Research On Rating Prediction Based On User Purchasing Power And Item Picture Characteristics

With the continuous development of Internet technology,data mining and artificial intelligence have been widely used in various fields.As one of the frontier applications of artificial intelligence,recommendation system plays an important role in providing personalized information for users and helping to deal with information overload.It has been a hot topic in academic research and industrial application.The application of recommendation system improves the efficiency of platform operation,not only facilitates the user's shopping and consumption choice,but also brings profit enhancement to the platform.Since the concept of recommendation system was proposed,many algorithms have been proposed and verified in practical application.At present,recommendation system algorithms are mainly divided into two directions,one is content-based recommendation algorithm,and the other is collaborative filtering algorithm.The two kinds of algorithms have their own advantages in different application scenarios.Current researches are used to treating rating data as a disordered data set,without considering the growth of user consumption experience at the individual level.In fact,users in their own consumption activities,according to their own learning ability,will continue to accumulate consumption experience.Therefore,the user's consumption history can not be regarded as disordered data to process,and the evolution of user experience level should be excavated from these data.On the other hand,from the personal level,users will be more or less influenced by herd psychology in the process of purchasing and evaluating.However,the extent to which users are affected depends on the experience level they are in.This paper tries to mine the evolution process of user experience level from the personal level,and at the same time,constructs a reasonable scoring model by integrating the factors of herd psychology.In addition,we have further explored the relationship between the personal experience level and herd psychology,and found that the higher the experience level,the smaller impact of conformity psychology on users.By studying the distribution of ratings at different experience levels,we can clearly find the differences between different experience level and different user ratings.In the process of purchasing,users need to compare different commodity information to choose.The most direct information that has the greatest impact on users is the price of goods and the picture feature.Firstly,this paper eliminates the price differences among different commodity categories and constructs a set of models to describe commodity price levels.At the same time,through matrix operation,we construct the characteristics of user's preference for commodity prices.In terms of visual features,we use Variational AutoEncoder to extract feature vectors of commodity pictures,and reduce the dimension of visual features through embedding layer operation.Then,we integrate the user's price preference features and visualization features into the standard recommendation system,and construct a rating model based on personal purchasing power level and visualization features.On the other hand,we add price factor feature and visual picture feature to click prediction model,which enriches the characteristics of prediction model.Experiments show that our idea can effectively improve the accuracy of model prediction,and achieve better results.Further,we explore the relationship between individual purchasing power and purchasing decision-making.We find that high purchasing power users prefer to buy goods beyond their purchasing power level.On the contrary,low purchasing power users usually make relatively rational purchasing decisions.We validate our models through the real Amazon data set.Experiments show that user's evolving experience level and herd mentality together affect user's consumption rating process.In addition,personal purchasing power and visual characteristics can directly affect the user's ratings decision-making.Building a reasonable model based on these two important factors can effectively improve the performance of the recommendation system.