Globally Optimized Mutual Influence Aware Ranking in E-Commerce Search

Tao Zhuang, Wenwu Ou, Zhirong Wang

Abstract

Traditional e-commerce ranking models often treat items in isolation, ignoring the "mutual influence" between items in a search result set. This paper proposes a globally optimized ranking framework that models ranking as a sequence generation task. By utilizing Recurrent Neural Networks (RNNs) with an attention mechanism, the model captures contextual dependencies and optimizes for Gross Merchandise Volume (GMV). Evaluation on Taobao search data demonstrates a significant 5% increase in GMV, highlighting the importance of context-aware ranking in commercial search engines.

Introduction

In traditional web search, relevance and diversity are the primary objectives. However, e-commerce search presents a unique challenge: users are not just looking for relevant results; they are actively comparing options. The purchase probability of an item is heavily influenced by its context. For instance, a mid-priced product may appear more attractive when surrounded by high-priced luxury items than when placed next to budget alternatives.

Most existing ranking models score items individually and sort them, neglecting these "mutual influences." Furthermore, the ultimate goal in e-commerce is maximizing Gross Merchandise Volume (GMV) rather than just clicks. This research addresses these gaps by treating ranking as a global optimization problem over sequences.

Methodology

The proposed framework decomposes ranking into estimating purchase probabilities within a sequence and finding the optimal permutation. Three primary approaches were explored:

  1. Global Feature Extension (miDNN): This model captures the context of the entire result set by calculating "global features," such as normalizing local features (e.g., price) against the minimum and maximum values in the current search result set.
  2. Ranking as Sequence Generation (miRNN): Modeling the top-to-bottom browsing behavior of users, this approach uses an RNN (specifically an LSTM) to estimate the purchase probability of an item based on the items already placed above it.
  3. Attention Mechanism: To prevent the model from "forgetting" the influential items at the top of the list, an attention mechanism was added. This allows the model to explicitly weigh the influence of top-ranked items on items further down the page.
Architecture of the miRNN+attention model, showing how previous items and their positions influence the current prediction.

To handle the computational complexity of finding the best sequence among $N!$ combinations, the authors employed a Beam Search algorithm, which maintains the top-$k$ best partial sequences at each step to efficiently find a high-quality global ranking in real-time.

Experimental Setup

The models were trained and evaluated using massive real-world datasets from Taobao Search, consisting of approximately 17 million search records and 850 million items. The feature set included 23 local features such as price, Click-Through Rate (CTR), Conversion Rate (CVR), and personalized user preference scores. Final validation was conducted via online A/B testing on Taobao’s production traffic over one month.

Results and Discussion

The miRNN+attention model significantly outperformed the baseline DNN, increasing the Area Under the ROC Curve (AUC) from 0.724 to 0.774. In live A/B tests, this translated to a 5% increase in GMV, a substantial lift for a platform of Taobao's scale.

GMV increase over baseline for different models as a function of rerank size.

Key Insights

  • Comparison Effects are Quantifiable: The performance jump from the baseline to miDNN confirms that relative value is a critical factor in consumer decision-making.
  • Long-Range Influence: Visualization of attention weights revealed that items at the very top of search results continue to influence purchase decisions for items much further down the page, justifying the use of attention-based sequence models.
  • Efficiency Trade-offs: While miRNN+attention provided the best accuracy, it increased search latency by 400% (from 21ms to 105ms). To mitigate this, the authors recommend using these complex models for "re-ranking" only the top 50 items rather than the entire catalog.

Conclusion

This work demonstrates that accounting for mutual influence through sequence modeling and attention mechanisms can significantly improve e-commerce outcomes. While computational cost remains a challenge for real-time deployment, re-ranking strategies provide a viable path forward. Future research will focus on developing more efficient attention mechanisms to capture these influences with lower latency overhead.

Summary updated

BibTeX
@inproceedings{zhuang2018globally,
  author = {Tao Zhuang and Wenwu Ou and Zhirong Wang},
  title = {{Globally Optimized Mutual Influence Aware Ranking in E-Commerce Search}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/518},
}
Figure