Reccomendation-systems-Implementation

Implementing Content based filtering (similarity measures) , collaborative filtering(Matrix Factorisation) and Neural network based methods.

Method 1 : Similarity Measures

1. Cosine similairty

• Time taken : 3-5 seconds on imdb data of 14332 movies
• Average rmse : 1.256

2. Modified Norm-Cosine similairty

• Time taken : 3 minutes on imdb data of 14332 movies
• Average rmse : 0.875

Method 2 : Matrix-Factroization (MF)

• Dataset : 100k ratings by 1000 users on 10000 movies (Sparse dataset)
• MF on toy data: RMSE = 0.0273
• MF on a subset of 1k ratings(data points) : Time taken = 1 min & RMSE = 0.667
• MF on a subset of 10k ratings(data points) : Time taken = 10-12 min & RMSE = 0.086
• Note : If we use pandas pivot function data conversion is done within secs. But if we use naive methods then timetaken for data conversion {entire dataset} : 14-15 mins.

Error Plot (RMSE vs No.of Ratings)

Method 3 : Neural networks based recommender system

• Dataset : 100k ratings by 1000 users on 10000 movies (Sparse dataset)
• NN on a subset of 1k ratings(data points) : Time taken = 1 min & RMSE = 0.453
• MF on a subset of 10k ratings(data points) : Time taken = 8-10 min & RMSE = 0.027
• Note : Keras functions have been used to increase the speed of the program.

Error Plot (RMSE vs No.of datapoints)