Deep Learning to Address Candidate Generation and

Cold Start Challenges in Recommender Systems: A

Research Survey

Kiran R, Pradeep Kumar, Bharat Bhasker

([email protected], [email protected], [email protected])
IT and Systems Department
Indian Institute of Management Lucknow
Lucknow 226013, India

Corresponding Author: Kiran Rama

Address: No 295, 3rd Block, 6th Cross, HMT Layout, Vidyaranyapura PO, Bangalore - 560097
Telephone : +91 9900601013
Email: [email protected]

Author: Prof Pradeep Kumar

Indian Institute of Management Lucknow , Off Sitapur Road , Lucknow 226013, India

Author: Prof Bharat Bhasker

Indian Institute of Management Raipur, GEC Campus, Sejbahar, Raipur, Chhattisgarh 492015,
India

Keywords: Recommender Systems, Challenges in Recommender Systems, Deep Learning,

Candidate Generation, Cold Start, Accuracy, CNN, RNN, GAN, Autoencoders, DNN, DBN
 ABSTRACT
Among the machine learning applications to business, recommender systems would take one

of the top places when it comes to success and adoption. They help the user in accelerating the

process of search while helping businesses maximize sales. Post phenomenal success in

computer vision and speech recognition, deep learning methods are beginning to get applied to

recommender systems. Current survey papers on deep learning in recommender systems provide

a historical overview and taxonomy of recommender systems based on type. Our paper addresses

the gaps of providing a taxonomy of deep learning approaches to address recommender systems

problems in the areas of cold start and candidate generation in recommender systems. We outline

different challenges in recommender systems into those related to the recommendations

themselves (include relevance, speed, accuracy and scalability), those related to the nature of the

data (cold start problem, imbalance and sparsity) and candidate generation. We then provide a

taxonomy of deep learning techniques to address these challenges. Deep learning techniques

including convolutional neural networks, recurrent neural networks, autoencoders, stacked multi-

layer perceptron, stacked Restricted Boltzmann Machines and Generative Adversarial Networks

are categorized as supervised and unsupervised, as well as based on their applicability to

different types of data. These techniques are mapped to the different challenges in recommender

systems providing an overview of how deep learning techniques can be used to address them.

We contribute a taxonomy of deep learning techniques to address the cold start and candidate

generation problems in recommender systems. Cold Start is addressed through additional

features (for audio, images, text) and by learning hidden user and item representations. Candidate

generation has been addressed by separate networks, RNNs, autoencoders and hybrid methods.

We also summarize the advantages and limitations of these techniques while outlining areas for

future research.
1 INTRODUCTION
The business world today is characterized by information overload and increased

competition for customer’s attention and wallet share. In this context, Recommender Systems

provide several benefits to both consumers and business firms. They help consumers with

reduced search costs, increased product fit and management of choice overload. They help

business firms with higher sales volume, higher web usage, higher customer retention and higher

margins (Lee, Huntsman, Fl, Huntsman, & Fl, 2014). Several successes of recommender systems

have been highlighted like 80% of movies watched on Netflix came from recommendation

(Gomez-Uribe & Hunt, 2015) and 60% of video clicks came from home page recommendations

on YouTube (Davidson et al., 2010). We define Recommender systems based on several

definitions in literature as

“Systems that seek to predict the future preference of a set of items for a user either as a

numeric rating for the item or as a list of recommendations or as a binary score indicating

preference for the item” (Handbook, RICCI, & ROKACH, 2011), (Portugal, Alencar, & Cowan,

2017) (Cheng et al., 2016) (S. Zhang & Yao, 2017)

Advances in graphical processing unit hardware and decreasing costs have made GPUs

accessible. With their multi-core nature, they enable highly parallel matrix computations that are

at the core of Deep Learning. GPUs have been found to speed up the learning of a deep learning

network by a factor of more than 50 (Schmidhuber, 2015). Consequently, Deep Learning has

seen tremendous success in computer vision and speech recognition. Deep Learning methods

employ neural networks with multiple hidden layers and are shown to be infinitely flexible

functions capable of solving any problem provided there is sufficient data and computing power.

