A learning-based model in clickstream-based collaborative filtering.

摘要

Clickstream-based Collaborative Filtering (CCF) is an area that has been well studied in research and widely adopted in industry due to its scalability. The commonly used prediction models for CCF are Markov models, sequential association rules, association rules, and clustering. These models have shown varying performance depending on the data set used and/or the parameters such as number of pages (or items) in recommendation, window size, minimum confidence and support. In order to increase recommendation effectiveness by addressing the trade-off relationship between precision and recall, some studies have combined two or more different models or applied multiorder models. The increase of recommendation effectiveness by these models, however, is at best marginal in that they mainly focus on precision and there is room for improving recall because of their first order (one model type) application. To increase recall while minimizing the loss of precision and therefore to increase overall performance, measured by the F value, we build a sequentially applied model (SAM) applying the individual models in an order determined through a learning process. We evaluate SAM over the individual models with both an objective measure and a subjective measure. For the objective measure, we tested SAM with the log data of several sites and it shows that SAM excels in the performance of recall and the F measure over the individual models, while its precision is comparable to the highest precision from the models. The subjective measure also shows that SAM is preferred to a model with the highest precision.; With the individual models and SAM, we propose also a novel model integration scheme (MIS) where a user selectively chooses a model, either an individual model or SAM, based on the user's utility function. The scheme MIS is open to any current and future optimized individual models so that they can be readily plugged into the scheme. The scheme is embodied into lookup tables in a batch process for efficient real-time recommendation.