rpasricha / transfm

TransFM

This repository contains an implementation of TransFM, as described in the paper:

Rajiv Pasricha, Julian McAuley, "Translation-based Factorization Machines for Sequential Recommendation", RecSys 2018.

This repository also includes implementations of vanilla FMs, as well as the proposed PRME-FM and HRM-FM models.

Please cite the paper above if you use or extend our models.

File formats

• Input dataset

  • One example per line
  • <user_id> <item_id> <rating> <timestamp>
  • Values separated by a space
  • No header row
  • Example row: User_12 Item_65 5.0 1376697600

• Item categories

  • CSV file, one item per line
  • Expected header: item_id,item_cat_seq
  • item_cat_seq: comma separated list of item category IDs, enclosed as a string.
  • Example row: 2643,"[165, 193, 442]"

• User features

  • CSV file with numeric features, one user per line
  • Header row expected, first column should be named idx

• Item features

  • CSV file with numeric features, one item per line
  • Header row expected, first column should be named idx

• Geographical features

  • CSV file with numeric features, one item per line
  • Header row expected, first column should be named place_id

Example command

python main.py \
      --filename ratings_Automotive.txt.gz
      --model TransFM
      --features categories
      --features_file item_cat_seq_Automotive.csv.gz
      --max_iters 1000000
      --num_dims 10
      --linear_reg 10.0
      --emb_reg 1.0
      --trans_reg 0.1
      --init_mean 0.1
      --starting_lr 0.02
      --lr_decay_factor 1.0
      --lr_decay_freq 1000
      --eval_freq 50
      --quit_delta 1000

transfm's Issues

How to set the filename and feature model?

But I cant go https://jmcauley.ucsd.edu/data/amazon/

user data

How do you get the csv file with numeric features of users? Thanks for your help.

Possible error in prediction operation

Hi,

I think there could be an error in the factorization machine prediction operations. Possibly also in the translation-based factorization machine prediction operations. Its seems that all the features are summed up already in the sparse_tensor_dense_matmul operation.

The operations for the positive predictions are for example:

pos_emb_mul = tf.sparse_tensor_dense_matmul(sparse_pos_feats, var_factors)
term_1_pos = tf.square(tf.reduce_sum(pos_emb_mul, axis=1, keep_dims=True))
term_2_pos = tf.reduce_sum(tf.square(pos_emb_mul), axis=1, keep_dims=True)
pos_preds = pos_linear + 0.5 * (term_1_pos - term_2_pos)

This would result in the following model:

Instead of the desired model:

The following changed operations might fix the problem:

term_1_pos = tf.sparse_tensor_dense_matmul(sparse_pos_feats, var_factors)
term_1_pos = tf.reduce_sum(tf.square(term_1_pos), axis=1, keep_dims=True)
term_2_pos = tf.sparse_tensor_dense_matmul(tf.square(sparse_pos_feats), tf.square(var_factors))
term_2_pos = tf.reduce_sum(term_2_pos, axis=1, keep_dims=True)
pos_preds = pos_linear + 0.5 * (term_1_pos - term_2_pos)

Am I right here or am I missing something?

Best regards,
Markus

Where can I download the experimental data set?

Where can I download the experimental data set? thanks~