Abstract:
Recommender systems assist individual users in ï¬nding the right items in large option space. Absolventen.at, an Austrian job board for graduates, uses such a system for rec- ommending appropriate jobs to applicants. So far, this system has only considered the resume as input for the user proï¬le, which is compared with the available jobs. How- ever, only around half of the registered job seekers ï¬ll out the resume, for the other half no personalized recommendations can be generated. To improve this, the recom- mender system has been enhanced with implicit relevance feedback and the impacts of this approach have been examined in this thesis. Implicit feedback can be captured in an unobtrusive way and allows the system to infer user preferences. Four different user actions for implicit feedback have been identiï¬ed on Absolventen.at, including reading of a job description, bookmarking, applying and searching for jobs. All of them provide different levels of evidence for interest, as an application is a more reliable indicator for interest than just reading a job description, which is taken into account with individual weighting parameters. In addition to that, gradual forgetting factors are used for adapt- ing the proï¬le over time. All of this information is included in the hybrid user proï¬le, which is represented as hyperdimensional vector and calculated by a linear combination of the resume and the preferred jobs. To evaluate the new approach, the preferred jobs of 46 job seekers were compared with the recommendations. The results show that including implicit feedback helps to increase the user coverage, as well as the accuracy of the recommendations.