They fit these several thousands of parameters using gradient descent. There have been
tremendous improvements in backpropagation and gradient descent since the seminal paper on

gradient descent in 1986 (Rumelhart, Hinton, & William, 1985). There are several methods for

optimization used like RMSProp (Dauphin, de Vries, & Bengio, 2015), AdaDelta, AdaMax,

Nadam, AMSGrad, Adam (Kingma & Ba, 2015) etc. The advancement of several frameworks

like PyTorch from Facebook (Ketkar, 2017), MXNet from Apache (Chen et al., 2015) and

TensorFlow (Google, 2018) from Google have increased the adoption of deep learning.

The motivation to write this survey paper is to not to communicate the list of results but

to provide an understanding into deep learning applications in recommender systems, providing

a taxonomy of how deep learning techniques have been used to address the candidate generation

and cold start challenges in recommender systems.

1.1 RESEARCH GAP AND OBJECTIVE

There exist several surveys in traditional recommender systems but to the best of our

knowledge, there are very few surveys on the application of deep learning to recommender

systems. (S. Zhang & Yao, 2017) classified the techniques into methods that solely relied on

deep learning and those that integrate deep learning with traditional recommender systems,

further breaking them as loosely coupled and tightly coupled. The paper provides a historical

course of development of techniques in the field. Another classification of deep learning for

recommender systems is bucketing them into content, collaborative and hybrid buckets (Singhal,

Sinha, & Pant, 2017). (Betru & Onana, 2017) is another survey paper on deep learning in

recommender systems. None of these survey papers provide a taxonomy of deep learning

methods based on the challenges of recommender systems that they address.

The research objective of this paper is to address the literature gaps mapping different deep

learning techniques against the recommendation system challenges they address and provide a
mapping of the techniques into applicable scenarios. The scope of this paper will be limited to

the application of deep learning techniques to address the cold start and candidate generation

problems in recommender systems.

1.2 RESEARCH METHODOLOGY

We survey different research papers on the application of deep learning to recommender

systems. We extract articles from Google Scholar Search for the period 2009-present using

combinations of the following keywords: deep learning, recommender systems, rec sys, recurrent

neural networks, convolutional neural networks, stacking Restricted Boltzmann Machines, Deep

Neural Networks, Autoencoders, Candidate Generation and Cold start. The resulting papers are

manually curated based off the paper title and abstracts to find the articles. The deep learning

techniques in the papers are then mapped to recommender system challenges building out a

taxonomy. We avoid taking the historical course of development approach or the taxonomy

based on content-based and collaborative filtering methods as these have already been addressed

by several other survey papers. We focus on the techniques used to address specific challenges,

their advantages and limitations and areas for further research.

1.3 ORGANIZATION OF THE PAPER

In the first section, we introduce recommender systems and deep learning topics. We also

highlight the research objective and methodology. In the second section, we provide a

background to the recommender system challenges and deep learning by providing a taxonomy.

In the next section, we contribute to the research on the topic by coming up with a taxonomy of

deep learning methods to handle the problems of cold start and candidate generation in

recommender systems. In the last two sections, we provide a summary of the limitations of the
research in the area and conclude by summarizing our contributions and areas for future

research.
2 BACKGROUND
In this section, we summarize the challenges associated with recommender systems. We then

provide a taxonomy of the deep learning methods, both based on the technique and based on the

area of applicability.

2.1 RECOMMENDER SYSTEM CHALLENGES

The challenges in recommender systems can be classified into those related to the

recommendations themselves, nature of the data and candidate generation. This is summarized in

Figure 2.1. Challenges related to the recommendations themselves include speed, scalability,

accuracy and relevance. Speed and Scalability are satisficing criteria while accuracy and

relevance are optimizing criteria. Recommender systems are expected to provide the outputs in a

reasonable amount of time while scaling to large amounts of data. Accuracy is measured by

different metrics based on the nature of the task like RMSE, RMSLE (for regression problems),

F-score, AUC, accuracy (for classification problems), NDCG and MAP (for ranking problems).

The key goal here in relevance to avoid noise and provide fresh recommendations. Challenges

related to the nature of the data include cold start problem, imbalance and sparsity. When there is

no history about the user to whom the recommendation is to be made or the item that is to be

recommended, the recommender systems run into the ‘cold start’ problem. Funnily, the origin of

the cold start problem was from cars in cold regions that had problems starting up and running

till it reaches its optimum temperature. For recommender systems, this means that there are not

enough historical interactions for item or user.

