Lu Qiao, Li Silin, Yang Tuo, Xu Chenheng
Taicu Music co Ltd Shenzhen China, Shenzhen, United Kingdom.
School of Economics, Tianjin University of Commerce, Tianjin, China.
PeerJ Comput Sci. 2021 Sep 17;7:e716. doi: 10.7717/peerj-cs.716. eCollection 2021.
Recent advances in communication enable individuals to use phones and computers to access information on the web. E-commerce has seen rapid development, e.g., Alibaba has nearly 12 hundred million customers in China. Click-Through Rate (CTR) forecasting is a primary task in the e-commerce advertisement system. From the traditional Logistic Regression algorithm to the latest popular deep neural network methods that follow a similar embedding and MLP, several algorithms are used to predict CTR. This research proposes a hybrid model combining the Deep Interest Network (DIN) and eXtreme Deep Factorization Machine (xDeepFM) to perform CTR prediction robustly. The cores of DIN and xDeepFM are attention and feature cross, respectively. DIN follows an adaptive local activation unit that incorporates the attention mechanism to adaptively learn user interest from historical behaviors related to specific advertisements. xDeepFM further includes a critical part, a Compressed Interactions Network (CIN), aiming to generate feature interactions at a vectorwise level implicitly. Furthermore, a CIN, plain DNN, and a linear part are combined into one unified model to form xDeepFM. The proposed end-to-end hybrid model is a parallel ensemble of models via multilayer perceptron. CIN and xDeepFM are trained in parallel, and their output is fed into a multilayer perceptron. We used the e-commerce Alibaba dataset with the focal loss as the loss function for experimental evaluation through online complex example mining (OHEM) in the training process. The experimental result indicates that the proposed hybrid model has better performance than other models.
通信领域的最新进展使个人能够使用手机和电脑访问网络信息。电子商务发展迅速，例如，阿里巴巴在中国拥有近12亿客户。点击率（CTR）预测是电子商务广告系统中的一项主要任务。从传统的逻辑回归算法到最新流行的遵循类似嵌入和多层感知器的深度神经网络方法，有几种算法被用于预测CTR。本研究提出了一种结合深度兴趣网络（DIN）和极端深度因子分解机（xDeepFM）的混合模型，以稳健地进行CTR预测。DIN和xDeepFM的核心分别是注意力和特征交叉。DIN遵循一个自适应局部激活单元，该单元结合了注意力机制，以从与特定广告相关的历史行为中自适应地学习用户兴趣。xDeepFM还进一步包括一个关键部分，即压缩交互网络（CIN），旨在隐式地在向量级别生成特征交互。此外，一个CIN、普通深度神经网络和一个线性部分被组合成一个统一的模型以形成xDeepFM。所提出的端到端混合模型是通过多层感知器的模型并行集成。CIN和xDeepFM并行训练，其输出被输入到一个多层感知器中。我们使用电子商务阿里巴巴数据集，以焦点损失作为损失函数，在训练过程中通过在线复杂示例挖掘（OHEM）进行实验评估。实验结果表明，所提出的混合模型比其他模型具有更好的性能。
