Research On The Theory And Key Technology Of Domain-oriented Recommendation System

Due to the rapid growth of information on the Internet,the users are faced with the problem of information overload.With the help of relevant technologies in data mining and artificial intelligence,recommendation system can help the users quickly find the information to meet their interest,which has been widely used in social network,e-commerce,online reading and advertising,etc.With the diversified development of Internet applications,it is difficult for the traditional recommendation models to be directly applied in new domains to solve the corresponding problems.Electronic products,such as smart phones and laptops,usually update frequently,while the consumption cycles of users are relatively long.The traditional recommendation systems will make recommendations based on the users' previous purchase records.On this condition,some purchase records may have lost their timeliness and cannot provide enough effective information to recommend new products.To overcome this problem,we propose to model on two perspectives,i.e.,item level and user level.The former reflects the level of the item compared with similar competing goods,while the latter reflects the level of items that the user likes.By analyzing the variation tendency of item attributes and the characteristics of users' consumption behaviors,we find that a user's preference is stable to some extent,that is,a user tends to buy items with similar attribute levels in different periods.Based on this point,we design two new similarity measure methods,and then extend the latent factor model by integrating them to propose two new methods: Item Level Matrix Factorization(ILMF)and User Level Matrix Factorization(ULMF).In addition,we further combine these two methods to study the contribution of these two methods in the final performance.The experimental results show that our methods perform better than the compared method on the real datesets.Different from the traditional recommendation task,point of interest recommendation presents the prediction of the most interesting place for target users in a specific spatial and temporal situation.Point of interest recommendation benefits not only the users for obtaining the better visit experience,but also the service providers for promoting theirservices with the users' feedbacks.Point of interest recommendation needs to fully consider the characteristics of users' visit behaviors,and the users' visit behaviors are influenced by geographical and time factors,and also show group characteristics.Therefore,it is a huge challenge to integrate these influential factors into the unified recommendation framework.To solve this problem,we first propose a group-based interest point recommendation method,GTSAR-RNN,which comprehensively considers time information,comment information,location category information and location geographic information.In order to improve the pertinence of the model,we divide the users into different groups based on their check-in data,and trained an independent neural network for each group of users to recommend points of interest.GTSAR-RNN adopts a flexible multi-group strategy to divide the users into multiple groups,and the neural network of each group will produce an independent recommendation result,while the final recommendation merges the recommendation results of different groups.A user who has several interest is divided into multiple groups,since we adopt the strategy of independent training,the corresponding neural network of each group captures part of his interest.This strategy both considers the personalization and diversity in recommendation.Experiments on the real datasets show that GTSAR-RNN perform significantly better than the compared methods.The traditional recommendation systems based on rating prediction calculate the matching score between the item attributes and the user preference,and then recommend the top-k items with the highest matching score to the user.However,this kind of top-k recommendation strategy mainly makes personalized recommendation from the perspective of the user.Manufacturers and sellers hope that the recommendation system can quickly find the potential customers for the item,so as to adopt targeted marketing methods to promote the sales of the item.To find the potential customers for the item,we introduce the reverse top-k query into the recommendation system.The existing recommendation models deal with the deterministic data,while in real life,a user do not always follow certain rules when selecting items,and he may select with some randomness.As a core task of recommendation system,the current research on capturing the user's preference only considers the variation of user preference in different scenarios,but ignores the uncertainty of user preference in the same scenario.Therefore,this paper first models the users' uncertain preferences,and then proposes a reverse top-k query on the uncertain preferences.In order to improve the query efficiency,RUI-tree is designed to index the users' uncertain preference data,and on this basis,the UPBBR algorithm is proposed.Experimental results show that the UPBBR algorithm is superior to other comparison algorithms in both the generated data set and the real data set,and shows good scalability.