Method For Commodity Recommendation Based On User-Browsing Tracks

With the rapid development of computer network, recommending system has been widely used in E-commerce web sites. Currently the collaborative filtering recommendation algorithm is most widely used. However this method has sparse matrix and cold-start problems. In order to alleviate these problems, this paper proposes two methods based on the user browsing history to predict a user’s preference. These methods are based on the users’ access logs of an e-commerce site, through mining user preference and establish the user browsing preference model based on browsing trajectories, which are extracted from the users’ browsing sequence, and recommend preference commodities for users. They are devoted to solve cold start problem, which is caused by poor purchase history and rating records.At present, the main methods for recommending commodities through analyzing browsing tracks are considered from the view of relationship between one commodity and the next browsed commodity. This study views from the relationship between browsing commodities and eventually bought commodity (commodities). The two recommending methods proposed in this paper, one is based on the purchase transfer relationship, and the other one is based on the commodities’feature trends in browsing tracks. The recommending method based on purchase transfer relationship, basically views from the relationship between the browsing commodities and eventually purchased commodity (commodities), at the same time considers the timeliness of users’browsing and purchase records, which means using time attenuation policy on data. In addition, we consider the order of the commodities in the browsing tracks, by using different weights on commodities of different track distance in this method. This recommending method builds purchase transfer probability model based on above considerations, and recommends commodities to a current user. The method based on commodities’ feature trends, builds Markov feature trend model based on commodities feature sets which are statistically gotten from the browsing tracks. When a new user browses the products online, to recommend the user’s preference commodities, firstly we predict the feature sets according to the Markov commodities feature trend model and variable commodities features in the user’s current browsing path. The union set of predicted feature sets and the stable feature sets is the feature set that the user wants. Then, recommend commodities that are most matched to the previous feature set which user wants. This method also considers the timeliness of the history data and the order of feature sets in the browsing tracks.Experiments show that these two methods achieved better results for recommendation. In other words, they have obvious improvement comparing with the existing methods based on user browsing tracks. They can play a certain role in solving the problem of cold start for new users and new commodities. The experimental results also show that these methods can enhance the accuracy and recall rate of the recommending method.