Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

1.5em 0pt

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 1 of 23

Algorithmic Exploration in Reading Behavior Analysis and
Recommendations Using Machine Learning
Mahatir Ahmed Tushera , Saket Choudary Kongara b and Gangavarapu Sreeram c
a VIT AP UNIVERSITY, AMARAVATHI, Vijayawada, 522241, Andhra Pradesh, India

ARTICLE INFO ABSTRACT

Keywords: In this study, we look into book recommendation algorithms, which are an important aspect of
Book Recommendations modern online retail and e-commerce systems. Most of the research on book recommendation
Machine Learning systems has relied on cosine similarity or K-Nearest Neighbors (KNN). Nevertheless, these
Cosine Similarity approaches frequently fail to deliver concise and relevant recommendations. Here, we have
Singular Value Decomposition (SVD) combined cosine similarities with Singular Value Decomposition (SVD) using content-based
Hybrid Approach and collaborative filtering strategies to solve this issue. Cosine similarity is an effective strategy
to recognize associated titles based on user ratings, however it neglects to account for all
factors impacting user choices. In order to overcome this, we apply SVD matrix factorization,
which uncovers underlying variables that influence ratings. Our recommender system seeks to
provide recommendations that are more relevant by combining these methods. To evaluate our
performance, we employ measures like Mean absolute error (MAE), recall, precision, Root Mean
Squared Error (RMSE), and Normalized Mean Absolute Error (NMAE). With a precision score
of 0.80, about 80% of our top-k recommendations are relevant, demonstrating the effectiveness
of our system. Our results highlight the possibility of hybrid algorithms for more accurate
recommendations, as they show that SVD and cosine similarity together outperforms systems
depending only on cosine similarity.

1. Introduction
The problem of information overload has emerged in the age of digital transformation due to the amount of data.
This is particularly evident in the realm of book recommendations, where users are often overwhelmed by the vast
array of choices available.The challenge at hand is to offer consumers customized and personal recommendations for
books so they may improve their reading experience and promote a reading culture. A significant number of recent
studies on book recommendation systems has been concentrated on collaborative filtering approaches. By examining
how comparable users have rated products in the past, collaborative filtering makes recommendations for users. Among
the widely used collaborative filtering methods are K-Nearest Neighbors (KNN) and cosine similarity [28] [29]. These
methods, although effective to a certain extent, have their limitations. For instance, cosine similarity and KNN often
struggle with issues of data sparsity and scalability [28].

With the help of Singular Value Decomposition (SVD), we provide in this study a unique method for book
recommendation systems that combines the advantages of matrix factorization and cosine similarity. SVD factors a
matrix by representing it as a product of many smaller matrices. Based on the learnt latent components, it can estimate
missing values, which has shown to be very useful in addressing the sparsity of user ratings data. By integrating
cosine similarity and SVD, we aim to provide more relevant and precise book recommendations. To help explain our
distinctive procedure, we have included Figure 1, which depicts the general structure of the proposed approach.

Our model stands out for the way it is able to effectively handle both item similarity and user choice. Through the use
of cosine similarity, we are able to identify the similarities between different books, which allows for more sophisticated
suggestions. Meanwhile, Singular Value Decomposition (SVD) integration enables us to reveal latent components that
capture abstract concepts that impact a user’s book evaluation. This dual approach improves suggestion accuracy and
provides additional insight into user preferences and the characteristics of the books being evaluated. Essentially, our
∗ Corresponding author
[email protected] (M.A. Tusher); [email protected] (S.C.K. ); [email protected]
(G.S. )
ORCID (s):

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 2 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 1: An overview of the framework for the Hybrid Book Recommendation System integrating Cosine Similarity
alongside SVD.

approach provides an elegant framework that effectively strikes a compromise between user desire and item similarity,
improving the recommendation process as a whole. Our model stands out for the way it is able to effectively handle
both item similarity and user choice. Through the use of cosine similarity, we are able to identify the similarities
between different books, which allows for more sophisticated suggestions. Meanwhile, Singular Value Decomposition
(SVD) integration enables us to reveal latent components that capture abstract concepts that impact a user’s book
evaluation.This two-pronged technique enhances the accuracy of the recommendations while also offering an expanded
awareness of the user’s preferences and the inherent attributes of the books under scrutiny. Essentially, our approach
provides an elegant framework that effectively strikes a compromise between user desire and item similarity, improving
the recommendation process as a whole.

Here are the key contributions of our research:

1. Developed a robust book recommendation system combining cosine similarity and Singular Value Decomposi-
tion (SVD).
2. Effectively addressed data sparsity, a common challenge in recommendation systems, by leveraging SVD.
3. Designed to be scalable, capable of handling large datasets while maintaining accurate recommendations.
4. Provides highly tailored book recommendations based on both item resemblance and individual preferences.
5. Performed an extensive scrutiny of older techniques, indicating superior effectiveness.
6. Integrating cosine similarity and SVD led to significant advancements in book recommendation systems,
ultimately enhancing user reading experiences and fostering a culture of reading.
The study will explore the details of our model, talk about how it is implemented, and provide a comparison
with conventional approaches. Our research will make significant contributions to the field of book recommendation
systems, paving pathways for greater accuracy and customization.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 3 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

1.1. The motivation and purpose of this study

The motivation behind this study is the issue of information overload in the age of digital transformation
and the amount of data that comes with it. It’s common for customers to feel swamped with the sheer quantity
of book recommendation options accessible to them. Providing customers with tailored and individualised book
recommendations to improve their reading experience and encourage a culture of reading is the current issue. Below
are the main reasons for conducting this study:
1. Handle Information Overload: In the digital age, consumers are usually overwhelmed by the vast amount
of book recommendations. The study offers tailored book choices in an attempt to alleviate this information
overload.
2. Improve Recommendation Accuracy: Scalability and data sparsity problems frequently plague existing
techniques like cosine similarity and KNN. This work aims to improve suggestion accuracy by integrating
Singular Value Decomposition (SVD) alongside cosine similarity.
3. Capture Latent factors: The latent factors underlying user evaluations can be effectively captured by SVD.
These latent factors provide a more profound understanding of user preferences since they symbolize fundamen-
tal properties of both the user and the item.
4. Handle User Choice and Item Similarity: The model is notable for its capacity to manage user choice and
item similarity in an efficient manner, enhancing the recommendation process overall.
5. Contribute to the area: By improving users’ reading experiences and encouraging a culture of reading, the
research hopes to significantly advance the area of book recommendation systems.
6. Open the Door for Future Improvements: The study establishes the foundation for subsequent advancements
in the customization and precision of book recommendation algorithms’ selections.
Section 2 offers a comprehensive overview of the literature with a focus on the advantages and disadvantages of
the existing methods. Section 3 provides a detailed description of the technique of the proposed method, including data
preprocessing. In Section 4, we perform an experimental analysis by contrasting the assessment metrics of our model
with the most recent methods. Future scope and limitations were discussed in Section 6 after the discussion in Section
5 is covered. The paper is concluded in Section 7.

