San Jose State University

SJSU ScholarWorks
Master's Projects Master's Theses and Graduate Research

Fall 2015

RECOMMENDATION SYSTEM USING

COLLABORATIVE FILTERING
Yunkyoung Lee
San Jose State University

Follow this and additional works at: https://scholarworks.sjsu.edu/etd_projects

Part of the Computer Sciences Commons

Recommended Citation
Lee, Yunkyoung, "RECOMMENDATION SYSTEM USING COLLABORATIVE FILTERING" (2015). Master's Projects. 439.
DOI: https://doi.org/10.31979/etd.5c62-ve53
https://scholarworks.sjsu.edu/etd_projects/439

This Master's Project is brought to you for free and open access by the Master's Theses and Graduate Research at SJSU ScholarWorks. It has been
accepted for inclusion in Master's Projects by an authorized administrator of SJSU ScholarWorks. For more information, please contact
[email protected].
RECOMMENDATION SYSTEM USING COLLABORATIVE FILTERING

A Thesis

Presented to

The Faculty of the Department of Computer Science

San José State University

In Partial Fulfillment

of the Requirements for the Degree

Master of Science

by

Yunkyoung Lee

December 2015
©2015 Yunkyoung Lee

All Rights Reserved

 The Designated Project Committee Approves the Project Titled

Recommendation System Using Collaborative Filtering

By

Yunkyoung Lee

APPROVED FOR THE DEPARTMENT OF COMPUTER SCIENCE

SAN JOSE STATE NIVERSITY

December 2015

Dr. T Y Lin Department of Computer Science

Dr. H. Chris Tseng Department of Computer Science

Dr. Thomas Austin Department of Computer Science

 ABSTRACT

Recommendation System Using Collaborative Filtering

by Yunkyoung Lee

Collaborative filtering is one of the well known and most extensive

techniques in recommendation system its basic idea is to predict which items a

user would be interested in based on their preferences. Recommendation

systems using collaborative filtering are able to provide an accurate prediction

when enough data is provided, because this technique is based on the user’s

preference. User-based collaborative filtering has been very successful in the

past to predict the customer’s behavior as the most important part of the

recommendation system. However, their widespread use has revealed some real

challenges, such as data sparsity and data scalability, with gradually increasing

the number of users and items.

To improve the execution time and accuracy of the prediction problem,

this paper proposed item-based collaborative filtering applying dimension

reduction in a recommendation system. It demonstrates that the proposed

approach can achieve better performance and execution time for the

recommendation system in terms of existing challenges, according to evaluation

metrics using Mean Absolute Error (MAE).

 ACKNOWLEDGEMENT

I am very thankful to my advisor Dr. Tsau Young Lin for his continuous

guidance and support throughout the project and having firm believe in me. Also,

I would like to thank the Committee members Dr. H. Chris Tseng and Dr.

Thomas Austin for monitoring the progress of the project and their valuable time.

Table of Contents

LIST OF FIGURES ................................................................................................ 3

LIST OF TABLES ................................................................................................. 4

CHAPTER 1 .......................................................................................................... 5

Introduction.......................................................................................................... 5

CHAPTER 2 .......................................................................................................... 8

RELATED WORK ................................................................................................. 8

2.1 Collaborative Filtering (CF) .................................................................................. 9
2.1.1 User-based Collaborative Filtering (UBCF) ................................................... 10
2.1.2 Item-based Collaborative Filtering (IBCF) ..................................................... 10
2.2 Collaborative Filtering Process ......................................................................... 11
2.2.1 User Rating Score Data Input ......................................................................... 12
2.2.2 The Formation of Neighbors ........................................................................... 13
2.2.2.1 Cosine Vector Similarity .............................................................................. 15
2.2.2.2 Pearson Correlation Coefficient .................................................................. 16
2.2.2.3 Euclidean Distance Similarity ...................................................................... 16
2.2.2.4 Tanimoto Coefficient ................................................................................... 17
2.2.3 Prediction generation ...................................................................................... 18
2.2.3.1 Prediction Computation of UBCF ................................................................ 18
2.2.3.2 Prediction Computation of IBCF .................................................................. 19
2.3 Existing Limitations of Collaborative Filtering................................................. 19
2.3.1 Data Sparsity .................................................................................................. 20
2.3.2 Data Scalability .............................................................................................. 20

CHAPTER 3 ........................................................................................................ 21

Item Based Collaborative Filtering Applying Dimension Reduction ............ 21

3.1 IBCF Applying Dimension Reduction ............................................................... 21

3.2 Architecture of IBCF Applying Dimension Reduction ..................................... 23

CHAPTER 4 ........................................................................................................ 25

Experiments And Evaluation Metrics .............................................................. 25

4.1 Experiments Dataset........................................................................................... 25
4.2 Performance Evaluation Criteria ....................................................................... 26
4.3 Experiment Environment.................................................................................... 27
4.4 Architecture of Apache Mahout ......................................................................... 28
4.5 Algorithm of IBCF Applying Dimension Reduction ......................................... 29
4.6 Benchmark UBCF................................................................................................ 31

CHAPTER 5 ........................................................................................................ 32

PERFORMACE RESULTS ................................................................................. 32

5.1 Optimum Similarity Measurement ..................................................................... 32
5.2 Optimum The Number of Ratings per Item ....................................................... 33
5.3 Optimum Training/Test Ratio ............................................................................. 35
5.4 Optimum The Neighborhood Size of UBCF ...................................................... 36
5.5 Comparison of Prediction Quality with Benchmark ........................................ 37
5.6 Comparison of Runtime with Benchmark......................................................... 38

CHAPTER 6 ........................................................................................................ 40

REFERENCES .................................................................................................... 41

LIST OF FIGURES

[Figure 1] Shopping cart recommendation at Amazon .......................................... 9

[Figure 2] User-based collaborative filtering ........................................................ 10

[Figure 3] Item-based collaborative filtering ........................................................ 11

[Figure 4] The Collaborative filtering process ...................................................... 12

[Figure 5] The neighborhood formation process ................................................. 13

[Figure 6] Item based similarity computation ....................................................... 14

[Figure 7] Diagram of IBCF applying dimension reduction .................................. 24

[Figure 8] Apache Mahout architecture ............................................................... 29

[Figure 9] The impact of the similarity computation on IBCF and R-IBCF ........... 33

[Figure 10] Comparison of Impact of the number of ratings on R-IBCF to IBCF . 34

[Figure 11] Sensitivity of the parameter x in IBCF and R-IBCF ........................... 35

[Figure 12] Sensitivity of neighborhood size in UBCF ......................................... 36

[Figure 13] Comparison of the prediction quality of IBCF, R-IBCF, and UBCF ... 38

[Figure 14] Comparison of runtime of IBCF, R-IBCF, and UBCF ........................ 39

LIST OF TABLES

[Table 1] User-Item ratings matrix ....................................................................... 12

[Table 2] User-Item matrix before dimension reduction ...................................... 22

[Table 3] User-Item matrix after dimension reduction ......................................... 22

[Table 4] Raw dataset of MovieLens ................................................................... 26

[Table 5] User-Item Matrix by raw dataset .......................................................... 26

CHAPTER 1

Introduction

E-commerce markets have been restructured into new markets revolving

around mobile commerce since the advent of smart devices. User has more

opportunity to access diverse information and the amount of information that can

be collected has exponentially increased. The immense growth of the World

Wide Web has led to an information overload problem. It is difficult for users to

quickly obtain what they want from massive information. In recent years, each

customer can actively share their review and get a discount based on customer

participation such as in social surveys on E-commerce sites. It has become

essential for E-commerce markets to effectively take advantage of these data by

evolving new marketing strategy based on such data.

Besides, E-commerce markets have actively introduced an automated

personalization service to analyze the customer’s behavior and patterns as

purchase factors. E-commerce sites try to collect various users’ interests, such

as purchase history, product information in the cart, product ratings, and product

reviews in order to recommend new relevant products to customers.

Collaborative filtering is the most commonly used algorithm to build personalized

recommendations on the website including Amazon, CDNOW, Ebay, Moviefinder,

and Netflix beyond academic interest [1, 14].

Collaborative filtering is a technology to recommend items based on

similarity. There are two types of collaborative filtering: User-based collaborative

filtering and Item-based collaborative filtering [8]. User-based collaborative

filtering algorithm is an effective way of recommending useful contents to users

by exploiting the intuition that a user will likely prefer the items preferred by

similar users. Therefore, at first, the algorithm tries to find the user’s neighbors

based on user similarities and then combines the neighbor user’s rating score by

using supervised learning like k-nearest neighbors algorithm and Bayesian

network or unsupervised learning like k-means algorithm [8, 9].

Item-based collaborative filtering algorithm fundamentally has the same

scheme with user-based collaborative filtering in terms of using user’s rating

score. Instead of the nearest neighbors, it looks into a set of items; the target

user has already rated items and this algorithm computes how similar items are

to the target item under recommendation [8, 9]. After that it also combines the

customer’s previous preferences based on these item similarities.

Collaborative Filtering has been effective in several domains, but their

widespread use has revealed some potential challenges, such as rating data

sparsity, cold-start, and data scalability [2, 6, 8, 9]. Therefore, to solve the

problems of sparsity and scalability in the collaborative filtering, in this paper, I

proposed collaborative filtering applying dimension reduction.

The rest of this paper is organized as follows: Chapter 2 summaries the

related work and their capabilities and limitation. The proposed approach is

described in Chapter 3. Chapter 4 describes the experimental configuration and

evaluation metrics. Experimental results are given in Chapter 5. Finally, Chapter

6 concludes this paper and provides directions for future work.

CHAPTER 2

RELATED WORK

Since the advent of the information age, the immense growth of the World

Wide Web gives rise to the difficulty for users to quickly find what they want given

a variety of applications. Recommendation systems have rigorously been used in

various applications as a way to suggest items that a customer would likely be

interested in by predicting customer preference. The most popular applications

using recommendation systems are movies, music, news, grocery shopping,

travel guides, online dating, books, restaurants, E-commerce sites and so forth.

Recommendation systems can be broadly categorized as contents-based

filtering, collaborative filtering, and hybrid approach [3]. Contents-based filtering

systems are used to recommend items based on a description of items the user

used to like before, or corresponding with pre-defined attributes of the user, such

a system having its roots in information retrieval techniques. Collaborative

filtering systems recommend items to user based on the past preferences of

items rated by all users. Hybrid techniques combine both these approaches. In

this paper, I will deal mainly with collaborative filtering (CF).

