Implementation of "A Neural Autoregressive Approach to Collaborative Filtering"

• Chainer 1.10.0

Download MovieLens 1M/10M dataset from http://grouplens.org/datasets/movielens/ and extract it.

$ python src/convert_dataset.py dataset/ml-1m/ratings.dat dataset/ml-1m/movies.dat dataset/ml-1m/ratings.pkl

For 10M dataset you have to specify minimum rating value and rating unit .

$ python src/convert_dataset.py -m 0.5 -u 0.5 dataset/ml-10M100K/ratings.dat dataset/ml-10M100K/movies.dat dataset/ml-10M100K/ratings.pkl

For MovieLens 1M dataset:

$ python src/train.py -g 0 -o model/test.model -d dataset/ml-1m/dataset.pkl -b 512 --lr 0.001

For MovieLens 10M dataset:

$ python src/train.py -g 0 -o model/test.model -d dataset/ml-10M100K/dataset.pkl -b 512 -e 50 --lr 0.0005

Optiions:

• -g (--gpu) <GPU device index>: Optional
  GPU device index. Negative number indicates using CPU (default: -1)
• -o (--output) <File path>: Required
  Output model file path
• -d (--data_file) <File path>: Required
  Dataset file path
• -b (--batch_size) <int>: Optional
  Mini batch size for training (default: 512)
• -l (--layer_num) <int>: Optional
  Number of neural network layers (default: 1)
• ordinal_weight float: Optional
  Ordinal loss function weight (default: 1)
• --iter <int>: Optional
  Iteration of training (default: 200)
• --save_iter <int>: Optional
  Iteration of saving model (default: 10)
• --lr <float>: Optional
  Learning rate: alpha of Adam (default: 1e-3)
• --lr_decay_iter <int>: Optional
  Iteration interval of learning rate decay. (default: 60)
• --lr_decay_ratio <int>: Optional
  Ratio of learning rate after decay. (default: 0.25)
• --weight_decay float: Optional
  Weight decay (default: 0.015)
• --random_seed <int>: Optional
  Random seed (default: 1)
• --item_base: Optional
  Do item-base prediction if set