It might appear that item-based collaborative and content-based recommendations are the same. In reality, they are not the same. Let's touch upon the differences.
ITCF is totally based on user-item rankings. When we compute the similarity between items, we do not include the item attributes and just compute the similarity of items based on all customers' ratings. So the similarity between items is computed based on the ratings instead of the metadata of item itself.
In content-based recommendations, we make use of the content of both the user and the item. Generally, we construct a user profile and item profile using the content of a shared attribute space. For example, for a movie, we represent it with the actors in it and the genre (using binary coding, for example). For a user profile, we can do the same thing based on the user, such as some actors/genres. Then the similarity of user and item can be computed using cosine similarity, for example. This cosine measure leads to the recommendations.
Content-based filtering identifies products that are similar based on the tags assigned to each product. Each product is assigned weights on the basis of term frequency and inverse document frequency of each tag. After this, the user's probability of liking a product is calculated in order to arrive at the final recommendation list.
While content-based recommendation systems are highly efficient and personalized, there is an inherent problem with this model. Let's understand the over-specialization problem of content-based recommendations with an example.
Assume there are the following five movie genres:
- Comedy
- Thriller
- Science fiction
- Action
- Romance
There is this customer, Jake, who generally watches thriller and science fiction movies. Based on this preference, the content-based recommendation engine will only recommend movies related to these genres and it is never going to recommend movies from other categories. This problem arises as content-based recommendation engine solely relies on the user's past behavior and preferences to determine the recommendation.
Unlike content-recommendation systems, in ITCF recommendations, similar products build neighborhoods based on positive preferences of customers. Therefore, the system generates recommendations with products in the neighborhood that a customer might prefer. ITCF does this by making use of the correlation between the items based on the ratings given them by different users, while collaborative filtering relies on past preferences or rating correlation between users and it is able to generate recommendations for similar products even from customer's interest domain. This technique can lead to bad predictions if the product is unpopular and very few users have given feedback about it.