2.1 Collaborative Filtering (CF)

Recommendation systems in various applications have tried to provide

users with an accurate recommendation to meet the needs of the user and to

bring higher benefits to companies. Collaborative filtering is an effective and well-

known technology in recommendation systems. Many web sites, particularly E-

commerce sites, have used collaborative filtering technology in their

recommendation systems to personalize the browsing experience for each user

as seen [Figure 1]. As successful use cases of collaborative filtering, Amazon

increased sales by 29% [11], Netflix increased movie rentals by 60% [12], and

Google news increased click-through rates by 30.9% [13].

[Figure 1] Shopping cart recommendation at Amazon

Collaborative filtering (CF) can be categorized into two main methods as

user-based collaborative filtering (memory-based) and item-based collaborative

filtering (model-based) [8].

2.1.1 User-based Collaborative Filtering (UBCF)

User-based collaborative filtering approach is to predict items to the target

user that are already items of interest for other users who are similar to the target

user. For example, as seen [Figure 2] [15], let User 1 and User 3 have very

similar preference behavior. If User 1 likes Item A, UBCF can recommend Item A

to User 3. UBCF needs the explicit rating scores of items rated by users [8] to

calculate similarities between users and exploits k-nearest neighbor algorithms to

find the nearest neighbors based on user similarities. And then, it generates

prediction in terms of items by combining the neighbor user’s rating scores based

on similarity weighted averaging [9].

1 2 3

[Figure 2] User-based collaborative filtering

2.1.2 Item-based Collaborative Filtering (IBCF)

Item-based collaborative filtering approach is to predict items by inquiring

into similarities between the items and other items that are already associated

10



with the user. For example, as seen in [Figure 3] [15], let’s say Item A and Item C

are very similar. If a User likes Item A, IBCF can recommend Item C to the User.

IBCF needs a set of items that the target user has already rated to calculate

similarities between items and a target item. And then, it generates prediction in

terms of the target item by combining the target user’s previous preferences

based on these item similarities [9]. In IBCF, users’ preference data can be

collected in two ways. One is that user explicitly gives rating score to item within

a certain numerical scale. The other is that it implicitly analyzes user’s purchase

records or click-through rate [8].

[Figure 3] Item-based collaborative filtering

2.2 Collaborative Filtering Process

In a fundamental scenario, collaborative filtering (CF) processing can be

mainly divided into three steps; Step 1) collecting user ratings data matrix, Step 2)

selecting similar neighbors by measuring the rating similarity, and then Step 3)

generating prediction as seen diagram [Figure 1] [4, 6, 7, 8, 9].

11



[Figure 4] The Collaborative filtering process

2.2.1 User Rating Score Data Input

Generally, input data in recommendation system based on the CF

technology consists of user, item, and user opinions on observed items as a

matrix m × n as shown in [Table 1]. Symbol m symbolizes the total number of

users and n symbolizes the total number of items. 𝑅!,! is the score of item In

rated by user Um.

Item
I1 I2 I3 … In
User
U1 𝑅!,! 𝑅!,! 𝑅!,! … 𝑅!,!

U2 𝑅!,! 𝑅!,! 𝑅!,! … 𝑅!,!

U3 𝑅!,! 𝑅!,! 𝑅!,! … 𝑅!,!

… … … … … …
Um 𝑅!,! 𝑅!,! 𝑅!,! … 𝑅!,!
[Table 1] User-Item ratings matrix

12



2.2.2 The Formation of Neighbors

The CF approaches use statistical techniques to analyze the similarity

between users and to form a set of users called neighbors. A set of similarity

measures is a metric of relevance between two vectors [9]. User-based similarity

is to compute the relevance between users as the values of two vectors. In

UBCF, after the similarity is calculated, it is used in building neighborhoods of the

current target user. For example, as seen in [Figure 5] [4], the distance between

the target node (black node) and every other node is calculated by a similarity

measure. And then, 5 users in the center are selected by k-nearest neighbor

algorithm (k = 5).

[Figure 5] The neighborhood formation process

13



[Figure 6] Item based similarity computation

In contrast, it should be noted that IBCF does not form neighborhoods

after the similarity is calculated. This is because IBCF already begins computing

the similarity between co-rated items only as the value of two vectors [9,14]. For

example, as seen in [Figure 6] [2], this item-based similarity is calculated by

looking into Item i and Item j rated by User 2, l, and n. Each of these pairs are

given by different users. This is a similar process to the formation of neighbors in

UBCF.

Since the similarity measure plays a significant role in improving accuracy

in prediction algorithms, it can be effectively used to balance the ratings

significance [9]. There are a couple of popular similarity algorithms that have

been used in the CF recommendation algorithms [8]. In this paper, I present four

similarity methods; Cosine vector similarity, Pearson correlation, Euclidean

distance similarity, and Tanimoto coefficient [14, 16].

14



2.2.2.1 Cosine Vector Similarity

Cosine vector similarity is one of the popular metrics in statistics. Since it

notionally considers only the angle of two vectors without the magnitude, it is a

very useful measurement with data missing preference information as long as it

can count the number of times that term appears in the data [17].

In the following formula, the cosine vector similarity looks into the angle

between two vectors (the target Item i and the other Item j) of ratings in n-

dimensional item space. 𝑅!,! is the rating of the target Item i by User k. 𝑅!,! is the

rating of the other Item j by user k. n is the total number of all rating users to Item

i and Item j.
!
𝚤×𝚥 !!! 𝑅!,! 𝑅!,!
𝑠𝑖𝑚 𝑖, 𝑗 = cos 𝚤, 𝚥 = ! !
=
𝚤 × 𝚥 ! ! ! !
!!! 𝑅!,! !!! 𝑅!,!

