Research On Cold-start Problem In Recommender System

With the rapid development of the Internet,social networks,personalized emails,advertisements,takeaway and e-commerce is more widespread and popular,which have led to the explosive growth of Internet information,resulting in the phenomenon of information overload.Losses caused by overload of information are exacerbated-people spend a lot of time,but it's hard to find the information they really need.In order to achieve mutual benefits for both users and merchants,the recommender system as an important tool for decision-makers arises at the historic moment.Collaborative filtering,as the most effective algorithm in the beginning,has been widely used and researched.which is used to analyze users' historical behavior,mine users' potential preferences,and on this basis to make recommendations.However,for new users and new items,because there is no historical data to depend on,the accuracy of recommendation decreases,and the recommendation effect is not good.We call it a cold-start problem.Aiming at alleviating the cold-start problems in collaborative filtering.The research work of this thesis is as follows:(1)According to the forgetting curve of Ebbinghaus,a time weight function that is in line with the change of people's interest is proposed.And then integrated it into the matrix factorization model based on the user's preference so that the model can predict the missing item that is closer to the change of the user's interest and improve the accuracy of the recommendation.(2)In order to alleviate new user cold-start problem,a matrix mapping recommender model that fuses user attributes is proposed.First of all,one-hot encoding of all users' attribute information to obtain users' attribute vector.Then,using the matrix factorization model to train the historical rating records offline,the mapping matrix of the user's attribute to the potential interest preference and related of bias can be obtained.Finally,calculate the target user's predictive rating for all items and recommend the top few items with high score to the user.The experiments show that this method effectively improves the real-time recommendation.(3)In order to alleviate new item cold-start problem,a collaborative filtering model that fuses item category attributes and image is proposed.First of all,the user-item rating matrix was predicted and populated using a matrix factorization on users' preferences and time weight.The VGG16 convolutional neural network was used to extract the features of the items' images,and the similarity of the image's features was calculated using the cosine similarity formula.Then,combine the similarity of item category attribute to find the nearest neighbors of the target item,Finally,the rating of the target item is predicted according to the rating of the neighbors to complete the recommendation.Experiments show that this method effectively improves the accuracy of the recommendation.(4)In order to alleviate the cold start problem existing in the online shopping mall recommendation system,the algorithm is applied thereto.First of all,the hierarchical structure of the recommender system is analyzed according to the architecture of the shopping mall.Then,the information available for the recommender algorithm is extracted based on the features of the mall.Finally,the algorithm is integrated into the recommender system,the function is displayed,and the recommendation result is analyzed.