 Recommender
Systems
Challenges

Recommendations
Nature of Data
Related

Candidate Performance
Cold Start Problem
Generation Related Related

Relevance Sparsity

Accuracy Imbalance

Scalability

Speed

Fig 2.1: Challenges in Recommender Systems

Larger share of the ratings matrix is empty with no ratings for a user-item pair and this leads to the

sparsity problem. Very few item-user combinations in the data matrix are minority class examples

leading to the problem of imbalance. The last bucket of problem relates to candidate generation.

Creating an entry for all user-item combinations is an onerous task and it becomes necessary to

generate pairs of users and items for training. For example: With 1 million videos and 100 million

users, YouTube has a candidate generation problem as it cannot create an entry for every user-

video pair. (Covington, Adams, & Sargin, 2016).

2.2 TAXONOMY OF DEEP LEARNING METHODS
In figure 2.2, we categorize the deep learning techniques based on their complexity and type.

Deep-Learning
Types

Discriminatory &
Stacked Hidden Convolutional Recurrent Neural
Generator Autoencoders
Layers Neural Network Network
Networks

Long Short Term Generative

Deep Neural Deep Belief
Memory Networks Adversarial Neural
Networks Networks
(LSTM) Networks

Stacked Restricted
Stacked Multi Layer Gated Recurrent
Boltzmann
Perceptrons Untis (GRU)
Machines

Figure 2.2: Taxonomy of Deep Learning Methods by Method and Architecture Type
Multi-layer perceptrons (MLP) are feedforward neural networks with multiple hidden

layers between the input layer and the output layer. Restricted Boltzmann Machines (RBM) are

two layered neural networks with a visible layer and a hidden layer. Both MLPs and RBMs are

traditional neural networks. By stacking many RBNs together, we get a Deep Belief Network

and by stacking several hidden layers in a MLP together, we get a Deep Neural Network (DNN).

CNNs have several convolutional layers that apply filters followed by pooling layers with non-

linear transformations and have found huge success in image processing. RNNs are suitable for

modelling sequential recommendations and they are a unique type of neural network that contain

loops that enable the networks to have memory storage. RNNs are of two types namely Long

Short Term Memory (LSTM) and Gated Recurrent Units (GRU). Generative Adversarial
Networks (GANs) are generative neural networks that consist of a discriminator and a generator

network with both trained simultaneously competing with one another. All the methods

discussed so far are supervised deep learning methods. Autoencoders are unsupervised models

that attempt to reconstruct the input data in the output layer and use the bottleneck layer as a

salient representation of the input. Figure 2.3 shows the classification of the deep learning

methods based on the popular types of data they are applied to. It is to be noted that these are not

the “only” applications but rather the most popular applications

Data Type & The

Dominant Deep
Learning Method

Joint Learning of
Computer Vision Speech Structured Data Text Data Sequential Data
Latent Features

Word
CNN CNN DNN RNN GAN
Embeddings

RNN

Figure 2.3: Taxonomy of Deep Learning Methods by Popular application type

3 PROPOSED TAXONOMY OF DEEP LEARNING
TECHNIQUES TO ADDRESS COLD START AND
IMBALANCE
In the section, we map the different deep learning techniques and variants to the

recommender system challenges. We then summarize the advantages and disadvantages of each

of these approaches.

3.1 COLD START

Cold Start Problem is one of “Making recommendations where there are no prior interactions

available for an user or an item” (Lam, Vu, & Le, 2008). (Bernardi, Kamps, Kiseleva, &

Mueller, 2015; Lika, Kolomvatsos, & Hadjiefthymiades, 2014) highlight the cold start problem

and classify it into user cold start and item cold start cases, referring to cases of insufficient

examples of items and users respectively.

Deep Learning
Methods for Cold
Start

User & Item

Audio Features Image Features Text Features Hidden
Representations

Latent Audio Sequential Audio Co-operative

CNN RNNs Autoencoders
Features Features Neural Networks

Generative
GRU based
CNN RNN LSTM Methods Adversarial
methods
Networks

Order Sensitive
Text Encoders

Collaborative
Sequence
Models

Fig 3.1: Taxonomy of Deep Learning Methods to address Cold Start Problem
The usual approach to solving the cold start problem is to include additional user features by

combining content based and collaborative filtering based methods (Chiliguano & Fazekas,

2016).

