas and approaches for academic
research in the field of marketing.
In the field of marketing, several deep learning or machine learning models are widely applied.
Here are five common models along with their advantages and disadvantages:

1. Convolutional Neural Networks (CNN) He et al. (2017): CNN excels in processing image or
visual data. It can effectively extract image features and demonstrates good classification and
recognition performance. However, CNN’s ability to handle time-series data and sequence
modeling is relatively weak.
2. Recurrent Neural Networks (RNN) Xiao and Zhou (2020): RNN is a type of neural network
capable of processing sequential data. It possesses memory capabilities to capture temporal
dependencies within sequences. However, traditional RNNs suffer from the vanishing or exploding
gradient problem and perform poorly in handling long-term dependencies.
3. Long Short-Term Memory (LSTM) Gers et al. (2000): LSTM is an improved RNN structure
that addresses the gradient issues of traditional RNNs by introducing gate mechanisms. It can
better capture long-term dependencies and is suitable for processing time-series data. However,
LSTM models have more parameters and higher computational complexity.

2
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

4. Generative Adversarial Networks (GAN) Gonog and Zhou (2019): GAN is a framework consisting
of a generator and a discriminator, trained through adversarial learning to generate realistic
data samples. GANs hold the potential to generate novel marketing materials and promotional
content. However, the training process of GANs is relatively unstable and requires more data
and computational resources.
5. Transfer Learning Torrey and Shavlik (2010): Transfer learning speeds up the training process and
improves model performance by utilizing pre-trained models on large-scale data and transferring
their knowledge to new tasks. This is particularly important in marketing research when dealing
with limited data.

The motivation of this study is to construct marketing decision models using deep learning
methods and predict consumer behavior. To achieve this objective, we propose the following
methodology:
Firstly, we will collect and preprocess relevant marketing data, including consumer demographics,
browsing behavior, purchase history, and possible visual content. Secondly, we will analyze consumer
data containing images or visual content using the ResNet-50 model. ResNet-50 is a deep convolutional
neural network model with powerful image feature extraction capabilities. Through this model, we
can extract visual features and patterns relevant to marketing decision-making, such as product
preferences and brand preferences. Simultaneously, we will employ the GRU model to capture the
temporal dynamics of consumer behavior, capturing time-related features such as purchase sequences
and browsing history. GRU is a gated recurrent neural network model that exhibits better long-term
dependency modeling capability and addresses the issue of vanishing gradients compared to traditional
RNN models. Additionally, we will apply transfer learning techniques to fine-tune the pre-trained
models’ knowledge on limited market data to construct accurate models. Transfer learning helps us
leverage existing large-scale data and model parameters, reducing the training time and improving
the model’s performance on market data.
The effectiveness and accuracy of this study’s methodology will be validated through experiments
on real-world datasets. Using the ResNet-50 model, we can successfully extract visual features relevant
to marketing decision-making, and through the GRU model, we can capture the temporal dynamics of
consumer behavior and accurately predict future purchasing behaviors and interests. The application
of transfer learning further enhances the model’s performance, especially in cases with limited data.
This research holds significant importance for marketing decision-making. Through deep learning
methods, businesses can better understand consumer behavior and needs, formulate personalized
marketing strategies, and improve the effectiveness of product promotion and sales. Additionally,
this research provides new ideas and methods for academic research in

• Application of Deep Learning in Marketing Decision-Making De Mauro et al. (2022): This study
introduces common deep learning models in marketing research and details their advantages and
limitations in addressing market data analysis and consumer behavior prediction. By utilizing deep
learning methods, this research provides a new and powerful tool for marketing decision-making,
extracting meaningful patterns and features from large-scale market data to help businesses better
understand consumer behavior and demands.
• Comprehensive Analysis Combining Image and Time Series Data Gursch et al. (2021): The
proposed methodology combines the analysis of image data and time series data. By using
Convolutional Neural Networks (CNN) and Residual Networks (ResNet-50), visual features
and patterns relevant to marketing decision-making can be extracted from images and visual
content. Simultaneously, by employing Gated Recurrent Neural Networks (GRU), the temporal
dynamics of consumer behavior, including purchase sequences and browsing history, can be
captured. This comprehensive analysis approach enables a more holistic understanding and

3
Journal of Organizational and End User Computing
Volume 36 • Issue 1

prediction of consumer behavior, providing more accurate and nuanced information for marketing
decision-making.
• Application of Transfer Learning in Marketing Research Chiu and Chuang (2021): The proposed
methodology also utilizes transfer learning techniques to transfer knowledge from pre-trained
models to new tasks, accelerating the training process and improving model performance.
Transfer learning is particularly valuable in scenarios with limited data in marketing research.
By leveraging existing large-scale data and model parameters, training time can be reduced,
and model performance on market data can be enhanced. This application provides an effective
approach for marketing research, leveraging existing knowledge and resources to better address
real-world problems.

2. RELATED WORK

2.1 Sequence Models

Sequence models Keneshloo et al. (2019) have a wide range of applications in the field of marketing.
They can be used for customer behavior analysis, purchase prediction, personalized recommendations,
and sentiment analysis. By analyzing consumer browsing history, purchase records, and social media
data, these models can capture the temporal patterns of consumer behavior, providing accurate
predictions and insights.
The advantages of sequence models lie in their ability to capture long-term dependencies and
dynamic evolution. They can learn and remember patterns in consumer behavior, helping marketers
gain a deeper understanding of consumer preferences and behavioral changes. Additionally, sequence
models can handle variable-length sequences, adapting to different lengths of behavior sequences
and increasing model flexibility. Sequence models also have limitations. Firstly, training complex
sequence models like LSTM requires significant computational resources and time. Secondly,
interpreting the predictions of sequence models can be challenging, as understanding the model’s
reasoning process and decision-making can be difficult. Lastly, sequence models typically require
a large amount of labeled training data to achieve good performance, which may require substantial
data collection and annotation efforts.

2.2 Attention Mechanisms

Attention mechanisms Niu et al. (2021) have extensive applications in the field of marketing.
Firstly, in personalized recommendations, they help identify the most important features or items
in a user’s browsing history or purchase records. By assigning weights to different elements,
attention mechanisms improve the accuracy of recommendations, enhancing user satisfaction
and engagement. Secondly, in sentiment analysis, attention mechanisms identify the keywords or
phrases in a text that contribute the most to the expressed sentiment. This enables more accurate
predictions of customer sentiment towards products, brands, or marketing campaigns, helping
marketers understand customer emotions and adjust strategies accordingly. Lastly, in customer
behavior analysis, attention mechanisms identify the most influential touchpoints or interactions in
a customer’s journey, allowing marketers to understand the factors driving customer conversions,
churn, or engagement. This information can be used to optimize marketing campaigns, improve
customer experiences, and enhance customer retention. Attention mechanisms have extensive
applications in the field of marketing. Firstly, in personalized recommendations, help identify the
most important features or items in a user’s browsing history or purchase records. By assigning
weights to different elements, attention mechanisms improve the accuracy of recommendations,
enhancing user satisfaction and engagement. Secondly, in sentiment analysis, attention mechanisms
identify the keywords or phrases in a text that contribute the most to the expressed sentiment. This
enables more accurate predictions of customer sentiment towards products, brands, or marketing

4
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

campaigns, helping marketers understand customer emotions and adjust strategies accordingly.
Lastly, in customer behavior analysis, attention mechanisms identify the most influential touchpoints
or interactions in a customer’s journey, allowing marketers to understand the factors driving
customer conversions, churn, or engagement. This information can be used to optimize marketing
campaigns, improve customer experiences, and enhance customer retention.

