Recommendation system research

Recommendation system is aim to recommend some items to user. It can be devided into:

1. User-based collaborative filtering
2. Item-based collaborative filtering

We use kNN to find the similar users/items.

Based on data, we will calculate the similarity among users/items for kNN by:

• Jaccard similarity
• Euclidean distance
• Manhattan distance
• Minkowski distance
• Cosine similarity

All of those measure methods were implemented in python (measure.py).

Example

We have a user-drink ratings matrix M below, where 6 users (rows) have rated 6 drink type (columns). Ratings can take up integer values from 1–10 and 0 indicates absence of rating. Now we have to decide that we should recommend coffee to user 2 (currently he is active user) or not. Assuming that user 2 will only like coffee if his rate is greater or equal to 5.

This problem can be resolved by 2 ways:

1. User-based collaborative filtering:

• Using kNN to find k most similar user with action user.

• In order to predict the coffee rate of active user, we use the function:

  Where:

  • r(x,s) is the prediction for active user x for item s
  • r(x) is the average rating of user x
  • r(y,s) is rate of user y with item s
  • sim(x,y) is the similarity between users a and u
  • S(xy) is the neighborhood of most similar users.

2. Item-based collaborative filtering:

• Using kNN to find k most similar items with coffee.
• Calculate average of active user's rating with other drink type. This will be rate of active user with coffee.

This example will be solved in coffee-recommend.py

Reference:

1. Recommender Systems — User-Based and Item-Based Collaborative Filtering
2. Collaborative Filtering based Recommendation Systems exemplified..
3. Item-Based Collaborative Filtering Recommendation Algorithms
4. User-Based Collaborative-Filtering Recommendation Algorithms on Hadoop