Modern Song Recommender Systems fail to recommend tracks that are less widely known

and suffer from the cold start problem. Deep convolutional networks have been used to generate

latent factors for songs from their audio when no usage is available (Oord, Dieleman, &

Schrauwen, 2013) and using audio content of the song to add some item based features alleviates

the cold start problem. CNNs have been applied to extract features from images (Oord et al.,

2013). The sequential property of audio has been used to build a LSTM-based recurrent neural

network architecture in combination with audio based features from a convolutional networks

have been attempted. (Balakrishnan, 2014). Content-Based Deep Learning methods include

extracting latent factors for songs from their audio (Liang, 2014), enhancing content based quote

recommendations(Tan, Wan, & Xiao, 2016), converting item text into latent features. GRU

based recurrent neural networks have also been used to convert item text into latent features to

improve the collaborative filtering results for the cold start problem (Bansal, Belanger, &

McCallum, 2016).

For textual data, order sensitive encoders in combination with latent factor models have been

used to address the cold start problem by employing RNNs to encode text sequences into latent

factors (Bansal et al., 2016). Extracted item representations are combined with user embeddings

to get predicted ratings for each user-item pair. Compared to traditional methods that ignore

word order, leveraging the word order in recurrent neural networks helps soften the cold start

problem. Collaborative Sequence models based on recurrent neural networks have also shown

good results. (Ko, Maystre, Grossglauser, Durrant, & Kim, 2016). (Devooght & Bersini, 2016)
show that collaborative filtering can be viewed as a sequence prediction problem and application

of LSTMs is very competitive in addressing the cold start problem.

Latent representations from the data are learnt in an unsupervised manner and the implicit

relationships between items and users are learnt from both the content and the rating (Xiaopeng

Li & She, 2017). The autoencoder is forced to learn a latent distribution for content in latent

space instead of observation space through an inference network and can easily be extended to

non-text type multimedia as well. Instead of learning the user latent factors and item latent

factors separately, Co-operative Neural Networks (CoNN) learn hidden latent features for users

and items jointly using two coupled neural networks. These are examples of Generative

Adversarial Networks (GANs) which consist of a discriminator and a generator networks and

both are trained simultaneously competing in a minimax game framework

3.2 CANDIDATE GENERATION

If there are 1 million users and 0.3 million items, a data matrix at a user-item level would

have 0.3 * 10^12 combinations as an input into a data mining algorithm to solve a classification

or prediction task. Such a huge data matrix lends itself unviable from a computational and

imbalance perspective. Hence candidate generation is an important problem in recommender

systems. A summary of the methods in literature to solve the problem of candidate generation in

deep learning is shown in Figure 3.2

 Deep Learning
Methods for Candidate
Generation

Separate Candidate Recurrent Neural

Autoencoders Hybrid Methods
Generation Network Network

Video
Question-Answer Heuristic Methods +
Recommendations on Word Embeddings Deep Learning + ML
Systems ML
YouTube

Deep Keyphrase
k-NN Bloom Filter Chains
generation

Association Rules Thumb Rules

Multi-class
Classification
Algorithms

Similarity Search

Fig 3.2: Taxonomy of Deep Learning Methods to address Candidate Generation Problem

YouTube used a candidate generation network (Covington et al., 2016) with inputs as events

from user history. The problem of returning hundreds of videos from millions was treated as an

extreme multi-class problem and is solved using a deep neural network that learnt the user

embeddings as a function of the user’s history and context. RNNs have been used for candidate

generation for sequential data in Question-Answer Systems and deep keyphrase generation. In

(Shen, Chen, & Huang, 2016), the authors turn to paraphrases as a means of capturing the way

natural language expresses the information needed and present a general framework using a

deep neural model which learns felicitous paraphrases for various tasks. In (Y. Zhang, Fang, &
Weidong, 2018), first a list of phrase candidates with heuristic methods are returned followed

by scoring each candidate phrase for its likelihood of being a keyphrase in the given document.

(Xiaonan Li, Li, & Yu, 2012) combined word embeddings with L2 regularized large scale

logistic regression to search sentences similar to a query from Wikipedia articles and directly

use the human-annotated entities in the similar sentences as candidate entities for the query.

Trivial business rule-based techniques and simple machine learning methods have been

used combined with deep learning extensively in literature to create hybrid systems to solve the

problem of candidate generation. In Google Apps, there are over a million apps in the database.

