The Evolution Characteristics Analysis Of Information Recommendation Networks

Information recommendation technology has attracted a lot of research and attention because of its huge application value. Current research has been penetrated into all aspects of recommendation, while most of them only concentrate on the single step’s recommendation or static network analysis, ignoring the evolution characteristic of online systems and long-term performance of recommendation algorithms. In this paper, we use bipartite network to depict the online recommendation system to study the long-term cumulative effects of recommendation on the evolution of online system and also the subsequent performance of the recommendation based on the network structure analysis. Main contents are:1) We study the long-term evolution of recommendation. We design a co-evolution process of recommendation and online system to study the long-term effects of common recommendation algorithms on the structure characteristics and recommendation function of online network systematically for the first time. We find that recommendation can strengthen the important structure characteristics of the online network to a large extent. More importantly, the corresponding recommendation next becomes poorer and poorer. The study makes some guidance and theoretical support to design a better recommendation algorithm for the long-term development of the system.2) We study the co-evolution of recommendation and online system. We follow the tracks of the network evolution process to reveal the causes of the problem above. From both the relationship between degrees of users and items on both ends of each edge and the relationship between each item’s initial degree and its increment, we see that there is only a few popular items gradually living in the system in the co-evolution process and eventually make recommendation fail. The current recommendation algorithms always tend to recommend popular products, or lose accuracy when concerning about diversity, causing recommender system can’t develop healthily in the long run. And rearranging recommendation lists can not improve the co-evolution fundamentally. It needs optimization of recommendation.3) We study the behavior features of most users in the online system. We come up with a simple controllable network evolution model to simulate the real online system. Considering that users may refuse recommendation, we set a probability that users receive recommendation. A user chooses an item from his valid recommendation list based on preferential attachment of recommendation scores when he receives recommendation, otherwise he will choose an item from all the items based on preferential attachment of their degrees. By adjusting the parameters of the probability receiving recommendation and the degree of preference attachment we can determine the optimal parameters. The experimental results show that the model better reproduce the statistical properties of the network. The results show that the model can reproduce the statistics of online network well.