Understanding and using contextual information in recommender systems

With the rapid development of information technology, the availability of huge amounts of online information makes retrieval a hard task for the average user. Recommender systems (RS) have been employed across several domains to ease this so-called “information overload” problem since the mid-1990s. Recently, context-aware recommender systems (CARS), aiming to further improve the performance accuracy and user satisfaction by fully utilizing contextual information (such as time, location, mood and company) into RS, has become one of the hottest topics [1]. Although a certain progress has been made, CARS still has to face to many challenges. This thesis investigates some key problems in CARS and then proposes some tested and untested approaches to mine the latent relationship among users, contextual information and items (such as movies, web pages and mobile services). In this thesis, the first task is how to elicit contextual user preferences implicitly. All of the existing CARS are based on the assumption that there are available explicit contextual user ratings (e.g., “Sam×Avatar×Morning×Home3”). However, it is hard to obtain sufficient contextual user preferences in practice. This thesis proposes a MAUT (multi attribute utility theory)-based approach to implicitly elicit contextual user preferences through analyzing contextual user behaviors. It considers every type of context as an attribute of items, elicit every unidimensional contextual user preferences based on a n ew context-based IF-IDF formula, and finally elicit multidimensional contextual user preferences after identifying different weights of different contexts. We design a personalized mobile services-oriented prototype system as a test bed to elicit contextual user preferences as well as generate contextual recommendations. I perform experimental comparison of this approach against the other baseline approaches, attaining significant improvements. Secondly, how to alleviate the sparsity problem in CARS is a key challenge. The data sparsity exists in any traditional RS. While incorporating contextual information, the problem of sparse in CARS becomes even more serious. I propose a HOSVD-based contextual recommendation approach, called TensorCARS [2]. It first constructs an N-order tensor to represent multidimensional contextual user preferences and decomposes it into (N-2) 3-order tensors according to different contexts, then uses the HOSVD technique to predict unknown unidimensional contextual user preferences, and then calculates every contextual influence coefficient that each context factor influences user preferences, and finally constructs a new N-order tensor using weighted linearization method. I perform experimental comparison using the prototype system, showing TensorCARS can help alleviate the sparsity problem and increase the prediction accuracy. Thirdly, I consider mood as an important context and design two mood-based hybrid collaborative filtering approaches. ACM CAMRa2010 [3] releases two datasets gathered by the Moviepilot and Filmtipset. I participate in the Moviepilot challenge track which addresses the contextual dimension related to a user’s mood. I first propose a new mood-based user-based collaborative filtering approach based on the assumption that users with similar moodpreference patterns may have similar user preferences to items, and then propose two new hybrid ones through fusing the former presented approach and the traditional user-based CF (i.e., a multiple-step KNN similarity fusion and a weighted predicted rating fusion strategy respectively). Both hybrid CF approaches outperform the other user-based CF ones in terms of all three evaluation metrics (i.e., MAP, P@N (N=5, 10) and AUC). Lastly, I focus on bring social network analysis into CARS in order to further improve recommendation accuracy. Intuitively, user needs are not only affected by his interests and environment, but also by many other factors such as social network relationship (due to the content and information contributed or distributed through the Social Web). Now I am designing a general model of social network-based context-aware recommender systems (SCRS) to search nearest neighbors in high dimensional spaces that contain both social networks and contextual user preferences data. I plan to perform experiments on the public Filmtipset dataset.