Every app cannot be scored for every query. Returning a short list of items that best match the

query using various signals is achieved using a combination of machine learned methods and

human defined rules (Cheng et al., 2016). Bloom Filter Chains have also been used for

candidate generation for real-time tweet search involving sorting of tweets using a simple

scoring function and then rescoring them with a separate component (Asadi & Lin, 2013). IBM

Watson, a commercial tool calls out “Hypothesis Generation” as a key component and the same

uses candidate generation to retrieve answers to a particular question (Chu-Carroll et al., 2012).

Association Rules, a classic machine learning technique has also been used for candidate

generation as Generalized Sequential Patterns (Srikant & Agrawal, 1996). Traditional methods

have also been combined with deep learning techniques to create hybrid systems.
4 OPPORTUNITIES
On candidate generation, while there have been novel implementations of deep learning,

they have been very domain focused like candidate generation network using Deep Neural

Network for YouTube, RNNs for Question-Answer Systems and Deep Keyphrase generation and

autoencoders for word embeddings. These methods have been shown to solve the candidate

generation problem in a specific domain, but they do not generalize to all scenarios. Also, the

candidate generation methods apart from the YouTube example have not combined deep

learning with traditional machine learning techniques like k-NN, association rules, bloom filters

and classification algorithms. These present an opportunity for future research.

On cold start, hand-crafting features has been reduced greatly in the cases of images, audio

and text where CNNs and RNNs have shown tremendous results. Still, for traditional structured

recommendations data, hand-crafting features is required in many cases, which means that

features need to be maintained and domain experts need to be deployed. Employing User and

item features effectively into deep network in an automated way into deep learning methods to

solve the cold start problem is very important. Common to both cold start and candidate

generation, there are very few examples of using the different nature of deep learning methods in

a way complementary to one another. Some examples include combining RNNs and

autoencoders. Deep Neural Networks on text ignore the word order and semantic meanings but

Recurrent Neural Networks can take the word order into account. Autoencoders can leverage the

semantic order. Combining these methods can potentially give better results. The actual

incremental improvement that can be expected from deep learning methods on candidate

generation remains an important research area. An immediate improvement in alleviating cold

start becomes possible where there have been several implementations taking one of the
techniques into account. In one of the papers, latent factors learned from Weighted Matrix

Factorization was not accurate enough to train a CNN, but could have easily adapted to a

different technique.

4.1 RESEARCH OPPORTUNITIES

Some of the research areas that we have identified include:

• Using Autoencoders to effectively automatically engineer features to solve the cold start

problem. The autoencoder approach to feature engineering that converts a user vector to a

dense vector could bring out a significant improvement in the performance on the

experimental datasets, but there is not much literature on it. Autoencoders seem to have

been used heavily for text recommendations but not on structured datasets

• Factorization machines can avoid tedious feature engineering for cold start problem and

could be promising to integrate them with deep learning techniques

• Exploring CNNs for structured data. Techniques from image processing to features like

filters, pooling, residual networks, one-pass object detections could find parallels in

structured data modeling. This is in addition to leveraging CNNs for extracting features

from images data

• Recommendations are a form of sequential data. There is not enough literature on

recurrent neural networks and their types LSTM and GRU to automate some feature

engineering

• All algorithms have tried to optimize for overall accuracy metrics like RMSE, RMSLE,

AUC, F-score, NDCG, MAP etc., which are global measures for the whole dataset. There

could be potential in coming up with a custom evaluation metric that gives higher weight

to cold start cases

• Coming up with hybrid methods that combine Deep learning with Association rules or k-

NN for a generalized framework for candidate generation is a promising area

5 CONCLUSION
Current survey papers on deep learning in recommender systems provide a historical

overview and taxonomy of recommender systems based on type. We have grouped the

recommender system challenges into those related to recommendations, those related to

candidate generation and those related to the nature of the data. We have focused on candidate

generation and cold start challenges in this paper. Our paper addresses the gaps of providing a

taxonomy of deep learning approaches to address recommender systems problems in the areas of

cold start and candidate generation. Cold Start problem is due to insufficient information about

users or items. Deep Learning techniques have tried to address this by generating features for the

audio (CNN, RNN), for the images (CNN), for the text (LSTM methods and GRU methods) and

for hidden user and item representations (Autoencoders and GANs). The candidate generation

problem has been addressed by creating separate candidate generation networks (DNN) and