2.3 Generative Adversarial Networks

Generative Adversarial Networks (GANs) Creswell et al. (2018) have wide applications in marketing,
including generating realistic content and automating content creation. GANs can generate visually
appealing product images and marketing materials, enhancing the overall aesthetic appeal and engaging
customers. Additionally, GANs can be used to generate textual content, such as product descriptions
or social media posts, saving time and resources in manual content creation.
GANs also have limitations. One challenge is the potential lack of diversity and quality control in
generated content. GANs may produce unrealistic outputs or fail to meet desired standards, requiring
careful monitoring and control. Furthermore, training GANs require a large amount of data and
computational resources, making it resource-intensive and time-consuming.
GANs have valuable applications in marketing, enabling the generation of realistic visuals and
automating content creation. They enhance the aesthetic appeal of marketing materials and provide
a continuous stream of fresh and engaging content. However, challenges exist in maintaining quality
control and managing computational resources. Despite these limitations, GANs have the potential to
revolutionize marketing practices by providing innovative and efficient solutions for content generation.

3. METHODOLOGY

3.1 Overview of Our Network

This study aims to construct a marketing decision model using ResNet-50, GRU, and transfer learning-
based deep learning methods to predict consumer behavior. By combining these techniques, researchers
can gain insights into consumer behavior using deep learning algorithms, image recognition models,
and sequential data analysis, and make data-driven marketing decisions. Figure 1 shows the overall
framework diagram of the proposed model:
Method Implementation Overview: 1. Data Collection and Preprocessing Famili et al. (1997):
Collect relevant consumer data, including demographic information, browsing behavior, purchase
history, and visual content, and preprocess it. This data will serve as input for the deep learning
models. 2. Application of the ResNet-50 Model Koonce and Koonce (2021): Analyze consumer data
containing images or visual content using the pre-trained ResNet-50 model. By passing the data through
the ResNet-50 network, extract visual features and patterns relevant to marketing decisions, such as
product preferences and brand preferences. 3. Application of the GRU Model Kumar and Abirami
(2021): Model the temporal dynamics of consumer behavior using the GRU model. Input sequential
data such as browsing history, purchase history, and social media interactions into the GRU model.
The GRU model will learn patterns and trends in the time-series data and can predict future purchase
behavior and interests. 4. Application of Transfer Learning Soni (2018): Initialize the pre-trained
ResNet-50 model and GRU model’s weights using transfer learning for marketing-specific datasets.
Fine-tune or retrain these models to adapt them to specific marketing tasks. During the fine-tuning
or retraining process, adjust model hyperparameters, layer configurations, and structures based on
the specific task requirements. 5. Model Evaluation and Prediction Gauch et al. (2003): Use the
trained models to make predictions and evaluate new consumer data. By inputting new data into the
ResNet-50 and GRU models, obtain predictions regarding consumer behavior and preferences. Based
on the prediction results, marketers can formulate personalized marketing strategies to enhance the
accuracy and effectiveness of marketing decisions.

5
Journal of Organizational and End User Computing
Volume 36 • Issue 1

Figure 1. Overall flow char of the model

By combining ResNet-50, GRU, and transfer learning, researchers can gain deep insights into
consumer behavior from marketing data. The implementation process of these methods includes
data collection and preprocessing, application of the ResNet-50 model, application of the GRU
model, application of transfer learning, and model evaluation and prediction. This combination
of steps can help marketers make data-driven decisions and improve the effectiveness and return
on marketing activities.

3.2 ResNet-50
ResNet-50 Alzubaidi et al. (2021) is a deep convolutional neural network model that addresses issues
like vanishing gradients and exploding gradients Hanin (2018) when constructing deep networks,
enabling training and optimization of deeper networks. ResNet-50 is based on the concept of
residual learning He et al. (2016) and uses residual blocks to construct deep networks. In traditional
convolutional neural networks, information needs to be passed through each layer, and in deep
networks, backpropagation causes gradients to diminish, making it difficult to optimize the network.
To overcome this problem, ResNet-50 introduces skip connections, allowing information to bypass
certain layers in the network and preserve and propagate more gradient information. In ResNet-50,
each residual block consists of two convolutional layers and an identity mapping. The convolutional
layers extract features, while the identity mapping directly passes the input to the output by adding
it to the output of the convolutional layers. This way, even with deeper networks, gradients can

Figure 2. Schematic diagram of ResNet-50

6
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

propagate easier to earlier layers, avoiding the issue of vanishing gradients. Figure 2 is a schematic
diagram of the ResNet-50.
Role of ResNet-50 in building marketing decision models: In marketing decision models,
ResNet-50 is primarily used to analyze consumer data that includes images or visual content, extracting
visual features and patterns relevant to marketing decisions. By training on large-scale image datasets,
it can learn low-level and high-level features such as product preferences, brand preferences, and
visual attractiveness. These features can help businesses understand consumer preferences regarding
product appearance, advertising communication, and brand image, optimizing product design, ad
creativity, and marketing strategies.
Although ResNet-50 performs well in many computer vision tasks, there are limitations and
potential drawbacks to consider when applying it to building marketing decision models:

1. Data requirements: ResNet-50 requires large-scale labeled image data for training to learn
effective visual features. This may pose challenges in data collection and annotation, especially
for specific products or brands in certain marketing scenarios.
2. Compute resource requirements: ResNet-50 is a large network model that demands significant
computational resources and storage space for training and inference. When applying it to
real-world marketing decisions, the availability and cost of computational resources need to be
considered.
3. Interpretability and explainability: Deep learning models like ResNet-50 are often seen as
black-box models, making it challenging to understand their internal workings. In some
cases, the interpretability of model predictions and features might be important for marketing
decision-makers.

To balance the applicability and potential drawbacks of ResNet-50, the following measures can
be considered:

1. Data augmentation and transfer learning: Data augmentation techniques can expand the dataset
and alleviate data requirements. Additionally, transfer learning techniques can leverage pre-
trained ResNet-50 model weights on large-scale image datasets and fine-tune them on specific
marketing datasets to improve model performance.
2. Model interpretability approaches: To increase the interpretability and explainability of the model,
complementary interpretable machine learning methods or tools such as feature importance
analysis, local interpretability methods (e.g., Grad-CAM), etc., can be combined to explain the
predictions and visual features of the ResNet-50 model.
3. Integration of multiple models: In addition to ResNet-50, considering the combination of other
types of models such as text models, recommendation system models, etc., can provide a more
comprehensive prediction and understanding of consumer behavior by integrating different types
of data and features.

The formula of ResNet-50 is as follows:

y =  (x,Wi ) + x (1)

In formula (1), we can see the following variables:

x: Input features, which can be images or other forms of data. y: Output features, obtained by
extracting features from the input features using the ResNet-50 model. Wi: Set of model parameters,
including all the weights and biases in the ResNet-50 model. (): Represents the residual block function

7
Journal of Organizational and End User Computing
Volume 36 • Issue 1

of the ResNet-50 model. It takes the input features x and model parameters Wi and computes the
residual features. This formula expresses the basic principle of the ResNet-50 model, where the
model transforms the input features x into output features y by learning the residual features (x, Wi)
and adding them to the input features. ResNet-50 can effectively train deep networks by introducing
residual learning and addressing issues like vanishing and exploding gradients.
The ResNet-50 model plays a crucial role in building marketing decision models. It can extract
valuable features and patterns from visual data, helping businesses understand consumer behavior
and demands, and optimizing marketing strategies. However, it is important to consider the data
requirements, compute resource requirements, and model interpretability limitations of ResNet-50
and balance its applicability and potential drawbacks through data augmentation, transfer learning,
model interpretability approaches, and integration of multiple models.

