Mithrillion/steam_app_frontend

Use deep learning and nearest neighbour search to find games similar to a user input, based on the similarity between their store page images.

This repository is the web frontend portion of my steam similar game finder project.

The live demo of the project is at https://mithrillion.github.io/steam_app_frontend/.

The web backend portion can be found at https://github.com/Mithrillion/steam_app_backend

The data processing code can be found at https://github.com/Mithrillion/steam_app_data_processing.

The project consists of the following steps:

1. Scrape Steam app titles and IDs from a saved version of steamdb.com (not using live scrapping according to terms of steamdb)
2. Scrape the store page image URLs and user-defined tags from store.steampowered.com for all app IDs obtained in step 1.
3. Download images from Steam's CDN.
4. Use PyTorch to train a deep neural network to predict the tags of a Steam game. The neural network is an adapted version of AlexNet, with the final prediction layer swapped with a two-layer RNN and a dense prediction layer in order to deal with multiple image inputs per game.
5. Use the network up to second last layer to generate a semantic encoding of top 10000 games on Steam, based on their store images.
6. Use an optimised nearest-neighbour algorithm to find closest games to a given input game.
7. Wrap the algorithm in step 6 in a web interface.

NEXT STEPs:

1. Use deeplearn.js to deploy image hashing model onto the web frontend.
2. Build a game explorer based on hierarchical clustering of games.