2. Related work and Literature survey

The industry of book recommendation systems has seen an extensive number of novel approaches aiming at
improving user experiences and responding to distinct interests. Mathew et al. proposed a very effective book
recommendation system (BRS) that included content-based filtering, collaborative filtering, and mining association
rules [1]. By customizing recommendations based on customers’ interests, this hybrid model increases user satisfaction
and boosts revenue for online book sellers. An opinion mining-based method is introduced by Sohail et al. [2] that
sorts through user reviews, weights features, and ranks books in different computer science fields. This approach helps
consumers find highly rated publications that are relevant to their interests quickly and easily. A hybrid recommender
system with temporal features and demographic factors integrated by Kanetkar et al. [3] goes beyond personalizing.
This method improves customer utility and pleasure by providing personalized recommendations based on variables
like location, gender, and age.

Customized recommendation systems are explored by Sarma et al. [4] through the use of machine learning
techniques.Their method provides more precise and targeted suggestions by employing the K-means Cosine Distance
function to categorize books and the Cosine Similarity function to detect commonalities. In contrast, a comprehensive
four-level recommendation system is presented by Mounika et al. [5]. Their methodology includes sentiment
analysis, collaborative filtering, K-nearest neighbor algorithm, semantic network grouping, and other techniques to
achieve unparalleled accuracy in the least period of time. Sivaramakrishnan et al. [6] explored neighborhood-based
collaborative filtering techniques and used several similarity measures to enhance recommendation accuracy. Their
strategy, based on local algorithms, seeks to improve users’ perceptions of the relevancy of book recommendations.
Zhou et al. [7] describe an innovative information recommendation system utilizing an enhanced Apriori algorithm.
This system offers effective book suggestions with little computational overhead by getting around performance issues
with conventional approaches.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 4 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Chen et al. [8] addresses the integration of mobile technology in library settings. Problem-based learning in actual
library settings is made easier by their Intelligent Mobile Location-Aware Book Recommendation System (IMLBRS).
By means of guidance and suggestion features based on maps, students are able to efficiently navigate the library space
and identify pertinent resources. Kuroiwa et al. [9] present a dynamic Book Utilization System (BUS) that makes use
of virtual library improvements and web services. By improving tailored suggestions and enabling user sharing, this
system encourages book aficionados to work together as a community.

Jia et al. [10] concentrated their investigation into the usage of recommendation systems in library management
on content and collaborative filtering procedures. This paper covers the architecture and functionalities of a
recommendation system customized for library contexts with the goal of improving information retrieval efficacy.
A hybrid recommender system based on characteristics and personality is put out by Hariadi et al. [11]. Their
technology generates personalized book suggestions by incorporating user personality aspects to the recommendation
process in addition to attribute-based methods. Rahutomo et al. [12] look on how online businesses provide book
recommendations. Their study emphasizes the value of user-generated evaluations in cooperative filtering techniques
and the part embedding models play in improving suggestion accuracy. By combining these several techniques, book
recommendation systems can be made more adaptable and accommodate a greater variety of user preferences and
demands. To solve the issue of an excessive amount of product selection on e-commerce platforms, Sharma et al.
[13] developed a recommendation model which is hybrid that combines collaborative-based filtering with content-
based filtering. A hybrid system for recommendation based on patterns and using semantic links to enhance book
recommendations was presented by Wayesa et al. [14]. In order to enhance top-N suggestions, Xin et al. [15] concentrate
on community detection algorithms, especially in academic library contexts.

Mariana et al. [16] use association rule mining to create a powerful online public access catalog (OPAC)
recommendation system. In order to successfully handle sparse data, Desai et al. [17] describe an enterprise-friendly
recommendation system based on biclustering.In order to generate personalized book suggestions, Tewari et al. [18]
presented a comprehensive recommendation system that makes use of content filtering, collaborative filtering, and
association rule mining. Using classification and opinion mining techniques, Tewari et al. [19] propose books to
reduce the amount of information available on e-commerce platforms. In order to protect user privacy while providing
tailored suggestions, Luo, Le, and Chen [20] put special emphasis on privacy issues raised by users and suggest
the Privacy-Preserving Book Recommendation System (PPBRS). Sariki et al. [21] proposed an enhanced framework
for book recommendation accuracy that incorporates modules for Named Entity Recognition (NER), stylometry and
visual feature extraction.A hybrid recommendation engine called NOVA is introduced by Pathak et al. [22] in order
to offer effective and efficient book recommendations. Personalized suggestions are highlighted in Zhu et al.’s [23]
investigation of collaborative filtering-based book recommendation algorithms. Lastly, based on rankings for college
libraries Verma et al. [24] offer a hybrid recommendation system that exhibits a significant improvement in suggestion
accuracy over earlier techniques.

The literature study includes an extensive spectrum of innovative techniques for book recommendation systems.
A number of approaches are looked into, including opinion mining-based systems, hybrid models that integrate
collaborative and content-based filtering, and temporal considerations, demographic information, and machine learning
algorithms-based personalized recommendation systems. Studies also explore mobile technology integration in library
settings, neighborhood-based collaborative filtering, and better algorithms like the improved Apriori algorithm.
Additionally, studies show how important personality features, community detection methods, semantic linkages, and
user-generated ratings are to improving suggestion accuracy. When considered collectively, these studies significantly
improve the subject of book recommendation systems by complying with the requirements and interests of a diverse
spectrum of users across several platforms and in a variety of situations. Table 1 is a comparative table of the literature
survey.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 5 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Table 1: Comparative analysis of systems for recommendation.

Paper Advantages Disadvantages Year

Mathew et al [1] - A hybrid model that gener- - Complexity associated with 2016
ates customized suggestions integration of multiple tech-
by integrating collaborative niques may pose implementa-
filtering,content-based filter- tion challenges
ing, and association rule min-
ing.
Sohail et al.[2] - Utilizes opinion mining for - Dependency on user re- 2013
ranking books based on user views may introduce biases
reviews, aiding users in dis- and inaccuracies in recom-
covering top-ranked books mendations
efficiently
Kanetkar et al. [3] - Integrates temporal aspects - Collection and management 2014
and demographic parameters of demographic data may
into recommendation system raise privacy concerns and re-
for personalized recommen- quire user consent
dations
Sarma et al. [4] - Leverages machine learning - Requires robust data prepro- 2021
algorithms for personalized cessing and feature engineer-
recommendations, enhancing ing to ensure effectiveness of
accuracy through clustering machine learning algorithms
and similarity metrics
Mounika et al. [5] - Offers a comprehensive - Integration of multiple al- 2021
recommendation system with gorithms may increase com-
multiple levels of processing, putational overhead and com-
aiming for unparalleled plexity of the system
accuracy within a minimal
time frame
Sivaramakrishnan et al. - Explores neighborhood- - Performance of neighbor- 2018
[6] based collaborative filtering hood algorithms may degrade
techniques to refine relevance with large datasets, requiring
of book recommendations optimization
for users
Zhou et al. [7] - Introduces an efficient infor- - Adoption of Apriori algo- 2020
mation recommendation sys- rithm may limit scalability
tem based on an improved and adaptability of the system
Apriori algorithm, reducing
computational overhead
Chen et al.[8] - Integrates mobile - Reliance on mobile technol- 2013
technology into library ogy may introduce accessibil-
environments, enhancing ity barriers for users with lim-
user experience and ited access to smartphones
navigation through map-
based guidance
Continued on next page

