Crossing Recommendation In Multi-B2CS Based On Transfer Learing

In the era of big data, the Recommendation System is an effective way toimplement the personalization technology. It aims to recommend valuable informationto user, which makes the user get information more efficiency. The research onRecommendation System has great social significance and economic value. Since1990,data mining researchers give out a number of solutions and improvement strategiesform different recommendation perspectives, which make the Recommendation Systemhas been successfully used in various commercial systems.Although Recommendation System has been widely used in the Internet, there arestill some main problems which should be solved, such as “data sparseness” and "coldstart" problem. At the same time, the behavior which user on the Internet often carriessome intention, but the system is lack of consideration these intention. This thesis aimsto solve existing problems using approaches such as transfer learning and intentionprediction. The main work is follows:1. Firstly, this thesis proposes a crossing recommendation algorithm in Multi-B2CSbased on transfer learning. This algorithm aims transfer the “knowledge” learned froman auxiliary domain to a target domain to help the target domain do more precisionrecommendation. The algorithm considers the influence causes by domain distributionand domain sparseness. Through off-line simulation experiments, the results show thealgorithm has a better representation than collaborative filtering in precision and recall.2. Then, this thesis proposes two kinds of user’s intention prediction algorithm, oneis called “prediction user’s intention based on threshold ", another is called “predictionuser’s intention based on logical regression ". Both algorithms are needed to extract thefeatures from the user’s behavior, and applied the features to corresponding algorithm.The algorithm predicts user’s intention based on threshold has strong flexibility and thealgorithm predicts user’s intention based on logical regression has high expansibility.Through off-line simulation experiments, the results show that the both algorithms havehigh accuracy in predicting user’s intention.3. Finally, this thesis proposes a recommendation system based on Scenario Engine. The system calculates user’s intention which is used for doing recommendation byScenario Engine. By introducing the scene conception, the real time user’s behaviormodel can turn into user’s intention. The Scenario Engine computes user’s intent byjudging the user’s scene.