Learning to Ask Critical Questions for Assisting Product Search

TL;DR

Product search in eCommerce often lacks direct modeling of a user’s evolving interest. DualSI combines implicit session signals with explicit, attribute-focused questions using a Transformer-based Contextualized Selection Net to compute a contextualized utility score $\alpha_i$ and a Dual Ranking Net to rank items after user responses. The approach introduces a novel utility score and a dual-learning framework that jointly leverages click-stream information and targeted questions, with synthetic-item generation guiding training. Experiments on the public Diginetica dataset show substantial improvements over strong baselines and reveal that asking the right questions at appropriate times can shorten exploration while improving ranking.

Abstract

Product search plays an essential role in eCommerce. It was treated as a special type of information retrieval problem. Most existing works make use of historical data to improve the search performance, which do not take the opportunity to ask for user's current interest directly. Some session-aware methods take the user's clicks within the session as implicit feedback, but it is still just a guess on user's preference. To address this problem, recent conversational or question-based search models interact with users directly for understanding the user's interest explicitly. However, most users do not have a clear picture on what to buy at the initial stage. Asking critical attributes that the user is looking for after they explored for a while should be a more efficient way to help them searching for the target items. 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. We first establish a novel utility score to measure whether a clicked item provides useful information for finding the target. Then we develop the dual Selection Net and Ranking Net for choosing the critical questions and ranking the items. It innovatively links traditional click-stream data and text-based questions together. To verify our proposal, we did extensive experiments on a public dataset, and our model largely outperformed other state-of-the-art methods.

Figures (5)

• Figure 1: The DualSI framework. The Selection Net calculates contextualized utility scores for clicked items, which helps to select the most critical questions. The dual Ranking Net learns to predict the target item based on user's feedback.
• Figure 2: The Selection Net architecture. The item sequence obtains contextualized representations via static dynamic item embeddings and self-attention blocks. The query signals attend over them for better utility score calculation. Details of the Static Dynamic Item Embedding module are further provided in Figure \ref{['fig:static_dynamic']}.
• Figure 3: The static item embedding concatenate with the dynamic attribute count feature to form a comprehensive representation for each item.
• Figure 4: The Dual Ranking Net. There are normalization layer, flatten layer, dropout layer and dense layer used in the $\delta$ network. After multiplying with the item pool $X'^{T}$, there are again a normalization layer, some dense layers and dropout layers. They are omitted here for a concise and clear illustration on the main concept.
• Figure 5: Training scheme for the Ranking Net. The one in the grey box is chosen as the input for Ranking Net Training if synthetic item 2 has the highest score among all synthetic items generated based on the target item.