ContextFair

Exploring the Impact of Temporal Bias in Point-of-Interest Recommendation (RecSys 22)

Note

The paper has been submitted to the RecSys conference and upon acceptance, the details of codes, datasets, and results will be made available.

Context Fairness

Contents

• Folders:
  • datasets: including datasets (Gowalla and Yelp)
    • In Yelp dataset we do not have the frequecny of ckeck-ins, we have the rating to each location. For example, in Yelp_checkins.txt file we have 39 check-ins for user 0 and the numebr of records in Yelp_train.txt + Yelp_test.txt + Yelp_tune.txt is equal to 39. Should we consdier them as a frequency score?
  • item_groups: groups of items (leisure and working items)
  • plots: plots of analysis and results
  • results: results of models
  • user_groups: groups of users
• Notebooks:
  • XXX:

How to run the pipeline?

1. We use the generated results from the previous experiments (from ESWA'21) on Gowalla (i.e., files in results/Gowalla)
   • Each folder indicates a model.
   • The main files are results_top_N.txt which are the user scores per metric.
2. Run temporalSplit.ipynb to split users based on the different timestamp and generate some plots to show the correlation between attributes.
   • Outputs: plots in \plots and user and item groups in \groups
3. compute_results.py
4. merge_results.ipynb