When the angle between two vectors is near 0 degree (they are in the

same direction), Cosine similarity value, sim(i,j), is 1, meaning very similar. When

the angle between two vectors is near 90 degree, sim(i,j) is 0, meaning irrelevant.

When the angle between two vectors is near 180 degree (they are in the

opposite direction), sim(i,j) is -1, meaning very dissimilar. In case of information

retrieval using CF, sim(i,j) ranges from 0 to 1. This is because the angle between

two term frequency vectors cannot be greater than 90 degrees [17].

15



2.2.2.2 Pearson Correlation Coefficient

Pearson correlation coefficient is one of the popularly used methods in CF

to measure how larger a number in one series is, relative to the corresponding

number. As following formula shows, it is used to measure the linear correlation

between two vectors (Item i and Item j).

!
𝑅!,! − 𝐴! , 𝑅!,! − 𝐴! !!!(𝑅!,! − 𝐴! ) (𝑅!,! − 𝐴! )
𝑠𝑖𝑚 𝑖, 𝑗 = = ! !
𝑅!,! − 𝐴! 𝑅!,! − 𝐴! !!!(𝑅!,! − 𝐴 ! )! × !!!(𝑅!,! − 𝐴! )!

It measures the tendency of two series of numbers, paired up one-to-one,

to move together [14]. When two vectors have a high tendency, the correlation,

𝑠𝑖𝑚(𝑖, 𝑗), is close to 1. When two vectors have a low tendency, 𝑠𝑖𝑚(𝑖, 𝑗) is close

to 0. When two vectors have opposite tendency, 𝑠𝑖𝑚(𝑖, 𝑗) is close to -1. As

mentioned above in [Figure 6], item-based similarity is computed with the co-

rated items where users rated both (Item i and Item j).

𝑅!,! is the rating of the target Item i given by User k. 𝑅!,! is the rating of the

other Item j given by User i. 𝐴! is the average rating of the target Item i for all the

co-rated users, and 𝐴! is the average rating of the other Item j for all the co-rated

users. n is the total number of ratings users gave to Item i and Item j.

2.2.2.3 Euclidean Distance Similarity

Euclidean distance method is based on the distance between items. It

forms coordinates to put preference values between items and measures

Euclidean distance between each point. When distance value between two

16



points, 𝑠𝑖𝑚 𝑖, 𝑗 , is large, it means the two points are not similar. When 𝑠𝑖𝑚(𝑖, 𝑗) is

small, it means two points are similar. This is Euclidean distance formula is given

below.

𝑠𝑖𝑚(𝑖, 𝑗) = (𝑅!,! − 𝑅!,! )!

!!!

𝑅!,! is the ratings of the target Item i given by User k. 𝑅!,! is the ratings of

the other Item j given by User k. n is the total number of rating users to Item i and

Item j.

2.2.2.4 Tanimoto Coefficient

Tanimoto coefficient is known as the Jaccard similarity coefficient. It does

not take into account preference values of an item rated by a user. It only

considers if users express a preference. Tanimoto coefficient is the ratio of the

size of the intersection, or overlap, in two users’ preferred items, to the union of

users’ preferred items [14]. When two items are completely overlapped,

Tanimoto coefficient, 𝑠𝑖𝑚 𝑖, 𝑗 , is 1. When two items are not completely

overlapped, 𝑠𝑖𝑚 𝑖, 𝑗 is 0.

As shown in the following formula, it examines the overlapped degree

between two sets to compare the similarity and diversity of two sets. 𝑓! is a set of

Item i for which users express preference. 𝑓! is a set of Item j for which users

17



express preference. 𝑓! ∩ 𝑓! is intersection of Item i and Item j for items where

preference is expressed by users.

𝑓! ∩ 𝑓!
𝑠𝑖𝑚 𝑖, 𝑗 =
𝑓! + 𝑓! − 𝑓! ∩ 𝑓!

In fact, many users tend not to rate items or recommendation system

might not have enough users’ information. This metric would be helpful to

compute similarity as long as at least preference information as Boolean type is

available.

2.2.3 Prediction generation

Once CF computes the similarity between users (in UBCF) or items (in

IBCF) and then finds the set of most similar user or similar items, it generates

prediction of the target user’s interest as the most significant step in CF.

2.2.3.1 Prediction Computation of UBCF

Since UBCF gets the neighborhood of user, UBCF can calculate the

predictive rating for the target User u on the target Item i. It is scaled by the

weighted average of all neighbors’ ratings on the target Item i as following [2, 4]:
!
!!!
𝑅!,! − 𝐴! × 𝑠𝑖𝑚(𝑢, 𝑤)
𝑃!,! = 𝐴! + !
!!! 𝑠𝑖𝑚(𝑢, 𝑤)

𝐴! is the average ratings of the target User u to all other rated items and

𝐴! is the average ratings of the neighbor User w to all other rated items. 𝑅!,! is

18



the rating of the neighbor User w to the target item i. 𝑠𝑖𝑚(𝑢, 𝑤) is the similarity of

the target User u and the neighbor User w. And n is the total number of

neighbors.

