CubeCobraRecommender

Recommender System for CubeCobra.

The web folder contains a Flask API that will return the Machine LearningCubeCobraRecommender recommendation in a json response. This is for integrating the machine learning algorithm on the website, which is currently in development as demonstrated in this tweet. The version live on CubeCobra will exist in the Production branch.

Machine Learning Algorithm

This recommender system is a Denoising AutoEncoder. Every data point (x) is a cube, represented as a binary vector such that if x[i] is 1, this means that card i is in the cube.

Noise is added by flipping some percentage of those 1s to zero. Additional noise is added by flipping zeros to 1 via negative sampling, where the sampling probability is proportional to the amount of times a card is seen in the data.

There is an additional level of regularization via a Conditional Probability Graph. This can be done with an external datasource, or the same collection datasource. But the idea is as follows:

1. Take a dataset of items, and generate a matrix M such that M[i,j] is the probability that item j is in a collection given item i is in the collection.
2. Convert this matrix into rows of probabilities by dividing each row by its sum (the sum of every row should now equal 1)

Then, use this matrix to regularize the AutoEncoder. The model is built of a couple parts:

1. Normal Encoder (E)
2. Normal Decoder (D1)
3. Second Decoder (D2)
4. Function for applying noise to the input (F)

Let I be a the identity matrix of the same shape as M. Then, we optimize the following loss function:

Loss = BinaryCrossEntropy( X, D1(E(F( X ))) ) + 0.1 * KL-Divergence( M, D2(E(I)) )

Note that 0.1 is the hyperparameter regularization coefficient.

Generating The Adjacency Matrix

running python src/scripts/create_mtx.py will create a local version of the adjacency matrix as well as a lookup dictionary. This will be stored in the outputs folder. It is in .gitignore, so make sure to create a local version.

Generating Recommendations

After generating the adjacency matrix, given any cube list, you can get the top N recommendations. To do this, run python src/scripts/recommend.py cube_id N. For example, if I wanted the top 50 recommendations for my Combat Cube, I would run python src/scripts/recommend.py combat 50.

If you would like a recommendation on cards to cut, rather than cards to add, run python src/scripts/cut_cards.py cube_id N.

Lastly, if you would like recommendations from the machine learning algorithm rather than the adjacency matrix, run python src/scripts/ml_recommend.py cube_id N

Git - LFS

In order to upload the data used in this project, it was zipped and tracked via git-lfs. You may need to install this in order to download the repo.

To Do

1. Add code for training the ML model (currerntly ml_files contains a wrapped pretrained version).
2. Clean up repository to have a better structure. Comment everything. Separate scripts from source.