How Alibaba’s Brand‑Level Ranking Boosts E‑Commerce Clicks with Attention‑GRU

1.1 Introduction

On e‑commerce sites such as Taobao, brand preference increasingly influences user clicks and purchases because users care about product quality, which brands signal. Existing ranking systems are not designed for brand preference, leading to sub‑optimal results or extra interaction costs. We propose the first brand‑level ranking system to address this gap.

Our solution first designs feature engineering for personalized brand ranking, then builds an Attention‑GRU model with three important enhancements to improve brand ranking accuracy. Experiments on Alibaba's Tmall platform validate the model's effectiveness and user feedback.

1.2 Related Work

1.2.1 RNN, GRU and Attention‑GRU

RNNs have proven effective for sequence modeling. The GRU mitigates gradient vanishing and is more compact than LSTM. Attention‑GRU incorporates an attention mechanism to weigh different inputs, which has been successful in tasks such as speech synthesis and machine translation.

1.2.2 RNN models for behavior modeling

Previous works have applied RNNs to session‑based recommendation and basket recommendation. The Time‑LSTM adds a time gate to capture long‑ and short‑term interests, which we also embed into our model.

1.3 Task Definition and Model Adaptation

We denote the set of M users as U and the set of N brands as B. For each user u∈U, their historical behavior sequence H_u consists of actions (click or purchase) on brands at specific timestamps. The task is to predict the probability that user u will perform an action on brand b at a given time.

We adapt traditional RNN models to this task by converting user‑brand interactions into suitable feature vectors, using either one‑hot encoding or learned embeddings, and incorporating time intervals.

1.4 Brand‑Level Ranking System

1.4.1 Feature Engineering

Analyzing massive user logs on Taobao and Tmall, we find price intervals crucial for brand modeling. We define seven price levels and, for each brand, construct a 56‑dimensional vector by concatenating eight e‑commerce metrics for each price level.

1.4.2 Ranking Model Design

We adopt an Attention‑GRU as the base model and introduce three improvements:

Integrate heuristic brand features with learned brand embeddings to better represent brands.

Model different behavior types (click vs. purchase) by assigning separate matrices and explicitly multiplying them with brand representations.

Incorporate a time gate (from Time‑LSTM) to model intervals between actions, capturing both short‑ and long‑term user interests.

These enhancements are illustrated in the accompanying figures.

1.4.3 Loss and Training

The model predicts a probability p̂ for each user‑brand pair. We use a binary cross‑entropy loss weighted by a factor w for negative samples and train with AdaGrad.

1.5 Offline Experiments

1.5.1 Dataset

We collected a large real‑world dataset from Tmall, containing 3,591,372 users, 90,529 brands, and 82,960,693 interactions after filtering low‑activity users and brands.

1.5.2 Baselines and Evaluation

We compare our Attention‑GRU‑3M (the model with three improvements) against GRU, vanilla Attention‑GRU, Time‑LSTM, Session‑RNN, and libFM. All RNN baselines use 256 hidden units; hyper‑parameters follow the original papers.

Evaluation metrics are AUC and F1.

1.5.3 Results and Analysis

Attention‑GRU‑3M outperforms all baselines, confirming the benefit of attention and our three improvements. Ablation studies show that removing any improvement degrades performance, with the first improvement having the largest impact, especially as training data size grows.

1.6 Online Experiments

We conducted a 7‑day A/B test on Tmall with a “Brand” button that switches users to the brand‑level ranking interface. The new version achieved higher CTR, ATIP, and a 3.51% increase in GMV, demonstrating real‑world impact.

1.7 Conclusion

We introduced a brand‑level ranking system that personalizes brand ordering by combining engineered brand features with an enhanced Attention‑GRU model. The system significantly improves user experience and commercial metrics. Future work will extend the framework to incorporate multi‑entity (user‑tag‑item) relationships.

References

[1] D. Zipser, Y. Chen, F. Gong. The modernization of the Chinese consumer. McKinsey Quarterly, 2016. [2] Y. Zhu et al. What to do next: Modeling user behaviors by Time‑LSTM. IJCAI‑17, 2017. [3] J. K. Chorowski et al. Attention‑based models for speech recognition. NIPS, 2015. [4] K. Cho et al. Learning phrase representations using RNN encoder‑decoder for statistical machine translation. arXiv:1406.1078, 2014. [5] D. Bahdanau, K. Cho, Y. Bengio. Neural machine translation by jointly learning to align and translate. arXiv:1409.0473, 2014. [6] J. L. Elman. Finding structure in time. Cognitive Science, 1990. [7] A. Graves. Generating sequences with recurrent neural networks. arXiv:1308.0850, 2013. [8] V. Mnih et al. Recurrent models of visual attention. NIPS, 2014. [9] B. Hidasi et al. Session‑based recommendations with recurrent neural networks. ICLR, 2016. [10] Y. K. Tan et al. Improved recurrent neural networks for session‑based recommendations. RecSys, 2016. [11] B. Hidasi et al. Parallel recurrent neural network architectures for feature‑rich session‑based recommendations. RecSys, 2016. [12] F. Yu et al. A dynamic recurrent model for next basket recommendation. SIGIR, 2016. [13] S. Hochreiter, J. Schmidhuber. Long short‑term memory. Neural Computation, 1997. [14] J. Duchi et al. Adaptive subgradient methods for online learning and stochastic optimization. JMLR, 2011. [15] J. Chung et al. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv:1412.3555, 2014. [16] R. Jozefowicz et al. An empirical exploration of recurrent network architectures. ICML, 2015. [17] J. O. Berger. Statistical decision theory and Bayesian analysis. Springer, 2013. [18] S. Rendle. Factorization machines with libFM. TIST, 2012. [19] R. Pan et al. One‑class collaborative filtering. ICDM, 2008. [20] X. Huang et al. Dynamic web log session identification with statistical language models. JASIST, 2004. [21] M. Kim, J. Leskovec. Nonparametric multi‑group membership model for dynamic networks. NIPS, 2013.