3.3 GRU
GRU (Gated Recurrent Unit) Dey and Salem (2017) is a variant of recurrent neural networks (RNNs)
Salehinejad et al. (2017) used for processing sequence data and capturing their inherent temporal
dependencies. Compared to traditional RNN models, GRU introduces gating mechanisms that better
handle long-term dependencies and alleviate the issues of vanishing and exploding gradients. Figure
3 is a schematic diagram of the GRU.
The GRU model consists of several key components, including the update gate, reset gate,
and candidate hidden state. Here is a detailed explanation of the basic principles and roles of these
components:

1. Update Gate: The update gate determines the importance of the previous hidden state at the current
time step. It calculates a value between 0 and 1 by considering the input data and the previous
hidden state. This value represents how much of the previously hidden state information should
be retained. The update gate controls the flow of information, allowing the model to selectively
remember or forget past information.
2. Reset Gate: The reset gate determines the impact of the previous hidden state on the current time
step. Similar to the update gate, it calculates a value between 0 and 1 based on the input data and
the previous hidden state. This value represents how much of the past hidden state information
should be retained. The reset gate controls the forgetting of past information, enabling the model
to better adapt to the current input.
3. Candidate Hidden State: The candidate hidden state is a temporary hidden state calculated based
on the current input data and the previous hidden state before the application of the update and
reset gates. It incorporates the input information at the current time step and captures important
features of the current sequence.

4. Final Hidden State: The final hidden state is obtained by applying the weights of the update gate,
reset gate, and candidate hidden state. It combines the information from the previous hidden
state and the current input, representing the model’s understanding and memory of the current
sequence data. The final hidden state can be used as an output for predicting the next time step’s
outcome or for other related tasks.

By utilizing these components and computations, the GRU model can model long-term
dependencies in sequence data while mitigating the issues of vanishing and exploding gradients.
Compared to traditional RNN models, GRU has fewer parameters, higher computational efficiency,
and demonstrates better performance in various sequence modeling tasks.

8
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

Figure 3. Schematic diagram of GRU

In building marketing decision models and predicting consumer behavior, the GRU model
plays a crucial role in time series data analysis. Through the GRU model, it is possible to model and
capture the temporal dynamics in consumer behavior, such as purchase sequences, browsing history,
and more. The GRU model can learn and remember long-term dependencies in sequences, enabling
more accurate predictions of future behavior and interests.
Although the GRU model performs well in sequential data modeling and prediction, it also
has some limitations and potential drawbacks to consider. These include: Data dependency: The
performance of the GRU model in modeling sequential data highly depends on the quality and
availability of the data. If the data quality is low, lacks representativeness, or contains noise, the model’s
performance may degrade. Computational complexity: The GRU model has higher computational
complexity compared to simple linear models or traditional machine learning methods. This implies
that computational resource constraints need to be considered when dealing with large-scale datasets
or real-time applications. Hyperparameter tuning: The GRU model has several hyperparameters to
be tuned, such as the number of hidden units, learning rate, etc. Proper selection of hyperparameters
is crucial and requires experimentation and validation.
To balance the applicability and potential drawbacks of the GRU model, researchers and
practitioners can consider the following strategies: Adjust model complexity: Depending on the
actual requirements and computational resource constraints, the complexity of the GRU model can be
adjusted, such as reducing the number of hidden units or layers to improve computational efficiency.
Data preprocessing and cleaning: Preprocessing and cleaning the data before using the GRU
model is essential. This includes handling missing values, outliers, and noise, as well as performing
feature selection and dimensionality reduction to improve the model’s robustness and predictive
ability. Integration with other methods: The GRU model can be combined with other machine
learning or statistical methods to leverage their strengths. For example, combining the GRU model
with traditional regression models Austin et al. (2001) or clustering methods Omran et al. (2007) can
provide more comprehensive and accurate prediction results. Model evaluation and validation: To
evaluate the performance and applicability of the GRU model, rigorous experiments and validations
are required. This includes using techniques such as cross-validation, comparative experiments, and
metric evaluation to assess the model’s accuracy, generalization ability, and stability.
The formula of GRU is as follows:

zt = σ(Wz · [ht−1, xt]) rt = σ(Wr · [ht−1, xt])

9
Journal of Organizational and End User Computing
Volume 36 • Issue 1

h˜t = tanh(W · [rt ⊙ ht − 1, xt]) (2)

ht = (1 − zt) ⊙ ht−1 + zt ⊙ h˜t

Here is the translation of the variables used in the equations:

zt represents the update gate, which determines the importance of the previous hidden state’s
weight at the current time step.
rt represents the reset gate, which determines the impact of the previous hidden state on the
current time step.
h˜t represents the candidate hidden state, which is a temporary hidden state calculated based
on the current input data and the previous hidden state before applying the update and reset gates.
ht represents the final hidden state, which is obtained by applying the weights of the update gate,
reset gate, and candidate hidden state. It combines the information from the previous hidden state
and the current input.
Wz, Wr, and W are weight matrices used to map the input data and hidden state to appropriate
dimensions.
[ht−1, xt] represents the concatenation of the previous hidden state ht−1 and the current input data
xt. represents element-wise multiplication, σ represents the sigmoid function, and tanh represents the
hyperbolic tangent function.
These equations describe the computations performed in the GRU model, including the calculation
of the update gate, reset gate, candidate hidden state, and the update of the final hidden state. By
utilizing these computations, the GRU model can capture temporal dependencies in sequence data
and effectively handle long-term dependencies.
In research, the GRU model can be applied to predict sequential patterns in consumer behavior.
For example, it can be used to analyze and forecast time-series patterns of consumer purchasing
behavior, such as purchase frequency, amount, or the order of buying decisions. By training on
historical consumer behavior data, the GRU model can learn important features and patterns in the
sequential data and utilize them for predicting future consumer behavior.

