الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 146 |  |  |
| | | from | 146 | | |
Abstract
The world wide web (WWW) is becoming the most accessed source for searching for information and performing day-to-day activities. It is also becoming an active medium for conducting business. With this proliferation of the internet, the amount of information and products available through it is increasing exponentially. The amount of information available made the customization of content and the recommendation of products of crucial importance.
All these challenges motivated the introduction of web recommender systems as a means for representing user preferences and recommending suitable objects. There are two approaches to recommender systems, memory-based and model-based methods. Memory-based methods store all ratings of users to generalize from them, while model-based methods develop a model of user behavior.
In this thesis, we are proposing a framework for the next page prediction that uses techniques from both memory-based and model-based recommender systems. The system builds a user-item matrix representing user preferences. Clustering is then applied to this matrix to group users with similar preferences together. The clustering results are then used to group the frequent access patterns mined from server logs into groups corresponding to clusters.
When a new user accesses the website, he is matched to his cluster, or the nearest cluster if he/she is an unknown user. To make a prediction the set of patterns assigned to the user’s cluster are searched for matching patterns. We suggested three different representations of the user-item matrix.
The first representation of user preferences is a user-page matrix showing the average time spent by each user on each page. We tested this approach on three different datasets, namely, the logs from the NASA, University of Saskatchewan, and Ain Shams University web servers. The clustering showed an average reduction in the prediction time by 22.2%, 44.4% and 69.8 % for the three datasets respectively. The approach also increased the overall prediction accuracy by 0.5%, 0.4% and 12.8% respectively.
We next suggested the introduction of semantic information to the process. A user-concept matrix was suggested to represent the set of concepts the user is interested in as extracted from the text of the pages he/she visited. We tested this approach on the Ain Shams University server logs. The introduction of semantic information offered further improvement in prediction accuracy by 33% without affecting the prediction time.
Finally, we suggested two approaches to combine the previous two methods. For the first, the average time spent on a page was used to adjust concept counts. This approach improved the accuracy even further by 47.3%. The second approach used an updated distance matrix combined from the two matrices obtained in the first two methods to represent the distances between users. This approach improved the prediction accuracy by 54.3%
The chapters of the thesis are organized as follows:
Chapter 1: Introduction
This chapter offers an introduction to the field as well as the motivation behind the work. The chapter also lists the objectives of the proposed research.
Chapter 2: Background
This chapter gives an overview of the techniques used in developing recommender systems. The chapter explores the two main categories of recommender systems, namely, memory-based and model-based systems. The chapter also presents a comparison between the two techniques for recommender systems. The concepts behind the semantic web as well as its organization are also discussed in this chapter. The chapter explores the importance of semantic information as a dimension added to the information available on the web.
Chapter 3: Related Work
The chapter explores the research done in the fields of recommender systems and semantic web mining. For recommender systems, the chapter presents the different techniques used to overcome the shortcomings of these systems. For semantic web mining, the focus is how semantic information can be integrated into recommendation systems specially web mining algorithms.
Chapter 4: Hybrid Framework for Next Page Prediction
This chapter introduces the hybrid recommender system we are proposing that incorporates memory-based and model-based techniques. Here clustering of user-item matrix was suggested for focusing the search for the user’s next page on relevant patterns. The chapter also explains the experiments conducted to test the proposed approach along with the obtained results.
Chapter 5: Semantic Data for Improved Prediction
In this chapter semantic information extracted from the text of web pages was suggested to be added to the proposed framework. Also a decision fusion approach from both usage and semantic data was introduced. The experiments conducted to test the proposed approaches were presented and compared to traditional approaches.
Chapter 6: Conclusions and Future Work
The final conclusions of our work are presented in this chapter. Also, this chapter presents the possible directions for future research.