Alibaba’s AI Team Sponsors CIKM 2025 and Launches the AnalyticCup Competition

1. Alibaba International Intelligent Technology Papers at CIKM 2025

1.1 Distribution‑Guided Auto‑Encoder for User Multimodal Interest Cross Fusion

Traditional recommendation methods compute similarity between candidate item embeddings and embeddings of items a user previously clicked, capturing implicit collaborative‑filtering signals. To alleviate item‑ID sparsity, current models fuse multimodal information such as text and images, but they often rely on early‑fusion and ignore the dynamic context provided by user behavior sequences, weakening the ability to capture multimodal user interests.

The proposed Distribution‑Guided Auto‑Encoder (DMAE) performs cross‑fusion of multimodal interests from the user‑behavior perspective. Extensive offline and online A/B experiments demonstrate that DMAE outperforms existing multimodal recommendation approaches.

Paper link: https://dl.acm.org/doi/10.1145/3746252.3761367

1.2 Global‑Distribution Aware Scenario‑Specific Variational Representation Learning Framework

Most recommendation systems use a single model for all scenarios, sharing underlying feature representations, which limits the model’s ability to capture scenario‑specific characteristics. Users and items exhibit distinct behaviors across scenarios, requiring scenario‑specific representation vectors, yet data sparsity is even more severe in multi‑scenario settings.

The proposed Global‑Distribution Aware Scenario‑Specific Variational Representation (GSVR) framework can be directly applied to existing multi‑scenario recommendation models. Large‑scale offline and online A/B results confirm GSVR’s effectiveness in learning more robust scenario‑specific representations.

Paper link: https://dl.acm.org/doi/10.1145/3746252.3760866

1.3 Dynamic Reserve Price Design with Distributed Solving Algorithm

In sponsored search, unexpected ads can erode user trust in organic results, causing hidden losses for e‑commerce platforms.

To address this, a Dynamic Reserve Price framework is proposed that incorporates implicit costs into the auction mechanism, enabling precise decisions on whether to sell traffic. The design aims to balance revenue and user experience while keeping advertisers’ long‑term incentives.

The framework includes a scalable distributed algorithm capable of handling billions of data points in production. Offline tests and online A/B experiments show strong computational efficiency and industrial deployability; real‑world deployment validates its effectiveness in increasing platform revenue and preserving user experience.

Paper link: https://arxiv.org/abs/2206.10295

2. AnalyticCup Competition and Workshop on Multilingual E‑commerce Search

The International Digital Commerce Group (AIDC) operates platforms such as AliExpress, Lazada, Daraz, Miravia, TaoJP, and the overseas version of Taobao, covering over 100 countries and supporting more than 20 languages, making high‑quality multilingual search essential.

In the past two years, rapid advances in large language models (LLM) have transformed information retrieval, reshaping core components from query intent understanding and rewriting to retrieval, ranking, and even annotation workflow automation.

AIDC’s AI team hosted the CIKM 2025 AnalyticCup competition, focusing on two fundamental multilingual search tasks: (1) query‑category relevance and (2) query‑product relevance. A total of 110 teams from many countries participated; after the preliminary round, 12 teams advanced to the finals, and the top‑10 rankings are shown below.

On November 14, the Alibaba International E‑commerce Product Search Competition Workshop was held, where winning teams presented technical reports and Alibaba AI engineers reported on LLM applications in international e‑commerce search, followed by in‑depth discussions.

Competition link: https://tianchi.aliyun.com/competition/entrance/532369?lang=en-us

Workshop link: https://alibaba-international-cikm2025.github.io/