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 6 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Table 1 – continued from previous page

Paper Advantages Disadvantages Year
Kuroiwa et al. [9] - Introduces a dynamic Book - Dependence on web ser- 2007
Utilization System (BUS) vices may introduce vulnera-
leveraging web services and bilities and security risks in
virtual library enhancements the system
for personalized
recommendations
Jia et al. [10] - Collaboration and content- - Adoption of System 2013
based filtering tools, along for Recommendation
with software systems may require significant
specifically designed for infrastructure and resource
library management, can investment
improve the efficiency of
information retrieval.
Hariadi et al.[11] - Proposed a hybrid attribute - Integration of personality 2017
and personality-based recom- traits into recommendation
mender system yielding su- process may raise privacy
perior results in personalized concerns and require user
book recommendations consent
Rahutomo et al. [12] - Emphasizes user-generated -Handling large volumes of 2019
ratings in collaborative filter- user data may require signifi-
ing. cant resources.
Sharma et al.[13] - Alleviates product overload - May still face limitations 2022
on e-commerce platforms; - ofContent-Based Filtering
Integrates Collaborative and and Collaborative Filtering
Content-Based Filtering approaches
Wayesa et al. [14] - Enhances book recommen- - Semantic relationships may 2023
dations using semantic rela- not always capture nuanced
tionships; - Utilizes a pattern- user preferences effectively
based hybrid system
Xin et al. [15] - Improves top-N recommen- - Community detection may 2014
dations, particularly in aca- be computationally intensive
demic library; - Utilizes com- and require large datasets
munity detection techniques
Mariana et al. [16] - Builds an effective recom- - Association rule mining 2017
mendation system for an On- may not capture complex user
line Public Access Catalog preferences accurately
(OPAC); - Utilizes associa-
tion rule mining techniques
Desai et al. [17] - Offers an enterprise- - Biclustering may require 2016
friendly recommendation significant computational re-
system; - Addresses sparse sources and may not be suit-
data effectively using able for all datasets
biclustering
Tewari et al. [18] - Provides personalized - Integration of 2014
book recommendations by multiple techniques may
integrating content filtering, increase complexity and
association rule mining, and computational overhead
Collaborative Filtering
Continued on next page

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 7 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Table 1 – continued from previous page

Paper Advantages Disadvantages Year
Tewari et al. [19] - Addresses information - Opinion mining may not al- 2014
overload on e-commerce ways accurately capture user
platforms; - Leverages sentiments and preferences
opinion mining and
classification techniques
for recommendations
Luo et al. [20] - Safeguards user privacy - Privacy-preserving 2009
while offering personalized techniques may
recommendations; - add complexity and
Addresses user privacy computational overhead
concerns to the recommendation
process
Sariki et al. [21] - Improves recommendation - Enhanced framework may 2022
accuracy using an enhanced require significant computa-
framework; - Utilizes Named tional resources and may not
Entity Recognition, visual always capture user prefer-
feature extraction, and ences accurately
stylometry
Pathak et al. [22] - Provides efficient and effec- - Hybrid recommendation en- 2013
tive book recommendations gines may require careful
using NOVA, a hybrid rec- tuning and optimization to
ommendation engine achieve optimal performance
Zhu et al. [23] - Emphasizes personalized - Collaborative filtering may 2016
recommendations using face challenges in capturing
collaborative filtering diverse user preferences ac-
algorithms curately
Verma et al. [24] - Demonstrates significant - Implementation of the pro- 2024
improvement in posed system may require
recommendation accuracy adaptation to different library
for college libraries; - environments and may not
Utilizes a ranking-based generalize well to all contexts
hybrid recommendation
system

3. Methodology
This research paper’s methodology section outlines the strategy and methods used to accomplish our objective of
creating a trustworthy book recommendation system. We employ a variety of methods, including collaborative filtering
and matrix factorization, and we ground our strategy in data science and machine learning concepts. First, we describe
the steps involved in prepping the data: loading the datasets, addressing missing values, and integrating the data. Next,
we explore the basis of our recommendation system, which is composed of two primary phases: Matrix Factorization
and Collaborative Filtering.

Using cosine similarity, we may discover comparable users based on how they have rated books during the
Collaborative Filtering step. In the Matrix Factorization stage, the latent elements underlying user evaluations are
extracted using Singular Value Decomposition (SVD). The subsequent subsections provide a detailed explanation of
each of these phases as well as the underlying mathematical reasoning. We also go over the assessment metrics that
we utilized to evaluate our model’s performance.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 8 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 2: Distribution of Book Ratings and Top 30 Users Based on Number of Ratings Given.

3.1. Data Preprocessing

3.1.1. Datasets
Books.csv: This dataset contains details about books that may be found using their unique ISBNs. Content-based
information from Amazon Web Services, such as publisher, year of publication, author, and title, is included. There
are additional URLs to cover photos that go to Amazon's website in small, medium, and large versions.
Ratings.csv: This dataset includes explicit feedback ratings for books. Higher ratings, on a scale of 1 to 10, indicate
greater appreciation. Users' implicit ratings, which indicate which books are rated and which are not, are represented
by the number 0.
Users.csv: This dataset provides user data that has been anonymized and mapped to numbers in order to protect
user privacy. If accessible, it contains demographic information such as age and location, which helps with tailored
suggestions.

3.1.2. Initial Data Exploration and Handling Missing Values

We carry out a preliminary study to comprehend the data's organization after data loading. A data frame’s number
of rows and columns are determined by the shape parameter, and the head() method is used to print the data frame’s
first few rows. This gives us information on the size and appearance of the data. Next, we use the isnull().sum() function
to search for missing values in our datasets; this returns the number of missing values for each column. Depending
on how many and what kind of missing data were discovered, we may decide to eliminate the columns or rows that
contain the missing values or replace them with a certain value (such the mean or median).

3.1.3. Data Integration and Refinement

Our three data frames are first merged using the merge() function to expedite the data preparation procedure.
First, we merge ratings with books according to their "ISBN" column. Then, we combine the resulting dataframe
with users according to their "User-ID." The result of this consolidation is a single dataframe that contains all the
information required by our recommendation algorithm. We next perform data transformation on our merged dataframe
by optimizing the 'Location' column. This involves removing the last section, which is separated by commas, from the
"Location" string in order to isolate and save only the country information. Our dataset is made simpler and less
dimensional by this improvement, which makes analysis and modeling more effective. Two key visualizations are
included in Figure 2, a bar plot displaying the top 30 most active users by rating count, which is vital for assessing
user behavior and enhancing the effectiveness of our recommendation system, and a rating distribution histogram for
comprehending user preferences.