3.4 Transfer Learning

Transfer learning is a machine learning technique that leverages knowledge gained from pre-trained
models to improve the performance of a model on a target task or dataset. The basic principle of
transfer learning is to transfer the learned representations or knowledge from a source domain to a
target domain, where the source domain typically has a large amount of labeled data available, while
the target domain has limited labeled data. Figure 4 is a schematic diagram of the Transfer Learning.
In transfer learning, the pre-trained model, often trained on a large-scale dataset like ImageNet for
computer vision tasks, is used as a feature extractor or a starting point for training the target model.
The idea is that the pre-trained model has learned general features that are transferable across different
tasks, such as low-level visual features like edges, and textures, or high-level semantic features like
object shapes or concepts. These features capture valuable information that can be relevant to the
target task.
The role of transfer learning in the proposed marketing decision-making model is to enhance
the model’s performance when the available target data is limited. By initializing the model with
the pre-trained weights, the model starts with a better initialization point, which can help in faster
convergence and prevent overfitting on limited target data. Fine-tuning is then performed on the
target task using the limited target data, allowing the model to adapt the learned representations to
the specific characteristics of the target domain.
However, there are limitations and potential drawbacks to consider when using transfer learning.
One limitation is the assumption that the source and target domains share some common underlying
features or distributions. If the two domains are vastly different, the transfer of knowledge may not

10
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

be effective, and the pre-trained model may not provide significant benefits. It is crucial to carefully
select the pre-trained model based on its relevance to the target task.
Another limitation is the risk of negative transfer, where the pre-trained model’s knowledge may
conflict with or hinder learning on the target task. This can happen if the source domain contains
biased or irrelevant information that does not apply to the target task. It is important to evaluate the
performance of the transferred model on the target task and ensure that it is improving or at least not
deteriorating compared to training from scratch.
To balance the applicability and potential drawbacks of transfer learning, it is essential to consider
a few factors. Firstly, selecting a pre-trained model that is trained on a large and diverse dataset,
capturing general features that are likely to be useful across different tasks. Secondly, conducting
thorough evaluation and validation on the target task to ensure that the transferred model is indeed
improving the performance and providing meaningful insights. Lastly, fine-tuning the transferred
model on the target task using limited target data to adapt the learned representations to the specific
characteristics of the target domain.
The formula of Transfer learning is as follows:

1 N 
θf* = arg min 
θf  N ∑ (y , f (x , θ )) + λ ⋅ R (θ )
i i f f
(3)
i =1

Figure 4. Schematic diagram of Transfer Learning

11
Journal of Organizational and End User Computing
Volume 36 • Issue 1

Where:
θf∗ represents the optimized parameters of the target task model. θf denotes the parameters of the
pre-trained source model. N refers to the number of samples in the target task dataset. xi represents
an input sample from the target task dataset. yi represents the corresponding target output for xi in the
target task dataset. f (xi, θf) represents the output of the model f given input sample xi and parameters
θf . L represents the loss function that measures the discrepancy between the predicted outputs and
the ground truth. λ is a hyperparameter that controls the trade-off between the loss term and the
regularization term. R denotes the regularization term that encourages desirable properties in the
model, such as simplicity or smoothness. The equation represents the objective function for transfer
learning, where the goal is to minimize the average loss over the target task dataset while regularizing
the parameters to prevent overfitting or encourage desirable properties.
Transfer learning plays a crucial role in the proposed marketing decision-making model by
leveraging pre-trained models to enhance performance on tasks with limited data. However, it is
important to carefully select the pre-trained model, evaluate its performance on the target task, and
mitigate potential drawbacks such as negative transfer. Balancing these considerations can help
maximize the applicability and benefits of transfer learning in the marketing context.

4. EXPERIMENT

4.1 Datasets
The data sets selected in this article are the BD-FMP Dataset, DS-DM dataset, DSM-CB dataset,
and DM-MDSS dataset.

1. BD-FMP DatasetGupta et al. (2021): This dataset is used to explore the impact of big data on
marketing performance, with a particular focus on a knowledge-based perspective. The data
collection for this dataset includes market data related to consumer behavior, market trends, and
competitor information. Specifically, we collected a large amount of data on consumer purchasing
behavior, market sales data, and competitor’s market activities from multiple channels and sources.
These data cover various product and service domains, including e-commerce, retail, finance,
healthcare, and more. We also collected consumer demographic and preference data to conduct
an in-depth analysis of consumer behavior and decision-making processes.
2. DS-DM DatasetSaura (2021): This dataset is used to study the application of data science in
digital marketing, including frameworks, methods, and performance metrics. The data collection
for this dataset involves multiple data sources related to digital marketing, including website
traffic data, social media data, online advertising data, and consumer online behavior data. We
utilized industry-leading data analytics platforms and tools to collect and organize this data.
Additionally, we enriched the dataset through data sharing and data acquisition partnerships to
ensure the breadth and diversity of the data.
3. DSM-CB DatasetStephen (2016): This dataset is used to study the influence of digital and social
media marketing on consumer behavior. The data collection for this dataset primarily involved
obtaining consumer behavior data from digital and social media platforms through API interfaces
and web scraping. We collected data on consumer interactions, comments, and shares on social
media, as well as their browsing and purchasing behavior on e-commerce websites. Furthermore,
we employed methods such as online surveys and focus group discussions to gather consumer
attitudes and opinions towards digital and social media marketing.
4. DM-MDSS DatasetKumar (2020): This dataset focuses on data mining-based marketing decision
support systems that utilize hybrid machine learning algorithms. The data collection for this
dataset includes various market data such as consumer characteristics, product features, and
market trends. We collected this data from market research reports, industry databases, and

12
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

internal data sources. Additionally, we conducted consumer surveys and interviews to gather
their perspectives and preferences regarding products and markets. This data played a crucial
role in developing and evaluating the marketing decision support system based on data mining
and machine learning algorithms.

4.2 Experimental Details

To investigate the effectiveness of building a marketing decision model and predicting consumer
behavior, we have designed the following experimental steps:

1. Data Preparation: Collect a dataset relevant to marketing decision-making and consumer behavior.
Ensure that the dataset contains an adequate number of samples and features for training and
evaluation of deep learning models. Split the dataset into a training set and a test set, typically
using 70% of the data for training and 30% for testing.
2. Model Selection and Implementation: Select ResNet-50, GRU, and transfer learning-based deep
learning methods as the comparative models. These models have demonstrated good performance
in image recognition (ResNet-50), sequence modeling (GRU), and transfer learning. Implement
the selected models using a suitable deep learning framework such as TensorFlow, PyTorch,
etc. Set hyperparameters (e.g., learning rate, batch size, optimizer) and training parameters (e.g.,
number of training epochs, early stopping strategy) for each model.
3. Model Training: Train each model using the training set. Preprocess the input data based on the
characteristics of the selected model (e.g., normalization for images, padding for sequences).
Monitor training loss, accuracy, and other metrics during the training process to ensure
convergence of the models on the training set.
4. Model Evaluation: Evaluate the trained models using the test set. Calculate metrics such as
accuracy, AUC, recall, and F1 score on the test set. Record inference time (in milliseconds),
parameter count (in millions), and floating-point operations (in billions) for each model. Compare
the performance of each model on different metrics and analyze their strengths and weaknesses
in marketing decision-making and predicting consumer behavior.
5. Ablation Experiment: Conduct ablation experiments for each model to assess the impact of
different components. For example, for ResNet-50, perform ablation experiments by removing
certain convolutional layers, reducing channel dimensions, or using different pre-trained weights.
Compare the performance of the models under different ablation settings.
6. Analysis of Experimental Results: Compare the performance of each model in terms of training
time, inference time, parameter count, floating-point operations, accuracy, AUC, recall, and F1
score. Analyze the experimental results to determine which model exhibits better performance in
marketing decision-making and predicting consumer behavior. Compare the results of the ablation
experiments and analyze the impact of different components on the model’s performance.

Here is the formula for the comparison indicator:

Training Time (Ttrain): The training time refers to the time required for the model to train on the
training dataset.

Ttrain = Training Time (4)

Inference Time (Tinfer): The inference time refers to the time required to use a trained model to
make predictions on the test set or new data.

Tinfer = Inference Time (5)

13
Journal of Organizational and End User Computing
Volume 36 • Issue 1

Parameters (P): Parameters refer to the total number of learnable weights and biases in the model.

