I found an interesting article on how to communicate risks and probabilities to the public.
Archive for the ‘Statistics’ Category
Ideas on communicating risks and probabilities to the general public
Saturday, December 4th, 2010Birthday Paradox
Sunday, October 17th, 2010Here’s an interesting real world example for the Birthday Paradox: Lottery number combination repeats itself. Obligatory XKCD link.
Elo Scores and Rating Contestants
Thursday, August 5th, 2010Kaggle 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(tm) 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.
Alternative measures to the AUC for rare-event prognostic models
Tuesday, February 16th, 2010How 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.
Strong profiling is not mathematically optimal for discovering rare malfeasors (on rare event detection)
Sunday, January 10th, 2010Just in time for the latest Christmas terror scare, I came across an interesting paper: “Strong profiling is not mathematically optimal for discovering rare malfeasors” (William H. Press; PNAS 106(6), p. 1716-1719 www.pnas.org/cgi/doi/10.1073/pnas.0813202106). In the paper, the author investigates whether profiling by nationality or ethnicity can be justified mathematically and tries to answer the question of how much screening must we do, on average, to catch the bad guys in the crowd. Rare events detection is hard as it is, and it’s interesting to see a look from the sampling perspective. It’s an interesting and short read. Long story short, it shows that using an indiscriminate feature like nationality or ethnicity is not optimal (as is any screening at least in proportion to a prior probability) and wastes resources.
Adversarial Scenarios in Risk Mismanagement
Sunday, January 11th, 2009I 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…
Credit Card companies adjusting Credit Scores
Monday, December 22nd, 2008I 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.
Deploying SAS code in production
Saturday, November 1st, 2008I 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 🙂
Taxons, Taxometrics and the Number of Clusters
Thursday, August 21st, 2008In 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…
Using Psychological Domain Knowledge for the Netflix Challenge
Thursday, February 28th, 2008I read an interesting article today about using psychological domain knowledge for improving recommender-system predictions. A very interesting idea…