Dual-Sequence Fusion for New‑User Cold‑Start Recall in Content Recommendation

Background: In Taobao’s content recommendation scenario, millions of users browse daily, and new‑user cold‑start is a classic challenge. The recall stage is a crucial optimization point, currently consisting of two types of recall: a precise but low‑diversity I2I (item‑to‑item) recall and a more diverse deep U2I (user‑to‑item) recall based on a dual‑tower architecture.

Base Dual‑Tower Model: The traditional vector‑based recall model contains separate towers for user and item. The user tower ingests both content‑click and product‑click sequences, learns user embeddings, and retrieves top‑N items via vector indexing.

DAN Dual‑Sequence Fusion Network: Inspired by the VQA task, the Dual Attention Network (DAN) jointly attends to content and product sequences, producing a fused representation. Experiments show a slight improvement over the Base model (HitRate@100: 0.169 vs 0.165).

Offline Evaluation: The DAN model marginally outperforms the Base model, but the gain is limited.

Online Results: For new users, the DAN model shows almost no net benefit (pctr –0.85%, pvr +0.74%). The limited improvement is attributed to the sparsity of content sequences and the noise introduced when fusing them directly with rich product sequences.

Contextual‑Gate DAN: To address the imbalance, a Contextual‑Gate is introduced, leveraging a contextual embedding (new‑user flag, activity statistics, top‑5 product and content category preferences) to dynamically control the contribution of each sequence and to weight individual elements.

Offline Evaluation of Contextual‑Gate DAN shows consistent improvements over DAN (HitRate@100: 0.175, HitRate@50: 0.119, HitRate@5: 0.021). Online experiments on new users demonstrate notable gains (pctr +0.38%, uctr +2.70%, duration +1.45%; single‑route pvr +28.65%, pctr +11.63%).

Conclusion and Outlook: By fusing content and product sequences with a Contextual‑Gate, the system better exploits multi‑domain user behavior, achieving higher recall performance for new users. Future work includes exploring more advanced sequence‑fusion techniques, extending fusion to all users, and combining with multi‑interest models such as MIND and MultCLR.

References: [1] Nam, H., Ha, J. W., & Kim, J. "Dual Attention Networks for Multimodal Reasoning and Matching." CVPR 2016. [2] Centauri‑batch engine. [3] Verma, S. et al. "Deep‑HOSeq: Deep Higher Order Sequence Fusion for Multimodal Sentiment Analysis." ICDM 2020. [4] Yu, Y., Kim, J., & Kim, G. "A Joint Sequence Fusion Model for Video Question Answering and Retrieval." ECCV 2018. [5] MIND: Multi‑Interest Network with Dynamic routing. [6] MultCLR: Multi‑Vector Retrieval Explore & Exploit.

Team Introduction: The authors are from Alibaba’s Taobao Content Recommendation team, highlighting large business scope, strong infrastructure, and a research‑oriented culture. They invite interested candidates to contact them via email.