2.2.3.2 Prediction Computation of IBCF

Since IBCF has got the neighborhood of items, IBCF tries to make sure

how the target user rates similar items. To check if the prediction is in the

predefined range [8], the predictive rating for the target User u on the target Item

i is scaled by the weighted average of all neighbor items’ ratings given by the

target User u according to the following formula [8, 10].

!
!!! 𝑅!,! ×𝑠𝑖𝑚(𝑖, 𝑗)
𝑃!,! = !
!!! 𝑠𝑖𝑚(𝑖, 𝑗)

𝑅!,! is the rating of the target User u to the target Item i. 𝑠𝑖𝑚(𝑖, 𝑗) is the weighted

similarity of the target Item i and the neighbor Item j, n is the total number of

neighbor items.

2.3 Existing Limitations of Collaborative Filtering

Since the number of users and items in each application has steadily

increased at the same time as the growth of World Wide Web, collected input

data has been a big problem in producing an accurate prediction and in running

19



recommendation system using collaborative filtering. There are two main

challenges in user-based collaborative filtering [8, 9, 10, 15].

2.3.1 Data Sparsity

User-based collaborative filtering depends on explicit feedback, such as

ratings given by user to item. User-item input data matrix could have a few rating

scores of the total number of items available, even though users are very active.

In addition, because users tend not to rate actively, calculating similarity over co-

rated set of items could be a challenge. These problems give rise to inaccurate

performance of the recommendation system.

Even the cold-start problem is caused by the data sparsity. Collaborative

Filtering predicts items based on user’s previous preference behavior. That is, it

could not predict recommendable items to new users unless new users rate

many items. Also, new items could be considered for recommendation, because

they have less rating scores by a sufficient number of users.

2.3.2 Data Scalability

For over millions of users and millions of items in user-item input data

matrix, the nearest neighbor algorithm is required for high scalability of

computation between users as the values of two vectors. Also, recommendation

systems could not quickly react to online requirements and immediately make

recommendations as it was a time-consuming job.

20



CHAPTER 3

Item Based Collaborative Filtering Applying Dimension

Reduction

UBCF is easy to implement and good to scale correlated items [2].

However, as stated previously above, it comes up against a couple of problems:

data sparsity and data scalability. Data sparsity problem could lead to a skewed

prediction and low reliability of predictions. Besides, data scalability requires low

operation time and high memory feature to scale with all users and items in the

database.

To address these issues in UBCF, this paper proposes IBCF approach

applying dimension reduction [6].

3.1 IBCF Applying Dimension Reduction

Enormous users and products have been added at E-commerce domains.

A typical example is Amazon. Amazon added 30 million new customers in 2013

and had had over 244 million active customers as Geekwire reported in 2014 [18].

Also, Amazon had sold over 200 million products as ReportX reported in 2013

[19]. Currently in 2015, it is expected that Amazon would have more than these

numbers of users and products. If the recommendation system using UBCF at

21



Amazon should look into all datasets similar to a 244 million × 200 million matrix,

it will encounter data scalability and data sparsity issues. In UBCF, more the

number of users and items increase, more the number of matrix dimensions

increase and runtime takes long to find nearest neighbor of users. Therefore, it is

assumed that using denser data having much more preference information given

by users with IBCF effectively addresses data scalability and data sparsity

problems. To focus on active items assuming that they have many ratings given

by users, matrix is required to reduce dimension in IBCF without regard to

passive items.

Item
I1 I2 I3 I4 I5
User
U1 2.0 4.0 3.0 3.0 3.0
U2 1.0 3.0
U3 5.0 1.0 5.0 5.0
U4 4.0 3.0
[Table 2] User-Item matrix before dimension reduction

Item
I1 I3 I4
User
U1 2.0 3.0 3.0
U2 1.0 3.0
U3 5.0 5.0 5.0
U4 4.0 3.0
[Table 3] User-Item matrix after dimension reduction

22



For instance, as seen in [Table 2], each item can get up to a maximum of

4 ratings by users. Item I2 has 2 ratings and Item I5 has 1 rating, which means

the number of ratings for Item I2 and Item I5 is not bigger than half of the total

number of ratings. We can assume that Item I2 and Item I5 do not carry much

weight with this matrix. Hence, when matrix has impactful items like Item I1, Item

I3, and Item I4 as seen in [Table 3], running time of the recommendation system

in computing similarity between items and to provide more accurate prediction is

expected to reduce.

3.2 Architecture of IBCF Applying Dimension Reduction

Here is a scenario of IBCF applying dimension reduction as seen in

[Figure 7]. This is mainly divided into four steps. This approach is based on

general collaborative filtering algorithm. To compute similarity between items, the

algorithm uses an optimized data by reducing dimention of items that have the

number of ratings less than a specific value. For example, if it needs to consider

items that have over 20 ratings from users, it extracts data in terms of items

having over 20 ratings. In other words, such items are rated by over 20 users.

23



Input dataset [User ID, Item ID, Rating]

Optimize data; which Items have over a particular number of ratings

Compute similarity between Items and find neighbors of an item

Recommend Items

[Figure 7] Diagram of IBCF applying dimension reduction

24



CHAPTER 4

Experiments And Evaluation Metrics

In this section, I describe dataset, evaluation metrics, and methodology to

optimize data by reducing dimension based on Apache Mahout.

4.1 Experiments Dataset

The data used in this experiment is the MovieLens 1m datasets by

GroupLens Research [20]. It contains 1,000,000 ratings by 6040 users on 3952

movies. Each user has rated at least 20 movies [21]. The range of ratings is from

1 (less interesting) to 5 (very interesting) as integer type.

