Recommendation system reliant on a user's other reviews in predicting their rating for a product.
- Brian Tsai - tsai209@purdue.edu
- Ryan Shue - rshue@purdue.edu
- Laura Zhou - lczhou@purdue.edu
Video_Games_5.json can be downloaded from https://cseweb.ucsd.edu/~jmcauley/datasets/amazon_v2/ , or via this direct url: https://jmcauley.ucsd.edu/data/amazon_v2/categoryFilesSmall/Video_Games_5.json.gz
Once downloaded, this file should be placed in the Preprocessing folder. Do not unzip the file.
FastText embeddings should be downloaded from this url: https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.en.300.bin.gz
Once downloaded, this file should be placed next to the ML_LSTM.ipynb notebook.
The data is already preprocessed and located at Preprocessing/dataset_operating.csv (if not, you can download from the Unavailable Data link the appendix). To remake this file, change your directory into Preprocessing and execute preprocess.py. Remember that the Video_Games_5.json.gz needs to be in that directory, still zipped! The preprocessing removes duplicate or empty reviews, and requires every user to have at least six reviews.
The environment assumes you installed PyTorch 2.2.0 or above compiled with the necessary CUDA installation. When using CUDA, ensure you have at least 8GB of memory in your GPU. If you want to use the CPU instead, replace existences of torch_device = torch.device("cuda") with torch_device = torch.device("cpu").
In addition, install the following packages:
jupyter-notebookorjpyterlabto run the code.pandasfor dataframe manipulationprettytablefor the part where the number of parameters is displayedscikit-learnfor thetrain_test_splitand product embedding visualizationtqdmfor progress bars.torchmetricsfor R2 metricstransformersOptional, for the transformer code (see appendix)
The relevant code to the project is located at ML_LSTM.ipynb.
If data is unavailable, you can tryi this shared onedrive link: https://purdue0-my.sharepoint.com/:f:/g/personal/tsai209_purdue_edu/Ens8h2sboOpCqxic5Mf0lvwB7ExrT5tq90c_pfgzscqi0Q?e=y3bgER
The Review Module component, visualized with a LSTM modeling of the text.
The transformer-based network is located at ML_Transformers.ipynb. This was an earlier iteration where a Bert Transformer instead of an LSTM was used instead to process the text data in reviews. Unfortunately, due to the complete model's 113 million parameters, limited compute and memory requirements prevented its full use. The program was tested and confirms to work a on Google Colab NVIDIA L4 Tensor core GPU, and its 24GB of memory was unable to handle batch sizes greater than four. Each epoch took around 90 minutes to complete.