De-Anchor: Mitigating Attention Polarization for Lifelong User Behavior Modeling in Click-Through Rate Prediction

User lifelong behavior sequences are essential for click-through rate (CTR) prediction tasks in industrial recommender systems. Attention-based module, especially multi-head target attention (MHTA), has been proven to be effective in aggregating behavior features given a certain target item. However, we found a common phenomenon that attention weights in MHTA tend to over-concentrate on merely a small subset of a user's historical behaviors, producing a sparse one-hot distributed attention weights in training gradually, which we callAttention Polarization (AP). These polarized weights on certain behaviors (which we call " anchor ") could make the model fail to capture a user's diversified interests, and harm the learning of behavior embedding as the gradients on these features are nearly zero. We introduce two indicators:anchor rate andattention entropy to measure the magnitude of AP, and proposeDe-Anchor , a novel method to alleviate it, which can serve as a stand-alone and parameter-efficient plug-in to existing CTR backbones. De-Anchor contains two modules:Anchor-aware gradient dropout (AGD) forcing the model to capture diversified interest information from behavior sequences by discarding gradients of non-behavior features, andTarget-aware attention anchor (TAA) providing a pseudo behavior to offload excessive weights of MHTA. Extensive offline experiments and industrial online A/B tests demonstrate the efficacy of our method.