Description Problem Description GNN has performed very well in node classification and link prediction of heterogeneous and isomorphic graphs. In many recommendation scenarios, the trust relationship between users and the essential attributes of products are used, combined with users' purchase records, and using the representation generated by GNN can also achieve good results. However, in the scenarios given by the academic platform, we can find that collaborator recommendations, paper recommendations, and reviewer recommendations for journals and conferences are the main tasks.
For most of the above problems, there are relatively good benchmarks and baselines. However, in the paper recommendation scenario, if it is for scholars conducting scientific research through their collaborators and the paper community where their papers are located, citation recommendation still lacks better datasets and models. A dataset we share here is used in a recommender system. When both the user and product have an associated network, we can extract group features for the user group and product category. Based on the relation between products and the connections between users, it provides users with better and more various recommendations, and this may help to solve the problem of less user behavior caused by the cold start of the recommender system.
Here is a link prediction problem in an academic network. We collected 6,611 authors and corresponding 79,937 papers from top journals in the field of GeoScience as well as citation information of their publications. The collected information is used to form an academic network, and there is a feasible way:
Build a heterogeneous network, which contains two types of nodes, one type of nodes represents authors, and the other represents papers. In this network, each edge between an author node and a paper node means that the authors have read the paper (connecting authors and the papers cited by the papers written by the authors), each edge between two author nodes denotes the co-authorship, and each directed edge between two paper nodes represents the citation relation.
You can propose some other way to form an academic network. Note that what we provide is the author and paper information so far, and it reveals the correlation between work of different authors. Consequently, suppose that you are designing an academic reading recommendation system, and you need to pick out papers related to the author's previous research. This problem can be modelled as a link prediction problem, and your task is to predict each author-paper pairs in the test set based on the information provided. If the paper is recommended to the author, mark it as 1, otherwise mark it as 0.
Attention Before submitting your predictions, please update your name with "GroupNumber_StudentName" before submitting your predictions. You can update your name in the "Team" tab.
For fairness, we set a few rules for you to obey --- violation against the rules will lead to score deduction:
Please do not copy someone else's code. We will run plagiarism check after submission. Please do not download the dataset somewhere else to train your model. We have shuffled the dataset and will reproduce your experimental results using the dataset. A violation would be considered if there is a big gap between your reported results and our reproduced results. Please do not use the pre-trained model. Everything should be built from scratch. Acknowledgements Data used in this project is provided by Acemap group.