Can Global Ranking Boost E‑Commerce GMV? A New AI Approach

Introduction

Conventional e‑commerce search ranking scores each item independently, assuming that the click‑through rate (CTR) and conversion rate (CVR) of a product are unaffected by other products shown alongside it. In reality, the surrounding items (the display context) influence user decisions—for example, a cheaper neighboring product reduces the likelihood of buying the target item, while more expensive neighbors increase it.

This work proposes the first global ranking method that explicitly models mutual influences among items. The goal is to maximize the expected Gross Merchandise Volume (GMV) of a search session, which requires estimating transaction probabilities that depend on the entire set and order of displayed items.

Global Ranking Method

The input to the global ranking stage is a set of N items; the output is a permutation of these N items. We denote the top‑N list produced by a baseline sorter as

. The set of all possible permutations is

. For an item i in permutation o , its position is

and its display context is

. The optimization objective is to find the permutation that maximizes the expected GMV:

This problem decomposes into two sub‑problems: (1) accurately estimating the transaction probability of each item under mutual influence, and (2) solving a combinatorial optimization to select the permutation with the highest expected GMV.

2.1 Global Feature Expansion

The first class of models incorporates only the set of items in the display context, ignoring their order. Each item’s local features are compared against those of other candidates, and the comparison results are concatenated as global features. For example, price ranks are normalized to a value between 0 (cheapest) and 1 (most expensive) and used as a global price feature. The final feature vector for item i combines its local and global features, enabling the model to predict transaction probability with context awareness.

Because the item set is static, the probability estimates can be computed independently, and a simple position‑bias correction is applied:

Sorting items by the corrected scores yields the optimal ordering under this model.

2.2 Sequence Generation

The second class of models also considers the exact order of preceding items. The transaction probability of item i is modeled as a sequence probability conditioned on the previously placed items, which is naturally estimated with a Recurrent Neural Network (RNN). To address long‑range dependencies, an attention mechanism with position embeddings is added, allowing the model to learn how much influence earlier positions should have on later items.

During inference, a beam search algorithm explores high‑scoring partial permutations, retaining the best sequences until a complete ordering is produced.

Experiments

3.1 Transaction Probability Estimation

Baseline DNN uses only local features, while reDNN adds the global features described above. The miRNN model employs the RNN‑based sequence generator, and miRNN+attention adds the attention mechanism.

3.2 Online A/B Test on Taobao Search

The inference time of miRNN and miRNN+attention grows with the number of items N and beam size k , which can be prohibitive for large‑scale search. In practice, we re‑rank only the top N items from the baseline, keeping computation tractable while still capturing most of the benefit.

Across various N and k settings, the proposed global ranking methods achieve up to a 5% increase in GMV with modest latency overhead compared to the baseline DNN.

Conclusion

We present the first global ranking framework that accounts for mutual item influences in e‑commerce search, demonstrating significant GMV gains on Taobao’s main search platform. The main remaining challenge is the computational cost of the RNN‑based approach, motivating future work on more efficient models.