Personalized Recommender System for Digital ...

18 International Journal of Web-Based Learning and Teaching Technologies, 9(1), 18-32, January-March 2014

Personalized Recommender System for Digital Libraries Omisore M. O., Department of Computer Science, Federal University of Technology Akure, Akure, Ondo, Nigeria Samuel O. W., Department of Computer Science, Federal University of Technology Akure, Akure, Ondo, Nigeria

ABSTRACT The huge amount of information available online has given rise to personalization and filtering systems. Recommender systems (RS) constitute a specific type of information filtering technique that present items according to user’s interests. In this research, a web-based personalized recommender system capable of providing learners with books that suit their reading abilities was developed. Content-based filtering (CBF) was used to analyze learners’ reading abilities while books that are found suitable to learners are recommended with fuzzy matching techniques. The yokefellow cold-start problem inherent to CBF is assuaged by cold start engine. An experimental study was carried out on a database of 10000 books from different categories of computing studies. The outcome tracked over a period of eight months shows that the proposed system induces greater user satisfaction and this attests users’ desirability of the system. Keywords:

Cold Start Problem, Content-Based Filtering, Digital Library, E-Learning System, Fuzzy Logic, Information Retrieval, Paired Samples Test, Recommender System

INTRODUCTION The tremendous growth and usage of information has led to the problem of information overload in which users find it difficult to locate right information at the right time (Resnick et al., 1994). Knowledge helps to shape readers’ perspectives about the content of a text, attention given to it, their interests about the content, and their judgments of importance regarding the content (Alexander et al., 1994). Hence, the complexity of knowledge is dependent on the forms and dimensions of available learning styles.

Learning is being conducted with new forms of technologies, yet the potential outcomes of students’ interactions with available technologies are uncertain (Alexander et al., 1994). The incessant availability of books makes it difficult to proffer suitable Text-Based Learning models for assessing the relationships between books and learners. However in recent time, researches have been conducted with the goal of providing learners with more precise and personalized services (Pazzani & Billsus, 2007). Linguistic quality of text is a critical factor used to quantify users’ reading ability, different approaches, such as Lexile Measures (Schnick & Knickelbine, 2004) were proposed to measure

DOI: 10.4018/ijwltt.2014010102 Copyright © 2014, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.

International Journal of Web-Based Learning and Teaching Technologies, 9(1), 18-32, January-March 2014 19

the Lexile rates of readers. (Pazzani & Billsus, 2007) presented several factors that contribute to the considerateness and quality of texts. As information and e-commerce burgeons, RS becomes inevitable incessant tool. A RS applies data analysis techniques to support users in identifying interesting items among large numbers (Ghauth & Abdullah, 2010). Among the popular RS techniques, CBF and CF are commonly used to recommend products, meticulously, to users. Various RS techniques had been explained in (Melville et al., 2002; Ojokoh et al., 2012; Resnick & Varian, 1997; Sarwar et al., 2001). Hybrid RSs combine two or more recommendation techniques to gain better performance with fewer drawbacks, such systems were emphasized in (Leyla & Olfa, 2010; Resnick & Varian, 1997). Personalization is a special form of differentiation which allows a RS to respond to users’ unique needs. Personalized recommender systems using profiles, content data, and feedbacks have been proposed (Mei-Hua Hsu 2008; Ojokoh et al., 2012). The objective of this study is to propose a web-based personalized recommender system based using CBF and Fuzzy Matching Techniques. The system examines the readability of a user and recommends books that are relative to the user’s level. The web based system was implemented on a database with 10000 books from different categories of computing studies. An experimental study was carried out within eight months and the result obtained was used to evaluate the performance of the system. The outcome of the experiment shows that the proposed system induces a greater user satisfaction and in turn, demonstrates users’ desirability of personalized recommender systems. The remaining part of this paper is organized with the background study and literature review presented in the next section, methodology of the proposed system in the section after, and an experimental study in the section following that. Also, the section after presents an evaluation of the proposed system while conclusion and area of future research are presented in the conclusion.