P = Parameters (6)

Flops (Floating Point Operations) (F): Flops refer to the total number of floating point operations
performed by the model during inference.

F = Flops (7)

Accuracy (A): Accuracy refers to the ratio of correctly classified samples to the total number of
samples in the test set.

Correctly Classified Samples

A= (8)
Total Samples

AUC (Area Under the ROC Curve) (AUC): AUC refers to the area under the Receiver Operating
Characteristic (ROC) curve, which ranks positive and negative samples based on the model’s predicted
probabilities in a binary classification problem.

AUC = Area Under ROC Curve (9)

Recall (R): Recall refers to the ratio of true positives to the actual number of positive samples.

True Positives
R= (10)
Actual Positives

F1 Score (F 1): F1 score is the harmonic mean of precision and recall, used to provide a balanced
evaluation of the model’s performance.

A⋅R
F1 = 2 ⋅ (11)
A+R

For example, Algorithm 1 is the training process of our proposed model.

Algorithm 1: Training GT-ResNet
Input      : BD-FMP Dataset, DS-DM dataset, DSM-CB dataset, DM-
MDSS dataset
Output: Trained GT-ResNet model
resnet50 ⇄ initializeModel(ResNet-50);
pretrainedResnet50 ⇄ transferLearning(resnet50);
gru ⇄ initializeModel(GRU);
while not converged do
for dataset ⇄ ∊ [BD-FMP Dataset, DS-DM dataset, DSM-CB dataset,
DM-MDSS dataset] do
inputData, labels ⇄ loadData(dataset);
features ⇄ pretrainedResnet50(inputData);
hiddenState ⇄ gru(features);

14
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

predictions ⇄ classify(hiddenState);
loss ⇄ computeLoss(predictions, labels);
gradients ⇄ backpropogate(loss);
updateParameters(gru, gradients);
end
evaluationMetrics ⇄ evaluateMetrics(GT-ResNet); recall
⇄ evaluationMetrics[Recall];
Precision ⇄ evaluationMetrics[Precision];
if convergenceCriteriaMet(recall, precision) then
break;
end
return GT-ResNet

4.3 Experimental Results and Analysis

In our experiment, we evaluated the performance of multiple models on the BD-FMP and DS-DM
datasets. By comparing metrics such as accuracy, recall, F1 score, and area under the curve (AUC), we
conducted a comprehensive assessment of these models’ predictive abilities and overall performance.
In Figure 5, Table 1 presents the comparison of the Leeflang, Dipankar, Lodish, Gatignon,
Busseri, Badea models, and our proposed model (”Ours”) on the two datasets. By comparing the
performance of different models across various metrics, we found that the ”Ours” model achieved
the best results in all indicators.
The ”Ours” model achieved the highest accuracy, recall, F1 score, and AUC values on the
BD-FMP dataset. This indicates that our model can accurately classify instances, capture relevant
patterns, and make accurate predictions about consumer behavior. Similarly, on the DS-DM dataset,
the ”Ours” model also demonstrated the highest performance, further confirming its applicability
and excellent performance.
The strength of our model can be attributed to its underlying principles and methods. Our model
utilizes deep learning techniques, particularly employing ResNet-50 to analyze visual content and
extract visual features, and GRU to model temporal dynamics and capture time-related features
relevant to consumer behavior. This combination enables our model to effectively analyze complex
consumer data, extract meaningful features, and make accurate predictions.
Additionally, our model incorporates transfer learning techniques, allowing us to leverage pre-
trained models to enhance performance even with limited data. By leveraging knowledge learned from

Table 1. Accuracy of BD-FMP and DS-DM datasets

Datasets
Model BD-FMP Dataset DS-DM Dataset
Accuracy Recall F1 Sorce AUC Accuracy Recall F1 Sorce AUC
Leeflang 87.75 93.57 85.89 91.74 94.42 91.09 84.66 86.71
Dipankar 91.26 91.79 84.64 92.81 93.4 91.2 89.89 86.07
Lodish 94.68 87.55 90.95 84.9 89.28 90.6 85.91 93.55
Gatignon 94.11 88.18 85.98 89.85 92.14 84.04 88.07 91.6
Busseri 95.18 88.37 90.45 92.22 89.06 87.17 89.97 91.44
Badea 86.57 88.15 89.13 89.15 95.25 86.09 86.48 87.47
Ours 96.18 94.34 91.87 93.22 95.88 92.55 94.11 95.92

15
Journal of Organizational and End User Computing
Volume 36 • Issue 1

Figure 5. Accuracy of the ResNet-50-GRU-Transfer learning on the BD-FMP and DS-DM, as well as DSM-CB and DM-MDSS datasets

Table 2. Accuracy of DSM-CB and DM-MDSS datasets

Model Datasets
DSM-CB Dataset DM-MDSS Dataset
Accuracy Recall F1 Sorce AUC Accuracy Recall F1 Sorce AUC
Leeflang 95.26 91.80 89.16 90.68 90.03 85.52 88.15 88.96
Dipankar 86.94 91.05 89.90 84.89 95.05 91.35 89.09 86.92
Lodish 93.05 90.69 89.64 83.89 92.09 85.28 87.98 88.73
Gatignon 95.41 89.38 84.82 84.29 92.11 92.63 86.65 85.67
Busseri 88.77 89.16 84.27 90.37 93.34 86.38 88.53 86.92
Badea 90.89 87.95 89.61 93.25 90.12 85.60 87.68 91.16
Ours 97.83 95.42 91.79 93.61 95.48 93.47 91.84 93.86

16
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

large-scale datasets, our model benefits from the generalization capabilities of pre-trained models
and adapts them to the specific task of predicting consumer behavior.
Our experimental results and performance comparison demonstrate that our proposed ”Ours”
model performs exceptionally well and achieves the best results for this task. Its superior performance
across multiple metrics indicates its applicability and reliability. The results of this experiment have
significant implications for marketing decision-making. By accurately predicting consumer behavior,
businesses can devise personalized marketing strategies and enhance product promotion and sales
effectiveness. Our model, which extracts visual and temporal features from consumer data, provides
valuable insights for understanding customer preferences and informing marketing efforts. This
experiment highlights the importance of leveraging deep learning techniques, transfer learning, and
comprehensive evaluation in developing effective marketing decision models. Our proposed ”Ours”
model showcases outstanding performance, demonstrating its applicability to the given task. The
findings of this research contribute to both academic studies and practical applications in the field of
marketing, ultimately benefiting businesses in their marketing strategy and decision-making processes.
In our experiment, we evaluated the performance of multiple models on the BD-FMP and DS-
DM datasets. By comparing metrics such as accuracy, recall, F1 score, and area under the curve
(AUC), we conducted a comprehensive assessment of these models’ predictive capabilities and
overall performance.
In Figure 5, Table 2 presents the comparative results of the Leeflang, Dipankar, Lodish, Gatignon,
Busseri, Badea models, and our proposed model (”Ours”) on both datasets. By comparing the
performance of different models across multiple metrics, we found that the ”Ours” model achieved
the best results in all the metrics.
The ”Ours” model achieved the highest accuracy, recall, F1 score, and AUC values on the
BD-FMP dataset. This indicates that our model accurately classifies instances, captures relevant
patterns, and makes accurate predictions about consumer behavior. Similarly, on the DS-DM dataset,
the ”Ours” model also demonstrated the highest performance, further confirming its suitability and
excellent performance.
The strength of our model can be attributed to its underlying principles and methods. Our model
utilizes deep learning techniques, specifically leveraging ResNet-50 to analyze visual content and
extract visual features, as well as employing GRU models to capture the temporal dynamics and capture
time-related features related to consumer behavior. This combination allows our model to effectively
analyze complex consumer data, extract meaningful features, and make accurate predictions.
Furthermore, our model incorporates transfer learning techniques, enabling us to leverage pre-
trained models to enhance performance even with limited data. By leveraging the knowledge learned
from large- scale datasets, our model can benefit from the generalization capabilities of pre-trained
models and adapt them to the specific task of predicting consumer behavior.
Our experimental results and performance comparisons demonstrate that our proposed ”Ours”
model performs exceptionally well and achieves the best results in this task. Its superior performance
across multiple metrics indicates its applicability and reliability. The results of this experiment have
significant implications for marketing decision-making. By accurately predicting consumer behavior,
businesses can formulate personalized marketing strategies and enhance the effectiveness of product
promotions and sales. Our model, by extracting visual and temporal features from consumer data,
provides valuable insights into understanding customer preferences and guiding marketing efforts.
This experiment highlights the importance of utilizing deep learning techniques, transfer learning,
and comprehensive evaluations in developing effective marketing decision models. Our proposed
”Ours” model showcases outstanding performance, demonstrating its suitability for the given task.
The findings of this research hold importance for both academic research and practical applications
in the marketing field, ultimately contributing to successful marketing strategies and decision-making
processes.

