Learning To Select Historical News Articles For Interaction Based Neural News Recommendation
Though some recent “multi-channel” methods (Liu et al., 2019; Chen et al., 2018) attempt to cover richer information by maintaining multiple representation vectors, they are still limited to modeling fine-grained interaction in an explicit and reliable manner. In order to capture fine-grained matching signals between the candidate news and the user, Wang et al. Wang et al., 2020) proposed an interaction-based model. It computes similarity matrices between the candidate news. Every historical news piece of the user at word level to derive the click probability. Despite the effectiveness improvement, the model is especially slow. It has to recompute term-level interaction matrices with every historical article when scoring each candidate, which is far more expensive than dot product used in representation-based methods. Intuitively, we shouldn’t involve all the historical articles in interaction. Tailoring the user history to several recent browsed news items seems to be a straightforwar...