.png){kind=icon}
.png){kind=icon}
Return
Behavior-Aware Consistent Distillation for Cold-Start Recommendation
.png){kind=icon}
PRE
AI
H
H
L
H
F
K
Y
F
DOI：10.1109/tkde.2026.3705082.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Cold-start recommendation for users and items with no prior interactions remains a fundamental challenge, as content-driven cold-start entities differ substantially from behavior-rich warm entities. Generative methods attempt to address this by inferring embeddings from content features, while dropout-based techniques improve robustness by randomly masking behavioral signals during training. However, these approaches primarily focus on limited cold-start scenarios and do not fully capture the differences between cold-start and warm recommendations. As a result, generative models may produce unbalanced predictions and dropout strategies can reduce recommendation quality for warm entities. To address these limitations, we propose <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><u>B</u>ehavior-<u>A</u>ware <u>C</u>onsistent <u>D</u>istillation (BACD)</b>, a unified framework for cold-start recommendation for both users and items. BACD employs a behavior-informed teacher-student paradigm together with a behavior-aware teacher-quality weighting mechanism to ensure reliable supervision. In addition, it introduces three consistency-oriented distillation objectives, including <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Rating Distribution Consistency</b>, <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Ranking Consistency</b>, and <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Representation Distribution Consistency</b>, which collectively reduce the gap between cold-start and warm entities. Extensive experiments on four benchmark datasets show that BACD consistently improves recommendation performance in both cold-start and overall settings and achieves substantial gains over state-of-the-art baselines across multiple backbone models.
Keywords:
Cold-start recommendation
consistent distillation
content features
Journal
IF:
10.4
Papers:
6.7K
Citations:
3.2W
Organization
Cited Papers