lhenry15/csce670

CSCE670

Lin 2019-04-17:

Include GRU4REC in repository.

I used GRU4REC framework to predict next item using current item in hope of finding some time relation between items.
I modified the way GRU4REC does negative sampling. Negative sample are selected randomly from training set. Unlike GRU4REC sampled them from the same batch.

How to run:

1. pip install -r requirements.txt
2. Use preprocess.py to separate train.csv to train_tr.csv and validation.csv
3. run main.py with desired hyper-param.

Further Improvement

1. Embed action_type and concate with reference embedding.
2. Sample more negative sample. As pointed out in the paper, more negative sample can improve performance a lot.
   But using all samples will need tremendous amount of memory.
3. Increase hidden unit number.
4. Try TOP1 max or BPR max loss.

Henry:

script/validate.py: use for offline checking the performance by sampling some sessions and evaluate on training data.

script/other_scrtipts: for feature extraction

script/rerank.py: reranking via xgboost

Note: For now, I only extract the user-item relations and construct the user-item graph(mixed.edgelist) by some heuristics, and apply some GNN algorithm for learning the item representations.

For recommendation scenario, I apply the concept of item-based KNN. Based on the learned item representation from GNN algorithm, I make recommendations for each session by calculating the similarity in the learned embedding space for the items in the session, and choose the top-k most similar items for the recommendation list.

Performance of Embedding Algorithm

HopRec: 0.4859 MF: 0.4752 APP: 0.4657 HPE: 0.4498 Deepwalk: 0.4151 Item-Features only: 0.4398

P.S: We need to find a way to share the features, they are too big to share on github.

Chao:

Metadata Preprocessing.ipynb: extracting the item features from metadata, and convert the features into a metrix for all items. It also extracts the locations and user interactions as the features for items.

2layer_30d_AutoEncoder.py: a two-layer autoencoder that can reduce the dimentionality of the item feature matrix into a 30-dimensional matrix.

Qing:

I used autoencoder method to extract 10 dimensional embedding features from item metadata, the loss is 0.03 which is acceptable.

The feature is stored in .csv file.