These are a few parts of MovieLen dataset. It consists of user ID, Item ID

and Rating as seen in [Table 4]. I consider it as the User-Item matrix as seen in

[Table 5].

25



User ID Item ID Rating
1 1035 5
1 1287 5
1 3408 4
6 1035 5
6 1380 5
6 3408 5
10 1035 5
10 1380 5
10 1287 3
10 3408 4
10 1201 2
26 1035 2
26 1380 4
26 3408 2
26 1201 2
… … …
[Table 4] Raw dataset of MovieLens

Item
1035 1380 1287 3408 1201 …
User
1 5 5 4 …
6 5 5 5 …
10 5 5 3 4 2 …
26 2 4 2 2 …
… … … … … … …
[Table 5] User-Item Matrix by raw dataset

4.2 Performance Evaluation Criteria

To evaluate the accuracy of a recommendation system, I use statistical

accuracy metrics. Mean Absolute Error (MAE) is a widely used metric in the

26



recommendation system using collaborative filtering to measure the deviation of

recommendations from their true user actual ratings.

In the following formula, N is the total number of actual ratings in an item

set. 𝑝! is the prediction of user’s ratings. 𝑞! is corresponding real ratings data set

of users.
!
!!! 𝑝! − 𝑞!
𝑀𝐴𝐸 =
𝑁

It computes the average of the absolute difference between 𝑁 pairs;

prediction scores of users’ ratings and actual user ratings for the user-item pairs

in the test dataset [2]. Lower the MAE value, better is the recommendation

system’s accuracy of prediction of user ratings.

4.3 Experiment Environment

• Processor: 2.6 GHz Intel Core i5

• Memory: 8 GB 1600 MHz DDR3

• Operation System: OS X El Capitan Version 10.11.1

• Language: Java

• Platform: Apache Mahout with a pseudo-distributed mode in Apache

Hadoop

27



4.4 Architecture of Apache Mahout

To implement this experiment, I used Apache Mahout, widely used in

recommendation system using collaborative filtering. Apache Mahout is one of

the most powerful open source platforms in supporting scalable machine learning

and distributing processing of a large dataset cluster of computers using the

Apache Hadoop system required for recommender development [15]. Developer

can effectively customize recommendation system with a rich set of modules

abstractions provided by Apache Mahout. There are five key abstractions to

define the Mahout interface shown in [Figure 8] [15].

1) DataModel is the interface to repository about users and their

associated preferences from any source. Users and items are identified

solely by an ID value as numeric type [14, 15]. A GernericPreference

object encapsulates the relation between an item and preference score. A

GenericUserPreference object stores preferences for all users. A

PreferenceArray object encapsulates the relation between item and

preferred items by users.

2) UserSimilarity measures similarity between users used in UBCF.

3) ItemSmilarity measures similarity between items used in IBCF.

4) UserNeighborhood finds K-nearest neighborhood of similar users near

a target user in UBCF. Since IBCF begins with a list of a user’s preferred

items, it does not need to find K-nearest neighborhood of items.

5) Recommender provides items with a target user given a DataModel.

28



[Figure 8] Apache Mahout architecture

4.5 Algorithm of IBCF Applying Dimension Reduction

Input: A: MovieLens dataset “UserID,ItemID,Rating,”

U: a target user U,

I: Item what user is interested,

M: Minimum ratings,

N: the number of items to recommend

29



1) Parse raw dataset and count the number of ratings per item

FastByIdMap C is a map of the total number of ratings per item.

ArrayList D is a list of all ratings.

For each line L in A Do

Array P ← Convert L to [UserID, ItemID, Rating]

Integer S = the number of ratings per item

If C(P[ItemID]) exists Then

C ← <P[ItemID], S+1 >

Else

C ← <P[ItemID], 1 >

D ← P

2) Create optimized data structure by checking if each item’s total number of

ratings is bigger than M.

FastByIdMap O = <P[UserID], R>

For each P in D Do

If (C(P[ItemID]) > M) Then

ArrayList<GenericPreference> R

← GenericPreference (P[UserID], P[ItemID], P[Rating])

O ← <P[UserID], R>

3) Convert users’ ArrayList<GenericPreference> to GenericUserPreferenceArray

For each <P[UserID], R> in O Do

FastByIdMap X ← <P[UserID], GenericUserPreferenceArray(R)>

4) Create W = GenericDataModel (X)

5) Compute Similarity between I and other items in W

6) Create GenericItemBasedRecommender

30



7) Select the most similar N items against I

Output: N recommended ItemIDs for U

4.6 Benchmark UBCF

To compare the performance of recommendation system using IBCF, I

also implemented UBCF with Apache Mahout.

31



CHAPTER 5

PERFORMACE RESULTS

In this chapter, I implemented item-based collaborative filtering applying

dimension reduction (R-IBCF). The goal of the proposed R-IBCF is to provide

better quality of prediction in terms of the MAE measure and to make faster

execution time. I compared the R-IBCF algorithm to IBCF in order to find an

optimal similarity algorithm and training/test ratio of the dataset. Also, I selected

an optimal value of the number of ratings per item on R-IBCF as I varied the

value of it.

In addition, I implemented UBCF as a benchmark to compare runtime of

R-IBCF and IBCF to UBCF and the quality of prediction with optimal parameters.

