NBCRS

Implementation for NBCRS: Neighborhood-based Collaborative Filtering for Conversational Recommendation (Recsys 24).

The take away is KNN is a fairly strong baseline for CRS, too, esp when you don't have KGs or have a large dataset!

Should I use NBCRS?

Hopefully the figure below sheds some light on when to use which class of models - do double check your current SOTA so you don't miss things!

datasets

Processed datasets (Inspired, Redial, Reddit) are in datasets folder.

training

• Training code is in train_knnlm.py, see modeling_nmf.py for the actual knnlm model.
• See train_knnlm.sh for bash commands for training the model.

inference

• see inference_knnlm.ipynb for the code for tuning number of neighbors to use for the KNN component and doing inference on test set for the datasets.

environment

• see requirements.txt which is exported via conda list -e > requirements.txt.

Procedure for running on a dataset:

• First, run generate_embeddings.py, to generate semantic embeddings by factorizing item-item co-occur matrix. (mainly to stablize training)
• Then, run train_knnlm.py. We use huggingface style training pipeline.
• Run inference_knnlm.ipynb to get evaluation results
  • Got way higher numbers for Inspired/Redial than in the paper? We found that the way we processed the data (including both movie and non-movie entities as target items during training while evaluating on predicting movies) results in lower numbers for the models compared to prior works.
  • Now, we exlucde non-movie entities during prediction by default. See usage of inspired/redial_eligible_entities variables for details in inference_knnlm.ipynb; should be fairly easy to switch of this behavior by commenting out the post-filtering line.