REASONER

REASONER is an explainable recommendation dataset. It contains the ground truths for multiple explanation purposes, for example, enhancing the recommendation persuasiveness, informativeness and satisfaction. In this dataset, the ground truth annotators are exactly the people who produce the user-item interactions, and they can make selections from the explanation candidates with multi-modalities. This dataset can be widely used for explainable recommendation, unbiased recommendation, psychology-informed recommendation and so on. Please see our paper for more details. The dataset contains the following files. REASONER-Dataset interaction.csv user.csv video.csv bigfive.csv tag_map.csv video_map.csv preview We joined them in the next way (images and video not included): interaction and users joined by the column "user_id" The resulting dataset joined with video_df by the column "video_id" The resulting dataset joined with bigfive by the column "user_id" In addition, we converted the tags in final_df from a list of tag IDs to a string of tag contents using the tag_map dataframe Finally, we delete the column "rating", since it can make things super easy for any classifier to predict the column "like" (target) paper_url = "https://papers.nips.cc/paper_files/paper/2023/file/2ebf43d20e5933ab6d98225bbb908ade-Paper-Datasets_and_Benchmarks.pdf" original_data_url = "https://reasoner2023.github.io/docs/dataset" We limited the columns name to less than 64 characters and we ensure that all the columns are unique without non-ASCII characters.