We have covered a lot of ground for both content-based and collaborative filtering approaches. We even did some hands-on examples. Now let's do a fair comparison of both these approaches. First we discuss content-based approach, followed by the collaborative-filtering approach to recommender systems.
Content-based recommends items that are similar to those that the user liked in the past. In content-based recommendation, we really don't need user ratings. We can start with global recommendations, and as the user interacts more with the system, we can make recommendations even without the user rating any items. A content-based recommender will not generalize to different kinds of items as we discussed earlier. Content-based recommender systems therefore are implemented as case-based recommender systems, which take into account items specific to a particular category. Also note that in content-based recommendation, the items don't change their features frequently. So once we have prepared a model, we can re-use until more items are added, or removed.
To summarize:
- We do not need ratings, or data on other users.
- We are able to tailor recommendations according to a user's taste using the content-based approach.
- We can recommend an item regardless of an item being new or unpopular. The only criteria is that the feature vector for that item comes in a close neighborhood of a user's pretend item.
- We can explain the reason why an item was recommended to a user because of his/her history.
Collaborative filtering recommends items that similar users like collaboratively. Filtering assumes that we need item ratings before we can proceed. Since the algorithm depends on ratings (either implicit or explicit), there are no notion item features. Therefore, collaborative filtering algorithms are applicable to datasets containing diverse categories of items.
Because collaborative filtering algorithms require ratings, we face a road-block. Since the interaction of users and items is very dynamic, that is users keep visiting and rating items all the time, we need to learn or update the models in real time. This poses a significant challenge in engineering a good recommender system.
To summarize:
- The key assumptions in collaborative filtering algorithms are:
- Past activity of a user will predict tastes of that same user in the future
- The tastes of two similar users now, will likely be the same in future too
- The new or unpopular items will likely be recommended later, when their ratings improve.
- Explaining recommendations is also very challenging because now the recommendations are based on a collaborative effort of many users, and not just one user.