Background Study and Related Works The Lexile framework is a scientific approach for measuring the difficulty levels and reading abilities of books and learners respectively (Schnick & Knickelbine, 2004). As the most widely adopted reading measure in use today, Lexile offers a scientific approach that facilitates learning and instruction by improving interpretation ability and informing educational decisions and instructional strategies (Collen & Hal, 2004). Using Lexile, it is possible to match students with appropriate texts and track their reading ability with common scale. The framework includes a Lexile Measure which represents the difficulty of textual materials and users’ reading ability (MetaMetrics, 2004). The term “reading ability” is used to describe a user’s capability of reading texts. This is measured and analyzed with the aid of certain quantitative factors which include: the number of unfamiliar words in a text; the complexity of sentences in the text; and the overall length of the text. MetaMetrics developed scientific measures of academic achievement and complementary technologies that link assessment results with instruction. It helps learners achieve their goals by providing unique insight of their ability and potential for growth. Lexile Text Measure is a measure of how difficult a book or an article is to comprehend and it is based on two strong predictors which are: word frequency and sentence length. The Lexile text measure is a good starting point in book-selection process however many other factors that affect the relationship between learners and books, such as content, age, and interests of the learners are yet to be fully elucidated (Kerkiri et al., 2007). RS is an Internet tool that helps users to navigate through huge amount of information available over the network and receive information relevant to users’ preferences (Pazzani & Billsus, 2007). RSs are widely accepted as critical tool in sustaining the Internet economy (Shapiro & Varian, 1999). Recent surveys (Gediminas & Alexander, 2010; Robin, 2006)

Copyright © 2014, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.


justify the importance of RSs to Internet community. Content Based Recommender Systems uses CBF approached to analyze items and identify those of particular interest to a user (Billsus & Pazzani, 2000). This approach has its roots in Information Retrieval (IR) and thereby employs many of its techniques. Textual based documents are recommended based on a comparison between their content and a user profile (Balabanovic & Shoham, 1997a). CB filtering techniques are best fitting where users’ preferences are predicted by analyzing the relationships between item ratings and corresponding attributes (Balabanovic & Shoham, 1997b). This technique has the ability to learn user’s previous actions about particular content type and make inferences for other types. Despite this vantage, a large set of key attributes is required as smaller sets have being insufficient to learn users’ profiles. Fuzzy logic is a probabilistic logic that deals with approximate reasoning rather than fixed or exact (Novák et al., 1999). Fuzzy Logic System (FLS) is a nonlinear mapping of an input data set to a scalar output data set (Mendel, 1995). Such systems have attracted growing attention and interest in modern IT, pattern recognition, data analysis, and decision making among others (Samuel et al., 2013). In fuzzy set theory, linguistic terms are used to illustrate how membership functions correlate. When linguistic variables are used, their degrees may be managed by specific functions such

Membership Functions. Triangular membership function is a particular case of membership function that shows the degree of membership of each class of linguistic term and it is often viewed as possibility distribution (Ojokoh et al., 2012). Figure 1 represents a typical membership function of input and output variables. Recent trends in e-learning recommender systems show that most of the researchers use data mining approach and information retrieval techniques as recommendation strategies (Kerkiri et al., 2007; Liang et al., 2006; Zaiane, 2002). Hence, RSs have been successfully applied in many domains. In Schafer et al. (1999), five stages of personalization were followed to recommend items that accurately match users’ preferences while Mooney & Roy (2000) describes a content-based book recommending system that utilizes information extraction and a machine-learning algorithm for text categorization. Most researches rely on key-based search or browsing through proceedings of top conferences and journals to find related works. As a result, Joonseok et al. (2002) proposed a personalized academic research paper recommendation system, which recommends articles that are found related to researchers’ topic using CF methods. Zaiane (2002) proposed the use of web mining techniques to build agents that could recommend online learning activities or shortcuts in a course website based on learners’ access histories to improve course navigation as well as assist with the online learning process.

