Globally Optimized Mutual Influence Aware Ranking in E-Commerce Search

Tao Zhuang, Wenwu Ou, Zhirong Wang

Introduction

Traditional Learning to Rank (LTR) models in e-commerce often treat items as independent entities, scoring them based on individual features like price or click-through rate (CTR). However, real-world consumer behavior is heavily influenced by context; an item's attractiveness is relative to its neighbors on a search results page. This "mutual influence" means a product might appear as a bargain next to expensive items or overpriced next to cheaper ones. This paper proposes a framework to explicitly model these interactions to maximize Gross Merchandise Volume (GMV).

Methodology

The authors introduce a two-stage global optimization framework designed to capture the holistic context of a ranked list.

Global Feature Extension

To provide the model with a sense of the overall candidate set, "local" features (e.g., price) are augmented with "global" features. These are derived by comparing an item's attributes to the minimum and maximum values within the entire set of returned items for a query. This allows the model to perceive relative values, such as whether a price is at the high or low end of the current search results.

miRNN with Attention

The ranking process is framed as a sequence generation task. A Recurrent Neural Network (RNN), specifically an LSTM, is employed to estimate the purchase probability of an item $o_i$ conditioned on the items ranked above it ($o_1, \dots, o_{i-1}$).

Architecture of the miRNN with attention mechanism

To address the "forgetting" problem inherent in standard LSTMs and to account for the disproportionate influence of top-ranked items, an attention mechanism is integrated. This allows the model to "look back" at influential items at the top of the list when calculating probabilities for items further down.

Beam Search Optimization

Given the $N!$ possible permutations of a ranked list, finding the global optimum is computationally intractable. The framework utilizes Beam Search to efficiently explore the search space, identifying a sequence that maximizes the total expected GMV.

Evaluation and Results

The model was evaluated using Taobao search query logs (17M queries, 850M items) and live online A/B testing.

  • Offline Performance: The miRNN+attention model achieved an AUC of 0.774, significantly outperforming a 5-layer DNN baseline (0.724).
  • Online Impact: Deployment on Taobao Search resulted in a 5% increase in GMV, a substantial gain for a platform of its scale.

Key Insights

The attention mechanism provides a window into how items influence one another. Visualization of the attention matrix confirms that items at the very top of the list continue to exert a strong influence on the purchase probability of items as far down as the 20th position.

Attention matrix visualization showing the influence of top-ranked items

By explicitly defining the objective function as the mathematical expectation of GMV, the ranking strategy aligns directly with business goals rather than proxy metrics like NDCG.

Limitations and Future Work

The primary trade-off for the increased accuracy of miRNN+attention is computational cost. The model increases search latency by approximately 401% compared to the baseline. To mitigate this, it is implemented as a re-ranking step for the top 50 items rather than for the entire catalog. Future work will focus on more efficient attention mechanisms to reduce this overhead while maintaining the gains in GMV.

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