3.2. Filtering and Pivot Table Creation

Our data preprocessing pipeline starts with Filtration of Data, involving people who have rated over 200 books
and novels with more than 50 ratings. By fostering recommendations based on solid data subsets, this curation seeks

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 9 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 3: Heatmap of User-Item Matrix for Top 100 Users and Books

to reduce data sparsity. We utilize groupby() and count() routines to count ratings for each user and book, then filter
based on the results.
After data filtering, we create a pivot table, which results in an organized matrix with rows denoting book titles and
columns denoting users. User-book ratings are wrapped within this framework to facilitate analysis. For dealing with
missing data, we use the fillna() function to create a dense matrix that improves the efficiency of our recommendation
system. The data preparation approach ensures our recommendation engine’s accuracy and reliability. We maximize
system speed by carefully preprocessing, providing users with relevant and trustworthy book recommendations. Figure
3 displays a heatmap of the user-item matrix for the top 100 users and their books. The heatmap displays user ratings
for each book. Darker hues indicate higher ratings, and the color intensity in each cell reflects the rating value.

3.3. Model Description

We provide a highly personalized book recommendation system based on a sophisticated framework that combines
the features of both matrix factorization and collaborative filtering. It starts with a collaborative user-to-user filtering
method that finds users who are similar to each other based on how they rate books. The cosine similarity measure
is used to assess the similarity. The system then use a popular matrix factorization technique called Singular Value
Decomposition (SVD) to find the latent components that underlie user assessments. Such implicit components,
representing both the user and the object, provide a deeper understanding of preferences of users.
By combining these methods, our system is able to manage user preference and item similarity in an efficient
manner, which enhances the relevance and accuracy of the suggestions.The system is made to be scalable, meaning it
can manage big datasets and yet function well. Figure 4 illustrates the proposed system's comprehensive architecture.

3.3.1. Collaborative Filtering

During the initial phase, the system utilizes a collaborative filtering approach between users. By comparing users
to one another, the cosine similarity method pairs people who have comparable rating behaviors. By measuring the
ratings of two users, two vectors create an angle that this metric calculates the cosine of. The cosine similarity measure
ranges from 0 to 1, with 0 representing total dissimilarity and 1 indicating user equivalence. To find out how similar
two users, u and v, are,
∑ we apply the cosine similarity metric:
𝑟𝑢𝑖 𝑟𝑣𝑖
𝑠𝑖𝑚(𝑢, 𝑣) = √∑ 𝑖∈𝐼 √ ∑ (1)
2 2
𝑖∈𝐼 𝑟𝑢𝑖 𝑖∈𝐼 𝑟𝑣𝑖
In this scenario, for item i, rui represents rating of "u" user, whereas I is the collection of objects rated by both u
and v. After identifying comparable users, the algorithm suggests products they have liked. The recommend method
returns a list of books that match the input book name.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 10 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 4: Comprehensive architecture of the proposed book recommendation system.

3.3.2. Matrix Factorization (SVD)

During the second phase, the Singular Value Decomposition (SVD) matrix factorization is used by the system to
provide recommendations. SVD can be used to factor a user-item rating matrix, resulting in a diagonal matrix, two
orthogonal matrices with more compact dimensions, and more. The SVD extracts the latent elements underlying user
ratings while reducing the dimension of the original data. The underlying user and item properties are represented by
these latent components.Books may have latent aspects such as author styles, duration, and genres. The SVD algorithm
splits the user-item rating matrix R into a diagonal matrix and two orthogonal matrices of lesser dimension:

𝑅 = 𝑈 Σ𝑉 𝑇 (2)

Where, Σ is the singular value diagonal matrix, which is essentially a weight matrix, U indicates the left singular
vectors (user "features" matrix), R is the user-item rating matrix and the right singular vectors are denoted by VT (item
"features" matrix). Users’ ratings for items j are implied by each element rij in R and R is divided into U, Σ, and VT by
the SVD in the following ways:
∑𝑛
𝑟𝑖𝑗 ≈ 𝑇
𝑘=1 𝑢𝑖𝑘 𝜎𝑘 𝑣𝑘𝑗 (3)

Where, in the kth column and ith row of U, the element is denoted by uik , k is the kth singular value in , and
𝑣𝑇𝑘𝑗is the element in the kth row and jth column of VT . All n latent factors are included in the sum. The rating rij is
roughly represented by this equation, which is the product of user features, singular values (weights), and item features.
This expresses how each latent factor affects the rating. Figure 5 illustrates matrix factorization, a crucial step in our
suggested design. On the right is the reduced-dimensional version of the original user-item interaction matrix, which
reveals latent variables driving user-item engagement.

3.4. Algorithm Implementation

Our proposition is an innovative algorithm for book recommendations that utilizes Singular Value Decomposition
(SVD) in conjunction with Cosine Similarity.The algorithm starts with randomly picked user and item profiles and
makes iterative adjustments to them in an effort to minimize the discrepancy between actual and anticipated ratings. To
handle data sparsity, it applies SVD to the user-item interaction matrix after each update, keeping just the top 'k' singular

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 11 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 5: Demonstration of Matrix Factorization in Recommendation Systems

values. After that, it uses Cosine Similarity to generate an item-by-item similarity matrix, which is then used to suggest
items to consumers. This adaptive and versatile method guarantees customized and relevant suggestions, enhancing
the reader's experience. Algorithm 1 presents an algorithmic representation of the proposed recommendation model.

Algorithm 1 Book Recommendation System

Require: User-Item Interaction Matrix 𝑅, User Profile Matrix 𝑈 , Item Profile Matrix 𝐼
Ensure : Updated User-Item Interaction Matrix 𝑅′
1 Training:
Initialize User Profile Matrix 𝑈 and Item Profile Matrix 𝐼 randomly
Compute the initial User-Item Interaction Matrix 𝑅 = 𝑈 × 𝐼 𝑇 for each epoch do
2 for each user 𝑢 do
3 for each item 𝑖 rated by user 𝑢 do
4 Compute the error 𝑒𝑢𝑖 = 𝑟𝑢𝑖 − 𝑢𝑖 × 𝑖𝑇𝑢
Update user and item profiles:
𝑢𝑖 ← 𝑢𝑖 + 𝛾 × (𝑒𝑢𝑖 × 𝑖𝑢 − 𝜆 × 𝑢𝑖 )
𝑖𝑢 ← 𝑖𝑢 + 𝛾 × (𝑒𝑢𝑖 × 𝑢𝑖 − 𝜆 × 𝑖𝑢 )
5 end
6 end
7 Calculate the revised User-Item Interaction Matrix’s SVD. 𝑅
Keep 𝑘 singular values to get the approximated interaction matrix 𝑅′ = 𝑈𝑘 × Σ𝑘 × 𝐼𝑘𝑇
Compute item-item similarity matrix 𝑆 = cosine_similarity(𝐼𝑘 )
for each user 𝑢 do
8 Compute the prediction score for unrated items Recommend the top-𝑁 items with the highest prediction
scores
9 end
10 end

3.5. Model Training

Our book recommendation system is based on two key algorithms: Cosine Similarity and Singular Value
Decomposition (SVD). To implement it, we made a pivot table and utilized cosine similarity. The rows of the table
represent books, the columns, users, and the cells, ratings. In order to address the issue of missing data, we have
completed the gaps in this user-item matrix. We have then computed cosine similarity scores between novels based
on these ratings. Our recommend feature uses these similarity scores to identify books that are most similar to a given
book. The SVD component is the setting in which we have utilized the surprise library. Using Surprise’s dataset format,
we first split the data into a training set and a testing set. The SVD method is then defined and trained using the training
set of data. After training, the model makes predictions on the test set. We adopt a hybrid strategy that combines matrix
factorization (using SVD) with item-item collaborative filtering (using cosine similarity). This enables us to offer more

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 12 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

accurate and tailored recommendations for books by utilizing the benefits of both approaches. Our model is designed
to learn from users' past behavior (rated books) and use this data to forecast users' future behavior (book choices they
would enjoy). Table 1 is a list of the hyperparameters that our book’s recommendation system employed. To improve
model performance, these parameters were chosen using grid search and cross-validation. The regularization term
(reg_all), learning rate (lr_all), and number of iterations (n_epochs) are all included.