Figure 1. Membership function of input and output variables



Tang et al. (2003) proposed an evolving web-based learning system for searching relevant personalized content on the web and adapt the content learner’s activities. The research applied Clustering Technique on previously accumulated ratings of learners and then CF to calculate learners’ similarities for content recommendation. Hence, learners do not interact directly with the open Web. Chen et al. (2005) proposed a personalized e-learning system based on Item Repository Theory. The experiment shows that the system can precisely provide personalized course material recommendations based on learners’ abilities and accelerate learners’ learning efficiency and effectiveness Soonthornphisaj et al. (2006) applied CF approach to predict the most suitable documents for learners, however the approach predicts new learning materials were recommended with a high degree of similarity. Liang et al. (2006) applies knowledge discovery technique and a hybrid of CBF and CF to make personalized recommendations for courseware selection. The algorithm used was able to reflect users’ interests with high efficiency but hybridized RSs suffer from portfolio effect characterized by data sparsity and cold start problems (Omisore et al., 2013). Kerkiri et al. (2007) proposed a framework that exploits the description and reputation of metadata to recommend personalized learning resources to learners. An experimental study of the model showed that the use of reputation metadata augmented the learner’s satisfaction by retrieving materials that were evaluated positively. Liu & shih (2007) also designed a material recommendation system based on association rule mining and collaborative filtering. The study applied LDAP and JAXB techniques to reduce the overload of search engine and the complexity of content parsing. Khribi et al. (2008) proposed an automatic recommender based on learners’ recent navigation histories by exploiting similarities and dissimilarities among users’ preferences and the contents of their learning resources. A

multi-technique approach that combines web usage mining together with CBF and CF was developed to compute relevant links to make recommendations for an active user. Tai et al. (2008) proposed e-learning course recommendation based on artificial neural network (ANN) and data mining techniques. ANN is used to classify learners based on groups of similar interests so that learners can obtain course recommendations from a group’s opinion. Leyla and Olfa (2010) presented the implementation of a hybrid recommender system that personalizes user’s experience on a real online learning repository. The repository contains educational content of courses, lectures, multimedia resources. The RS is driven by content-based (domain ontology model) and rule-based (learner’s interest based and cluster-based) recommendation techniques. The hybrid model proposed by this study provides users with influenced recommendations by retrieving and ranking of items based on weights. An experimental study was carried out on HyperManyMedia semantic search engine at Western Kentucky University, top-n-Recall and Top-n-Precision were used to measure effectiveness of re-ranking based on the semantic profile of learners. The results demonstrated the effects of personalization during recommendation process. Chhavi and Sanjay (2012) proposed a book RS that gives diverse recommendation by combining users’ choices with similar and other users’ choices. The RS recommends items that changes with time using CBF. A new dimension, called temporal dimension, was embedded with a counter to get update on certain time intervals and thereby improves the recommendation process. This study presents a web-based personalized system that utilizes content-based filtering and fuzzy matching technique to recommend books for learners in a digital library environment. The effect of cold-start problem is assuaged by adding an additional component that considers all new users as a beginner and adjusts their reading level upon subsequent interactions.



Architecture of the Recommender System The architecture of the web based personalized recommender system is presented in Figure 2. The architecture comprises of interface agent (IA), personalized agent (PA), and database system (DBS). Each component of the architecture provides some basic functions towards the success of the recommendation process.

Interface Agent The Interface Agent serves as the communication channel between the learners (users) and the RS. Interactions between users and systems include: profile registration and updating, displaying recommendations, and feedbacks in form of comments.

Personalized Agent The personalized agent is the core of the proposed system. It has three major components which are the Cold Start Engine, the Content Based Inference Engine, and the Fuzzy Matching Engine.

Cold Start Engine One of the major challenges faced by RSs built, singly, on CB filtering technique is Cold Start, inability of the RS to recommend products for new users. In this research, the effects of cold start are assuaged by introducing cold start engine which recommends books for every new user that accesses the system. Literature show that RSs built on CBF has an inability to make recommendations for new users. This is because such user has no history with the system and CBF technique basically works by taking two forms of variables in order to give toothsome recommendation. These variables are user’s history, and if necessary, inputs supplied during recommendation process (Ghauth & Abdullah, 2010; Mooney & Roy 2000). Assuaging cold-start effects is a major improvement and a contribution of this study to RSs because earlier studies do not have means of recommending items to new users due to unavailability of information. Moreover, in this study, the cold start engine considers new users as Beginner and recommends textual materials at moderate difficulty level to them. Once a first-timer encounters the system, two variables: Lexile Book Measure (LBM) and Lexile Learner Measure (LLM); are set to pre-defined default values. The first

Figure 2. Architecture of the proposed recommender system



recommendation made for any user is termed as initial recommendation while the initial LBM and LLM values for the learner and materials are taken as the pre-defined values. An LBM value dictates the level of books that can be recommended for a user at anytime, but the default LBM is Moderate whose range varies between 1.84 and 2.74. Also, the default LLM for new users is 1.84. LLMs are processed on a Likert scale such that the Lexile Measure of a learner is always correlated to a set of books in the database. The LLM of a user is stored in his/her profile and updated upon subsequent with the proposed RS.