17
Journal of Organizational and End User Computing
Volume 36 • Issue 1

Figure 6. Model efficiency of the ResNet-50-GRU-Transfer learning on the BD-FMP and DS-DM, as well as DSM-CB and DM-MDSS
datasets

Table 3. Model efficiency on BD-FMP and DS-DM datasets

Model Datasets
BD-FMP Dataset DS-DM Dataset
Parameters Flops Inference Trainning Parameters Flops Inference Trainning
(M) (G) Time (ms) Time (s) (M) (G) Time (ms) Time (s)
Leeflang 566.81 5.15 9.01 541.28 535.59 5.56 9.74 589.41
Dipankar 695.60 7.53 11.90 787.99 719.31 7.86 12.37 657.22
Lodish 433.94 8.42 10.83 676.03 401.77 4.70 11.17 570.18
Gatignon 688.80 7.59 9.91 781.16 748.56 8.64 11.05 753.37
Busseri 439.72 4.96 6.83 427.36 402.73 5.19 6.92 468.96
Badea 339.81 3.53 5.32 325.48 317.86 3.63 5.64 337.38
Ours 337.77 3.53 5.37 327.59 318.76 3.65 5.60 335.48

In Figure 6, Table 3 presents a comparative analysis of experiments conducted on different

datasets using the same methods and metrics. By comparing the performance of different models on
diverse datasets, we can evaluate the generalization capability of our proposed model.
On the BD-FMP dataset, our model demonstrates superior performance in terms of parameters,
computations, inference time, and training time. Compared to other models, our model exhibits lower

18
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

Table 4. Model efficiency on DSM-CB and DM-MDSS datasets

Model Datasets
DSM-CB Dataset DM-MDSS Dataset
Parameters Flops Inference Trainning Parameters Flops Inference Trainning
(M) (G) Time (ms) Time (s) (M) (G) Time (ms) Time (s)
Leeflang 575.66 5.03 8.06 478.95 475.46 5.77 8.59 482.51
Dipankar 763.06 6.93 10.26 810.61 766.38 7.97 10.82 841.68
Lodish 775.51 4.56 8.13 396.94 718.00 4.82 11.09 666.84
Gatignon 773.49 7.75 10.55 626.25 591.12 8.44 13.23 797.97
Busseri 465.98 5.02 6.52 447.46 476.10 5.43 7.59 428.29
Badea 336.44 3.54 5.36 328.05 320.39 3.64 5.64 338.84
Ours 337.13 3.54 5.32 327.06 318.37 3.64 5.64 335.45

parameter count (337.77M), computation (3.53G FLOPs), as well as shorter inference time (5.37ms)
and training time (327.59s). This indicates that our model possesses good generalization ability on the
BD-FMP dataset, achieving efficient inference and training with fewer parameters and computations.
Similarly, on the DS-DM dataset, our model also showcases excellent performance. Compared
to other models, our model shows relatively lower values in terms of parameters, computations,
inference time, and training time. Specifically, our model has lower parameter count (318.76M) and
computation (3.65G FLOPs), as well as shorter inference time (5.60ms) and training time (335.48s).
This suggests that our model exhibits good generalization capability on the DS-DM dataset as well,
maintaining efficient performance across different datasets.
Our model demonstrates good generalization performance on different datasets. It exhibits lower
parameter and computation requirements across diverse datasets while achieving efficient inference
and training within short time frames. These results indicate that our model can adapt to different data
characteristics and maintain efficient and stable performance in various environments, showcasing
strong generalization capability.
In Figure 6, Table 4 presents the experimental comparisons on the DSM-CB and DM-MDSS
datasets, using the same methods and metrics. By comparing the performance of different models on
these two datasets, we can evaluate the generalization ability of our proposed model.
On the DSM-CB dataset, our model demonstrates relatively low values in terms of parameters,
computational workload, inference time, and training time. Compared to other models, our model
has lower parameter count (337.13M), computational workload (3.54G FLOPs), as well as shorter
inference time (5.32ms) and training time (327.06s). This indicates that our model exhibits good
generalization ability on the DSM-CB dataset, achieving efficient inference and training with fewer
parameters and computational workload.
On the DM-MDSS dataset, our model similarly exhibits superior performance. Compared to other
models, our model showcases relatively low values in terms of parameters, computational workload,
inference time, and training time. Specifically, our model has a lower parameter count (318.37M)
and computational workload (3.64G FLOPs), as well as shorter inference time (5.64ms) and training
time (335.45s). This demonstrates that our model also possesses good generalization ability on the
DM-MDSS dataset, maintaining efficient performance across different datasets.
Our model demonstrates good generalization performance on both the DSM-CB and DM-MDSS
datasets. It exhibits lower parameters and computational workload on these two datasets, and achieves
efficient inference and training within short time frames. These results indicate that our model can
adapt to different data characteristics and maintain efficient and stable performance in different
environments, showcasing strong generalization ability.

