A collaborative filtering system filters suggestions based on a particular user’s likeness to others. Collaborative recommender systems rely on explicit and implicit data and assume that users with comparable preferences will likely be interested in the same items and potentially interact with them in similar ways in the future.

Amazon (link resides outside ibm.com), for instance, uses collaborative filtering for its product recommendations, just like Spotify (link resides outside ibm.com) does for its audio content.

Collaborative filtering recommenders can offer effective suggestions and won’t usually need detailed item descriptions. However, collaborative filtering is also prone to the cold start problem, which happens when the system has limited historical data to draw from, especially for new users.

There are 2 main kinds of collaborative filtering systems: memory-based and model-based.

Memory-based

Memory-based systems represent users and items as a matrix. They are an extension of the k-nearest neighbors (KNN) algorithm because they aim to find their “nearest neighbors,” which can be similar users or similar items. Memory-based systems are further subdivided into 2 types:

• User-based filtering computes similarities between a particular user and all other users in the matrix. User-based similarity functions are calculated between rows in the user-item matrix.

• Item-based filtering computes item similarity through user behavior (how users interact with items and not item features). Item-based similarity functions are calculated between columns in the user-item matrix.

Model-based

Alternatively, model-based systems create a predictive machine learning model of the data. The user-item matrix serves as the training data set for the model, which then yields predictions for missing values, that is, items that a user has not yet found and will therefore be recommended.

One of the most commonly used model-based collaborative filtering algorithms is matrix factorization. This dimensionality reduction method decomposes the often large user-item matrix into two smaller matrices—one for users and another for items—having a select few dimensions. The 2 matrices are then multiplied together to predict the missing values (or the recommendations) in the larger matrix.

A more advanced implementation of matrix factorization harnesses deep learning neural networks. Other model-based systems employ machine learning algorithms such as Bayes classifiers, clustering and decision trees.

A content-based filtering system filters recommendations based on an item’s features. Content-based recommender systems assume that if a user likes a particular item, they will also like another similar item. Content-based filtering considers item descriptions such as color, category, price and other metadata assigned by keywords and tags, along with explicit and implicit data.

Content-based filtering systems represent items and users as vectors in a vector space. Proximity is used to determine the similarity between items. The closer 2 vectors are in space, the more similar they’re considered to be. Vectors similar to previous items according to their supplied features will be recommended to the user.

Content-based recommenders apply a user-based classifier or regression model. Descriptions and features of items a user is interested in act as the model’s training data set, which then yields predictions for recommended items.

Content-based recommendation systems can be further improved by using natural language processing tags. However, this tagging process can be tedious for huge volumes of data.

Unlike collaborative filtering, the cold start problem is less of an issue since content-based filtering is based on metadata characteristics rather than past user interaction. However, content-based filtering can be limited in exploring new items, as it often suggests those similar to what users liked previously.

As its name implies, a hybrid recommendation system merges collaborative filtering and content-based filtering.

As a result, a hybrid approach can greatly enhance the performance of a recommendation engine, but it requires advanced architectures and intensive computational power.

Netflix, for instance, uses a hybrid recommendation system (link resides outside ibm.com) for its movie and TV show recommendations.