I\u2019m working on a temporal knowledge graph (TKG) model for link prediction and graph generation. Basically, I have snapshots of a persistent knowledge graph over time (subject, relation, object) triplets, and I want to train the model to autoregressively predict the next graphs over a sequence of timesteps. For training, it takes in a graph at timestep t and predicts the graph at timestep t+1.<\/p>\n
Unfortunately, I’m running into a pretty severe issue: the model overfits almost immediately, and Hits@K stays basically random.<\/p>\n
Current dataset:<\/p>\n
I’m currently using wikidata12k, which is a pretty small dataset, which I think may be causing some of the issues. It gives me about 200 knowledge graphs, one for each year from 1800 to 2020, each about 500 nodes.<\/p>\n
I would actually love a bigger dataset, but it has to be in a persistent knowledge graph format, which means the graph changes slowly over time, and the graph at timestep t is similar to the graph at timestep t+1. This unfortunately rules out a lot of popular TKG datasets like ICEWS.<\/p>\n
I’ve also looked at YAGO11k, but it suffers from the same lack of scale as wikidata12k.<\/p>\n
I’ve made another post in r\/learnmachinelearning<\/a> with details about the architecture and other issues I’m facing, which you can check out if you want more details.<\/p>\n https:\/\/www.reddit.com\/r\/learnmachinelearning\/comments\/1sjl7ck\/temporal_gnn_gat_pernode_lstm_overfitting\/<\/a><\/p>\n Thank you so much for the help, and I’m happy to answer any additional questions <\/p>\n<\/div>\n submitted by \/u\/Divine_Invictus <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
