content-based filtering

content-based filtering is an approach to recommender system that recommends items to you based on features of you as a user and item.

or find similar items to item by minimizing squared L2 norm of their distance

notation	description
	a vector of features of user
	a vector of features of item
or	a vector of values calculated from (i.e. values of user features)
or	a vector of values calculated from

Process

Picking features

To speed up the response time of your recommendation system, you can pre-compute the vectors for all the items you might recommend (Retrieval). This can be done even before a user logs in to your website and even before you know the or

Retrieval

1. Generate a large list of plausible item candidates, e.g.:
   • for each of the most 10 recent movies, find 10 similar movies
   • for most viewed 3 genres, find the top 10 movies
   • top 20 movies in the country
2. Combine retrieved items in list, removing duplicates or past items

Number of items to be retrieved?

Retrieving more item results in better performance, but slower recommendations

Ranking

1. Take list retrieved and rank using learned model
2. Display ranked items to users

Architecture

(Neural Network Architecture 0:30 - 3:50)

Neural networks for users’ features and items’ features can be combined into a SINGLE artificial neural network, as long as the structure for the last layer is the same

Cost function

When there’s user rating for item (), the cost function is calculated as

The cost can include artificial neural network regularization term to reduce the value of parameters

Implementation