Unsupervised Context-Aware User Preference Mining

In pervasive environments, users are situated in rich context and can interact with their surroundings through various services. To improve user experience in such environments, it is essential to find the services that satisfies user preferences in certain context. Thus the suitability of discovered services is highly dependent on how much the context-aware system can understand users’ current context and preferred activities. In this paper, we propose an unsupervised learning solution for mining user preferences from the user’s past context. To cope with the high dimensionality and heterogeneity of context data, we propose a subspace clustering approach that is able to find user preferences identified by different feature sets. The results of our approach are validated by a series of experiments. Pervasive environments are rich in context information and services. Users in such environments need services that suit to their current context and preferred activities. As more and more sensors are deployed to collect context information, the task of interpreting context becomes increasingly challenging (Lim and Dey 2010). Consequently, the usefulness of service discovery (Li, Sehic, and Dustdar 2010; Rasch et al. 2011) or service recommendation (Adomavicius et al. 2005) is largely limited by the ability to understand user preference. In the literature, context attributes used for learning user preferences are chosen by either empirical assumption (Munoz-Organero et al. 2010) or dimension reduction (Krause, Smailagic, and Siewiorek 2006) that renders a small set of features. However, these approaches are infeasible in a broad and ever-growing spectrum of context information. They fail to acknowledge the large variety of features needed to describe different user preferences. Therefore, they often struggle to find services that fulfill user requirements accurately. Context modeling (Baldauf, Dustdar, and Rosenberg 2007; Strang and Linnhoff-Popien 2004) can define context in an extensible and dynamic way. These models are useful when reasoning based on a priori knowledge, but fall short in acquiring implicit knowledge from past context. For identifying user preferences, most previous work applies superCopyright c © 2013, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. vised learning approaches that are targeted at identifying a set of known classes, usually activities (Ferscha et al. 2004). These approaches, however, are also challenged in a servicerich environment. Because of the highly diverse service invocation possibilities, they require a large amount of training data and a long learning process. This paper addresses the aforementioned challenges by modeling and analyzing context and services in a highdimensional data space. Our main contribution is a subspace clustering approach that is specialized to work on highdimensional context data. In the clustering process, we accommodate for different data types, different densities and context attributes of varying importance. Different to most known solutions for learning from context, our solution is an unsupervised learning process that requires minimal a priori knowledge. The output of clustering is a set of user preferences presented as subspace clusters, which can then be used for service discovery. Our experimental results demonstrate that the proposed approach is able to achieve very good clustering results, and in turn, provide highly reliable understanding of users’ preferred context and activities. The paper is organized as follows: We first give a background introduction to modeling services and preferences in high dimensional context spaces. Next, we discuss the challenges of applying subspace clustering to context data, followed by a detailed presentation of our preference mining approach. Afterwards, the proposed approach is evaluated by series of experiments. The paper concludes with discussions and future work.