Research On Key Technologies For Context-aware Recommender Systems

Context-Aware recommender systems (CARS), aiming to further improve performance accuracy and user satisfaction by fully utilizing contextual information (such as time, location, weather, activity, device, companion and mood), have recently become one of the hottest topics in the domain of recommender systems. Recent years have seen a growing interest and development in CARS not only because of its scientific significance but also its relative novelty and potential applications, which is closely correlated with its two major characteristics and inherent advantages:ubiquitous computing and personalization. Ubiquitous computing means that the access to information and computing resources can occur "at any time, in any place and by any way", while personalization aims at overcoming the problem of "information overload" to free people from massive redundant information. This kind of learning is based on the following theories:personalization technology and recommender systems, ubiquitous computing and context-aware computing, user modeling and user behavioral analysis, cognitive psychology and human decision-making, cloud computing and big data analysis, data mining and machine learning. Nowadays, this area has found many potential applied fields, including e-commence, information retrieval, multimedia content dilivery, mobile applications, tourism and intelligent home living/transportation/office and so on. However, there are still many challenges for this new field.This thesis aims at studying some key problems including implicit contextual user preferences elictition, data sparsity problem of context-aware recommender systems, mood-based recommendation, and how to introduce socia network information in CARS. Our proposed methods are based on the mathematic and experimental analysis. The main contributions of thie thesis are as follows:(1) Implicit contextual user preferences elictition. Inspired by the model of level of processing and the theory of distributed cognition, a cognitive psychology-based approach to contextual user preferences elicitation for mobile network services is proposed. It uses a six-tuple to describe the data structure of user preferences information, analyzes the level of processing of services for users to elicit context-free user preferences, then identifies valid types of contexts as well as their influences on user preferences, and finally elicits converged user preferences. Experimental comparisons of this approach against some baseline methods with a synthetic data set have been conducted, showing improvements in performance.(2) HOSVD-based context-aware recommender system (TensorCARS). After incorporating contextual information into RS, CARS has to face to the more serious multidimensional data sparsity problem. This thesis presents a new approach called TensorCARS to alleviate such problem by applying the higher order singular value decomposition (HOSVD) technique. It first constructs an N-order tensor to represent multidimensional contextual user preferences and decompose it into (N-2)3-order tensors according to different types of context (such as time, location). Then it uses HOSVD to discover latent associations among these different3-dimensional objects: users, items and unidimensional context (such as user-item-time), and predict unknown unidimensional contextual user preferences. Finally, it calculates every contextual influence coefficient that each type of context factor influences user preferences, and constructs a new N-order tensor using weighted linearization method in order to provide recommendations. We perform experimental comparison of this approach against the other baseline approaches in a simulated personalized mobile services environment. The experimental results suggest a significant improvement in performance.(3) Mood-based Recommendation. Recently mood has proved to be an important contextual feature in recommender systems by some studies. This thesis proposes some new approaches to mood-based recommender systems. We first propose a new mood-based user-based collaborative filtering (CF) which detects user preferences to each emotion through introducing a new mood-based rating volatility factor in the traditional user-based CF. On this basis, we then propose two hybrid CF approaches using multiplestep nearest neighbors search and predicted ratings fusion strategies respectively. Finally, we propose another HOSVD-based approach to mine the latent relationships between users, movies and moods so as to solve the "user-item-mood" data sparsity problem. We perform experimental comparisons of these new approaches against some baselines on the Moviepilot dataset. The experimental results illustrate that the new mood-based rating volatility factor affects mood-based recommendations; the mood-based hybrid CF approaches perform better than the traditional user-based CF that takes no consideration of mood, but the pure mood-based user-based ones does not; the HOSVD-based approach has the best performance.(4) Social network-based context-aware recommender systems. This thesis provides an initial study on introducing social network information into CARS. It first gives a definition of the social network-based context-aware recommender systems (SCRS) through combining CARS and social recommender systems (SRS). Then, this thesis proposes a heuristic approach called SCMSR to find nearest neighbors for each mobile user by using both social network information and those given contextual user preferneces, and finally predict unknown contextual user preferences as well as generate contextual recommendations. We evaluate the SCMSR approach on a real-world data. The experimental results show that SCMSR performs better than some classical baseline methods (including user-based CF, SVD, contextual pre-filtering) as well as the pure social recommendation and context-aware recommendation approaches.