Motivation and Goals
Personalization received much attention last years. Our possibilities to adapt and personalize has increased since people became part of the Web. Personalization itself mostly involves data mining techniques as proposed in the book on Adaptive web in the chapter on data mining for web personalization. Besides preprocessing and integrating data sources, they mainly discuss patterns discovery based on clustering, association rules mining or sequential pattern mining.
Current evolution of mobile devices indicates our need to be equipped by something smart all the time and everywhere. Smart devices are becoming our accompanies what also brings advantages we had never thought about. These devices are learning and adapting using advanced methods. They know who we are, what we do, where we are and even what we need. Everything should be understood as profitable for the user. Mobile devices could be used as a tool for observation. Even more, talking about pervasive computing, we are able to use every surrounding device to observe the user and his activities. Starting with surveillance cameras through credit cards and ending with Web browsers we are monitored and our actions could be stored. And this is only beginning of what we can do nowadays. Step by step, new an new technological inventions surround us and more options to observe users emerge.
Even if we use advanced techniques to observe user and put him into context of the situation we are not capable to capture all aspects of the particular moment. These aspects could be very useful (e.g. in information retrieval or recommendation). Thus seeking for other options to cover this missing space is important research problem.
To cover missing information on user behavior which we are not able to acquire in common manner (by direct user observation), we need to bring in implicit information acquisition.
In our work we focus on information associated with a user which describes the user and the state of his environment. Basically we recognize two sets of aspects which affect user behavior:
Our goal is to acquire these attributes. In general, user attributes are explicitly expressed by users. Users are usually asked to complete forms where this information is required. Acquired attributes have usually long-term nature. For instance, date of birth never changes, so does not gender. Other attributes such as weight or height are changing rarely or in long periods.
We focused on the context as a combination of user attributes and environment attributes. We analyzed trends and research which has been done in the area of user modeling, information retrieval, recommendation or adaptive systems. This led us to recognition of the potential research interest in this field. We show state-of-art approaches and link them with modern approaches, trends and perspectives stated during panel discussion on context and context-awareness. This emerged in the particular goals which we are focused on.
We proposed a method for context inference which enriches logs of user activity with information which could be relevant. Our aim is not to acquire available information but to estimate context components which are not available directly. We reduce the sparsity of information which leads to better recommendations or even other information retrieval processes. We work with visible and hidden user information. We also consider the availability for specific user.
We recognized that besides easily acquired contextual information (time, location) we need to acquire more complex information on context. The difficulty differs from domain to domain. It means that we can not define exact information type to be visible or hidden but we can split information types by the way they occur in specific domains.
We present evaluation of our method on several datasets. We work with news reading from SME.sk, social network azet.sk and dataset with movie ratings CoMoDa.
We presume that the inferred information correctly enriches existing data. We proved this presumption by the inference of visible context. We compared inferred and real contextual information and calculated relative error, precision and recall. We presume that inferred information improves the prediction of user activity. We repeat the same prediction with and without inferred context and compare the results of both predictions. Our intention was to present context inference as positive contribution to the quality of the data used for further personalization. We experimented with the domain of news, social network and movie ratings.
We discovered that the domain itself is not that important in case of precision. What is more important is the quality of the data before we apply our method for context inference and secondly the context types for particular domains. In other words, it means that very sparse data is less likely to be enriched with correct information. Our method for context inference is basically based on the extrapolation of original information. Very sparse datasets have very poor information on user behavior and since our method is based on user behavior, the quality lowers with dataset sparsity.
Second observation is that we could not say which context could be inferred with higher precision because it changes from domain to domain. It eventually means that particular context type could have different influence on user behavior in different domains. On the other hand, we are able to say what is the influence of a particular context for specific domain when necessary data is provided.
Regarding the recommendation itself we worked with simple recommendation based on predicting user interest based on the user history. Our added value, which we demonstrated in experiment is in aforementioned context inference. It basically means, that we are able to improve the quality of the dataset used for recommendation. We also showed that using context could be almost transparent for the recommendation technique. We are able to incorporate context directly to the user model thus recommendation could work with context as effective as with content (tags associated with user in the user model).
We proposed our method for context inference which follows these tasks. Our method is a contribution in the field of recommender systems. Our method reduces the sparsity in context-based user model for recommender system. We proved that predicting behavior considerably improves information retrieval or recommendation since behavior has profound effects on the information need. We also proved that our approach where we used inferred information outperformed the same recommender system which suffered by missing information, since user model was very sparse. We showed that our method and its precision changes with the domain and context type.
Our approach combines more techniques available in data mining with intention to infer context which is not directly available. We decided not to use standard association rules mining because it seriously suffers from discretized values. We work with continuous values instead of nominal or discretized. We showed how this approach is designed and what is the precision in the domain of news to contribute to the field of context-based recommender systems.
Another interesting fact about our method for context inference is that it uses rules which are assigned to incomplete user models respecting that our behavior is either mainstream, individual or stereotyped.The thesis extended abstract is available in the Bulletin of the ACM Slovakia.