Research On Personalized Recommendation Algorithm Based On Online Social Network

With the rapid development of the Internet, we gradually step into the era of ‘data explosion' due that the exponentially growing of data had gone beyond our processing ability. The overwhelming information brought the information overload problem. So how to rapidly mine valuable information from the massive data is an urgent problem for us. Meanwhile, it promoted the development of personalized recommendation technology.Personalized recommendation mainly used to build the model of users' preferences by utilizing the historical activities and personal interests. The model was able to predict the users' favorite products or projects, and consequently provide the personalized service for users. Currently, the more popular recommended algorithms included the collaborative filtering algorithm, content-based algorithm, network-based algorithm and so on. In this paper, the mainly research contents and achievements were as follows:(1) This paper particularly analyzed the researches' significance and purpose for personalized recommendation. And we introduced the more popular recommended algorithms and analyzed the advantages and disadvantages for different algorithms.It was laid a solid theoretical basis for this paper.(2) This paper took the deficiency of the present algorithm as the starting point of our research and put forward a new algorithm to improve the performance of recommended algorithm by effectively fusing various information.(3) Considering the impact were caused that transitional network-based method neglected user-object ratings. So weighted bipartite network projection for personalized recommendations was proposed. To effectively mine users' preferences and disgusts, we took the rating level and rating distribution into consideration. Firstly, the surface ratings were translated into users' preferences by half cumulative distribution method for each user. And we used the new weight to build weighted bipartite network. Then the similarity between objects was able to be calculated according to the diffusion algorithm. Calculate the probability which each object was recommended to the target user, and make the Top_n products to recommend the target user. Finally, experimental results on Movie Lens data set demonstrated the diversity of the proposed method was much more superior to competitive methods.(4) Considering only using the user-object binary relation brought the cold-start and data sparse problems. How to effectively fuse user's properties or user's social attributes(i.e., @, comments, re-tweet) was the key point for our research. This paper made the possibility of giving the target users to recommend others translate into the probability forecast problem by Logistic Regression. Firstly, many features reflecting the correlation between users were extracted from micro-blog. Secondly, we made the features as the input parameter of Logistic Regression model, then use the gradient descent method to obtain optimal regression parameters. The final recommended list in descending order byprediction score and made the Top_n to recommend to the target users. Finally,experimental results indicate that the recommendation model improves the precision of recommendation effectively.