Content Based Inference Engine Content based inference engine plays a major role while recommending books for existing learners. The personalized RS presented in this research suitably determines the LLM of a learner and recommends books found within a Lexile Range (LR) of the learner. LR is the reading comprehension “sweet spot” of a learner such that books recommended at the range is slightly above or below his LLM. A major feature of the content based inference engine is its ability to determine the LLM of a learner.

The LLM of a learner is determined by inquiring learners about the difficulty level of books recommended to him by the system. During the interaction, learners are required to give their feedbacks based on three factors, which are used to determine the new LLM of the learner. These factors are the: content of the book, structure of sentences in the book, and text structure of the book. These factors are assessed based on the following attributes: 1. Content of the book: Sight vocabulary, Meaning vocabulary, and Coherence; 2. Structure of sentences: Length of the sentences, Level of sentential embedded clauses, Punctuation standards, and Sentence constructs; and 3. Text structure: Text organization and Text elaboration. The attributes highlighted makes it much simpler to express the difficulty level of a book. The effects of each attribute are obtained from the learner. In order to do this, the system interacts with learners via the designed interface presented in Figure 3. The interface consists of well-structured qualitative questions rounded on the factors that influence the readability of a

Figure 3. User interface for interacting with learner(s)



book. Some sets of linguistic terms, presented in Table 1, enable learners to express their opinions regarding the level of ease or difficulty of the recommended book. The linguistic terms and their associated weights are presented in Table 2. The learner’s Targeted Reading Level (TRL) which is used to recommend books is computed by the content based inference engine. The TRL is a LR such that the LLM is within the Lexile Measure of some books. At this level, a learner will find recommended books normal, a bit easier, or a bit harder, i.e. the learner will likely encounter some level of difficulty with the text, but not enough to get frustrated. The net weight of all factors considered is one (1.0) however, each factor has different level of influence on the reading ability of a learner. Table 1 shows how each factor influences learners’ readability and how the quotas of each attribute of the factors are observed. Weights attached to each attribute of the factors, as shown in Table 1, signifies the contribution level of the attribute. Learner’s feedback on an attribute is dependent on the option selected by the learner. The influence of such attribute is calculated using:

Where W(Ai)is the score weight of ith attribute in jth factor, LW(Ai) is the weight of the linguistic term chosen by a learner for attribute Ai, AWi is the weight that signifies the attribute’s contribution, and Fj is the weight that signifies how jth factor influences the reading ability of the learner. The feedbacks are used to compute the learner’s LLM at each session, this is done as:

LWA * AWi i W (Ai ) = Fj

Where LLMAve is the TRL of a learner, and n is the number of all (previous and current) sessions

(1)

n

LLM o =∑

W (Ai ) n

j =1

(2)

Where LLMo is the Lexile Learner Measure computed during any session, and n is the number of factors used. Previous LLMs are the LM computed in preceding interactions and stored numerically in the database while current LLM is computed during the interaction.

LLM Ave

∑ =

n

LLM i

i =1

n

(3)

Table 1. Factors and attributes considered and the corresponding linguistic terms and weights S/N

Factor

Factors Weight

1 2

Linguistic Terms

Sight Vocabulary

0.20

Low/ Medium /High

Meaning Vocabulary

0.55

Easy/Moderate/Hard

Coherence Level

0.25

Bad/Average/Good

4

Length of Sentence

0.20

Short/Moderate/Long

Level of Embedded Clauses

0.35

Low/Medium/High

Punctuation Standards

0.15

Low/ Medium /High

Sentence Constructs

0.30

Bad/Average/Good

Organization

0.40

Bad/ Average/ Good

Elaboration

0.60

Low/ Medium /High

6

Structure of Sentence

0.40

Attributes Weight

3 5

Content of Book

Attribute

0.45

7 8 9

Text structure

0.15



Table 2. Linguistic terms and associated weights S/N

Linguistic Terms

Associated Weights

1

Bad/ Easy/Low/ Short

0.15

2

Average/ Medium/ Moderate

0.58

3

Good/ Hard/High/Long

1.00

Table 3. Book classifications using normalized difficulty level Normalized Difficulty Level

