Personalized Type-based Facet Ranking For Faceted Search

Abstract

Faceted Search Systems (FSSs) have gained prominence as one of the dominant search approaches in vertical search systems. They provide facets to educate users about the information space and allow them to refine their search query and navigate back and forth between resources on a single results page.

Type-based facets (aka t-facets) help explore the categories associated with the searched objects. In a structured information space, t-facets are usually derived from large hierarchical taxonomies. When the information available in the collection being searched increases, so does the number of associated t-facets. This makes it impractical to display at once, the entire t-facet taxonomy to the user.

To tackle this problem, facet ranking is implemented in the FSS. Ranking methods can take advantage of the information structure, the textual queries issued by the users, or the user logs. However, existing methods neglect both the hierarchical structure of the taxonomies, and hence of the t-facets, as well the user's historical preferences. As a consequence, users obtain a t-facet list that is difficult to read and irrelevant to their interests.

This thesis focuses on the task of personalizing t-facet ranking in precision-oriented FSSs. It investigates to what extent personalizing t-facet ranking, using user historical feedback, can minimize the user effort to reach the intended search target. This work proposes a two-step approach to solve this problem. The first step scores each individual leaf-node t-facet. In the second step, this score is used to re-order and select the sub-tree to present to the user. The final ranked tree reflects the t-facet relevance both to the query and the user profile.

For the first step, three scoring methods are developed: A probabilistic, a Vector Space Model (VSM), and a Deep Neural Network (DNN) scoring method. These methods combine different types of information in different ways to recommend the most relevant t-facets to the user. A common element to all of them is the individual user profile collected from the previous user's ratings in the system. The second step suggests three strategies to utilize this generated score in order to produce the final t-facet sub-tree to the user. The strategies group the relevant t-facets by their parent-nodes and order them by aggregating the t-facet scores from the earlier step.

Whilst evaluation protocols have been developed for the general facet ranking, the problem of personalizing the facet rank, based on user profile, has lagged behind due to the lack of appropriate datasets. To fill this gap, this thesis introduces a framework to reuse and customise existing real-life data collections. The framework outlines the eligibility criteria and the data structure requirements needed for this task. It also details the process to transform the data into a ground-truth dataset. We apply this framework to two existing data collections in the domain of Point-of-Interest (POI) suggestion: TREC-CS 2016 and Yelp Open Dataset.

In order to assess the performance of the proposed approach, the proposed framework combines the widely adopted user-simulation model and metrics proposed in the INEX 2011 Data Centric Faceted Search task. The evaluation approach aims to capture the user s effort required to fulfill their search needs, by using the ranked t-facet tree. Our experiments have found that the proposed DNN based scoring methods significantly minimize the number of actions the user need to perform in order to reach the intended search target. The evaluation results also demonstrated that using different tree construction strategies have a significant impact on the same number of actions metric. We conclude that the proposed personalized approach leads to better t-facet rankings and minimizes user effort.