Skip to content
Dario Loi
Search
Menu

Next-POI recommendation

A transformer-based hierarchical graph neural network for next point-of-interest recommendation, built in PyTorch Lightning for my M.S.

Next-POI recommendation. The model as a diagram. Seven inputs, the last visited place, the user, and the place at five coarser levels labeled G@2 to G@6, enter an embedding layer, then a block labeled GRN with an attention box inside that also takes two small graphs, then a feed-forward layer that outputs the next place at each level.

This M.S. project trains and tests a transformer-based hierarchical graph neural network that predicts the next point of interest a user will visit, from check-in sequences in the New York Foursquare dataset. It is written in PyTorch Lightning.

It improves on prior methods by +2.3% top-1 and +13.7% top-20 accuracy. The repository holds the end-to-end training and evaluation code and the written report.

  • PyTorch Lightning
  • Graph neural networks
  • Recommendation