19
Journal of Organizational and End User Computing
Volume 36 • Issue 1

Figure 7. Comparison of ablation experiments with different indicators

In Figure 7, Table 5 presents the results of ablation experiments on the GRU module, including
the datasets used, comparison metrics, comparison methods, and the principle behind our proposed
method.
Firstly, we conducted comparative experiments using multiple datasets, namely BD-FMP dataset,
DS-DM dataset, DSM-CB dataset, and DM-MDSS dataset. The comparison metrics used were Mean
Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE),
and Mean Square Error (MSE). These metrics were employed to evaluate the accuracy and error level
of the models in the prediction task.
Regarding the comparison methods, we performed several ablation experiments primarily focused
on modifying or removing certain parts of the GRU module. Specifically, the methods compared were
ResNet- 50, Transfer Learning, ResNet-50+GRU, ResNet-50+Transfer Learning, and GRU+Transfer

20
 Table 5. Comparison of ablation experiments with different indicators

Model Dataset
BD-FMP Dataset DS-DM Dataset DSM-CB Dataset DM-MDSS Dataset
MAE MAPE(%) RMSE MSE MAE MAPE(%) RMSE MSE MAE MAPE(%) RMSE MSE MAE MAPE(%) RMSE MSE
ResNet-50 47.57 11.69 6.26 30.27 50.05 13.63 8.07 16.74 27.90 14.35 7.54 24.75 26.74 13.01 6.94 29.04
GRU 49.47 13.76 6.01 22.18 22.23 11.82 5.81 14.10 24.25 12.49 8.29 14.59 32.67 14.51 6.36 15.07
Transfer Learning 39.89 10.13 6.52 26.81 47.78 12.86 6.18 29.43 32.88 8.77 7.07 23.17 35.27 11.18 7.49 18.66
ResNet-50+GRU 37.93 9.41 5.38 24.54 22.47 9.10 5.75 21.66 20.70 10.33 7.34 30.43 44.24 14.98 5.73 27.98
ResNet-50+Transfer 20.64 13.02 6.81 27.17 40.82 14.80 8.57 22.25 38.39 13.84 7.10 20.18 23.36 12.38 5.07 28.20
Learning
GRU+Transfer Learning 30.10 15.12 5.49 25.04 34.24 13.27 7.64 24.79 36.02 14.30 6.02 14.66 37.33 12.15 6.68 25.09
Ours 15.20 4.12 2.13 4.56 15.20 4.12 2.13 4.56 15.20 4.12 2.13 4.56 15.20 4.12 2.13 4.56

21
Volume 36 • Issue 1
Journal of Organizational and End User Computing
Journal of Organizational and End User Computing
Volume 36 • Issue 1

Learning. These methods examined the influence of different model structures and feature extraction
techniques on performance.
Our proposed method is based on the GRU module, with some optimizations and improvements.
Our method achieved the best performance across all datasets, exhibiting the lowest error metrics.
Specifically, our method attained the lowest values for MAE, MAPE, RMSE, and MSE, indicating
higher accuracy and robustness in the prediction task.
Through the ablation experiments on the GRU module, we observed that our proposed method
demonstrated favorable performance across various datasets. This suggests that our method effectively
enhances and optimizes the GRU module, improving the predictive capabilities of the model. Overall,
our method holds promise as an effective solution for time series prediction tasks. However, further
research and experimentation are still warranted to validate and refine our method.

5. CONCLUSION AND DISCUSSION

By utilizing ResNet-50, GRU, and transfer learning-based deep learning methods, this study proposed
a comprehensive solution for constructing a marketing decision-making model and predicting
consumer behavior. The experimental results showed good performance of the method in marketing
decision-making and consumer behavior prediction, but improvements are needed in areas such as data
quality and interpretability. Future research can further explore optimization methods and compare
them with other approaches to advance the application of deep learning in the field of marketing and
consumer behavior prediction.
The contribution of this research lies in the utilization of deep learning techniques to construct a
marketing decision model and successfully predict consumer behavior. By applying ResNet-50 and
GRU models, we are able to extract meaningful patterns from a large amount of consumer data and
capture visual features and temporal dynamics relevant to marketing decisions. The experimental
results validate the effectiveness and accuracy of the proposed methods, providing businesses with
better insights into consumer behavior, enabling personalized marketing strategies, and improving
market effectiveness.
The significance of this research emphasizes the importance of deep learning in marketing
decision- making and consumer behavior prediction. Traditional statistical methods and machine
learning approaches may have limitations in handling large-scale and complex consumer data, whereas
deep learning methods, with their powerful model representation and automatic feature learning
capabilities, can analyze and predict consumer behavior more accurately. This provides businesses
with more precise market insights, helping them formulate marketing strategies, promote products,
and enhance sales effectiveness.
Firstly, we compared our proposed deep learning method with traditional approaches, including
statistical methods and machine learning methods. Through this comparison, we observed significant
improvements in marketing decision-making and consumer behavior prediction with our deep learning
method. Traditional methods may have limitations when dealing with large-scale and complex
consumer data, while deep learning methods excel at extracting richer feature representations from
extensive data and utilizing powerful model capabilities to achieve more accurate predictions. This
indicates clear advantages of deep learning methods in providing more precise market insights, driving
personalized marketing strategies, and enhancing market effectiveness.
Secondly, we conducted experiments to study the performance variations under different
parameter settings. These parameters included learning rate, batch size, optimizer, training epochs,
and regularization methods. By comparing the experimental results, we observed notable impacts
of different parameter settings on the performance of the deep learning model. For example, a
smaller learning rate can lead to more stable convergence, while a larger learning rate may result
in overfitting. Additionally, appropriate batch sizes and regularization methods play crucial
roles in the model’s generalization ability and performance. Through comprehensive analysis

22
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

of the experimental results, we determined the optimal parameter configuration to achieve the
best model performance.
Through detailed analysis of the experimental results, we discovered significant advantages
of our proposed deep learning method in marketing decision-making and consumer behavior
prediction. Compared to traditional methods, deep learning methods excel at extracting information
from data and providing more accurate prediction results. Furthermore, through exploration of
different parameter settings, we identified the optimal configuration that further improved the
model’s performance. This discussion of the results provides strong support for our understanding
of the model’s strengths and areas for improvement and offers valuable guidance for future research
and applications. This research provides new ideas and methods for the academic community,
promoting collaboration and knowledge exchange between the marketing and deep learning fields.
Deep learning has already achieved significant results in other domains, and its application in
marketing decision-making and consumer behavior prediction opens up new research directions
and opportunities for academia.
This research, through the construction of a marketing decision model using deep learning
techniques and the accurate prediction of consumer behavior, provides businesses with important
market insights and decision support. It also offers new research directions for the academic
community. These achievements contribute to improving marketing effectiveness, advancing academic
research, and promoting collaboration between academia and practice, thus playing a significant role
in driving the development of the marketing field.

CONFLICT OF INTEREST STATEMENT

The authors declare that the research was conducted in the absence of any commercial or financial
relationships that could be construed as a potential conflict of interest.

ACKNOWLEDGMENT

This work was supported by 1. National Natural Science Foundation of China (No. 72174045); 2.
Social science Foundation of Heilongjiang Province (21GLB064); 3. Key Laboratory of Island Tourism
Resource Data Mining and Monitoring, Ministry of Culture and Tourism (KLITRDMM 2024).

