Colombus: providing personalized recommendations for drifting user interests

QA76 Computer software;T Technology (General)

Psarras, Ioannis ; Jose, Joemon

University:University of Glasgow

Department:School of Computing Science

关键词: personalization, recommendations, user interests, concept drift, colombus, adaptive;   

Others  :  http://theses.gla.ac.uk/651/1/2009psarrasmscr.pdf

来源: University of Glasgow

【 摘 要 】

The query formulationg process if often a problematic activity due to the cognitive load that it imposes to users. This issue is further amplified by the uncertainty of searchers with regards to their searching needs and their lack of training on effective searching techniques. Also, given the tremendous growth of the world wide web, the amount of imformation users find during their daily search episodes is often overwhelming. Unfortunatelly, web search engines do not follow the trends and advancements in this area, while real personalization features have yet to appear. As a result, keeping up-to-date with recent information about our personal interests is a time-consuming task. Also, often these information requirements change by sliding into new topics. In this case, the rate of change can be sudden and abrupt, or more gradual. Taking into account all these aspects, we believe that an information assistant, a profile-aware tool capable of adapting to users’ evolving needs and aiding them to keep track of their personal data, can greatly help them in this endeavor. Information gathering from a combination of explicit and implicit feedback could allow such systems to detect their search requirements and present additional information, with the least possible effort from them. In this paper, we describe the design, development and evaluation of Colombus, a system aiming to meet individual needs of the searchers. The system’s goal is to pro-actively fetch and present relevant, high quality documents on regular basis. Based entirely on implicit feedback gathering, our system concentrates on detecting drifts in user interests and accomodate them effectively in their profiles with no additional interaction from their side. Current methodologies in information retrieval do not support the evaluation of such systems and techniques. Lab-based experiments can be carried out in large batches but their accuracy often questione. On the other hand, user studies are much more accurate, but setting up a user base for large-scale experiments is often not feasible. We have designed a hybrid evaluation methodology that combines large sets of lab experiments based on searcher simulations together with user experiments, where fifteen searchers used the system regularly for 15 days. At the first stage, the simulation experiments were aiming attuning Colombus, while the various component evaluation and results gathering was carried out at the second stage, throughout the user study. A baseline system was also employed in order to make a direct comparison of Colombus against a current web search engine. The evaluation results illustrate that the Personalized Information Assistant is effective in capturing and satisfying users’ evolving information needs and providing additional information on their behalf.