Class

0.00 – 0.91

Very Easy

0.92 – 1.83

Easy

1.84 – 2.74

Moderate

2.75 – 3.66

Hard

3.67 – 4.57

Very Hard

The Fuzzy Matching Engine The fuzzy matching engine makes use of fuzzy matching technique to recommend books for learners. It establishes matches between the LBM of stored books and LLM of a learner. The LBM which is the difficulty or grade level of a book is computed using the Flesch-Kincaid formula (Powers et al., 1998). Flesch-Kincaid is used to compute the average number of words per sentence, and the average number of syllables per word in a given book. The equation is given as:  Words  no  Diff . Lev. = 0.39 *  +  Sentenceno   Syllable   no   11.0 *  −15.59  Wordsno 

learners are between 0.00 and 4.57 therefore, the difficulty level of a book is normalized for proper comparison using equation 5.The books in the database are classified for all levels of learners using the Normalized Difficulty Level as shown in Table 3. Normalized Diff .Lev. = Difficulty Level * 4.57 12.0

(5)

Finally, fuzzy rules, made up of If Then constructs, are used to match the current LLM against each class of books. A class of books that correlates the current LLM is recommended for learners.

Database (4)

where Wordsnois the number of words in a randomly selected sentence, Sentencesno is the number of sentences in the book, Syllableno is the average number of syllables per word and Diff. Lev is the derived difficulty level of a material. The LBM is a numeric value between 1 and 12. However, analysis on all combinations of the model shows that the possible LLMs of

The database stores both structured and unstructured data about the problem domain and serves as a repository for operational data is processed recommendation processes. Structured data includes learners’ profile, books’ profile, and uploaded books that are to be recommended. The unstructured data consists of the users’ response or feedbacks about recommended books. The database adopts the relational model proposed in (Codd, 1970).



Experimental Study and Results In this section we report the results from our initial experiments regarding the application of CBF to analyze a learner’s reading ability and fuzzy matching techniques to recommend suitable electronic books to the learner. First, we present the environmental features upon which the experiment is performed and lastly, the results of the experiment are presented.

Experiment Settings The proposed system is implemented on a three-tier architecture with front-end engine, middle engine, and back-end engine. An experimental study was observed on Windows Vista Home Premium using APACHE web server. The system was implemented on a relational database with electronic books that could be recommended. The relational database was hosted on the Internet and profiled with 10000 electronic books from seven major areas of computing. The areas include: Software Engineering, Computer Security, Web Design & Development, Computer Networks, Database, Game Theory, and Hardware. The choice of the back-end is My Structured Query Language (MySQL) Database

Management System; Hypertext Preprocessor (PHP) performs the interactions between learners and the system at the middle end, while HTML and Java scripts are used for determining the structure and behaviour of the interface agent respectively. The experimental study of the proposed system was carried out in two phases; the first study was conducted between August, 2012 and March, 2013 with the systems’ database profiled with 5000 books. The second phase was conducted between May and July, 2013 as an extension to the first phase. This was used to check the effects of database’s size on the proposed model. During this phase, the size of the database was doubled. To perform the experiments, we randomly choose twenty students with half of them currently running their post graduate studies and the other half are final year students at the Federal University of Technology, Akure, Nigeria. The system was implemented with double access types, one for learners and the other for the administrator. We created learners account for each student so that they can use the system to recommend books that match their standards. However, all learners are given a similar set of books on their first encounter with the system.

Figure 4. Books recommended for learner L-01 with LLM 1.84



The first set of ratings gotten from the post graduate students was used as the training set i.e. to train the personalized information filtering agents. The final sets are evaluation results gotten from the undergraduate students and are set aside as the test set so as to evaluate the functionality of the proposed system. In each experiment, the test ratings of the target learner were withheld and compared against the recommendation values produced by the system. These studies were used to evaluate the effectiveness and efficiency of the system. The analysis of the results and evaluations are presented below.

Results A key consideration in the design of the system was to support a natural interactive retrieval process. After the login exercise, the user is only required to provide his/her response to the questions in shown on the recommendation page as presented in Figure 3. No more action follows this save a single click for making new book selection. Figure 4 shows the typical recommendation results made to learner L-01 when he encountered the system for the second time. In this result, 15 books were recommended by the system using the previous LLM of the learner.

