原论文：Search-based User Interest Modeling with Lifelong Sequential Behavior Data for Click-Through Rate Prediction
It is mainly to solve the computational complexity problem brought by long sequences.
The solution is the advanced search in the first stage, and there are two ways of softsearch+hardsearch.
Then use the idea of mult-head, and use the candidate target item as the key to query the sequence of the topk after the search.
