Includes reference implementations for:
- BMR: https://arxiv.org/pdf/1205.2618.pdf
- VBMR: https://arxiv.org/pdf/1510.01784.pdf
- TVBMP: https://arxiv.org/pdf/1602.01585.pdf
recommender systems trained on an Amazon Reviews dataset:
Dataset: http://jmcauley.ucsd.edu/data/amazon/
Original Code: https://sites.google.com/a/eng.ucsd.edu/ruining-he/ ( Ruining He)
Traditional gcc tool-chain setup, i.e. not a mac (LLVM). * See Docker notes below *
- compiler support for openmp (ubuntu:
sudo apt install libomp-dev) - armidillo lib (ubuntu:
sudo apt-get install libarmadillo-dev)
All the models need the reviews dataset in a simplified format (CSV), use the convert_to_simple.py util to do this:
python convert_to_simple.py reviews_Clothing_Shoes_and_Jewelry.json.gz reviews_simple.gz
It is now possible to train w/ this output file which is all reviews across all categories of clothing. It is often useful to segment across categories:
It is useful for evaluation to segment the clothing dataset into:
- women
- men
- boys
- girls
- baby
using the getClothingSubReviews script.
For input it takes the gzipped simplified output from above and a meta data file which has some category mappings:
https://storage.googleapis.com/sharknado-recsys-public/productMeta_simple.txt.gz
The command takes the simplified dataset file and the metadata as input:
./getClothingSubReviews ../data/clothing_full.gz ../data/productMeta_simple.txt
This will generate these segmented review files (which you need to gzip to use in the training program):
reviews_Baby.txt
reviews_Boys.txt
reviews_Girls.txt
reviews_Men.txt
reviews_Women.txt
To build the suite, run the Makefile, by typing make in the project directory. This will build a binary named train.
To train all the models, pass in the following args:
- Review file path — the simplified dataset
- Img feature path — the image features
- Latent Feature Dim. (K) — hyperparameter
- Visual Feature Dim. (K') — hyperparameter
- alpha (for WRMF only) — hyperparameter
- biasReg (regularizer for bias terms) — hyperparameter
- lambda (regularizer for general terms) — hyperparameter
- lambda2 (regularizer for "sparse" terms) — hyper-parameter
- Epochs (number of epochs)
- Max iterations
- Corpus/Category name under "Clothing Shoes & Jewelry" (e.g. Women)
Although some parameters don't apply to some of the models, they are just passed in bulk.
To start the training routine:
./train simple_out.gz image_feat.gz 20 k2 alpha 10 10 lambda2 epoch 10 "Clothing/women"
Example Output:
{
"corpus": "simple_out.gz",
Loading votes from simple_out.gz, userMin = 5, itemMin = 0 ....
Generating votes data
"nUsers": 39387, "nItems": 23033, "nVotes": 278677
<<< BPR-MF__K_20_lambda_10.00_biasReg_10.00 >>>
Iter: 1, took 0.266870
Iter: 2, took 0.247799
Iter: 3, took 0.250260
Iter: 4, took 0.248863
Iter: 5, took 0.262415
[Valid AUC = 0.358993], Test AUC = 0.360745, Test Std = 0.305036
Iter: 6, took 0.245542
Iter: 7, took 0.245433
Iter: 8, took 0.236978
Iter: 9, took 0.236842
Iter: 10, took 0.234738
[Valid AUC = 0.610926], Test AUC = 0.611459, Test Std = 0.294283
<<< BPR-MF >>> Test AUC = 0.611459, Test Std = 0.294283
<<< BPR-MF >>> Cold Start: #Item = 11453, Test AUC = 0.554613, Test Std = 0.300705
Model saved to Clothing__BPR-MF__K_20_lambda_10.00_biasReg_10.00.txt.
}
The training routine heavy make use of threading. In experimentation we have been able to utilize 15 cores during training. This is especially important during the expensive AUC calculation.
##Docker
To run the docker image:
docker run -v ~/Development/DSE/capstone/UpsDowns:/mnt/mac -ti updowns /bin/bash
the v flag will mount the source repot to /mnt/mac in the container.
On Amazon, regularization hyper-parameter lambda=10 works the best for BPR-MF, MM-MF and VBPR in most cases.
../tools/getClothingSubImgFeatures _image_features_Clothing_Shoes_and_Jewelry.b productMeta_simple.txt.gz
./train data/reviews_Women.txt data/image_features_Women.b 10 10 na 1 1 1 na 10 "women"