Hyperparameters Values
Frequency of episodes(n_epochs) 20
Rate of learning(lr_all) 5 × 10−3
Regularization term (reg_all) ∼0.4

Table 1: Optimal Hyperparameters for the SVD Model in the Book Recommendation System.

4. Experimental Analysis
We assess the performance of our book suggestion system in detail in this section. Reasoning, parameters, and
algorithms are all included in the methodology explanation. Metrics like RMSE, MAE, NMAE, Precision at k, and
Recall at k are defined for assessing the findings. Results are presented in an understandable manner using tables and
graphs. Expectations that were met, unexpected results, and parallels with earlier research are all discussed. We talk
about constraints affecting findings and possible explanations for variations in results. The results provide directions
for future investigation.

4.1. Evaluation Matrics

We combine error measures with error metrics to create assessment metrics. Three error metrics are used to quantify
this difference between the actual predicted results: Root Mean Squared Error (RMSE), Normalized Mean Absolute
Error (NMAE), and Mean Absolute Error (MAE).Two evaluation standards that assess the caliber of the model’s top-K
suggestions are Recall at K and Precision at K. Together, these measures offer a thorough assessment of our model’s
effectiveness, accounting for the accuracy of the predicted ratings as well as the relevancy of the recommended books.

4.1.1. Root Mean Square Error (RMSE)

The Root Mean Square Error (RMSE) statistic can be used to determine the average difference between the
predicted and actual values in a given dataset. RMSE can be calculated by squaring, adding, and dividing the difference
between the known and unknown points by the total number of test points. This is how we can mathematically express
it: .
√
∑𝑁
𝑦(𝑖)∥2
∥𝑦(𝑖)−̂
𝑅𝑀𝑆𝐸 = 𝑖=1
𝑁
(4)

Here, y(i)= ith measurement, ŷ(i)= ith prediction, N= Number of data points

4.1.2. Mean Absolute Error (MAE)

In many domains, such as statistics and machine learning, assessing how well a model’s predictions correspond with
actual data is essential. For this kind of evaluation, Mean Absolute Error (MAE) is an invaluable tool. MAE measures
how much the actual and anticipated values differ from one another. It ignores the direction of the errors (positive or
negative), in contrast to certain other error measures. In terms of math, MAE is represented as:
∑𝑛
𝑎𝑏𝑠(𝑦𝑖 −𝜆(𝑥𝑖 ))
MAE = 𝑖=1
𝑛
(5)

Here, 𝜆(𝑥i ) = 𝑖th 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛, 𝑦i = 𝑖th 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑚𝑒𝑛𝑡, 𝑛 = 𝑑𝑎𝑡𝑎𝑝𝑜𝑖𝑛𝑡𝑠′ 𝑞𝑢𝑎𝑛𝑡𝑖𝑡𝑦.

4.1.3. Normalised Mean Absolute Error (NMAE)

The median number of mistakes in a series of forecasts is determined using the mean absolute error matrices, which
do not take the direction of the errors into consideration. The absolute inaccuracy of every projection is the discrepancy

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 13 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

between the actual and expected data. The Normalized Mean Absolute Error (NMAE) can be obtained by taking the
mean of the absolute errors.

1 ∑𝑁
𝑁 𝑖=1 |𝑦𝑖 − 𝑦̂𝑖 |
NMAE = (6)
max(𝑦) − min(𝑦)
𝑁 is the number of predictions. 𝑦𝑖 is the actual value. 𝑦̂𝑖 is the predicted value. max(𝑦) and min(𝑦) are the maximum
and minimum values in the actual data, respectively. NMAE provides a normalized measure of the error of a regression
mode. It helps assess how well the model predicts the target variable relative to its scale. NMAE is useful in scenarios
where the scale of the target variables varies significantly or when comparing models trained on different datasets. By
normalizing the error relative to the range of the target variable, NMAE provides a standardized measure of prediction
accuracy.

4.1.4. Precision and Recall

The effectiveness of classification algorithms is evaluated using two key measures, precision and recall, especially
in the context of recommendation systems. By dividing the total number of expected positives by the ratio of correctly
predicted positive observations, one can calculate the accuracy of the model’s positive predictions. A low false-positive
rate for the model is indicative of good accuracy. By dividing the total number of accurately predicted positive
observations by the total number of actual positive observations, recall evaluates the model’s ability to identify all
pertinent events. It is also known as true positive rate or sensitivity at times. A high recall rate in the model implies a
low false-negative rate. Both precision and recall are crucial variables to consider when evaluating recommendation
systems, as they ensure that most recommendations are relevant. It’s important to keep these aspects in balance because
improving one usually means sacrificing the other. One popular method for striking this equilibrium is the F1 score,
or harmonic mean of precision and recall. The following is the precision and recall scores’ mathematical definition:

𝑇𝑃
Precision = (7)
𝑇𝑃 + 𝐹𝑃
𝑇𝑃
Recall = (8)
𝑇𝑃 + 𝐹𝑁
Where, 𝑇 𝑃 (True Positives) is the number of correct positive predictions. 𝐹 𝑃 (False Positives) is the number of
incorrect positive predictions. 𝐹 𝑁 (False Negatives) is the number of actual positives that were not identified by the
model.

4.2. Experimental Setup and Configuration

We go into depth on the design and setup that we employed in this section of the study. The environment, hardware
specs, software and libraries, dataset, algorithms and models, and assessment measures are all included in this. Here
is a detailed summary of our experimental setup in Table 2.

Category Details
Environment Python,Google Colab (a Jupyter notebook environment hosted on the cloud)
Hardware Specifications Core i5-1135G7, RAM8GB, GPU-NVIDIA GeForce MX330 with 2GB GDDR5
Software and Libraries Python, pandas, NumPy, scikit-learn, scikit-surprise, matplotlib, seaborn
Dataset Book.csv, Ratings.csv, Users.csv from Goodreads
Algorithms and Models Collaborative Filtering (using Cosine Similarity), Matrix Factorization (using Singular
Value Decomposition (SVD)), Hyperparameter tuning (using GridSearchCV)
Evaluation Metrics RMSE, MAE, NMAE, Precision at k, Recall at k

Table 2: Overview of the Experimental Setup and Configuration for the Book Recommendation System.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 14 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

4.2.1. Evaluation of the System

We have used several evaluation matrics to evaluate the performance of our approach. The findings are demon-
strated in the Table 3. Better accuracy is indicated by lower MAE values; the range is 0 to infinity. Values smaller
than one in the RMAE scale, which goes from 0 to infinity, indicate superior performance over the benchmark. NMAE
normally has a range of 0 to 100 or 0 to 1, with values nearer 0 denoting greater accuracy. A higher percentage of actual
positive forecasts among all positive predictions is shown by higher values on the accuracy scale, which runs from 0 to
1. Greater recall values signify a greater proportion of precise positive predictions among all actual positives. Recall
also ranges from 0 to 1.

Assessment Measure Value

RMSE 0.5035
MAE 0.6747
Precision at k 0.80
NMAE 0.278
Recall at k 0.60

Table 3: The book recommendation system's performance

The performance of prediction algorithms may be compared on different scales with the help of this normalized
metric. Our computed accuracy at k is 0.80, which indicates that 80% of the top k books that are suggested to the
user are pertinent. Given a recall of 0.60 at k, we may deduce that 60% of the pertinent books are included in the
top-k recommendations. These results corroborate each other and show how well our hybrid recommendation system
works to provide accurate book recommendations. In order to enhance comprehension and facilitate comparison of
the performance metrics of our book recommendation system, we have employed bar charts (Figure 6) and radar
charts (Figure 7) to visually represent the data. Figure 6 displays the values of a number of performance metrics,
including RMSE, Precision at k, MAE, NMAE, and Recall at k. While greater values of Precision and Recall imply
better performance, lower values of RMSE and MAE indicate higher accuracy. A normalized comparison of these
measurements is shown in Figure 7, where each axis corresponds to a distinct statistic. The radar chart’s full area
represents the system’s performance across all criteria; a larger area denotes higher overall performance. Lastly, we
have examined the correlation between the anticipated and actual book ratings using the scatter plot displayed in
Figure 8. Each point represents a book's rating, where y is the predicted rating and x is the actual rating. For flawless
predictions, points should ideally line up along the diagonal from bottom left to top right.

4.2.2. Comparison with State-of-the-art Methods

Our hybrid book recommendation system has been compared and contrasted with a number of different research
projects. More specifically, we have chosen to compare our work with cosine similarity-based recommendation systems
for books, independent of the model. A variety of metrics were looked at, such as normalized mean absolute error
(NMAE), precision, root mean square error (RMSE), recall and mean absolute error (MAE). We have chosen a number
of studies for this comparison, including those by Rohit et al. [25], Amer et al. [26], Anwar et al. [27], and Monika
et al. [24]. These studies offer a solid foundation for comparison since they cover a wide range of recommendation
systems techniques and methodology. The comparison has been shown in Table 4.

Models MAE RMSE NMAE Precision Recall

Rohit et al. [25] 2.631 2.99 0.292 - -
Amer et al. [26] 0.77 - - 0.0316 0.2045
Anwar et al. [27] 0.72 0.92 - - -
Monika et al. [24] 0.074 1.0246 - 0.6834 -
Proposed Model 0.6747 0.5035 0.278 0.80 0.60
Table 4: Comparison with the existing mehods.

This assessment is meant to demonstrate our hybrid recommendation system’s precision and efficiency.We can
more clearly grasp the advantages of our strategy and pinpoint areas in need of future development by contrasting

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 15 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 6: Bar Chart of Performance Matrics

our findings with those of these earlier investigations. The portions of this article that follow will provide a thorough
comparison and debate.

4.2.3. Comparative Performance Analysis of Recommendation Models Using MAE and RMSE
In this part, we have evaluated our hybrid recommendation system’s performance against recommendation models
that used Singular Value Decomposition (SVD) or K-Nearest Neighbors (KNN) to see how effective it is. We focused
on Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) because they are commonly used metrics to
assess the accuracy of predictive models. We have selected the following studies for the KNN comparison: The studies
of Anagha et al.[32], Krishnan et al.[35], Li et al.[36], and Anwer et al.[27] have been selected for the SVD comparison.
In particular, we have taken the average RMSE and average MAE of three datasets (Hindi Movie, Book Cross, and
Movielens) for Krishnan et al. [35]. Table 5 shows this comparison. Nguyen et al. [31], Anagha et al. [32], Rohit et al.
[25], Esmael et al. [33], S. G. K. Patro et al. [34], and Anwer et al. [27]. This comparison has been demonstrated in
Table 6.

The outcomes showed that the hybrid model—which integrates cosine similarity and SVD—performed better
in both criteria. Our hybrid model efficiently tackles the drawbacks of employing KNN or SVD alone by utilizing
cosine similarity’s connection strengths to find underlying variables and produce more accurate and pertinent book
recommendations.

Models RMSE MAE

Anagha et al. [32] 1.4346 1.2152
Krishnan et al. [35] 0.88 0.68
Li et al. [36] 0.9 -
Anwer et al. [27] 0.979 0.7732
The proposed model 0.5035 0.6747
Table 5: Comparative Analysis of Performance Metrics (MAE and RMSE) for SVD based Models and our proposed
method.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 16 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 7: Radar Chart of Performance Metrics

Models RMSE MAE

Nguyen et al. [31] 1.087 0.856
Anagha et al. [32] 1.4346 1.2152
Rohit et al. [25] 2.99 2.631
Ahmed et al. [33] 1.0535 -
S. G. K. Patro et al. [34] 0.73 0.7165
Anwer et al. [27] 0.979 0.7732
The proposed model 0.5035 0.6747
Table 6: Comparative Analysis of Performance Metrics (MAE and RMSE) for KNN-based Models and the Proposed
Hybrid Model..

4.2.4. Recommendation Process

The matrix factorization and collaborative filtering approaches are combined in our recommendation system. User
and book ratings are amalgamated, missing values are filled in, and users with more than 200 ratings and books with
more than 50 ratings are removed throughout the data preparation process. Next, a pivot table showing user ratings is
built, with rows for each book and columns for each user. The zeros in missing ratings are used.

The cosine similarity between novels is then determined by taking into account each reader’s rating. To suggest
books that are related to a certain book, we utilize the similarity score matrix that has been constructed. The suggest
function organizes a list of relevant books based on similarity scores after obtaining a book title as input. We
simultaneously use the Surprise package’s SVD algorithm for forecasting. There are two kinds of data: training and test
sets. The SVD model is trained using the training set, and predictions are made on the test set. The model is evaluated
using a variety of measures, including RMSE, MAE, NMAE, accuracy at k, and recall at k.

The top ten books that have been recommended for two users (User IDs: 2313 and 276727) are shown in Figures
9 and 10. The "recommend books" feature generates these recommendations by predicting the user’s ratings for all
books they haven’t yet rated, arranging those forecasts in a decreasing order, and then going back to the top ten books.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 17 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 8: Scatter plot of Actual vs. Predicted Book Ratings

Figure 9: Top 10 Books Suggested Reading for "User 2313"

The user gets personalized recommendations determined by their previous ratings as well as the ratings of other people
who are similar to them. These serve as a stand-in for the novels that our model anticipates readers would like highly.
By providing these customized suggestions, our algorithm assists users in discovering new books that are relevant to
their interests. In summary, our hybrid recommendation system uses both item similarity and user behavior to provide
tailored book suggestions.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 18 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 10: Top 10 Books Suggested Reading for "User 276727"

5. Discussion
Our hybrid recommendation system, which fused collaborative filtering and matrix factorization, produced good
results. Recommendations for books were made by the algorithm using both user-equivalent ratings and ratings from
prior interactions. Performance metrics such as MAE, NMAE, RMSE, recall at k, accuracy at k, and others were
employed to evaluate the system. With an RMSE of 0.5035 and an MAE of 0.6747, the system can make reasonably
accurate predictions. It is crucial to remember that such metrics provide only a limited overview of the system’s
functionality. For instance, the recall at k of 0.60 implies that there might be relevant books that are removed from the
top-k recommendations, even though the precision at k of 0.80 indicates that a significant portion of the top k books
recommended are relevant to the reader. This points up a possible place where our system needs to be improved.

