Learning to Ask Critical Questions for Assisting Product Search

In this paper, we propose a dual-learning model that hybrids the best from both implicit session feedback and proactively clarifying with users on the most critical questions.

Hence, there are two broad branches of research trying to address this problem. On one hand, session-aware recommendation makes use of the user’s behavior data within the session to capture user’s current interest, i.e., predicts the successive item(s) that an anonymous user is likely to interact with. Two major classes of approaches are applied: Markov-based models and DNN-based models [19]. The former emphasizes on capturing sequential patterns such as in the popular FPMC method [23]. Such models often emphasize the last click in history while ignoring previous behaviors. Hence, they often adopt static representations of user intentions, which may result in sub-optimal performance. The later one employs deep learning models such as RNNs to capture users’ general preferences and current interests together. They leverage recurrent deep neural networks to encode historical interactions into hidden vector states, which leads to general and more informative representations. Nonetheless, there is still room for improvement since such models over-emphasize on monotonic behavior chains while failing to consider the more complex transition patterns among items.

On the other hand, interactive methods allow users to directly specify their needs in details. It is often realized under a conversational search and recommendation framework, where user preferences are clarified via aspect-value pairs [2, 35] or even unstructured text [33]. There are also efforts prompting users with clarification questions under the traditional search setting with a query [34, 37]. These methods largely enhance the accuracy of user preference modeling and flexibility of interaction. However, these conversational product search methods require intensive user involvement. They assume that the user has a clear target in mind, which does not hold true especially in early browsing stages [10].

In this work, we propose DualSI – a dual-learning model to hybrid the best from both session-aware recommendation and interactive methods. It makes use of the implicit session feedback to model user interests and proactively ask users the most critical questions to shorten the exploring session.