Machine Learning: Clustering & Retrieval

Clustering and retrieval are some of the most high-impact machine learning tools out there. Retrieval is used in almost every applications and device we interact with, like in providing a set of products related to one a shopper is currently considering, or a list of people you might want to connect with on a social media platform. Clustering can be used to aid retrieval, but is a more broadly useful tool for automatically discovering structure in data, like uncovering groups of similar patients.<p>This introduction to the course provides you with an overview of the topics we will cover and the background knowledge and resources we assume you have.

We start the course by considering a retrieval task of fetching a document similar to one someone is currently reading. We cast this problem as one of nearest neighbor search, which is a concept we have seen in the Foundations and Regression courses. However, here, you will take a deep dive into two critical components of the algorithms: the data representation and metric for measuring similarity between pairs of datapoints. You will examine the computational burden of the naive nearest neighbor search algorithm, and instead implement scalable alternatives using KD-trees for handling large datasets and locality sensitive hashing (LSH) for providing approximate nearest neighbors, even in high-dimensional spaces. You will explore all of these ideas on a Wikipedia dataset, comparing and contrasting the impact of the various choices you can make on the nearest neighbor results produced.

In clustering, our goal is to group the datapoints in our dataset into disjoint sets. Motivated by our document analysis case study, you will use clustering to discover thematic groups of articles by "topic". These topics are not provided in this unsupervised learning task; rather, the idea is to output such cluster labels that can be post-facto associated with known topics like "Science", "World News", etc. Even without such post-facto labels, you will examine how the clustering output can provide insights into the relationships between datapoints in the dataset. The first clustering algorithm you will implement is k-means, which is the most widely used clustering algorithm out there. To scale up k-means, you will learn about the general MapReduce framework for parallelizing and distributing computations, and then how the iterates of k-means can utilize this framework. You will show that k-means can provide an interpretable grouping of Wikipedia articles when appropriately tuned.

In k-means, observations are each hard-assigned to a single cluster, and these assignments are based just on the cluster centers, rather than also incorporating shape information. In our second module on clustering, you will perform probabilistic model-based clustering that provides (1) a more descriptive notion of a "cluster" and (2) accounts for uncertainty in assignments of datapoints to clusters via "soft assignments". You will explore and implement a broadly useful algorithm called expectation maximization (EM) for inferring these soft assignments, as well as the model parameters. To gain intuition, you will first consider a visually appealing image clustering task. You will then cluster Wikipedia articles, handling the high-dimensionality of the tf-idf document representation considered.