likely

Likely - is a python application that implements various recommendation methods like KNN, Bandit, and SVD in plug and play style.

We are surrounded by content and our experience depends on various factors and algorithms incorporated together. Maybe you stumbled on this repo by the GitHub recommendation system. This application is an attempt to understand various recommendation methods.

How to run DEMO

This application uses python 3.x and Flask.

1. Clone the repo using git clone https://github.com/amifunny/likely.git in the command line.
2. In the command line, go inside the folder and use set FLASK_APP=backend for windows or to avoid setting this, change backend.py to app.py.
3. Start the local server using flask run
4. Go to http://127.0.0.1:5000/ to interact with the demo.

Content

• likely.py : Contains class implementation for recommendation objects.
• backend.py : Handle routes and call recommender object for movie indices and rendering templates.
• static/dataset : Contains Movie Dataset and User Rating dataset.
• keywords.py : Auxilary File to create static/keywords.txt to get genre for MultiArmedBandit

Feedback System

We used LIKE and DISLIKE buttons to mark seen movies and assign ratings. As the dataset is for ratings between 0 to 5, dislike is assumed as 0 and like a 5-star rating.

Dataset

For User Ratings and corresponding Movie Titles, Movie lens Small dataset was used. But it had lots of genres missing and also posters of movies were needed to create the UI, so additional dataset from Kaggle was used from here. So finally we had a dataset of about ~6000 Movies and ~600 users rating.

Algorithms

• Multi-Armed Bandit - Classic Reinforcement Learning method to get most reward from probabilistic setting. We used Thompson sampling for exploration purpose. We use number of genres as arms. This method estimates average reward, or how likely users are to click on the item with certain keyword.

• SVD User-Item - SVD is a matrix decomposition method which in our case, decomposes User-Item matrix filled with rating values, into Three matrices. One contains Users' latent vectors, another Items' latent vectors, and a weights diagonal matrix. Latent Vectors are nothing but float values determining characteristics of user and Item.

  Doing a simple dot product of a particular user with all movies gives us predicted ratings of that user for all movies. Also, each user has their way of rating, so the User-Item matrix is normalized by the average rating of that user.

• SVD+KNN Item-Item - As Latent Vector of similar items are also similar, we use KNN to get nearest similar movies to the ones that the user has already watched.

• KNN User-User - Use CSR matrix of User-Item to get similar users on basis of their ratings. Use movies rated good by other users to give recommendations.