Copy the page URI to the clipboard
Vu Thanh Tien
(2017).
DOI: https://doi.org/10.21954/ou.ro.0000c5ad
Abstract
The performance of a personalised search system largely depends upon the ability to build user profiles which accurately capture the user's search interests. However, many approaches to user profiling have neglected the dynamic nature of the user's search interests. That is, a user's search interests typically change in response to their interactions with the search system during the search period. Therefore, a profile built for previous searches might not reflect that user's current search interests.
A widely used type of profile represents the topical interests of the user. In these cases, a typical approach is to build a user profile using topics discussed in documents which the user has found relevant, and where the topics are obtained from a human-generated ontology or directory. However, a key limitation of these approaches is that many documents may not contain the topics covered in the ontology. Moreover, the human-generated ontology requires manual effort to determine the correct categories for each document.
In this research, we address these problems by proposing novel techniques for dynamically building user profiles which capture the user's search interests changing over time. Instead of using a human-generated ontology, we use a topic modelling technique (Latent Dirichlet Allocation) for unsupervised extraction of the topics from documents. To dynamically build user profiles, we make two important assumptions. First, that the group of users with whom a user shares a set of common interests may be different depending upon the particular topic of interest. Second, the more recently clicked/relevant documents tell us more about the user's current search interests.
To test these assumptions, we develop and implement dynamic user profiles, and then evaluate them on two search personalisation tasks. Our first chosen task is personalising search results returned by a Web search engine, and the second is the task of personalising query suggestions made by an Intranet search engine. We found that dynamic user profiles can significantly improve the ranking quality over well-established baselines.