【摘要】 在大数据时代的点击率(Click-Through Rate, CTR)预测任务中，输入数据不仅数量多而且特征维度很高，在特征选择时容易出现信息干扰或丢失，在进行特征交互时不同的交互方式也会影响预测性能。针对该问题，文中提出了一种基于特征加权与自动交互的预测模型，用于学习原始特征权重并进行自动交互。首先，引入ECANet模块提出一种不降维的特征加权方法，该方法可以通过对k个相邻特征进行一维卷积有效实现。然后，分别用多头自注意网络和深度神经网络(DNN)去自动学习显式和隐式的特征交互。最后，将两者相结合进行预测，弥补了单一模型的缺陷。一方面，它能对输入特征进行重要性选择；另一方面，它能同时以显式和隐式的方式自动学习任意低阶和高阶的特征交互。通过在四个真实数据集上的实验，验证了其比以往的预测模型获得了更好的准确度。更多还原

【Abstract】 In the click-through rate(CTR) prediction task in the era of big data, the input data is not only large in quantity but also has a high feature dimension, which is prone to information interference or loss during feature selection. Different interaction modes during feature interaction will also affect the prediction performance. To solve this problem, a prediction model based on feature weighting and automatic interaction is proposed, which is used to learn the original feature weights and interact automatically. Firstly, we introduce an ECANet module and propose a feature-weighting method without dimensionality reduction, which uses one-dimensional convolution of k adjacent features to learn feature weights. Then, multi-head self-attention network and deep neural network(DNN) are used to automatically learn explicit and implicit feature interactions. Finally, the two are combined to predict, solving the defects of a single model. On the one hand, it can select the importance of input features. On the other hand, it can automatically learn arbitrary low-and high-order feature interactions in both explicit and implicit ways. And experimental results on four real data sets have proved that the prediction model is more accurate than the previous model.更多还原