5.1 Optimum Similarity Measurement

I implemented four different similarity measurements: Cosine vector

similarity, Pearson correlation coefficient, Euclidean distance, and Tanimoto

coefficient as described in Section 2.2.2. For each similarity algorithms, I

measured MAE to find an optimal similarity on IBCF and R-IBCF for this dataset.

32



[Figure 9] The impact of the similarity computation on IBCF and R-IBCF

[Figure 9] shows the experiments results. I observed that IBCF applying

dimension reduction generally produced better quality of predictions more than

IBCF with four similarity measurements. In particular, Tanimoto coefficient has a

clear advantage, as MAE is the lowest on IBCF and R-IBCF. Therefore, I select

Tanimoto coefficient similarity for the rest of my experiments.

5.2 Optimum The Number of Ratings per Item

If a user does not rate at least one item, the system cannot recommend

any items to the user. In this dataset, each item needs to have at least 627

ratings in order to recommend items to all users. That is, each item has to be

rated by at least 627 users. For example, if I use the optimized data with items

33



having at least 628 ratings, after dimension reduction based on it, one person

among 6040 users cannot get item recommendation. This is because that user

did not give items any ratings at all. Therefore, to prevent non-recommendation

from happening, I performed these experiments to find the optimal number of

ratings per item ranging from 50 (similar to raw data) to 627 (smaller dataset) in

mostly increments of 50, and computed MAE.

[Figure 10] Comparison of Impact of the number of ratings on R-IBCF to IBCF

My results are shown in [Figure 10]. I observed that the quality of

prediction increases as I apply reducing dimensions on IBCF based on the

number of ratings, x. When I reduce a few dimensions (x = 50), the quality of

prediction is almost same with IBCF (MAE = 0.786). On the other hand, when I

reduce lots of dimensions (x = 627), the quality of prediction is the best (MAE =

34



0.745). Therefore, I select 627 as an optimal value in terms of the number of

ratings per item for the rest of my experiments.

5.3 Optimum Training/Test Ratio

To determine the sensitivity of density of the dataset, I ran these

experiments where I varied the value of training/test ratio ranging from 0.2 to 0.9

in an increment of 0.1 and computed MAE. For instance, x is 0.2 means that my

experiments run with 20% of dataset as training data and 80% of dataset as test

data.

[Figure 11] Sensitivity of the parameter x in IBCF and R-IBCF

35



The results are shown in [Figure 11]. I observed that applying dimension

on IBCF generally makes the quality of prediction better than IBCF. When the

training ratio is 0.7, R-IBCF tends to be flat. Hence, I select 0.7 as optimal choice

of training/test ratio for the rest of my experiments.

5.4 Optimum The Neighborhood Size of UBCF

The size of neighborhood on UBCF plays an important role in affecting the

prediction quality [8]. To find the sensitivity of neighborhood size, I ran an

experiment where I varied the number of neighbors ranging from 10 to 50 in an

increment of 5 to be used, and measured MAE on UBCF.

[Figure 12] Sensitivity of neighborhood size in UBCF

36



[Figure 12] shows the experimental results. I observed that neighborhood

size has an effect on the quality of prediction. The quality of prediction gets better

by increasing the number of neighbors and when the rate is 30, the quality of

predictions tends to be flat. Therefore, I select 30 as an optimum value of the

neighborhood size for subsequent experiments with UBCF.

5.5 Comparison of Prediction Quality with Benchmark

Once I obtained the optimal values of the parameters, I compared both

IBCF and R-IBCF approaches with the benchmark UBCF. The purpose of this

experiment was to determine how each similarity algorithm influences the quality

of prediction accuracy of IBCF, R-IBCF, and UBCF. It is a critical step in

collaborative filtering to compute the similarity between each item or each user in

selecting the most similar neighbors of them.

I present the results in [Figure 13]. I performed them with selected values:

0.7 as the optimum training on three CFs, 627 as an optimal value of the number

of ratings per item on R-IBCF, and 30 as an optimal size of a neighbor on UBCF.

Overall, R-IBCF provides better quality of predictions than IBCF and UBCF. It

can be observed that data sparsity and data scalability problems affect the

quality of predictions in computing similarity between items. Reducing

dimensions means that it does not take into account items having fewer ratings

by users, but considers typical or representative items. Therefore, the results

37



show that reducing dimensions on IBCF contributes greatly to improve the quality

of predictions in terms of data sparsity and data scalability.

[Figure 13] Comparison of the prediction quality of IBCF, R-IBCF, and UBCF

5.6 Comparison of Runtime with Benchmark

Runtime of performance is also an important point in terms of data

scalability. I implemented R-IBCF consuming memory. I ran each experiment

with four similarity algorithms 30 times and got the average of their runtime

excluding the first 5 times.

These results are shown in [Figure 14]. Even though it takes more time to

filter data based on the number of ratings per item, I observed that it is faster

38



than computing similarity between all co-rated items or all users. Therefore,

reduction of dimension on IBCF has considerable impact on runtime being fast in

terms of data scalability. In addition, because IBCF and R-IBCF only consider co-

rated items to compute similarity, they do not take finding the nearest neighbors

step. Therefore, it generally influences on better runtime of IBCF and R-IBCF by

comparison with UBCF.

