Introducing Content-Based Filtering Models
- These models rely on non-rating data or very limited rating data
-
The main idea of content filtering is quite straightforward:
- Filter on items that a user positively rated (explicitly or implicitly) in the past
- Recommend other items similar to these examples
- Note, this approach is completely centered around how we define similarity
- Similarity is based on the content of a given item
-
And, similarity doesn't include behavioral data
- E.g. information about items that are frequently purchased or rated together by other users
Defining Content-Based Recommenders
-
When measuring similar items for a given user, a content-based recommender essentially only uses one row of the rating matrix
- Specifically, the profile of the user for whom the recommendations are prepared
-
Roughly, we predict a missing item for a given user by using:
- Known ratings for other, already rated items
- The user's behavioral data
- Optionally, the recommendations can be ranked according to their similarity scores or rating values
-
Typically, we can generate recommendations for each item for a user using:
- A neighborhood-based approach
- A regression (or other classification) model
- This model will predict whether a user will like an item or not
Advantages of Content-Based Filtering
-
User Independence:only uses ratings provided by the user for whom recommendations are prepared- This is a great advantage when the total number of collected ratings is small
- In other words, we can use this property to our advantage when dealing the with cold-start problem for collaborative filtering models
- The second advantage of user independence is the ability to generate interesting recommendations for users with ununsual tastes
-
New and Rate Items:able to recommend new and rare items more easily-
Since collaborative filtering suffers from the cold-start problem, it's harder for them to recommend newer or rate items
- This is because they only rely on ratings, and there's usually fewer ratings than content data
- Content filtering doesn't have this problem since it relies on content similarity
-
Disadvantages of Content-Based Filtering
Trivial Recommendations:can often produce trivial recommendations that aren't novel or interesting (i.e. serendipitous)Feature Engineering:requires feature engineering for content features included in any recommender model-
New Users:not able to provide recommendations for new users without history of content features (but can mitigate cold-start problem)- This is a direct result of the scoring method used by the model
- Since, it will only really favor and recommend extremely similar items
Defining Methods for Content-Based Filtering
-
Generally, there are two forms of content-based models:
- Neighborhood-based filtering
- Model-based filtering
- A k-Nearest Neighbor algorithm is usually used for neighborhood-based methods
- As a reminder, a kNN algorithm relies on similarity metrics
- Whereas, a model focuses on estimating the probability of an item receiving a positive rating
- Our model could include a naive-bayes classifier, XGBoost model, etc.