Douban Dataset with Multiple Downstream Tasks for Network Embedding

Our new network dataset is crawled from Douban Movies{https://movie.douban.com}, which is a website providing users comments on movies. Each node in the network represents a movie, and each edge represents that the movies on both ends of it are co-preferenced by audiences, which is provided by Douban. The network contains 31,761 nodes and 179,924 edges. We use the movie profiles to form the attributes of the node. Firstly, we use ``jieba''{https://github.com/fxsjy/jieba}, a widely used Chinese word segmentation tool, to segment movie profiles and filter common stop words and words that appear less than three times in the corpus. Then, we build a TF-IDF vector for each movie using scikit-learn and reduce the dimension to 500 via SVD. We build three downstream tasks for this Douban dataset, including movie genres prediction, rating score level prediction, and popularity level prediction. Genre predicting task is a multi-classification task, we directly use the genres of the movie provided by Douban as the label and each movie has at least one genre. To build the label of the rating score prediction, we rank movies by rating scores and divide them into 10 classes of the same size. Similarly, we rank all the movies according to the number of comments and divide them into three classes of the same size. For each task, we randomly sample 70\% nodes as the train set, 10\% as the validation set, and the rest as the test set.