Understanding user goals by analyzing logged interactions and asking the right questions

People use search engines and recommender systems for various information needs every day. The task that these systems solve is to provide users with the right information. This means that these systems must show information that is relevant to the submitted query, is up-to-date and satisfies a user’s preferences. In this thesis, we focus on the last aspect: personalization. The common way to make a result personalized consists of two steps: 1. infer a user’s profile; and 2. provide information that is relevant to people with this profile. There are two types of signal about users that are used to generate their profile: explicit and implicit. To collect explicit signals, users are encouraged to take the initiative and explicitly provide information about their preferences. Implicit signals are signals from user interaction with the system. Moreover, a user profile consists of two components: 1. long-term user preferences; and 2. short-term user preferences. Long-term preferences are preferences that represent lasting user interests. Short-term preferences are preferences that are present only in the current session. In this thesis, we investigate how to infer a user profile from her interactions with an interactive system and how to ask users informative and interpretable questions. In particular, in Chapter 2 we infer long-term user preferences using implicit signals. Then, in Chapter 3 we propose a new method for creating informative questionnaires consisting of pairwise preference questions. In Chapter 4 we propose a reinforcement learning method that combines implicit and explicit signals in order to select elements to display on the result page and to increase user satisfaction with the session. In Chapter 5 we look deeper into the problem of creating preference questionnaires: we would like the questionnaire to be not only informative but also interpretable. To this end, we need a disentangled representation of the items and consequently a metric of disentanglement. That is why in Chapter 5 we analyze existing metrics of disentanglement and propose a new one. Finally, in Chapter 6 we conclude the thesis and discuss limitations and future directions