| 摘要: |
| 在推荐系统领域，了解电商平台中在线用户的行为意图至关重要。目前的一些方法通常将用户与商品之间的交互历史数据视为有序的序列，却忽视了不同交互行为之间的时间间隔信息。另外，一个用户的在线行为可能不仅仅包含一种意图，而是包含多种意图。例如，当一位用户在浏览运动品类下的商品时，其可能同时有购买足球和运动衫这两种商品的意图。但是现有的一些电商平台用户意图预测方法很难有效对用户-商品交互对时间间隔信息进行建模，也难以捕捉用户多方面的购物意图。为了解决上述问题，我们提出了一种时间感知分层自注意力网络模型THSNet，以更有效对电商平台的用户意图进行预测。具体而言，THSNet模型采用一种分层注意力机制来有效地捕获用户-商品交互历史中的时间跨度信息以更有效建模用户的多种意图。THSNet模型的注意力层分为两层，底层的注意力层用于建模每个会话内部的用户-商品交互，上层的注意力层学习不同会话之间的长期依赖关系。另外，为了提高预测结果的鲁棒性和准确度，我们采用BERT预训练的方法，通过随机遮盖部分会话的特征表示，构造了一个完形填空任务，并将该任务与用户意图预测任务耦合成为多任务学习模型，这种多任务预测方法有助于模型学到一个具有鲁棒性和双向性的会话特征表示。我们在两个真实数据集上对所提方法对有效性进行了验证。实验结果表明，我们所提出的THSNet模型要明显优于目前最先进的方法。 |
| 关键词: 意图预测 商品推荐 序列预测 注意力机制 深度学习 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2021.09.13 |
| 投稿时间：2021-04-29修订日期：2021-08-09 |
| 基金项目:本课题得到中央高校基本科研业务经费“人工智能+”专项（No.NZ2020014）和广东省自然科学基金（No.2021A1515012239）资助。 |
|
| Time-Aware Hierarchical Self-Attention Networks for User Intent Prediction on E-Commerce Platforms |
| WANG Senzhang,LIU Yi,ZHANG Jiaqiang,YIN Chengyu |
| School of Computer Science and Engineering, Central South University, Changsha 410083, China;College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China |
| Abstract: |
| Understanding the behavior intent of online users on E-commerce platforms is critically important in many recommender systems. Current approaches generally regard the behavior interations between the users and the items as ordered sequences, which may largely ignore the time lag length between the behavior interactions. Meanwhile, instead of having only one intent, a user's online behavior on E-commerce platforms may have multiple intents. For example, when a user is browsing the sport equipment, she may want to buy a soccer and a sweatshirt simultaneously. It is difficult for existing approaches to both model the time lag length between the behavior interations and capture the multi-facet user intents on E-commerce platforms. To address these issues, we propose a Time-Aware Hierarchical Self-attention Networks model named THSNet to more effectively predict the user intents on E-commerce platforms. Specifically, THSNet uses a novel hierarchical attention mechanism to effectively capture the time span length between user-item interactions and a user's multi-facet intents. The hierarchical attention mechanism contains two layers. The bottom attention layer focuses on capturing the user-item interaction within each session, and the upper layer attention aims to learn the long term dependencies among the sessions. In addition, to learn a more robust and bidirection session embedding, motivated by the pre-training method in BERT we propose to add a Cloze task which aims to predict the randomly masked session embeddings. The Cloze task is jointly conducted with the user intent prediction task under a multi-task learning framework. We conduct extensive experiments on two real-world datasets. The results show that the proposed THSNet outperforms multiple current state-of-the-art methods. |
| Key words: intent prediction product recommendation sequence prediction attention mechanism deep learning |
