https://baijiahao.baidu.com/s?id=1726181167689858723&wfr=spider&for=pc
Graph structure: user - micro video -> concept (semantic information extracted from short video titles or comments) Heterogeneous three parts
Model operation: warm-up propagation -> graph denoising -> preference refinement
Preheating treatment: add information aggregation of attention mechanism, update embedding, and transfer information to concept->micro video->user (after this step, concept information is integrated into the embedding of user and item)
Graph denoising: Users may have erroneous behaviors to generate noise. The method used in this paper is user-centered BFS sampling, which refers to removing the noise nodes of micro-video and concept to generate a new graph that can better express user preferences. Due to the use of sampling for denoising, the entire learning model becomes non-differentiable. -> Model is not trainable. Introducing Gumble-softmax to solve discrete choice inseparability.
Preference refinement: Repeat the first step on the new graph to generate new and more accurate embedding vectors.