Figure 5 shows the learner’s timeline following his seventeenth time of using the system within eight months. It shows the relationship between learner’s readability and the difficulty level of recommended books. Also, it is observed from the figure how the learners’ learning rate adjusts with respect to the difficulty level of recommended books. Figure 6 shows a chart for the Reading Abilities of Learner L-01. In the Figure, it is clear that the LLM of the learner dropped from Moderate down to Easy after the first session with the system. This shows that the difficulty levels of electronic books recommended for learner L-01 were above the learner’s reading ability. Also, it is clear from the figure that the learner’s LLM keeps increasing after the seventh session showing that the system recommended books that perfectly suited his learning ability, and in turn increases his LLM

System Evaluation To evaluate the performance of the proposed system, the learners are presented with an evaluation interface. The performance of the system is observed using either the appropriateness of the books recommended by the system or the satisfactory level of a learner based on the service(s)

Figure 5. Timeline of learner’s lexile measure for L-01 within eight months of study



Figure 6. Learner L-01 reading ability chart

provided by the personalized system. The mean responses of 10learners are retrieved as shown in Table 4, and analyzed using Paired-Sample T test. Learners are mandated to evaluate the system after each session using the interface presented in Figure 7. The Paired-Sample T test is used to check if the effects of the two variables are not different in order to estimate the efficiency of the system using the variables. The Paired Samples Test Statistics are computed as follows:

∑ Sample Mean (x ) =

n

PSD

i =1

n

13.0 = 1.30 10

∑ (x −µ) Sample Variance =

=

2

n

=

156.62 = 15.62 10

Table 4. Learner’s evaluation and ratings Learner Id

APP

SAT

PSD (APP – SAT)

L-01

98.5

95.0

3.5

L-02

94.5

93.0

1.5

L-03

95.5

92.5

3.0

L -04

96.0

95.5

0.5

L -05

90.5

85.5

5.0

L -06

87.0

83.0

4.0

L -07

92.5

95.0

-2.5

L -08

88.5

96.5

-8.0

L -09

96.5

92.0

4.5

L -10

90.0

88.5

1.5



Figure 7. System evaluation form

Sample Standard Deviation (σ ) = SampleVariance =

15.62 = 3.95

where APP = Appropriateness; SAT = Satisfaction; and PSD = Paired Sample Difference. Therefore, the Confidence Interval for Mean (CIM) is determined by Equation 6.  2  CIM =x ∓ t (σ ) / n   

(6)

CIM = 1.30 ∓t (1.56) The 0.95-quantile of a t-variate with nine degrees of freedom is 1.812. Therefore, at 95% confidence interval, the CIM is given as follows: CIM t =0.95 =(−1.52745, 4.12745) Since the lower -1.52745 and upper 4.12745 bounds includes zero, the two variables are not different. Therefore, we estimate the efficiency of the system in terms of the Cumulative Average of the two variables as:

∑ Efficiency =

n

x

i =1 i

n

∑ +

n

y

i =1 i

n

2

* 100% (7)

Efficiency =

92.25 + 91.65 *100% 2 = 91.95% ~ 92%

(183.9 / 2) * 100%

The result of evaluation shows that the proposed system is 92.3% efficient in recommending books that are apposite to learners. Hence, it is clear that the proposed system is suitable for recommending books for users irrespective of user’s experience with the system or the user’s level of understanding materials.

CONCLUSION In this study, we proposed a personalized recommender system for learners of digital libraries. The system helps users search for suitable books that are best bet for their learning level. The recommendation procedures take the difficulty level of books and learner’s abilities as input, and then recommend appropriate books. Recommendation is done with content-based filtering and fuzzy matching techniques. The content-based filtering is used to determine learner’s current ability based on his/her recommendation history while a fuzzy matching technique of If-Else constructs is used to correlate the current learner’s ability with the difficulty level of books in the database. Furthermore, the general cold-start challenge faced by content-based recommender systems is taken care of by introducing a predefined starting point for new learners; such recommendations might not be appropriate for such new learners however, experimental study shows that the proposed system was able to recommend suitable books after first n-sessions. The system studies the learners’ rate in those n-sessions and adjusts its recommendation to precisely suit learners’ reading ability with respect to the difficulty level of books. The



experimental results show that proposed system is 92.3% efficient in recommending books to learners. This attests that the introduction of Cold Start Engine in the RS minimized the limitations of CBF and at the same time produced quality recommendations. Researchers in the field of opinion mining and recommender systems use several different measures for the quality of recommendations produced. Coverage metrics were used to evaluate the number of items for which a recommender system could provide recommendations. In many systems, coverage decreases as a function of accuracy i.e. the system can produce fewer accurate recommendations or more inaccurate ones depending on the number of candidate sets found in the database. Lastly, we should point out that these experiments tested the quality of the materials recommended by the web based system and not the performance or economics of such a system. The content based approach employed by the study was based on a series of user questions (see Figure 3) but there were no provisions for dealing with users’ frustration during recommendation process. Literature has pointed this as a major drawback of such RSs but could not be observed under our controlled experiments. We thereby recommend this as a research direction in future works.