The scatter plot of actual vs. predicted book ratings visually represented the relationship between the ratings.
Every point on the plot denotes a book rating; the y-coordinate displays the rating that our model predicted, and the
x-coordinate shows the actual rating provided by a user. This visual tool facilitates comprehension of our model’s
performance. Moreover, the top 10 book suggestions that the system generated for a particular user (User ID: 276727)
show how personalized it is.These recommendations, which are books the user is likely to rank highly based on our
algorithm, show that they are pertinent to the user’s interests. Lastly, our study has shown that a hybrid recommendation
system may effectively provide personalized book recommendations.

Nevertheless, there are lots of chances for growth and more research. Further research and development will focus
on enhancing the model, exploring other recommendation methodologies, and incorporating additional features to
increase the amount of customization offered by the suggestions. An illustration of how each factor—RMSE, MAE,
Precision at K, and Recall at K—affects the system’s performance is given in the pie chart in Figure 9. Readers will
have a better idea of the aspects that determine the accuracy of our book suggestions by looking at this chart.

6. Future Scope and Limitations

We have shown that our hybrid recommendation system performs promisingly when it comes to book recom-
mendations using a blend of collaborative filtering and matrix factorization approaches. Still, there are a number
of directions that future study may go in order to improve our system’s functionality and performance. The "cold
start" problem—which describes the challenge of proposing new users or things in the lack of past data—is one of
these; evaluating long-term user satisfaction via longitudinal studies; guaranteeing the scalability and efficacy of the
recommendation system as the dataset grows; and investigating the applicability of our recommendation system to
different item types, such as products, music, or movies. Additionally, these include exploring other recommendation
algorithms or ensemble methods to improve model performance and incorporating additional features like book genres,
author information, or user demographic data. We think that by investigating these avenues, future studies can further
the area of recommendation systems and provide consumers with more tailored and fulfilling experiences.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 19 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Figure 11: The relative contributions of several factors to the recommendation system’s overall performance

7. Conclusion
We summarize the main conclusions and future directions of our research in the closing remarks. With the use
of matrix factorization and collaborative filtering, our team has successfully developed a hybrid recommendation
system that offers tailored book recommendations. Metrics like RMSE (0.5035) and MAE (0.78), which measure the
system’s performance, show a respectable degree of accuracy. Although our approach has yielded encouraging results,
we recognize that it may be improved. To guarantee that more pertinent books are included in the top-k suggestions,
possible areas for improvement are indicated by the accuracy at k of 0.80 and recall at k of 0.60. Our work advances the
continuous creation of efficient and customized recommendation systems. In order to improve personalization in future
work, we plan to investigate other recommendation strategies, add new features, and further develop our model. This
dedication to ongoing development highlights our goal of providing very precise and customized book suggestions.

Credit Authorship Contribution Statement

Mahatir Ahmed Tusher: Conceptualization, Methodology, Software, Writing - Original Draft (lead), Writing -
Review Editing. Saket Choudary Kongara: Writing - Review Editing, Validation. Gangavarapu Sreeram: Writing
- Review Editing, Validation.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Data Availability
Data are available upon request.

Acknowledgments
No acknowledgments.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 20 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

Reference
1. P. Mathew, B. Kuriakose and V. Hegde, "Book Recommendation System through content based and collaborative
filtering method," 2016 International Conference on Data Mining and Advanced Computing (SAPIENCE),
Ernakulam, India, 2016, pp. 47-52, doi: 10.1109/SAPIENCE.2016.7684166.
2. S. S. Sohail, J. Siddiqui and R. Ali, "Book recommendation system using opinion mining technique," 2013
International Conference on Advances in Computing, Communications and Informatics (ICACCI), Mysore,
India, 2013, pp. 1609-1614, doi: 10.1109/ICACCI.2013.6637421.
3. S. Kanetkar, A. Nayak, S. Swamy and G. Bhatia, "Web-based personalized hybrid book recommendation
system," 2014 International Conference on Advances in Engineering & Technology Research (ICAETR - 2014),
Unnao, India, 2014, pp. 1-5, doi: 10.1109/ICAETR.2014.7012952.
4. Sarma, D., Mittra, T., & Shahadat, M. (2021). Personalized Book Recommendation System using Machine
Learning Algorithm. International Journal of Advanced Computer Science and Applications, 12, 212-219.
5. Mounika, A., Saraswathi, S. (2021). Design of Book Recommendation System Using Sentiment Analysis.
In: Suma, V., Bouhmala, N., Wang, H. (eds) Evolutionary Computing and Mobile Sustainable Networks.
Lecture Notes on Data Engineering and Communications Technologies, vol 53. Springer, Singapore. https:
//doi.org/10.1007/978-981-15-5258-8_11.
6. Sivaramakrishnan N, Subramaniyaswamy V, Arunkumar S, Renugadevi A, Ashikamai Kk. Neighborhood-based
approach of collaborative filtering techniques for book recommendation systems. International Journal of Pure
and Applied Mathematics, 2018. ⟨hal-01826674⟩
7. Zhou, Y.W. (2020) Design and Implementation of Book Recommendation Management System Based on
Improved Apriori Algorithm. Intelligent Information Management, 12, 75-87
8. Chen, C. M. (2013). An intelligent mobile location-aware book recommendation system that enhances problem-
based learning in libraries. Interactive Learning Environments, 21(5), 469–495. https://doi.org/10.1080/
10494820.2011.593525
9. T. Kuroiwa and S. Bhalla, "Dynamic Personalization for Book Recommendation System Using Web Services and
Virtual Library Enhancements," 7th IEEE International Conference on Computer and Information Technology
(CIT 2007), Aizu-Wakamatsu, Japan, 2007, pp. 212-217, doi: 10.1109/CIT.2007.72.
10. Jia, F., Shi, Y. (2013). Library Management System Based on Recommendation System. In: Yang, Y., Ma, M.,
Liu, B. (eds) Information Computing and Applications. ICICA 2013. Communications in Computer and Infor-
mation Science, vol 392. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-53703-5_
50
11. I. Hariadi and D. Nurjanah, "Hybrid attribute and personality based recommender system for book recommen-
dation," 2017 International Conference on Data and Software Engineering (ICoDSE), Palembang, Indonesia,
2017, pp. 1-5, doi: 10.1109/ICODSE.2017.8285874.
12. R. Rahutomo, A. S. Perbangsa, H. Soeparno and B. Pardamean, "Embedding Model Design for Producing Book
Recommendation," 2019 International Conference on Information Management and Technology (ICIMTech),
Jakarta/Bali, Indonesia, 2019, pp. 537-541, doi: 10.1109/ICIMTech.2019.8843769.
13. Sharma, S., Rana, V. & Malhotra, M. Automatic recommendation system based on hybrid filtering algo-
rithm. Educ Inf Technol 27, 1523–1538 (2022). https://doi.org/10.1007/s10639-021-10643-8
14. Wayesa, F., Leranso, M., Asefa, G. et al. Pattern-based hybrid book recommendation system using semantic
relationships. Sci Rep 13, 3693 (2023). https://doi.org/10.1038/s41598-023-30987-0
15. Xin, L., E, H., Song, J., Song, M., Tong, J. (2014). Book Recommendation Based on Community Detection. In:
Zu, Q., Vargas-Vera, M., Hu, B. (eds) Pervasive Computing and the Networked World. ICPCA/SWS 2013. Lec-
ture Notes in Computer Science, vol 8351. Springer, Cham. https://doi.org/10.1007/978-3-319-09265-2_
37
16. S. Mariana, I. Surjandari, A. Dhini, A. Rosyidah and P. Prameswari, "Association rule mining for building book
recommendation system in online public access catalog," 2017 3rd International Conference on Science in Infor-
mation Technology (ICSITech), Bandung, Indonesia, 2017, pp. 246-250, doi: 10.1109/ICSITech.2017.8257119.
17. T. Desai, S. Gandhi, P. Murlidhar, S. Gupta, M. Vijayalakshmi and G. P. Bhole, "An enterprise-friendly book
recommendation system for very sparse data," 2016 International Conference on Computing, Analytics and
Security Trends (CAST), Pune, India, 2016, pp. 211-215, doi: 10.1109/CAST.2016.7914968.

