Date on Master's Thesis/Doctoral Dissertation

5-2011

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Committee Chair

Nasraoui, Olfa

Author's Keywords

Swarm intelligence; Flock of agent; Clustering; Recommender system; Dynamic clustering; Data mining and visualization

Subject

Swarm intelligence; Web usage mining; Data mining

Abstract

Swarm Intelligence (SI) techniques were inspired by bee swarms, ant colonies, and most recently, bird flocks. Flock-based Swarm Intelligence (FSI) has several unique features, namely decentralized control, collaborative learning, high exploration ability, and inspiration from "dynamic social" behavior. Thus FSI offers a natural choice for modeling dynamic social data and solving problems in such domains. One particular case of dynamic social data is online/web usage data which is rich in information about user activities, interests and choices. This natural analogy between SI and social behavior is the main motivation for the topic of investigation in this dissertation, with a focus on Flock based systems which have not been well investigated for this purpose. More specifically, we investigate the use of flock-based SI to solve two related and challenging problems by developing algorithms that form critical building blocks of intelligent personalized websites, namely, (i) providing a better understanding of the online users and their activities or interests, for example using clustering techniques that can discover the groups that are hidden within the data; and (ii) reducing information overload by providing guidance to the users on websites and services, typically by using web personalization techniques, such as recommender systems. Recommender systems aim to recommend items that will be potentially liked by a user. To support a better understanding of the online user activities, we developed clustering algorithms that address two challenges of mining online usage data: the need for scalability to large data and the need to adapt cluster sing to dynamic data sets. To address the scalability challenge, we developed new clustering algorithms using a hybridization of traditional Flock-based clustering with faster K-Means based partitional clustering algorithms. We tested our algorithms on synthetic data, real VCI Machine Learning repository benchmark data, and a data set consisting of real Web user sessions. Having linear complexity with respect to the number of data records, the resulting algorithms are considerably faster than traditional Flock-based clustering (which has quadratic complexity). Moreover, our experiments demonstrate that scalability was gained without sacrificing quality. To address the challenge of adapting to dynamic data, we developed a dynamic clustering algorithm that can handle the following dynamic properties of online usage data: (1) New data records can be added at any time (example: a new user is added on the site); (2) Existing data records can be removed at any time. For example, an existing user of the site, who no longer subscribes to a service, or who is terminated because of violating policies; (3) New parts of existing records can arrive at any time or old parts of the existing data record can change. The user's record can change as a result of additional activity such as purchasing new products, returning a product, rating new products, or modifying the existing rating of a product. We tested our dynamic clustering algorithm on synthetic dynamic data, and on a data set consisting of real online user ratings for movies. Our algorithm was shown to handle the dynamic nature of data without sacrificing quality compared to a traditional Flock-based clustering algorithm that is re-run from scratch with each change in the data. To support reducing online information overload, we developed a Flock-based recommender system to predict the interests of users, in particular focusing on collaborative filtering or social recommender systems. Our Flock-based recommender algorithm (FlockRecom) iteratively adjusts the position and speed of dynamic flocks of agents, such that each agent represents a user, on a visualization panel. Then it generates the top-n recommendations for a user based on the ratings of the users that are represented by its neighboring agents. Our recommendation system was tested on a real data set consisting of online user ratings for a set of jokes, and compared to traditional user-based Collaborative Filtering (CF). Our results demonstrated that our recommender system starts performing at the same level of quality as traditional CF, and then, with more iterations for exploration, surpasses CF's recommendation quality, in terms of precision and recall. Another unique advantage of our recommendation system compared to traditional CF is its ability to generate more variety or diversity in the set of recommended items. Our contributions advance the state of the art in Flock-based 81 for clustering and making predictions in dynamic Web usage data, and therefore have an impact on improving the quality of online services.

Share

COinS