Cluster-based Recommendation with Mahout

Mahout includes a few new experimental recommenders that are weakly documented at the moment. One of them is TreeClusteringRecommender which clusters your model into a set of groups and makes recommendations based on distances between your users and items in these clusters.

A clustered-based recommendation may be a good choice if your data is sparse and rarely correlated to find obvious patterns. Another advantage is clustering may help you to provide recommendations to users even with very tiny data available. Yet, it decreases the level of personalization and output is not unique to users but to clusters.

Here is a quick start to create and run one:

UserSimilarity similarity = new LogLikelihoodSimilarity(model);
ClusterSimilarity clusterSimilarity = new FarthestNeighborClusterSimilarity(similarity);
Recommender rec = TreeClusteringRecommender(model, clusterSimilarity, 20);

rec.getCluster(1); // gets the cluster of userId=1
rec.recommend(1, 10); // recommends 10 items to userId=1

What about user similarity and cluster similarity?

• You basically have to provide a similarity function to be able to measure distances between different users. You may like to represent each user with a vector and calculate euclidean, Pearson, cosine, etc.  distance by looking at these features. Or LogLlikelihoodSimilarity may work as well. You may want to look at Surprise and Coincidence to understand what’s under the hood of this similarity.
• Cluster similarity is newly represented here. It’s a place to customize the measurement of the similarity between two clusters. There are already two implementations: NearestNeighborClusterSimilarity and FarthestNeighborClusterSimilarity. Beware that clusters are dynamic. As new data comes in, old clusters may be merged and new ones may be introduced.

And the rest is about experimental work to find a method fit for your data by analyzing the nature of your input,  plotting the results and evaluating. The initial clustering takes a lot of time compared to item or user similarity based recommenders of Mahout, yet it works to be OK online once you start working with pre-computed data. Even it’s not very mature, for a primitive start, you may still like to consider this recommender if what you want to achieve fits in clustering.