Machine Learning: A Love Story

The video from my keynote at Strange Loop 2010 is up!

You can watch the video here: Machine Learning: A Love Story

The original abstract:

Machine learning has come a long way in recent years — from a long-marginalized field so old it still has the word “machine” in the name, to the last, best hope for making sense of our massive flows of data.

The art of ‘data science’ is asking the right questions; the answers are generally trivial or impossible. This talk will focus more on questions than on answers. I’ll give a brief history of the field with a focus on the fundamental math and algorithmic tools that we use to address these kinds of problems, then walk through several descriptive and predictive scenarios.

Finally, I’ll show one example system using bit.ly data in-depth, from the backend infrastructure through the algorithms and data processing layer to show a functioning product.

Attendees should expect to hear some good stories of data gone right and data gone awry, and walk away with a few new clever tricks.

The presentation was calibrated for the audience in the room, but I’ll be happy to answer any questions in the comments below!


Conference: Web2 Expo SF

I gave a talk called A Data-driven Look at the Realtime Web Ecosystem at the Web2Expo SF conference in May in San Francisco. I attempted to highlight some of the interesting facets of the bit.ly data set, and it appeared to be well-received (showing up on TechCrunch, ZDNet, and a few other places).

I attended the full conference, and it was great. The attendees were extremely international and I met a ton of fascinating people.

I’m still getting a couple of e-mail requests per week for my slides and materials, so they’re posted below for posterity.

The slides:

And the video:

As always, I welcome your questions or comments.


Conference: Search and Social Media 2010

I recently attended the Third Annual Workshop on Search and Social Media, an academic workshop with very strong industry participation. The workshop was packed, and had some of the most informative and interesting panel discussions I’ve seen (not counting the one I spoke on!).

Daniel Tunkelang did a great job of writing up the specific presentations on his site and on the ACM blog, so I won’t attempt to re-create the presentations line by line at this late date. Rather, I’d like to highlight a few open problems and research questions that came out of the discussions that I hope to see developed in the next year.

Social search consists of a set of problems including (but hardly limited to) search of social content like status updates, real-time search, generating, labeling, and finding user-generated content, ‘long-tail’ events and interests, finding vs re-finding, and trend identification.

What data is available to social search? There are many kinds of social data, from e-mail (private) to blogs (public) and tweets (mostly public) — what is and should be searchable? How do we handle issues of privacy and identity management?

How do we compute relevance, taking into account freshness, accuracy, and degrees of social separation?

Will the architecture of these search engines look like the search engines we’re currently familiar with?

How do we evaluate accuracy and truthiness of social data?

How do we characterize social connections, around concepts like strong vs weak ties, and friend-of-a-friend vs friend-of-a-friend’s-friend? Can we converge on a single social graph representation?

How do we best filter social data to lead to accurate recommendations for content discovery? How do we accommodate the fact that as we move beyond static factual data, two people using the same query may be looking for very different results?

Finally, how do we deal with the chasm between the industry participants (who have LOTS of data) and the academic participants, who suffer from a lack of public (and publishable) data?

Thanks again to the organizers – Eugene Agichtein, Marti Hearst, Ian Soboroff, and Daniel Tunkelang – who put together a fantastic event.

For more on this and a cool demo, check out Gene Golovchinsky‘s look at the SSM2010 twitter coverage.


IgniteNYC: The video!

The video of my IgniteNYC presentation is up, and has gotten a great response!

I’m working on removing the me-specific bits from the code and I’ll be posting it as open-source very soon!


My code is on TV (and so am I)!

FoxNY did a piece featuring me and Diana as hackers who use our technical powers for good, not evil.

There are way too few female technologists on television, and I’m happy to do what I can to show that women kick ass with code! Look for my mischievous I’m-writing-infinite-nested-loops grin in the clip where I’m programming.

If this looks like fun to you, come join us at NYC Resistor (where the segment was filmed!) for Thursday night craft nights or for one of many awesome classes.


Hadoop World NYC

Yesterday, I attended the first Hadoop World NYC conference. Hadoop is a platform for scalable distributed computing. In essence, it makes analyzing large quantities of data much faster, and analyzing very large quantities of data possible.

Cloudera did a great job organizing the conference, and managed to assemble a diverse set of speakers. The sessions covered everything from academic research to fraud detection to bioinformatics and even helping people fall in love (eHarmony uses Hadoop)!

I’m not going to review every session, but I saw several themes emerging from the content and conversations.

Hadoop is Getting Easier

New integrated UIs like Cloudera Desktop and Karmasphere mean that developers will no longer be required to use a command-line interface to configure and execute Hadoop jobs. IBM’s M2 project hides Hadoop behind a spreadsheet metaphor, making the collection, analysis and visualization of data as easy as using Excel.

This doesn’t just speed up development time, it puts the tools for manipulating the data directly in the hands of the people who need the results, without requiring them to talk to a database programmer.

Hadoop is a Utility

The only organizations that talked about building their own Hadoop clusters are those who deal with very sensitive data (VISA) and those who deal with very very large quantities of data (Yahoo, Facebook, eBay). Organizations with more manageable data sets, such as eHarmony and the New York Times, use EC2 and Amazon’s Elastic Map-Reduce. Amazon, Rackspace, and Softlayer have offerings in this area and were all event sponsors.

Yes, you can turn on a cluster of nodes from your living room in your PJs!

Hadoop Can Talk to Your Existing Systems

Hadoop has an ecosystem of supporting products that allow organizations to adapt their existing infrastructure. Cloudera’s Sqoop (which is just fun to say out loud) is a tool for importing data from SQL databases, HBase is a Hadoop database, and Pig lets you talk to the system in a SQL-like language.

I expect we’ll see more information available in the near future to clarify which systems are more appropriate for which kinds of users (an ecosystem decision tree?).

Hadoop is Changing Things

I heard the phrase “an order of magnitude improvement in speed” so many times that I lost count. Speaking from personal experience, the difference you see in productivity between waiting minutes and hours for results and waiting days is immense. When you can see the answer to a question shortly after you ask it you can preserve the context you need to act on that answer immediately without having to spend the time to figure out why you were asking that question in the first place.

Most of the projects were doing fairly simple analysis over data like web user sessions or transactions. I was intrigued by Deepak Singh’s talk on bioinformatics and genome sequencing (slides) and Jake Hofman‘s talk on social network analysis (slides). More and more massive datasets are becoming available and will drive techniques for new analysis. I do wish there had been a talk about Mahout, which is a very promising approach to developing machine learning algorithms on the Hadoop platform.

I left the event more excited about the technology and very enthusiastic about the community. Thanks for a great day!

Update: A few other people have written up their notes and impressions from the event:

http://jakehofman.com/talks/hadoopworld_20091002.pdf

http://www.slideshare.net/mndoci/hadoop-for-bioinformatics


Do you do human subject research?

Dear friends and colleagues,

Do you do research that involves gathering data from human participants? This can be anything from marketing surveys to psychology experiments to medical science. If so, please take a short (5 to 10 minute) survey:

research tool survey

The results of the survey will help us design a new platform for online human research!

I’m very excited about this project and would very much appreciate your input. If you have colleagues who do this kind of work, please pass it on.

Thank you!