View source for Machine Learning with Large Datasets 10-605

== Information ==


* Course Number: ML 10-605
* Prerequisites: a machine learning course (e.g., 10-701 or 10-601) must be taken either before, or concurrently with, this course.
* Sample syllabus: [[Syllabus for Machine Learning with Large Datasets 10-605 in Spring 2012]]
* ''See [[Machine_Learning_with_Large_Datasets_10-605_in_Spring_2012|the page for the course as it will be taught in spring 2012]] for details on instructor, time and location, etc.''

== Description ==

Large datasets are difficult to work with for several reasons.  They are difficult to visualize, and it is difficult to understand what sort of errors and biases are present in them.  They are computationally expensive to process, and often the cost of learning is hard to predict - for instance, and algorithm that runs quickly in a dataset that fits in memory may be exorbitantly expensive when the dataset is too large for memory.  Large datasets may also display qualitatively different behavior in terms of which learning methods produce the most accurate predictions.

This course is intended to provide a student practical knowledge of, and experience with, the issues involving large datasets.  Among the issues considered are: scalable learning techniques, such as streaming machine learning techniques; parallel infrastructures such as map-reduce; practical techniques for reducing the memory requirements for learning methods, such as feature hashing and Bloom filters; and techniques for analysis of programs in terms of memory, disk usage, and (for parallel methods) communication complexity.  

The class will include programming assignments, and a one-month short project chosen by the student.  The project will be designed to compare the scalability of variant learning algorithms on datasets.  

== Prerequisites ==

An introductory course in machine learning, like 10-601 or 10-701, is a prerequisite or a co-requisite.  If you plan to take this course and 10-601 concurrently please tell the instructor.

The course will include several substantial programming assignments, so an additional prerequisite is 15-210, or 15-214, or comparable familiarity with Java and good programming skills.

Undergraduates need permission of the instructor to enroll.

== Planned Topics ==

* Overview of course, and overview lecture on probabilities.
* Streaming learning algorithms
** Naive Bayes for discrete data.  
** A streaming-data implementation of Naive Bayes.
** A streaming-data implementation of Naive Bayes assuming a larger-than-memory feature set, by using the 'stream and sort' pattern.
** Discussion of other streaming learning methods.
*** Rocchio
*** Perceptron-style algorithms
*** Streaming logistic regression
* More complex programs using stream-and-sort.
** Lecture topics:
*** Finding informative phrases in a corpus, and finding polar phrases in a corpus.
*** Using records and messages to manage a complex dataflow.
* The map-reduce paradigm and Hadoop.
* Reducing memory usage with randomized methods.
** Feature hashing tricks.
** Bloom filters for counting events.
** Locality-sensitive hashing.
* Nearest-neighbor finding and bulk classification.
** Using a search engine to find approximate nearest neighbors.
** Using inverted indices to find approximate nearest neighbors or to perform bulk linear classification.
** Implementing soft joins using map-reduce and nearest-neighbor methods.
** The local k-NN graph for a dataset.
* Week 8-10. Working with large graphs.
** PageRank and RWR/PPR.
** Special issues involved with iterative processing on graphs in Map-Reduce: the schimmy pattern.
*** Formalisms/environments for iterative processing on graphs: GraphLab, Sparks, Pregel.
** Extracting small graphs from a large one:
*** LocalSpectral - finding the meaningful neighborhood of a query node in a large graph.
*** Visualizing graphs.
** Semi-supervised classification on graphs.
** Clustering and community-finding in graphs.
* Stochastic gradient descent and other streaming learning algorithms.
** SGD for logistic regression.
** Large feature sets SGD: delayed regularization-based updates; projection onto L1; truncated gradients.
* Additional topics that may be covered.
** Scalable k-means clustering.
** Gibbs sampling and streaming LDA.
** Stacking and cascaded learning approaches.
** Decision tree learning for large datasets.

== Datasets ==

Some datasets will be provided by the instructors to use in the course.
* RCV2 - text classification dataset.
* Wikipedia links - page-page links for Wikipedia.
* Wikipedia pages - text of pages for Wikipedia.
* Geographical names and places - data on places from GeoNames, Wikipedia, and Geo-tagged Flikr images.
* NELL all-pairs data - NPs and the contexts they appear in on the web.
* Google Books n-gram data
* ?Million Song Database - audio signatures of songs with tags and meta-data.
* ?KDD search-engine queries.