Balabanovic, M., & Shoham, Y. (1997b). An adaptive web page recommendation system. Journal of the American Society for Information Science American Society for Information Science, 46(2).

REFERENCES

Hsu, M.-H. (2008). A personalized English learning recommender system for ESL students. Expert Systems with Applications, 34, 683–688. doi:10.1016/j. eswa.2006.10.004

Adomavicius, G., & Tuzhilin, A. (2010). Towards the next generation of recommender systems: A survey of the state-of-the-art and possible extensions (p. 55455). Minneapolis, MN: Carlson School of Management, University of Minnesota. Alexander, P. A., Kulikowich, J. M., & Jetton, T. L. (1994). The role of subject-matter knowledge and interest in the processing of linear and nonlinear texts. Review of Educational Research, 64, 201–252. doi:10.3102/00346543064002201 Balabanovic, M., & Shoham, Y. (1997a). Fab: Content-based, collaborative recommendation. Communications of the ACM, 40(3), 66–72. doi:10.1145/245108.245124

Billsus, D., & Pazzani, M. (2000). User modeling for adaptive news access. User-modeling and useradapted interaction. Communications of the ACM, 10(2), 147–180. Burke, R. (2006). Hybrid recommender systems: Survey and experiments. Fullerton: Department of Information Systems and Decision Sciences California State University. Chen, C. M., Lee, H. M., & Chen, Y. H. (2005). Personalized e-learning system using item repository theory. Computers & Education, 44(3), 237–255. doi:10.1016/j.compedu.2004.01.006 Chhavi, R., & Sanjay, K. (2012). Building a book recommender system using time based content filtering. WSEAS Transactions on Computers, 11(2). Codd, E. F. (1970). Relational model of data for large shared data banks. Communications of the ACM, 13(6), 377–387. doi:10.1145/362384.362685 Collen, L., & Hal, B. (2004). The Lexile framework as an approach for reading measurement and success. A white paper from The Lexile framework for reading. Retrieved March 6, 2013, from http:// www.lexile.com Ghauth, K. I., & Abdullah, N. A. (2010a). Learning materials recommendation using good learners’ ratings and content-based filtering. Educational Technology Research and Development. doi:10.1007/ s11423-010-9155-4

Kerkiri, T., Manitsaris, A., & Mavridou, A. (2007). Reputation metadata for recommending personalized e-learning resources. In Proceedings of the Second International Workshop on Semantic Media Adaptation and Personalization, Uxbridge, UK (pp. 110-115). Khribi, M., Jemni, M., & Nasraoui, O. (2008). Automatic recommendations for e-learning personalization based on web usage mining techniques and information retrieval. In Proceedings of the Eight IEEE International Conference on Advanced Learning Technologies (ICALT’08), Santander, Spain (pp. 241-245).



Lee, J., Lee, K., & Kim, J. G. (2002). Personalized academic research paper recommendation system. Atlanta, GA: Georgia Institute of Technology. Liang, G., Weining, K., & Junzhou, L. (2006). Courseware recommendation in e-learning system. Advances in Web Based Learning (ICWL2006) (pp. 10–24). Berlin, Germany: Springer. Liu, F., & Shih, B. (2007). Learning activity-based e-learning material recommendation system. In Proceedings of the Ninth IEEE International Symposium on Multimedia Workshops 2007 (ISMW’07), Beijing, China (pp. 343-348). Melville, P., Mooney, R. J., & Ramadass, N. (2002). Content boosted collaborative filtering for improved recommendations. In Proceedings of the Eighteenth National Conference on Artificial Intelligence (AAAI02) (pp. 187–192). Mendel, J. (1995). Fuzzy logic systems for engineering. A Tutorial. Proceedings of the IEEE, 83(3), 345–377. doi:10.1109/5.364485 MetaMetrics Inc. (2004). Lexiles: A system for measuring reader ability and text difficulty, scholastic reading counts (pp. 175–186). New York, NY: Scholastic Inc. Mooney, R. J., & Roy, L. (2000). Content-based book recommending using learning for text categorization. In Proceedings of Digital Libraries, San Antonio, TX (pp. 195–204). Novák, V., Perfilieva, I., & Močkoř, J. (1999). Mathematical principles of fuzzy logic. Dodrecht, Netherlands: Kluwer Academic. doi:10.1007/9781-4615-5217-8 Ojokoh, B. A., Omisore, M. O., Samuel, O. W., & Ogunniyi, T. O. (2012). A fuzzy logic based personalized recommender system. International Journal of Computer Science, Information Technology, &. Security, 2(5), 1008–1015. Omisore, M. O., Samuel, O. W., & Ogunniyi, T. O. (2013). A multi-technique approach for recommender systems. International Journal of Advanced Trends in Computer Science and Engineering, 2(5), 112–118. Pazzani, M. J., & Billsus, D. (2007). Lecture Notes in Computer Science: Vol. 4321. Content-based recommendation systems. The Adaptive Web (pp. 325–341). Springer. Powers, R. D., Sumner, W. A., & Kearl, B. E. (1998). A recalculation of four adult readability formulas. Journal of Educational Psychology, 49(2), 99–105. doi:10.1037/h0043254

Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., & Riedl, J. (1994). GroupLens: An open architecture for collaborative filtering of netnews. Proceedings of ACM 1994 Conference on Computer Supported Cooperative Work, Chapel Hill, NC (pp. 175-186). Resnick, P., & Varian, H. R. (1997). Recommender systems. Communications of the ACM, 40(3), 56–58. doi:10.1145/245108.245121 Samuel, O. W., Omisore, M. O., & Ojokoh, B. A. (2013). A web based decision support system driven by fuzzy logic for the diagnosis of typhoid fever. Expert Systems with Applications, 40, 4164–4171. doi:10.1016/j.eswa.2013.01.030 Sarwar, B. M., Karypis, G., Konstan, J. A., & Riedl, J. (2001). Item-based collaborative filtering recommendation algorithms. In Proceedings of the 10th International World Wide Web Conference (WWW10). Schafer, J. B., Konstan, J., & Riedi, J. (1999). Recommender systems in e-commerce. In Proceedings of the 1st ACM Conference on Electronic Commerce (pp. 158-166). Schnick, T., & Knickelbine, M. (2004). The Lexile framework: An introduction for educators. MetaMetrics. Shapiro, C., & Varian, H. R. (1999). Information rules: A strategic guide to the networked community. Boston, MA: Harvard Business School Press. Soonthornphisaj, N., Rojsattarat, E., & Yim-ngam, S. (2006). Smart e-learning using recommender system. Computational Intelligence (pp. 518–523). Berlin, Germany: Springer. Tai, D. W., Wu, H., & Li, P. (2008). Effective e-learning recommendation system based on self-organizing maps and association mining. The Electronic Library, 26(3), 329–344. doi:10.1108/02640470810879482 Tang, T. Y., & McCalla, G. (2003). Smart recommendation for an evolving e-learning system: Architecture and experiment. International Journal on E-Learning, 4(1), 105–129. Zaiane, O. R. (2002). Building a recommender agent for e-learning systems. In Proceedings of the International Conference on Computers in Education (ICCE’02), Auckland, New Zealand (pp. 55-59). Zhuhadar, L., & Nasraoui, O. (2010). A hybrid recommender system guided by semantic user profiles for search in the e-learning domain. Journal of Emerging Technologies in Web Intelligence, 2(4), 272–281. doi:10.4304/jetwi.2.4.272-281



Omisore Mumini Olatunji is a research student at the Federal University of Technology, Akure, Nigeria. He studied Computer Science at undergraduate level and he is currently about to complete the Masters of Technology Degree in Computer Science in the same institution. He has special interests in Software Engineering, Digital Libraries, and Database Administration. He has worked for more than three years as System Analyst at High Technology Research and Development Group Computer Limited, Nigeria. Samuel Oluwarotimi Williams has B.Sc. degree in Computer Science and he is currently pursuing a Maters of Technology Degree in Computer Science at the Federal University of Technology, Akure, Nigeria. He has over five years’ experience in teaching and research in the field of computing. He has special interest in Computational Intelligence, Cloud Computing, and Programming. He currently works with High Technology Research and Development Group (HTRDG) Computer LTD, Nigeria as a Research Fellow/Software Engineer.
