Alibaba's AKT-Rec uses LLM semantic IDs to boost long-tail recs by 3.5% GMV

Long-tail recommendation remains a tough challenge for e-commerce platforms due to severe data imbalance. Standard approaches that blend multimodal content features with collaborative signals often stumble because noisy signals from tail items can degrade representation learning for head items. A new paper from Alibaba researchers introduces AKT-Rec (Asymmetric Knowledge Transfer for Recommendation), which tackles this asymmetry directly. The framework first uses a Multimodal LLM (MLLM) with supervised fine-tuning to align content representations (text, images) with collaborative information for both users and items. These learned representations are then discretized into semantic IDs via a Residual-Quantized VAE (RQ-VAE), creating semantic clusters of similar entities.

AKT-Rec's core innovation lies in two components. First, Cluster-Guided Adaptive Embedding decomposes each ID into a cluster-level embedding (capturing shared semantics) and an individual embedding. An asymmetric contrastive objective, combined with an activity-aware gating mechanism, ensures knowledge flows from head to tail while preventing tail noise from contaminating head representations. Second, Hierarchical Feature Aggregation builds parallel feature views and adaptively fuses them to optimize predictions for items with varying activity levels. Extensive experiments on a large-scale industrial dataset and online A/B testing on Alibaba Tmall show AKT-Rec outperforms strong baselines: offline AUC improved by 0.35%, GAUC by 1.53%, and online CTR increased 2.76% with a 3.47% GMV lift. The paper is available on arXiv with the code expected to follow.