Content-aware Neural Hashing for Cold-start Recommendation

Citation: Casper Hansen, Christian Hansen, Jakob Grue Simonsen, Stephen Alstrup, and Christina Lioma. 2020. Content-aware Neural Hashing for Cold-start Recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’20). Association for Computing Machinery, New York, NY, USA, 971–980.

This link contains a downloadable version of all data+code used in the paper: https://www.dropbox.com/s/cyxgbgxvigbdw2h/cacf_data_code.zip?dl=0

The code is written for TensorFlow v. 1.12 and 1.14

Datasets

We use datasets from Yelp and Amazon with the following naming scheme:

• amazon: amazon dataset for the standard (in-matrix) setting
• amacold: amazon dataset for the cold-start (out-of-matrix) setting
• amacold_10p: amazon dataset for the cold-start setting with 10% training data. We also have _20, _30, _40 versions for 20-40% training data.

• Similar naming scheme for yelp (yelp and yelcold).

  Generating hash codes

  main.py is the main file of the project and can be called with a number of parameters, where the most important ones are:

  • --bits: the number of bits in the hash code (16-64 is used in the paper)
  • --dataset: name of the dataset (e.g., amacold)

While training, a pickle file is saved containing evaluation results and the hash codes. The hash codes can be gotten from this file like:

user_codes = filecontent[-3]
item_codes = filecontent[-2]