Document modelling using topics based on meaningful-interesting patterns for information filtering

In information filtering systems (IF) systems, modelling users’ information needs is the key component. The use of patterns-based topic models (PBTMs) yields higher performance than that of single topical terms and phrases for modelling the component. The topic ranking-based topic concept matching (TRbTCM) and semantic-based topic evaluation (SbTE) models are said to boost the performance of IFs by applying topic evaluation on the trained topics. In this study, these two methods of topic assessments were reviewed. A method to discover meaningful and interesting patterns called meaningful-interesting patterns (MIPs) to represent user interests in IFs was developed. Two IF models, namely MIP and topic evaluation for IF (MIPnTE), and MIP and SbTE for IF (MIPnSbTE) were developed using the MIPs to represent user preferences and a document relevance model to determine the relevance of a document in the filtering component. Experiments applying the proposed models including and some existing baseline models to the large benchmark RCV1 dataset were conducted. Results showed that MIPnSbTE outperformed MIPnTE and the existing baseline models such as SbTE, TRbTCM and PBTM, indicating that modelling user preferences on the basis of MIPs with disambiguation yielded higher performance than using the existing models for IF.