The Research On Recommendation Algorithm Based On Learning To Rank And Convolutional Neural Network

With the rapid development of internet techniques, especially E-Commerce like Taobao and Amazon. The data in the internet grows much faster than our human beings can receive, "Information overload" problem is becoming more and more serious. The information filtering technique, which helps us filter useful information from mass of data is more and more important. Personalized recommendation technique is just one of the ideal methods, which aims to find user's interest according to user's behaviors from large scale data. Recommender system plays an important role in improving sales of e-commerce platform and user's purchase satisfaction.User's behavior in the internet can be divided into two classes, one is implicit feedback behavior, and another is explicit feedback behavior. In the implicit feedback behaviors, users don't express preference to a specific item explicitly, including users'click, cart and collect behaviors etc. However, in the explicit feedback behaviors, users express explicit preference to a specific item, one of the most common behavior is users'ratings to items. There are different recommendation techniques for different types of user's feedback behaviors. In this paper, we do detailed analysis and mining for these two types of behaviors respectively, and propose two methods to improve the performance of recommender systems.For explicit feedback behavior, like rating behavior, we choose Top-K recommendation as our research target. We introduce learning to rank approaches into recommender system field and incorporate user social influence and item tag information. We extend a list-wise learning to rank-based matrix factorization method to make trusted users'preference vectors as close as possible. On one hand, the method fully considers the influence of social networks. At first compute trust values between users based on users'focus relationship, then add trust matrix into the original loss function as a social penalty term. On the other hand, this paper represents each item as a vector with tags, compute the tag similarities between items, and then add the item tag penalty term to the loss function to train our model. Experimental results on the real Epinions and BaiduMovie datasets show that our proposed method outperforms several traditional methods, especially on the NDCG value, improving the recommendation accuracy effectively.For implicit feedback behavior, we choose next basket recommendation as our research target. Firstly, we divide user's behavior into several time windows according to the timestamp of user's behaviors, and model user's preference from different dimensions for each time window. Secondly, we utilize the convolutional neural network model to train our classifier. Compared to traditional linear models and tree models, on the real Alibaba Mobile Recommendation Contest dataset, our proposed model has more powerful feature extraction ability and generalization ability, especially on the precision, recall and fl-value, improving the user satisfaction of our recommender system.