[Figure 14] Comparison of runtime of IBCF, R-IBCF, and UBCF

39



CHAPTER 6

CONCLUSION AND FUTURE WORK

Recommendation systems have been an important in E-commerce on the

web for the customer to suggest items what they would be interested. With the

increasing number of users and items, recommendation systems encounter the

main shortcoming: data sparsity and data scalability problems, which bring out

the poor quality of prediction and the inefficient time consuming.

In this paper, I have proposed item-based collaborative filtering approach

applying dimension reduction to improve the predictive accuracy and

recommendation quality in overcoming the existing limitations. By reducing the

noise of dimensional data, it focuses on typical and popular items to compute the

similarity between them and to predict the most similar items to users. The

experimental results show that this approach makes a considerable impact on

providing better accuracy of prediction and much faster execution time in

comparison with traditional UBCF and IBCF. It results in improving the quality of

recommendation system using collaborative filtering.

The potential limitation would use this approach with dataset widely

consisting of not enough ratings by users, expecting less accuracy. Therefore, to

overcome this challenge, I propose an approach to mix both explicit and implicit

ratings to alleviate the data sparsity problem further in this aspect.

40



REFERENCES

[1] Schafer, J. Ben, Joseph Konstan, and John Riedl. 1999. “Recommender

Systems In E-Commerce.” In 1St ACM Conference On Electronic Commerce,

158-166.

[2] Su, Xiaoyuan, and Taghi M. Khoshgoftaar. 2009. “A Survey Of Collaborative

Filtering Techniques.” Advances In Artificial Intelligence 2009: 1-19.

doi:10.1155/2009/421425.

[3] Melville, Prem, and Vikas Sindhwani. 2010. “Recommender Systems.”

Encyclopedia Of Machine Learning.

[4] Xingyuan Li.2011 “Collaborative Filtering Recommendation Algorithm Based

on Cluster”, International Conference on Computer Science and network

Technology(ICCSNT), IEEE, 4: 2682-2685.

[5] Francesco Ricc, Lior Rokach, Bracha Shapira.2011. Recommender Systems

Handbook. NY:Springer, 1-35

[6] Sarwar, George Kaypi, Joseph Konstan and John Riedl.2000. "Application of

Dimensionality Reduction in Recommender Systems -- A 6 Study." In ACM

WebKDD Workshop.

[7] Sarwar, George Kaypi, Joseph Konstan, John Riedl.2000. “Analysis of

recommendation algorithms for E-commerce.” In the Second ACMConference on

Electronic Commerce, 158–167

41



[8] Sarwar, George Kaypi, Joseph Konstan, John Riedl.2001. “Item-based

Collaborative Filtering Recommendation Algorithms.” In the 10th International

World Wide Web Conference, 285-295

[9] Gong, Songjie. 2010. “A Collaborative Filtering Recommendation Algorithm

Based On User Clustering And Item Clustering.” JSW 5 (7).

doi:10.4304/jsw.5.7.745-752.

[10] Yan Shi, Xiao, HongWu Ye, and SongJie Gong. 2008. “A Personalized

Recommender Integrating Item-Based And User-Based Collaborative Filtering.”

ISBIM '08 International Seminar On Business And Information Management 1

(2008): 264-267.

[11] Mangalindan, JP. 2012. “Amazon’S Recommendation Secret.” Fortune.

http://fortune.com/2012/07/30/amazons-recommendation-secret/.

[12] Koren, Yehuda.2009. “Collaborative Filtering with Temporal Dynamics.” 15th

ACM SIGKDD Int’l Conf. Knowledge Discovery and Data Mining (KDD 09), ACM

(2009): 447-455.

[13] Liu, Jiahui, Dolan, Peter, Pedersen, Elin Rønby.2010. “ Personalized news

recommendation based on click behavior”, In: Rich, et al. (eds.) In the 14th Int.

Conf. on Intelligent User Interfaces (IUI), ACM, (2010): 31–40.

[14] Owen, Sean, Anil, Robin, Dunning, Ted, Friedman, Ellen. 2011 . Mahout in

action. Shelter Island NY: Manning. ()

42



[15] Walunj, Sachin, Sadafale, Kishor. 2013. “An online recommendation system

for e-commerce based on apache mahout framework.” Proceedings of the 2013

annual conference on Computers and people research, ACM (2013): 153–158.

[16] Verbert, Katrien, Drachsler, Hendrik, Manouselis, Nikos, Wolpers, Martin,

Vuorikari, Vuorikari, Riina, Duval, Erik.2011. “Dataset-driven Research for

Improving Recommender Systems for Learning.” 1st

International Conference Learning Analytics & Knowledge, ACM (2011): 44-53.

[17] Wikipedia,. 2015. “Cosine Similarity.”

https://en.wikipedia.org/wiki/Cosine_similarity.

[18] Duryee, Tricia. 2014. “Amazon Adds 30 Million Customers In The Past Year

- Geekwire.” Geekwire. http://www.geekwire.com/2014/amazon-adds-30-million-

customers-past-year/.

[19] Grey, Paul. 2013. “How Many Products Does Amazon Sell? | Exportx.”

Exportx. https://export-x.com/2013/12/15/many-products-amazon-sell/.

[20] Resnick, Paul, Iacovou, Neophytos, Suchak, Mitesh, Bergstrom, Peter, Riedl,

John.1994. “ GroupLens: an open architecture for collaborative filtering of

netnews.” CSCW conference, ACM (1994).

[21] GroupLens,. 2013. “Movielens'. http://grouplens.org/datasets/movielens/.”

43