Text-attributed Temporal Graph Benchmark Datasets

These are six real-world temporal graph datasets with text attributes, varying in size and originating from different domains. Specifically, they are collected from culinary recipe feedback, movie reviews, book reading records, beer rating data, and online shopping interactions. In these temporal graphs, users and items are represented as nodes, while the interactions between them — in the form of user reviews or comments — serve as edges. Each edge is associated with both a timestamp and raw textual content. Additionally, each item node is accompanied by a descriptive text attribute. The files under the $dataset directory are as follows: 1、raw_node.npy and raw_edges.csv store the raw text attributes of nodes and edges, respectively. 2、ml_$dataset.csv records the temporal edges of the dataset, where each row in the format (u, i, ts) represents a user u interacting with an item i at timestamp ts. 3、$dataset_unique_labels.json contains the complete set of human-readable labels for the dataset. 4、Both $dataset_labels_text.json and $dataset_labels.json correspond to the labels associated with each edge in ml_$dataset.csv, where the former provides the textual form of the item labels that users are interested in, and the latter provides their corresponding numeric labels.