IAT: Instance-As-Token Compression for Historical User Sequence Modeling in Industrial Recommender Systems

A research team from Alibaba, led by Xinchun Li, has introduced a novel framework called Instance-As-Token (IAT) to overcome a fundamental bottleneck in industrial recommender systems. While advanced sequence models exist, their performance is often limited by the information capacity of manually crafted sequential features. IAT addresses this with a two-stage process: first, it compresses all features from a single user interaction (like a click or purchase) into a unified, dense 'instance embedding'—essentially turning a complex event into a compact, informative token. The paper proposes two compression schemes, with a 'user-order' method proving most effective for downstream tasks.

In the second stage, downstream recommendation tasks fetch a fixed-length sequence of these compressed instance tokens based on timestamps. This allows the system to use standard, powerful sequence modeling techniques (like Transformers) to learn long-range user preference patterns from a much richer, condensed history. The research demonstrates that IAT significantly outperforms current state-of-the-art methods in experiments and shows excellent transferability across different domains. Crucially, it's not just an academic exercise; IAT has already been deployed across several of Alibaba's major business lines, including e-commerce advertising, shopping mall marketing, and the high-growth live-streaming e-commerce sector, where it has delivered substantial improvements in core business metrics.