An Analysis Of Supervisory Signal Enhancement Based Recommender System Using Wide&Deep Learning

Due to information overload and people's pursuit of personalization,the recommender system has attracted people's attention in recent years.Our demand for personalized recommender systems is increasing gradually in fields like news acquisition,audio/video entertainment,and e-commerce etc.At the same time,Google's internal machine learning solutions have gradually transitioned to deep learning based on TensorFlow,meanwhile,artificial intelligence technology represented by deep learning is gradually infiltrating the recommender system field.Faced with massive data and features which are difficult to be structured,the traditional recommender algorithm often requires a huge amount of manpower in feature engineering.Even so,with the increase of the data size,it is more and more difficult to improve the recommend performance.The wide & deep learning recommender system,which was first proposed by Google,combines the logistic regression's capability of dealing with high-dimensional sparse inputs and the deep learning neural network's strength about generalization.It memorizes and generalizes better with less effort of feature engineering.Facing with the rough problem in the model's click-through rate prediction,this thesis proposes a supervisory signal enhancement based recommender system using wide and deep learning.The fusion of prior knowledge enhances the information entropy of the supervisory signal and helps the model to learn more information from the data.The model's performance of prediction then get improved.The main work of this thesis is as follows:1.This thesis researches and analyzes the traditional classical algorithms in the recommender system field,and summarizes the advantages and disadvantages of user-based collaborative filtering algorithm,item-based collaborative filtering algorithm,and latent factor model etc.2.We model and analyze the wide & deep learning recommender system in this thesis.In contrast of the original wide and deep recommender system model,we propose a supervisory signal enhanced wide and deep learning recommender system.We help the model gain more information from the data by combining prior knowledge such as popular effect of items,user behavior categories,and human forgotten curve which may improve the information entropy of supervisory signal.3.We conduct experiment on a real e-commerce platform's desensitization dataset in this thesis.The experiments show that the supervisory signal enhanced wide & deep recommender systems achieve better performance than other baseline methods.