search

yingtaoluo / Spatial-Temporal-Attention-Network-for-POI-Recommendation

Codes for a WWW'21 Paper. POI recommender system for location/trajectory prediction.
https://doi.org/10.1145/3442381.3449998
170 stars 38 forks source link

Some questions about model training #12

Closed chenliang-zhang closed 2 years ago

chenliang-zhang commented 3 years ago

Hi！In train.py， you use first 100 sequences as sample to train model and get great performence. But if I use the whole dataset instead of the first 100, I can't get similar recall in the paper. Is there any settings other than initial learning rate should be changed for bigger dataset?

yingtaoluo commented 3 years ago

Yes. As we have different user embedding in the paper, and each user is modeled individually, it is fine to train on many small proportions of users and then average the performance. Otherwise, we need to increase the model capacity/# hyperparameters such as embedding size.

yingtaoluo commented 3 years ago

When comparing with other baselines, the same (proportion) setting should be used, otherwise, there is a fairness problem. When training other baselines, we find it difficult to reach similar results reported in their original papers. It is a compromise.