A Novel CTR Prediction Model Based On DeepFM For Taobao Data

CTR(click through rate) prediction is a useful tool for enterprises to get the customer's preferences and usually applied in recommender system and advertisement. With the development of technology, there are many machine learning algorithms are proposed to predict CTR, such as generalized linear model, factorization machines and deep neural network. However, all of these models owns disadvantages. And in our paper, we utilize the DeepFM model, which is an end to end model and do not need manual feature engineering. The model is the combination of FM Component and Deep Component. In experiments process, we use the focal loss that could solve the imbalance problem of samples as the loss function. The data is from Taobao platform in eight days. And we divide the data into training data and text data. And AUC is the index to evaluate the prediction model's performance. The result shows that our model's AUC is 0.044 and 0.013 higher than the logistic model and neural network model. The higher AUC is, the better performance the model will gain.