Chapter 4. Making recommendations
This chapter covers
- User-based recommenders, in depth
- Similarity metrics
- Item-based and other recommenders
Having spent the last chapter discussing how to evaluate recommenders and represent the data input to a recommender, it’s now time to examine the recommenders themselves in detail. That’s where the real action begins.
Previous chapters alluded to two well-known styles of recommender algorithms, both of which are implemented in Mahout: user-based recommenders and item-based recommenders. In fact, you already encountered a user-based recommender in chapter 2. This chapter explores the theory behind these algorithms, as well as the Mahout implementations of both, in detail.
Both algorithms rely on a similarity metric, or notion of sameness between two things, whether they’re users or items. There are many ways to define similarity, and this chapter will introduce in detail your choices within Mahout. These include implementations based on the Pearson correlation, log likelihood, Spearman correlation, Tanimoto coefficient, and more.
Finally, you’ll also become familiar with other styles of recommender algorithms implemented within Mahout, including slope-one recommenders, SVD-based recommenders, and clustering-based recommenders.