Exploiting context information in recommendation systems
Pham, Nguyen Tuan Anh
Date of Issue2017-11-06
School of Computer Science and Engineering
With the rapid growth of data in recent years, especially online and user-generated data, the role of recommendation systems becomes more important. Many recommendation methods have been proposed, among which context-aware recommendation systems have received significant interests from researchers due to their superior performance. Different from traditional recommendation systems, which only consider users and items, context-aware recommendation systems exploit additional information that affects users' decisions on items. Context information is very diverse, from explicit information, such as user profiles (e.g., age, gender) or item description (e.g., weight, price), to implicit one, such as item-item dependency or users' requirements. In this thesis, we consider three context-aware recommendation problems, in which the context information has an essential role on improving recommendation accuracy. First, we consider the problem of recommendation in heterogeneous networks. Heterogeneous networks are information networks containing different types of entities (e.g., users, items, locations) and relationships between entities (e.g., a user rates an item). Heterogeneous networks can be found in many domains, such as movies (Netflix), music (Last.fm) and social networks (Facebook). Although many context-aware recommendation systems have been proposed for heterogeneous networks, one common shortcoming of those systems is that they cannot explicitly capture and model the interaction between different types of entities. Characterizing and utilizing this information is crucial to gain a good recommendation accuracy because the way each entity type behaves and interacts with others greatly affects the recommendation results. As a result, in this thesis, we propose a general-purpose recommendation model, which is able to solve general recommendation tasks in heterogeneous networks. Different from existing methods, our model is able to explicitly model the interaction between entity types and automatically learn the strength of those interactions. This not only helps our model achieve better performance than state-of-the-art methods, but also enables us to understand roles of entity types in different recommendation problems. Second, we propose and address a novel problem, namely out-of-town region recommendation. When traveling to new cities or countries, users usually have very limited time to visit places; hence, they tend to visit places in a small region. Therefore, it is more beneficial to recommend a region of point-of-interest (POI) than a list of POIs that locate far away from each other. When users choose a region to visit, one important consideration is whether the POIs in the region are attractive to users as a whole. In particular, users' decisions to visit a POI are also affected by how they are interested in nearby POIs, in addition to the user preference to the POI. Therefore, we propose a general framework for region recommendation in which influences among POIs are taken into consideration. Experiments on real world datasets validate that our proposed model outperforms baseline methods, in which POIs are considered independently. Third, when submitting recommendation requests, users usually have clear requirements or intention, e.g., having dinner or hanging out. To exploit this information, we propose a model for the problem of requirement-aware POI recommendation. Our model is able to precisely understand user intention by applying attention model, a deep learning model that has been used in many natural language processing tasks, so that we can provide better recommendation for users. Moreover, our model can be easily extended to incorporate additional information such as geographical influence. Empirical studies demonstrate our proposals achieve significant improvement over baseline methods. In summary, this thesis focuses on exploiting context information in recommendation. Three context-aware recommendation systems are proposed for some specific problems. The solutions are not limited in these problems. They can be applied in other recommendation tasks without much modification. We also discuss some promising directions to improve and extend our techniques in future work.