23
Journal of Organizational and End User Computing
Volume 36 • Issue 1

REFERENCES

Alon, I., Qi, M., & Sadowski, R. J. (2001). Forecasting aggregate retail sales: A comparison of artificial neural
networks and traditional methods. Journal of Retailing and Consumer Services, 8(3), 147–156. doi:10.1016/
S0969-6989(00)00011-4
Alzubaidi, L., Zhang, J., Humaidi, A. J., Al-Dujaili, A., Duan, Y., Al-Shamma, O., Santamaría, J., Fadhel, M.
A., Al-Amidie, M., & Farhan, L. (2021). Review of deep learning: Concepts, cnn architectures, challenges,
applications, future directions. Journal of Big Data, 8(1), 1–74. doi:10.1186/s40537-021-00444-8 PMID:33816053
Austin, P. C., Goel, V., & van Walraven, C. (2001). An introduction to multilevel regression models. Canadian
Journal of Public Health, 92(2), 150–154. doi:10.1007/BF03404950 PMID:11338155
Badea, L. M. (2014). Predicting consumer behavior with artificial neural networks. Procedia Economics and
Finance, 15, 238–246. doi:10.1016/S2212-5671(14)00492-4
Busseri, M. A., Lefcourt, H. M., & Kerton, R. R. (1998). Locus of control for consumer outcomes: Predicting
consumer behavior 1. Journal of Applied Social Psychology, 28(12), 1067–1087. doi:10.1111/j.1559-1816.1998.
tb01668.x
Chakravarti, D., Mitchell, A., & Staelin, R. (1979). Judgment based marketing decision models: An experimental
investigation of the decision calculus approach. Management Science, 25(3), 251–263. doi:10.1287/mnsc.25.3.251
Chiu, M.-C., & Chuang, K.-H. (2021). Applying transfer learning to achieve precision marketing in an omni-
channel system–a case study of a sharing kitchen platform. International Journal of Production Research, 59(24),
7594–7609. doi:10.1080/00207543.2020.1868595
Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B., & Bharath, A. A. (2018). Generative
adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1), 53–65. doi:10.1109/
MSP.2017.2765202
De Mauro, A., Sestino, A., & Bacconi, A. (2022). Machine learning and artificial intelligence use in marketing:
A general taxonomy. Italian Journal of Marketing, 2022(4), 439–457. doi:10.1007/s43039-022-00057-w
Dey, R., & Salem, F. M. (2017). Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th
international midwest symposium on circuits and systems (MWSCAS) (IEEE), 1597–1600 Famili, A., Shen,
W.-M., Weber, R., and Simoudis, E. (1997). Data preprocessing and intelligent data analysis. Intelligent Data
Analysis, 1, 3–23.
Gatignon, H. (2000). Commentary on peter leeflang and dick wittink’s “building models for marketing decisions:
Past, present and future”. International Journal of Research in Marketing, 17(2-3), 209–213. doi:10.1016/
S0167-8116(00)00019-7
Gauch, H. G., Hwang, J. G., & Fick, G. W. (2003). Model evaluation by comparison of model-based predictions
and measured values. Agronomy Journal, 95(6), 1442–1446. doi:10.2134/agronj2003.1442
Gers, F. A., Schmidhuber, J., & Cummins, F. (2000). Learning to forget: Continual prediction with lstm. Neural
Computation, 12(10), 2451–2471. doi:10.1162/089976600300015015 PMID:11032042
Gonog, L., & Zhou, Y. (2019). A review: generative adversarial networks. In 2019 14th IEEE conference on
industrial electronics and applications (ICIEA) (IEEE), 505–510 doi:10.1109/ICIEA.2019.8833686
Gupta, S., Justy, T., Kamboj, S., Kumar, A., & Kristoffersen, E. (2021). Big data and firm marketing performance:
Findings from knowledge-based view. Technological Forecasting and Social Change, 171, 120986. doi:10.1016/j.
techfore.2021.120986
Gursch, H., Ganster, H., Rinnhofer, A., Waltner, G., Payer, C., Oberwinkler, C., et al. (2021). Ki- waste-combining
image recognition and time series analysis in refuse sorting. Mensch und Computer 2021-Workshopband
Hanin, B. (2018). Which neural net architectures give rise to exploding and vanishing gradients? Advances in
Neural Information Processing Systems, 31.
He, K., Gkioxari, G., Dolla’r, P., & Girshick, R. (2017). Mask r-cnn. In Proceedings of the IEEE international
conference on computer vision. (pp. 2961–2969). IEEE.

24
 Journal of Organizational and End User Computing
Volume 36 • Issue 1

He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of
the IEEE conference on computer vision and pattern recognition, (pp. 770–778). IEEE.
Keneshloo, Y., Shi, T., Ramakrishnan, N., & Reddy, C. K. (2019). Deep reinforcement learning for sequence-to-
sequence models. IEEE Transactions on Neural Networks and Learning Systems, 31, 2469–2489. PMID:31425057
Koonce, B. and Koonce, B. (2021). Resnet 50. Convolutional Neural Networks with Swift for Tensorflow: Image
Recognition and Dataset Categorization, 63–72
Kumar, D. T. S. (2020). Data mining based marketing decision support system using hybrid machine learning
algorithm. Journal of Artificial Intelligence and Capsule Networks, 2(3), 185–193. doi:10.36548//jaicn.2020.3.006
Kumar, J. A., & Abirami, S. (2021). Ensemble application of bidirectional lstm and gru for aspect category
detection with imbalanced data. Neural Computing & Applications, 33(21), 14603–14621. doi:10.1007/s00521-
021-06100-9
Leeflang, P. S., & Wittink, D. R. (2000). Building models for marketing decisions: Past, present and future.
International Journal of Research in Marketing, 17(2-3), 105–126. doi:10.1016/S0167-8116(00)00008-2
Lodish, L. M. (2001). Building marketing models that make money. Interfaces, 31(3_supplement), S45–S55.
doi:10.1287/inte.31.3s.45.9681
Niu, Z., Zhong, G., & Yu, H. (2021). A review on the attention mechanism of deep learning. Neurocomputing,
452, 48–62. doi:10.1016/j.neucom.2021.03.091
Omran, M. G., Engelbrecht, A. P., & Salman, A. (2007). An overview of clustering methods. Intelligent Data
Analysis, 11(6), 583–605. doi:10.3233/IDA-2007-11602
Salehinejad, H., Sankar, S., Barfett, J., Colak, E., & Valaee, S. (2017). Recent advances in recurrent neural
networks. arXiv preprint arXiv:1801.01078
Saura, J. R. (2021). Using data sciences in digital marketing: Framework, methods, and performance metrics.
Journal of Innovation & Knowledge, 6(2), 92–102. doi:10.1016/j.jik.2020.08.001
Soni, A. N. (2018). Application and analysis of transfer learning-survey. International Journal of Scientific
Research and Engineering Development, 1, 272–278.
Stephen, A. T. (2016). The role of digital and social media marketing in consumer behavior. Current Opinion
in Psychology, 10, 17–21. doi:10.1016/j.copsyc.2015.10.016
Torrey, L., & Shavlik, J. (2010). Transfer learning. In Handbook of research on machine learning applications and
trends: algorithms, methods, and techniques (pp. 242–264). IGI global. doi:10.4018/978-1-60566-766-9.ch011
Xiao, J., & Zhou, Z. (2020). Research progress of rnn language model. In 2020 IEEE International Conference
on Artificial Intelligence and Computer Applications (ICAICA) (IEEE), (pp. 1285–1288). IEEE. doi:10.1109/
ICAICA50127.2020.9182390

Praveen Kumar Donta, Currently working as Postdoc at Distributed Systems Group, TU Wien (Vienna University
of Technology), Vienna, Austria since July 2021. He is received his Ph. D. at Indian Institute of Technology (Indian
School of Mines), Dhanbad from the Department of Computer Science and Engineering in May 2021. From July
2019 to Jan 2020, he is a visiting Ph.D. fellow at Mobile & Cloud Lab, Institute of Computer Science, Faculty of
Science and Technology, University of Tartu, Estonia, under the Dora plus grant provided by the Archimedes
Foundation, Estonia. He received his Master in Technology and Bachelor in Technology from the Department of
Computer Science and Engineering at JNTUA, Ananthapuramu, with Distinction in 2014 and 2012.

25