A new technique for intelligent web personal recommendation

Abstract

Personal recommendation systems nowadays are very important in web applications because of the available huge volume of information on the World Wide Web, and the necessity to save users’ time, and provide appropriate desired information, knowledge, items, etc. The most popular recommendation systems are collaborative filtering systems, which suffer from certain problems such as cold-start, privacy, user identification, and scalability. In this thesis, we suggest a new method to solve the cold start problem taking into consideration the privacy issue. The method is shown to perform very well in comparison with alternative methods, while having better properties regarding user privacy. The cold start problem covers the situation when recommendation systems have not sufficient information about a new user’s preferences (the user cold start problem), as well as the case of newly added items to the system (the item cold start problem), in which case the system will not be able to provide recommendations. Some systems use users’ demographical data as a basis for generating recommendations in such cases (e.g. the Triadic Aspect method), but this solves only the user cold start problem and enforces user’s privacy. Some systems use users’ ’stereotypes’ to generate recommendations, but stereotypes often do not reflect the actual preferences of individual users. While some other systems use user’s ’filterbots’ by injecting pseudo users or bots into the system and consider these as existing ones, but this leads to poor accuracy. We propose the active node method, that uses previous and recent users’ browsing targets and browsing patterns to infer preferences and generate recommendations (node recommendations, in which a single suggestion is given, and batch recommendations, in which a set of possible target nodes are shown to the user at once). We compare the active node method with three alternative methods (Triadic Aspect Method, Naïve Filterbots Method, and MediaScout Stereotype Method), and we used a dataset collected from online web news to generate recommendations based on our method and based on the three alternative methods. We calculated the levels of novelty, coverage, and precision in these experiments, and we found that our method achieves higher levels of novelty in batch recommendation while achieving higher levels of coverage and precision in node recommendations comparing to these alternative methods. Further, we develop a variant of the active node method that incorporates semantic structure elements. A further experimental evaluation with real data and users showed that semantic node recommendation with the active node method achieved higher levels of novelty than nonsemantic node recommendation, and semantic-batch recommendation achieved higher levels of coverage and precision than non-semantic batch recommendation.