Design And Implementation Of User Interest Adaptive Personalized Recommender System

Recommender systems are widely used in e-commerce systems which aim at helping a user find the right items user may like through the massive amounts of goods. The personalized recommender system in this paper was developed for a price comparisons shopping site. In order to reflect the dynamic change of every user’s interest, a method is given that mixes implicit feedback and explicit feedback. The system combines the user-based and the model-based CF algorithms which caters for the requirement of recommending goods for user’s potential interest. The system consists mainly of the following parts: data preprocessing, preference modeling, data mining, recommendation generation and client agent. The author took part in the design and development of the first four modules. The work is summarized as following:(1) Requirements analysis.(2) The general design and realization of related core modules.Data preprocessing filters the data from web log. CNZZ is used as assistant tool in this processing, and the tokenizer is ICTCLAS.Preference modeling includes initial preference modeling and adaptive model update. They’re updated using time widow and forgetting-factor function respectively.Data mining clusters users using k-means algorithm.Recommendation generation finds neighbors and recommends. It predicts ratings for those items which haven’t been rated and recommends Top-N items.(3) Finished system testing. An experiment was conducted to test and evaluate the performance of the recommender system. The experiment measured the precision of recommendation with MAE (Mean Absolute Error). The results show that the recommender precision differs from the size of the neighbor set and the functionality of the recommender system is satisfactory.For the reasons of limit data and sparsity, the MAE values are large, between0.861and0.958according to the size of the different neighbors set. The recommender system recommends not only items that are relative with the items user likes, but also potential unrelated items user may like.