Item-Based Collaborative Filtering in Ecommerce Data Science

Having learnt about the user-based collaborative filtering, you will now learn about item-based collaborative filtering.

To reiterate the item-based filtering procedure, the following steps are followed in sequence to predict the rating a user u will give to a movie m:

1. Find items similar to the movie m (often called peer group of items) using a similarity measure like cosine. 
2. Calculate the rating that the user will give to the movie m using the weighted average of the ratings given to the nearest movies by the user.
3. Recommend the top-n movies to the user.

To elaborate the example taken in the lecture, the similarity of the movie Casino Royale to Pulp Fiction and The Fast and The Furious was:

• sim(Casino Royale, Pulp Fiction) = 0.99
• sim(Casino Royale, The Fast and The Furious) = 0.98

To predict the rating Nikhil will give to Casino Royale, we take the weighted average of the nearest movies as follows:

pred(Nikhil, Casino Royale) = 

Similarly, the algorithm will compute the ratings Nikhil is likely to give to all the other movies he has not seen and recommend the list of top-n movies, where the choice of 'n' is up to the algorithm designer. 

Additional Reading

The cosine measure does not take into account the average ratings provided by users, which is important in predicting the ratings. Therefore, you can use the adjusted cosine similarity measure, which is exactly similar to the correlation measure we had used.

You can read more about the adjusted cosine similarity and how to predict the rating for a test user using these similarities at this link.