摘要：

传统单一模型对广告点击率(CTR)的预测效果较片面,且广告日志数据存在稀疏性问题。为此,将高斯混合模型(GMM)与因子分解机(FM)相结合,建立一种广告点击率预测模型GMM-FMs。对原始数据集进行高斯聚类,在聚类后得到的数据子集上分别建立CTR预测模型,将多个预测模型的结果进行有效加权得到最终结果。实验结果表明,相比传统逻辑回归与FM模型,该模型在训练集与测试集比例不同时均能取得较高的AUC值。

关键词: 广告点击率, 高斯混合模型, 因子分解机, 数据稀疏性, 逻辑回归模型

Abstract:

The traditional single model is one-sided in predicting the Click-Through Rate(CTR) of advertisement,and the data of advertisement log is sparse.To solve this problem,a GMM-FMs model for predicting advertisement CTR is established by combining the Gaussian Mixture Model(GMM) with Factorization Machine(FM).The original data set is clustered based on the Gaussian mixture model,and the CTR prediction models are built on the clustered data subset.The results of multiple prediction models are weighted effectively to get the final results.Experimental results show that compared with the traditional Logistic Regression(LR) model and FM model,the proposed model achieves higher AUC values regardless of the training-set to test-set ratio.

Key words: advertisement Click-Through Rate(CTR), Gaussian Mixture Model(GMM), Factorization Machine(FM), data sparsity, Logistic Regression(LR) model

中图分类号: 

• TP311