recurrent neural networks. We have summarized the limitations of each of the methods and the

important research opportunities

Our key contribution in this survey paper is the taxonomy of deep learning methods to

address cold start and candidate generation with a summary of the limitations and opportunities

for research
6 REFERENCES
Asadi, N., & Lin, J. (2013). Fast candidate generation for real-time tweet search with bloom filter chains.
ACM Transactions on Information Systems, 31(3), 1–36. https://doi.org/10.1145/2493175.2493178
Balakrishnan, A. (2014). DeepPlaylist : Using Recurrent Neural Networks to Predict Song Similarity.
Stanfort University, 1–7. Retrieved from https://cs224d.stanford.edu/reports/BalakrishnanDixit.pdf
Bansal, T., Belanger, D., & McCallum, A. (2016). Ask the GRU. In Proceedings of the 10th ACM Conference
on Recommender Systems - RecSys ’16 (pp. 107–114). https://doi.org/10.1145/2959100.2959180
Bernardi, L., Kamps, J., Kiseleva, J., & Mueller, M. J. I. (2015). The continuous cold start problem in e-
commerce recommender systems. In CEUR Workshop Proceedings (Vol. 1448, pp. 30–33).
Betru, B. T., & Onana, C. A. (2017). Deep Learning Methods on Recommender System: A Survey of State-
of-the-art. International Journal of Computer Applications, 162(10), 975–8887.
https://doi.org/10.5120/ijca2017913361
Chen, T., Li, M., Li, Y., Lin, M., Wang, N., Wang, M., … Zhang, Z. (2015). MXNet: A Flexible and Efficient
Machine Learning Library for Heterogeneous Distributed Systems. eprint arXiv:1512.01274.
https://doi.org/10.1145/2532637
Cheng, H.-T., Ispir, M., Anil, R., Haque, Z., Hong, L., Jain, V., … Chai, W. (2016). Wide & Deep
Learning for Recommender Systems. In Proceedings of the 1st Workshop on Deep Learning for
Recommender Systems - DLRS 2016 (pp. 7–10). https://doi.org/10.1145/2988450.2988454
Chiliguano, P., & Fazekas, G. (2016). Hybrid music recommender using content-based and social
information. In ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing -
Proceedings (Vol. 2016–May, pp. 2618–2622). https://doi.org/10.1109/ICASSP.2016.7472151
Chu-Carroll, J., Fan, J., Boguraev, B. K., Carmel, D., Sheinwald, D., & Welty, C. (2012). Finding needles in
the haystack: Search and candidate generation. IBM Journal of Research and Development, 56(3.4),
6:1-6:12. https://doi.org/10.1147/JRD.2012.2186682
Covington, P., Adams, J., & Sargin, E. (2016). Deep Neural Networks for YouTube Recommendations. In
Proceedings of the 10th ACM Conference on Recommender Systems - RecSys ’16 (pp. 191–198).
https://doi.org/10.1145/2959100.2959190
Dauphin, Y. N., de Vries, H., & Bengio, Y. (2015). RMSProp and Equilibrated adaptive learning rates for
non-convex optimization. NIPS ’15, 10. Retrieved from http://arxiv.org/abs/1502.04390
Davidson, J., Livingston, B., Sampath, D., Liebald, B., Liu, J., Nandy, P., … Lambert, M. (2010). The
YouTube video recommendation system. In Proceedings of the fourth ACM conference on
Recommender systems - RecSys ’10 (p. 293). https://doi.org/10.1145/1864708.1864770
Devooght, R., & Bersini, H. (2016). Collaborative Filtering with Recurrent Neural Networks. CoRR,
abs/1608.0. Retrieved from http://arxiv.org/abs/1608.07400
Gomez-Uribe, C. A., & Hunt, N. (2015). The Netflix Recommender System: Algorithms, Business Value,
and Innovation. ACM Trans. Manage. Inf. Syst., 6(4), 13:1--13:19. https://doi.org/10.1145/2843948
Google. (2018). TensorFlow.
Handbook, R. S., RICCI, F., & ROKACH, L. (2011). Recommender Systems Handbook. Recommender
 Systems Handbook, 54, 39–72. https://doi.org/10.1007/978-0-387-85820-3
Ketkar, N. (2017). Introduction to PyTorch. In Deep Learning with Python (pp. 195–208).
https://doi.org/10.1007/978-1-4842-2766-4_12
