An Improved Recommendation System Algorithm Based On Choice Model

With the development of information technology,the massive data generated by users in the process of using online services has become an important resource of various companies.User data can help enterprises enhance their understanding of users,and personalized recommendation services based on these data are developing rapidly.Learning to rank is widely used in recommendation system.It is of great significance in the field of recommendation system to accurately estimate the click-through rate of users and sort the products according to the probability for users to choose.In previous studies,pointwise learning to rank models,such as Factorization Machine,which are commonly used in recommendation and ranking,generally only consider the probability of users choosing a single item,ignoring the interaction between candidate items.Discrete choice model considers item candidate set as a whole,but it can only learn linear relationship between features,which has some limitations.Because of its strong non-linear learning and feature generalization ability,deep learning is gradually applied in the field of recommendation system.In this paper,the hypothesis of discrete choice model is introduced into deep learning model to predict user preferences.In this model,the hypothesis of discrete choice model is introduced into the deep learning model through feature concatenation.By designing relative feature layer,introducing feature extraction layer,attention mechanism,factorization machine layer and other structured network modules,we can help the deep learning model to compare the differences between different items.In the experimental part,taking the online air ticket purchase data as an example,it is verified that the performance of this model is better than the representative model MNL of discrete choice model,the representative model of pointwise learning to rank models,the Gradient Boosting Decision Tree and Factorization Machine,and the representative model of listwise learning to rank models,ListNet.The results show that introducing the discrete choice model hypothesis into the deep learning model and designing a reasonable network structure have certain practical and research value.