【作者】 张伟；

【导师】 胡雯蔷；

【作者基本信息】 华中科技大学 ， 软件工程， 2019， 硕士

【摘要】 计算广告是一种新兴的互联网盈利手段,已经成为互联网公司主要的的流量变现手段之一。广告点击率预估,作为计算广告领域的核心问题,对在线广告的精准投放起到了重要作用。通过对近年来一系列基于传统机器学习和深度学习的广告点击率预估模型的研究总结,发现将传统机器学习模型与深度学习模型相结合,可以较大程度上提高广告点击率预估问题的AUC指标。在现有的广告点击率预估模型中,Deep FM模型是一种常用的基于深度学习的点击率预估模型。其中,Deep Fm模型的因子分解机模块用于提取数据的低阶特征,神经网络模块用于提取数据的高阶特征,但是因子分解机模型用于提取二阶组合特征时,同一特征在不同的特征组合中映射为相同的隐向量,显然是不合常理的。基于此,论文设计的点击率预估模型在进行低阶特征的提取时,将因子分解机模型修改为场感知因子分解机模型以解决同一特征在不同的特征组合中映射为相同的隐向量这一问题,从而使得低阶特征的提取更为合理。对于神经网络模块,在Deep FM模型中全连接网络的基础上,引入了正则化方法,增强了模型鲁棒性,防止网络过拟合。此外,对于用户历史行为特征的提取,论文设计的点击率预估模型添加了Word2Vec模块对用户历史点击行为序列进行训练,以反映不同用户的偏好性。基于以上三点改进和优化,论文设计了基于神经网络与场感知因子分解机的广告点击率预估算法模型。在特征工程方面,设计了广告点击率预估领域一套完整的特征工程解决方案,涵盖了原始特征、历史统计特征、预测当天统计特征、用户历史行为序列特征四大特征群。通过分析对比不同模型在2019年华为DIGIX算法创新大赛上提供的广告点击率预估数据集上的实验结果,发现在AUC指标上,论文设计的模型取得了较大的提升。可以认为,论文提出的广告点击率预估模型具有一定的应用前景,可以有效提高广告点击率预估系统的准确性。更多还原

【Abstract】 Calculating advertising is an emerging method of Internet profit,which has become one of the main traffic monetization methods of Internet companies.Advertising CTR estimation,as a core issue in the field of calculating advertising,has played an important role in the precise delivery of online advertising.Through a recent study and summary of a series of ad click rate estimation models based on traditional machine learning and deep learning,it is found that combining traditional machine learning models with deep learning models can greatly improve the problem of ad click rate estimation.AUC indicator.Among existing advertising C TR estimation models,the Deem model is a commonly used deep learning-based CTR estimation model.Among them,the factorization machine module of the Deep Fm model is used to extract the low-order features of the data,and the neural network module is used to extract the high-order features of the data,but when the factorization machine model is used to extract the second-order combined features,the same feature has different features.It is obviously unreasonable to map the same hidden vector in the combination.Based on this,when the click-through rate prediction model designed in the paper is used to extract low-order features,the factor decomposition machine model is modified to a field-aware factor decomposition machine model to solve the problem that the same feature is mapped to the same hidden vector in different feature combinations.This problem makes the extraction of low-order features more reasonable.For the neural network module,based on the fully connected network in the Deep FM model,a regularization method is introduced to enhance the robustness of the model and prevent the network from overfitting.In addition,for the extraction of user historical behavior characteristics,the click rate estimation model designed in the paper adds the Word2 Vec module to train the user’s historical click behavior sequence to reflect the preferences of different users.Based on the above three improvements and optimizations,the paper designs an ad click rate estimation algorithm model based on neural network and field perception factor decomposition machine.In terms of feature engineering,a complete feature engineering solution was designed in the field of advertising click-through rate estimation,covering four major feature groups: original features,historical statistical features,statistical features of the prediction day,and user historical behavior sequence features.By analyzing and comparing the experimental results of different models on the ad click rate estimation data set provided by the 2019 Huawei DIGIX Algorithm Innovation Contest,it is found that the model designed in the paper has achieved a significant improvement on the AUC index.It can be considered that the advertising click-through rate estimation model proposed in the paper has certain application prospects and can effectively improve the accuracy of the advertisement click-through rate estimation system.更多还原