This is the source codes of Recsys2023 best short paper "Interpretable User Retention Modeling in Recommendation".
- Dataset for offline training: a small dataset from zhihurec "https://github.com/THUIR/ZhihuRec-Dataset".
- The dataset is zhihurec, a public dataset released by THUIR.
- This is the old version of zhihurec released by THUIR. And IURO used this old version. 链接:https://pan.baidu.com/s/1dKUln3FX5KDkr3rGdLtxuw 提取码:aecr
- Then download three file in this directory, including "answer_infos.txt", "user_infos.txt" and "zhihu1M.txt" Of course, the old version is the same as the new version, only with a small difference in ID. You can replace this old version with the new version by some small changes (including the process of xx_ID and the structure of log).
-
online serving We generate a small dataset for online serving evaluation, including candidate user pool and candidate item pool. The goal of online serving is to dynamically recommend high-quality aha items (selected from candidate item pool) to candidate users, empowering industrial recommender systems together with traditional models, such as CTR models.
-
We have made some improvements to the original IURO to make it more suitable for online serving. Although this has a negative impact on offline evaluation slightly, it is well known that online retention improvements in industry recommender systems should be more of a concern than offline evaluation. Some of the latest online evaluation will be presented in our subsequent work.