Abstract: In the current business scenario of online advertising, the linear model does not take into account these essential characteristics including the high-dimensional sparsity of advertising data and highly nonlinear association for features. Aiming at these problems, this paper introduces a feature extraction method based on the gradient boosting decision tree algorithm, and presents a factorization machine model based on the FTRL （Follow-The-Regularized-Leader） optimization algorithm. The FTRL algorithm can effectively learn the nonlinear relationship between features, enables different parameters to adapt to different learning rates and adds a mixed regularization term. The experimental results show that the factorization machine model based on the FTRL optimization algorithm can effectively improve the accuracy of advertising click through rate prediction.

Key words: advertising click through rate, logistic regression, factorization machine, Follow-The-Regularized-Leader（FTRL） algorithm

摘要： 当前在线广告的业务场景下，线性模型没有充分考虑到数据高维、稀疏性、非线性等特点。针对这些问题，引入了基于梯度提升决策树算法的特征提取方法，提出了基于FTRL（Follow-The-Regularized-Leader）优化算法的因子分解机模型。FTRL优化算法能有效地学习到特征之间存在的非线性关系，使不同参数可以自适应不同学习率，并加入了混合正则项。实验结果证明基于FTRL优化算法的因子分解机模型能有效提高广告点击事件的预测准确率。

关键词: 广告点击率, 逻辑回归, 因子分解机, FTRL算法