Home

Awesome

Time-LSTM on RecSys

Introduction

This project is the implementation of the paper "What to Do Next: Modeling User Behaviors by Time-LSTM".

Abstract of The Paper

Recently, Recurrent Neural Network (RNN) solutions for recommender systems(RS) are becoming increasingly popular. The insight is that, there exist some intrinstic patterns in the sequence of users' behaviors, and RNN has been proved to perform excellently tasks such as language modeling, RNN solutions usually only consider sequential order of objects from a sequence without the notion of interval. However, in RS, time intervals between users' behaviors are of significant importance in capturing the relations of users' behaviors and the traditional RNN architectures are not good ar modeling them. In this paper, we propose a new LSTM variant, i.e. Time-LSTM, to model users' sequential behaviors. Time-LSTM equips LSTM with time gates to model time intervals. These time gates are specifically designed, so that compared to the traditional RNN solutions, Time-LSTM better captures both of users' short-term and long-term interest, so as to improve the recommendation performance. Experimental results on two real-world datasets show the superiority of the recommendation method using Time-LSTM over the traditional methods.

Model

Time-LSTM proposed in the paper has three different architectures.

Requirments

The code is tested in the following envirenment.
theano=0.9.0
lasagne=0.2.dev1
pandas=0.18.1
cudnn=5.1
cuda=8.0

pip install -v theano==0.9.0
pip install --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip
pip install -v pandas==0.18.1

Data Preprocess

Download and put the original data file into data/music or data/citeulike.
For the citeulike dataset, you need run awk -F "|" '{print $1"|"$2"|"$3}' citeulike_dataset | uniq > citeulike-origin-filtered to get citeulike-origin-filtered and filter users and items with few interactions.
Run the python file in the preprocess folder (e.g. lastfm.py) to get three files: user-item.lst, user-item-delta-time.lst and user-item-accumulate-time.lst.
Use head/tail command (e.g. head -800 user-item.lst > tr_user-item.lst and tail -192 user-item.lst > te_user-item.lst) to generate the following 6 files for each(in data/{data_source}/).

tr_* is traing data. te_* is testing data. The content is something like:

user1, item1 item2 item3 ...
user2, item1 item2 item3 ...

Files

plstm.py: phased-lstm
tlstm1.py: TLSTM1
tlstm2.py: TLSTM2
tlstm3.py: TLSTM3
lstm.py: LSTM
utils.py: load data, save module, and some other method...
main.py: main process to train.

train.sh: command samples.

To train the model, run:

THEANO_FLAGS="${FLAGS}" python main.py --model TLSTM3   --data ${DATA} --batch_size ${BATCH} --vocab_size ${VOCAB} --max_len ${MLEN} --fixed_epochs ${FIXED_EPOCHS} --num_epochs ${NUM_EPOCHS} --num_hidden ${NHIDDEN} --test_batch ${TEST_BATCH} --learning_rate ${LEARNING_RATE} --sample_time ${SAMPLE_TIME}

or just run (the hyperparameters in train.sh are not carefully tuned):

bash train.sh

Important arguments: