August 5, 2010 1:06 am by Markus.

Kaggle has a new and interesting competition on building a chess rating algorithm that performs better than the official Elo rating system currently in use. Entrants build their own rating systems based on the results of more than 65,000 historical chess games and then test their algorithms by predicting the results on a holdout set of 7,800 games.

Looks like an interesting problem. The only other thing that comes to my mind literature-wise is that Microsoft built and published their TrueSkill™ Ranking System for the XBox in order to match players with similar skills in online games. In the original paper at NIPS, the authors had shown that TrueSkill outperformed Elo.

Posted in Statistics, Machine Learning | Print | No Comments »

July 11, 2010 8:56 pm by Markus.

Interesting article:  GraphLab: A New Framework for Parallel Machine Learning

From the abstract:

Designing and implementing efficient, provably correct parallel machine learning (ML) algorithms is challenging. Existing high-level parallel abstractions like MapReduce are insufficiently expressive while low-level tools like MPI and Pthreads leave ML experts repeatedly solving the same design challenges. By targeting common patterns in ML, we developed GraphLab, which improves upon abstractions like MapReduce by compactly expressing asynchronous iterative algorithms with sparse computational dependencies while ensuring data consistency and achieving a high degree of parallel performance. We demonstrate the expressiveness of the GraphLab framework by designing and implementing parallel versions of belief propagation, Gibbs sampling, Co-EM, Lasso and Compressed Sensing. We show that using GraphLab we can achieve excellent parallel performance on large scale real-world problems.

Given all the talk about Map-Reduce, Hadoop etc. this seems like a logical next step to make scaling data mining to large data sets a lot easier.

Posted in Data Mining, Machine Learning | Print | 1 Comment »

February 16, 2010 11:56 pm by Markus.

How can one evaluate the performance of prognostic models in a meaningful way? This is a very basic and yet an interesting problem especially in the context of prediction of very rare events (base-rates <10%). How reliable is the model’s forecast? This is a good question and of particular importance when it matters - think criminal psychology where models forecast the likelihood of recidivism for criminally insane people (Quinsey 1980). There are a variety of ways to evaluate a model’s predictive performance on a hold out sample, and some are more meaningful than others. For example, when using error-rates one should keep in mind that they are only meaningful when you consider the base-rate of your classes and the trivial classifier as well. Often this gets confusing when you are dealing with very imbalanced data sets or rare events. In this blog post, I’ll summarize a few techniques and alternative evaluation methods for predictive models that are particularly useful when dealing with rare events or low base-rates in general.

The Receiver Operator Characteristic is a graphical measure that plots the true versus false positive rates such that the user can decide where to cut for making the final classification decision. In order to summarize the performance of the graph in a single, reportable number, the area under the curve (AUC) is generally used.

Read the rest of this entry »

Posted in Statistics, Classification, Data Mining, Machine Learning | Print | 2 Comments »

January 26, 2010 9:54 pm by Markus.

The New Scientist describes a new method for spam detection by learning patterns. This new method exploits the spammers most powerful weapon - the automatic generation of many, similar messages by automated means (i.e., some grammar in a formal language) - and turns it against them. The article reports that a pattern can reliably be learned from about 1000 examples captured from a bot, allowing the method to classify new messages accurately and with zero false positives. This sounds really exciting given my full spam-folder.

However, I’m a bit cautious. The article is a bit sparse on technical details, so I might make some wrong assumptions here. First, zero false positives reported is the discrimination of spam from that particular spam-grammar versus other messages. At least that’s how I understand it. Second, it seems from the article that they only learn from positive examples. Overall the technique sounds to me like they are learning a pattern language. Pattern languages are a class of grammars that overlap with linear and context-sensitive grammars (Chomsky hierarchy). Unfortunately they don’t have a real Wikipedia page so I’ll try to give a bit of background. The closest I can give for an example right now would be regular expressions with back-references. I’m not sure if this is an accurate description for all possible patterns, but it’s close enough for an example.

I don’t know how the specific technique mentioned in the article works in detail, but I’ve learned two things about learning grammars from text: (a) we can’t learn all linear or context-sensitive languages, only all pattern language grammars; (b) learning patterns without negative examples leads to over-generalization really really fast.

While I haven’t worked with learning grammars in a long while, the only algorithm of which I’m aware is the Lange-Wiehagen algorithm (Steffen Lange and Rolf Wiehagen; Polynomial-time inference of arbitrary pattern languages. New Generation Computing, 8(4):361-370, 1991). This algorithm is not a consistent learner, but can learn all pattern languages in polynomial time. There might be better ones available by now, but learning grammars is not that popular in the machine learning community right now. I’m sure there are some other interesting applications besides spam filtering. Maybe it’s time for a revival.

Overall, it sounds like a promising new anti-spam technique, but I’d like to see some more realistic testing done. There are some obvious ways for spammers to make learning these patterns harder, but either way I’m curious - maybe the inventors of this technique discovered a better way to learn patterns? Maybe by using some problem-specific domain knowledge?

Posted in spam, Machine Learning | Print | No Comments »

March 1, 2009 5:18 pm by Markus.

Interesting post over at hunch.net about reviewers bidding for papers in order to shoot them down. Make sure to read the comments… That state of mind of some reviewers might explain why the least-informative and most negative reviews always come with the highest confidence rating in ML conferences (specifically NIPS).

Posted in Machine Learning | Print | 2 Comments »

January 11, 2009 4:31 pm by Markus.

I just read another article discussing weather Risk Management tools had an impact on the current financial crisis. One of the most commonly used risk management measures is the Value-at-Risk (VaR) measure, a comparable measure that specifies a worst-case loss for some confidence interval. One of the major criticisms is (e.g. Nassim Nicholas Taleb, the author of the black swan) that the measure can be gamed. Risk can be hidden “in the rare event part” of the prediction and not surprisingly this seems to have happened.

Given that a common question during training with risk assessment software is “what do I do to get outcome/prediction x” from the software it should be explored how to safeguard in the software against users gaming the system. Think detecting multiple model evaluations with slightly changed numbers in a row…

Edit: I just found an instrument implemented as an Excel Spreadsheet. Good for prototyping something, but using that in practice is just asking people to fiddle with the numbers until the desired result is obtained. You couldn’t make it more user-friendly if you tried…

Posted in Predictive Modeling, Society, Statistics, Data Mining, Machine Learning | Print | No Comments »

December 22, 2008 10:27 am by Markus.

I just read that Credit Card Companies are adjusting Credit Scores based on shopping patterns in addition to  credit-score and payment history. It seems they also consider which mortgage lender a customer uses and whether the customer owns a home in an area where housing prices are declining. All that to limit the growing credit card default rates.

That’s an interesting way to do it (from a risk modeling point of view) and I wonder how well it works in practice. There might also be some legal ramifications to this if it can be demonstrated that this practice (possibly unknowingly to them) discriminates e.g. against minorities.

Posted in Statistics, Data Mining, Machine Learning | Print | No Comments »

November 1, 2008 9:48 pm by Markus.

I had written a post about the issues of converting models into something that is usable in production environments as most stats-packages don’t have friendly interfaces to integrate them into the flow of processing data. I worked on a similar problem involving a script written in SAS recently. To be specific, some code for computing a risk-score in SAS had to be converted into Java and I was confronted with having to figure out the semantics of SAS code. I found a software to convert SAS code into Java and I have to say I was quite impressed with how well it worked. Converting one language (especially one for which there is no published grammar or other specification) into another is quite a task - after a bit of back and forth with support we got our code converted and the Java code worked on the first try. I wish there would be similar converter for STATA, R and SPSS

Posted in Predictive Modeling, Coding / Programming, Statistics, Data Mining, Machine Learning | Print | No Comments »

September 1, 2008 8:19 pm by Markus.

Recently I had a fun discussion with Bill over lunch about intellectual property and how that might apply to statistical modeling work. Given that there are more and more companies making a living from forming predictions with a model they have built (churn-prediction, credit-scores and other risk-models) we were wondering if there were any means of protecting them as intellectual property. For example, the ZETA-model for predicting corporate bankruptcies is a closely guarded secret with having published only the variables being used (Altman E. I. (2000); Predicting financial distress for companies: revisiting the Z-Score and ZETA models). Obviously this model is useful for lending and can make serious money for the user. Making decisions guided by a formula is becoming more popular. This might be something over which legal battles will be fought in the future.

Copyrighted works and patents often count towards what a company would be worth should somebody acquire it. This means there would be motivation for start-up companies to protect their models. A mathematical formula (e.g. a regression equation) cannot be patented, and copyright probably won’t apply either; even if copyright would apply, it’s trivial to build a formula that does essentially the same thing (e.g. multiply all the weights in the formula by 10). This leaves only trade secret protection and means there is no recourse once the cat is out of the bag. Often it’s also the data-collection method that is kept secret - a company called Epagogix developed a method to judge the success of movies from a script by scoring it against some scales that they keep secret.

Currently, I don’t see any legal protections with the exception of trade-secrets for this. And given that there is infinitely many ways to express the same scoring rules in a different way, this would be a fairly hard problem for lawyers and politicians to formulate sensible rules for establishing protection for this kind of intellectual property.

Posted in Society, Politics, Data Mining, Machine Learning | Print | 1 Comment »

August 21, 2008 9:17 pm by Markus.

In a survey-paper various methods for finding the number of clusters were compared (Dimitriadou et.al, An Examination of Indexes for Determining the Number of Clusters in Binary Data Sets, Psychometrika, 2002) - and there are plenty of methods. None of them work all the time. Finding the right number of clusters has been an open problem for quite a while and also depends on the application, e.g. if more fine or course grained clusters are of interest.

A similar problem occurs in psychopathology. Imagine some measurements taken from several people - some with and some without a mental illness. The question then becomes: are there two clusters or just one? Is the data simply continuous or generated by a latent Bernoulli distribution? There is a whole bunch of literature out there dealing with the same problem from the psychology standpoint (for example: Schmidt, Kotov, Joiner, “Taxometrics - Towards a New Diagnostic Scheme for Psychopathology“, American Psychological Association) One of the more famous researchers is Paul Meehl, who developed a couple of methods to detect a taxon in data. The MAXCOV-HITMAX invented by Paul Meehl is for the detection of latent taxonic structures (i.e., structures in which the latent variable is not continuously, but rather Bernoulli, distributed).

My problem with Meehl’s methods (MAMBAC, MAXCOV, MAXEIG etc.) is that in all the articles only an intuitive explanation is given. Despite being a mathematical method there were no clear definitions of what the method will consider to be a taxon, or any necessary/sufficient conditions on when the algorithm will detect a taxon. Zoologists for example have entire conferences on how to classify species and go through a lot of painful details on how to properly classify species. They have, it seems, endless debates on what constitutes a new species in the taxonomy. However, I still wasn’t able to find a mathematical definition of what constitutes a taxon.

In addition to that, there seems to be some problems when using MAXCOV with dichotomous indicators (Maruan et.al, An Analysis of Meehl’s MAXCOV-HITMAX Procedure for the Case of Dichotomous Indicators, Multivariate Behavioral Research, Vol. 38, Issue 1 Jan. 2003); in this article they pretty much take the entire procedure apart and show that it often fails to indicate taxons when they are there or indicates taxons when there is nothing.

I think the question of finding a taxon is strongly related to clustering, because it simply tries to answer if clusters exist in the data. However, from all the clustering literature I’ve read so far, clusters are generally defined as dense areas in a space and are found in various ways by maximizing or minimizing some criterion (mutual information etc.). What constitutes a cluster is often conveniently defined so it fits the algorithm at hand. And then you still have to deal with or at least acknowledge the fact that the current notion of clustering has been proven to be impossible (An Impossibility Theorem for Clustering; Kleinberg; NIPS 15).

In a new paper in Machine Learning called Generalization from Observed to Unobserved Features by Clustering (Krupka&Tishby; JMLR 9(Mar):339–370, 2008) the authors describe an idea that might change the way we view clustering. In the paper they show that (under certain conditions) given a clustering based on some features the items will be implicitly clustered by yet unobserved features as well. As an intuitive example, imagine apples, oranges and bananas clustered by shape, color, size, weight etc. Once you have them clustered, you will be able to draw conclusions about a yet unobserved feature, e.g. the vitamin content. The work, because it is oriented on the features, might even be a way around the impossibility-theorem.

This is half-way there for a nicer definition of a taxon or what should constitute a cluster for that matter: can we draw conclusions about features not used for the clustering process? If you are clustering documents by topic (using bag-of-words), can you predict which other words will appear in the article? If you cluster genes, can you predict other genes from the cluster-membership?

Re-clustering on only a subset of the features should also be a sanity check for clustering solutions (I had written about the McIntyre-Blashfield procedure and validating clustering solutions before). I think strong patterns should replicate with less features; at least they did in a clustering-study I did recently .

I’ll be pondering this idea and try formalizing it. Maybe I can come up with something usable for taxometrics and a means to get the number of clusters…

Posted in Statistics, Clustering, Data Mining, Machine Learning | Print | No Comments »