Kingma, D. P., & Ba, J. L. (2015). Adam: a Method for Stochastic Optimization. International Conference
on Learning Representations 2015, 1–15.
https://doi.org/http://doi.acm.org.ezproxy.lib.ucf.edu/10.1145/1830483.1830503
Ko, Y.-J., Maystre, L., Grossglauser, M., Durrant, R. J., & Kim, K.-E. (2016). Collaborative Recurrent Neural
Networks for Dynamic Recommender Systems. JMLR, 63, 366–381.
https://doi.org/10.1109/ICDE.2016.7498326
Lam, X., Vu, T., & Le, T. (2008). Addressing cold-start problem in recommendation systems. Proceedings
of the 2nd International, 208–211. https://doi.org/10.1145/1352793.1352837
Lee, D., Huntsman, J. M., Fl, H., Huntsman, J. M., & Fl, H. (2014). Impact of Recommender Systems on
Sales Volume and Diversity. Icis, 1–15.
Li, X., Li, C., & Yu, C. (2012). Entity-Relationship Queries over Wikipedia. ACM Transactions on Intelligent
Systems and Technology, 3(4), 1–20. https://doi.org/10.1145/2337542.2337555
Li, X., & She, J. (2017). Collaborative Variational Autoencoder for Recommender Systems. In Proceedings
of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD
’17 (pp. 305–314). https://doi.org/10.1145/3097983.3098077
Liang, D. (2014). Content-Aware Collaborative Music Recommendation Using Pre-trained Neural
Networks. ISMIR 2015: Proceedings of the 16th International Society for Music Information
Retrieval Conference, 295–301.
Lika, B., Kolomvatsos, K., & Hadjiefthymiades, S. (2014). Facing the cold start problem in recommender
systems. Expert Systems with Applications, 41(4 PART 2), 2065–2073.
https://doi.org/10.1016/j.eswa.2013.09.005
Oord, A. van den, Dieleman, S., & Schrauwen, B. (2013). Deep content-based music recommendation.
Electronics and Information Systems Department (ELIS), 9.
https://doi.org/10.1109/MMUL.2011.34.van
Portugal, I., Alencar, P., & Cowan, D. (2017). The Use of Machine Learning Algorithms in Recommender
Systems: A Systematic Review. Expert Systems with Applications.
https://doi.org/10.1016/j.eswa.2017.12.020
Rumelhart, D. E., Hinton, G. E., & William, R. J. (1985). Learning internal representation by back-
propagation errors. In Nature (Vol. 323, pp. 533–536).
Schmidhuber, J. (2015). Deep Learning in neural networks: An overview. Neural Networks.
https://doi.org/10.1016/j.neunet.2014.09.003
Shen, Y., Chen, J., & Huang, X. (2016). Bidirectional long short-term memory with gated relevance
network for paraphrase identification. In Lecture Notes in Computer Science (including subseries
Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 10102, pp. 39–50).
https://doi.org/10.1007/978-3-319-50496-4_4
Singhal, A., Sinha, P., & Pant, R. (2017). Use of Deep Learning in Modern Recommendation System: A
 Summary of Recent Works. International Journal of Computer Applications, 180(7), 17–22.
https://doi.org/10.5120/ijca2017916055
Srikant, R., & Agrawal, R. (1996). Mining sequential patterns: Generalizations and performance
improvements. Advances in Database Technology — EDBT ’96, 1057, 1–17.
https://doi.org/10.1109/ICDE.1995.380415
Tan, J., Wan, X., & Xiao, J. (2016). A Neural Network Approach to Quote Recommendation in Writings.
Proceedings of the 25th ACM International on Conference on Information and Knowledge
Management - CIKM ’16, 65–74. https://doi.org/10.1145/2983323.2983788
Zhang, S., & Yao, L. (2017). Deep Learning based Recommender System: A Survey and New Perspectives.
ACM J. Comput. Cult. Herit. Article, 1(35), 1–36. https://doi.org/10.1145/nnnnnnn.nnnnnnn
Zhang, Y., Fang, Y., & Weidong, X. (2018). Deep keyphrase generation with a convolutional sequence to
sequence model. In 2017 4th International Conference on Systems and Informatics, ICSAI 2017 (Vol.
2018–Janua, pp. 1477–1485). https://doi.org/10.1109/ICSAI.2017.8248519