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 21 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

18. A. S. Tewari, A. Kumar and A. G. Barman, "Book recommendation system based on combine features of
content based filtering, collaborative filtering and association rule mining," 2014 IEEE International Advance
Computing Conference (IACC), Gurgaon, India, 2014, pp. 500-503, doi: 10.1109/IAdCC.2014.6779375.
19. A. S. Tewari, T. S. Ansari and A. G. Barman, "Opinion based book recommendation using Naive Bayes
classifier," 2014 International Conference on Contemporary Computing and Informatics (IC3I), Mysore, India,
2014, pp. 139-144, doi: 10.1109/IC3I.2014.7019672.
20. Y. Luo, J. Le and H. Chen, "A Privacy-Preserving Book Recommendation Model Based on Multi-agent," 2009
Second International Workshop on Computer Science and Engineering, Qingdao, China, 2009, pp. 323-327,
doi: 10.1109/WCSE.2009.822.
21. Sariki, T. P. ., & Guntur, B. K. (2022). AN AGGRANDIZED FRAMEWORK FOR ENRICHING BOOK
RECOMMENDATION SYSTEM. Malaysian Journal of Computer Science, 35(2), 111–127. https://doi.
org/10.22452/mjcs.vol35no2.2
22. D. Pathak, S. Matharia and C. N. S. Murthy, "NOVA: Hybrid book recommendation engine," 2013 3rd
IEEE International Advance Computing Conference (IACC), Ghaziabad, India, 2013, pp. 977-982, doi:
10.1109/IAdCC.2013.6514359.
23. Y. Zhu, "A book recommendation algorithm based on collaborative filtering," 2016 5th International Confer-
ence on Computer Science and Network Technology (ICCSNT), Changchun, China, 2016, pp. 286-289, doi:
10.1109/ICCSNT.2016.8070165.
24. Monika Verma, Pawan Kumar Patnaik,An automatic college library book recommendation system using
optimized Hidden Markov based weighted fuzzy ranking model,Engineering Applications of Artificial Intelli-
gence,Volume 130,2024,107664,ISSN 0952-1976, https://doi.org/10.1016/j.engappai.2023.107664.
25. Rohit, Sabitha, S., Choudhury, T. (2018). Proposed Approach for Book Recommendation Based on User k-
NN. In: Bhatia, S., Mishra, K., Tiwari, S., Singh, V. (eds) Advances in Computer and Computational Sciences.
Advances in Intelligent Systems and Computing, vol 554. Springer, Singapore. https://doi.org/10.1007/978-981-
10-3773-3_53
26. Amer, A. A., Abdalla, H. I., & Nguyen, L. (2021). Enhancing recommendation systems performance using
highly-effective similarity measures. Knowledge-Based Systems, 217, 106842. ISSN 0950-7051. https://
doi.org/10.1016/j.knosys.2021.106842.
27. Anwar, T., Uma, V. Comparative study of recommender system approaches and movie recommendation
using collaborative filtering. Int J Syst Assur Eng Manag 12, 426–436 (2021). https://doi.org/10.1007/
s13198-021-01087-x
28. Hong, H., Guo, J., & Wang, B. (2012). An Improved KNN Algorithm Based on Adaptive Cluster Distance
Bounding for High Dimensional Indexing. International Journal of Creative Research Thoughts (IJCRT).
https://doi.org/10.1109/gcis.2012.86
29. Singh, Ramni & Maurya, Sargam & Tripathi, Tanisha & Narula, Tushar & Srivastav, Gaurav. (2020). Movie
Recommendation System using Cosine Similarity and KNN. International Journal of Engineering and Advanced
Technology. 9. 2249-8958. 10.35940/ijeat.E9666.069520
30. Bhardwaj, Aditya. (2023). Movie Recommendation System Using SVD (Letterboxd). IJARCCE. 12. 10.17148/IJAR-
CCE.2023.121013.
31. Nguyen, L. V., Vo, Q.-T., & Nguyen, T.-H. (2023). Adaptive KNN-Based Extended Collaborative Filtering
Recommendation Services. Big Data and Cognitive Computing, 7(2), 106. https://doi.org/10.3390/
bdcc7020106
32. Vaidya, A., & Shinde, S. (2019). Hybrid Book Recommendation System. International Research Journal of
Engineering and Technology (IRJET), 6(7).
33. Ahmed, E., & Letta, A. (2023). Book Recommendation Using Collaborative Filtering Algorithm. Applied
Computational Intelligence and Soft Computing, 2023, 1–12. https://doi.org/10.1155/2023/1514801
34. Patro, S. G. K., et al. (2020). A Hybrid Action-Related K-Nearest Neighbour (HAR-KNN) Approach for Rec-
ommendation Systems. IEEE Access, 8, 90978-90991. https://doi.org/10.1109/ACCESS.2020.2994056
35. Yadav, K. K., Soni, H. K., & Pathik, N. (2023). Recommendation System Based on Double Ensemble Models
using KNN-MF. International Journal of Advanced Computer Science and Applications (IJACSA), 14(5).
http://dx.doi.org/10.14569/IJACSA.2023.0140566

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 22 of 23

 Algorithmic Exploration in Reading Behavior Analysis and Recommendations Using Machine Learning

36. Li, G., & Zhang, J. (2018). Music personalized recommendation system based on improved KNN algorithm.
In 2018 IEEE 3rd Advanced Information Technology, Electronic and Automation Control Conference (IAEAC)
(pp. 777-781). https://doi.org/10.1109/IAEAC.2018.8577483

Mahatir Ahmed Tusher: Preprint submitted to Elsevier Page 23 of 23