Can web-based recommendation systems afford deep models: a context-based approach for efficient model-based reasoning

Web-based product and service recommendation systems have become ever popular on-line business practice with increasing emphasis on modeling customer needs and providing them with targeted or personalized service solutions in real-time interaction. Almost all the commercial web service systems adopt some kind of simple customer segmentation models and shallow pattern matching or rule-based techniques for high performance. The models built based on these techniques though very efficient have a fundamental limitation in their ability to capture and explain the reasoning in the process of determining and selecting appropriate services or products. However, using deep models (e.g. semantic networks), though desirable for their expressive power, may require significantly more computational resources (e.g. time) for reasoning. This can compromise the system performance. This paper reports on a new approach that represents and uses contextual information in semantic net-based models to constrain and prune potentially very large search space, which results in more efficient reasoning and much improved performance in terms of speed and selectivity as evidenced by the evaluation results.