recommender-system

Implementing different approaches for recommendation systems

Usage

Collaborative Filtering

run: python cf.py Returns the RMSE and MAE loss metrics on test data using three different approaches of collaborative filtering namely:

1. user-user filtering
2. item-item filtering
3. baseline approach

SVD

run: python svd.py

Performs Singular Value Decomposition on the given utility matrix and report the reconstruction error (RMSE and MAE loss) at the specified energy.

CUR

run: python cur.py

Similar to SVD, perform decomposition and reports reconstruction error for the specified r value. r is the parameter which specifies the number of columns and rows in C and R matrix respectively in CUR.

Latent Factor model

run: python main.py

Predicts user-movie rating using the latent factor model. Implemented using Stochastic Gradient Descent learns the latent (hidden) factors for each user and movie and along with baseline approximation computes the prediction.

Hyperparameters:

alpha (learning rate) = 0.01

beta (regularisation coefficient) = 0.05

epochs = 50

Results

Tuning Latent Factors

Latent Factors	RMSE (test)	MAE (test)
10	0.836	0.654
20	0.840	0.657
50	0.829	0.649
100	0.833	0.653

Collaborative Filtering

CF Approach	RMSE (test)	MAE (test)
Baseline	0.904	0.724
user-user filtering	1.147	0.831
item-item filtering	0.921	0.730

SVD

Energy	RMSE (test)	MAE (test)
100	0	0
90	0.243	0.132

CUR

r	RMSE (test)	MAE (test)
3000	0.615	0.205
2000	2.109	0.276

Loss curve for Latent Factor model

Using 50 latent factors: