Many users are probably aware of Shazam, one of the few mobile applications that really seems to live up to the notion "Wow! It's Magic." When you are listening to music, you can run the application (presumably on your phone), click "tag it", and after a few tens of seconds of listening and processing, Shazam will tell you what you are listening to, the artist and other details.
Recently, I found an interesting article describing how it works. A presentation, with more pictures and less text, is available here. Kudos to the company and to the author, Avery Wang, for providing technical detail.
The paper does a very good job explaining the details of the algorithm. My goal here is to describe the algorithm from a higher perspective, because it is an interesting example of a memory-based reasoning algorithm. That is, an algorithm that combines information from "nearest neighbors" to arrive at a prediction.
Assume that we have a database of songs and an excerpt that we want to find in a database of millions of songs. A first approach might be to do an exhaustive search of the database to find a match. This would take a long time.
Alternatively, we can frame the problem as follows: for all songs in the database, what is the longest period of time where the excerpt overlaps part of a song. The nearest neighbors are the ones with the longest period, and, in general, we would choose the single one with the longest overlap.
Simple problem to describe. However, the real world hits quickly. The songs are probably quite clean acoustically. However, the excerpt is subject to numerous problems: background noise, loss of fidelity due to compression as the excerpt is transmitted, poor (or at least different) equipment for recording the excerpt, and so on.
Fortunately, the world of acoustics has something of a solution for this, called the "frequency domain". This is a map of all the sound frequencies, taken at a periodic interval -- say, one second. If "frequency domain" conjures up memories of things like Fourier Transforms, then you really do understand the subject.
However, for our purposes, it is enough to say that the frequency domain for a song produces a very, very bumpy curve for each second of the song -- each point is the strength of a particular acoustic frequency at that point in the song. The song can be thought of as a collection of all these curves. Taken together, these curves might resemble a map of a very hilly area. This would be a three-dimensional map of the song.
This map has peaks anologous to the tops of hills (or perhaps the tops of buildings in a city). These peaks are called a constellations, and they pretty much uniquely identify the song, regardless of all the problems mentioned above. That is, the constellations are resilitient to background noise, loss of fidelity, and so on.
Of course, we can do this for the songs in the database in advance. And, we can do this processing for a single excerpt pretty quickly.
So, the problem of finding the song with the longest overlap in seconds with the excerpt is now handled by finding consecutive seconds in a song where the frequency domain peaks match the frequency domain peaks from the excerpt. This is still a daunting problem, because there are so many peaks available. In other word, comparing one excerpt to millions of songs requires comparing hundreds of peaks in the excerpts to the many, many billions in the database -- very time consuming.
Shazam takes a very clever approach to this problem. The algorithm treats each peak as an anchor, and creates peak-pairs with other peaks "close" to the anchor. Here, "close" means that the other peaks are within a few seconds of the first and not too different in frequency. These peak-pairs are then calculated for both the song and the excerpt. The pairs are used to find sets of anchors that match between each song and the excerpt. Because the algorithm is looking for exact matches, it can use some programming tricks to make things even faster (these are described in the paper).
In the end, there is a set of anchors for each song matched by a given excerpt. For each song, these are scanned to find consecutive seconds where the anchors in each second overlap. The longest period of overlap is the distance between the excerpt and the song.
The algorithm is quite clever on several different levels. I do think that understanding it at a high level is valuable, especially since it can provide guidance to other very difficult recognition problems. On the other hand, when I use Shazam and it identifies a song, I still think it's magic.
Friday, August 28, 2009
Tuesday, August 25, 2009
Neural Networks, Predicting Continuous Values
Hi!
Very good blog...
I'm doing some stuff with Clementine... and I have an issue...
My target for NN train dataset is a continuos value between 0 and 100... the problem is that is a normal/gaussian distribution and makes the NN predict bad...
How can I resolve the unbalancing data? split into classe with same frequency!?
Regards,
Pedro
Pedro,
I am not aware that neural networks have a problem with predicting values with normal distributions. In fact, if you randomize the weights in a neural network whose output layer has a linear transfer function, then the output is likely to follow a normal distribution -- just from the Central Limit Theorem of statistics.
So, you have a neural network that is not producing good results. There can be several causes.
The first thing to look for is too many inputs. Clementine has options to prune the input variables on a neural network. Be sure that you do not have too many inputs. I would recommend a variable reduction technique such as principal components, and advise you to avoid categorical variables that have many levels.
A similar problem can occur if your hidden layer is too large.
Whatever the network, it is worthwhile looking at the number of weights in the network (or a related measure called the degrees of freedom). Remember, you want to have lots of training data for each weight.
Another problem may be that the target is continuous, but bounded between 0 and 100. This could result in a neural network where the output layer uses a linear transfer function. Although not generally a bad idea, it may not work in this case because the range of a linear function is from minus infinity to positive infinity, which far exceeds the range of the data.
One simple solution would be to divide the output by 100 and treat it as a probability. The neural network should then be set up with a logistic function in the target layer.
Your idea of binning the results might also work, assuming that bins work for solving the business problem. Equal sized bins are reasonable, since they are readily understandable as quantiles.
Good luck.
Very good blog...
I'm doing some stuff with Clementine... and I have an issue...
My target for NN train dataset is a continuos value between 0 and 100... the problem is that is a normal/gaussian distribution and makes the NN predict bad...
How can I resolve the unbalancing data? split into classe with same frequency!?
Regards,
Pedro
Pedro,
I am not aware that neural networks have a problem with predicting values with normal distributions. In fact, if you randomize the weights in a neural network whose output layer has a linear transfer function, then the output is likely to follow a normal distribution -- just from the Central Limit Theorem of statistics.
So, you have a neural network that is not producing good results. There can be several causes.
The first thing to look for is too many inputs. Clementine has options to prune the input variables on a neural network. Be sure that you do not have too many inputs. I would recommend a variable reduction technique such as principal components, and advise you to avoid categorical variables that have many levels.
A similar problem can occur if your hidden layer is too large.
Whatever the network, it is worthwhile looking at the number of weights in the network (or a related measure called the degrees of freedom). Remember, you want to have lots of training data for each weight.
Another problem may be that the target is continuous, but bounded between 0 and 100. This could result in a neural network where the output layer uses a linear transfer function. Although not generally a bad idea, it may not work in this case because the range of a linear function is from minus infinity to positive infinity, which far exceeds the range of the data.
One simple solution would be to divide the output by 100 and treat it as a probability. The neural network should then be set up with a logistic function in the target layer.
Your idea of binning the results might also work, assuming that bins work for solving the business problem. Equal sized bins are reasonable, since they are readily understandable as quantiles.
Good luck.
Sunday, August 9, 2009
Pharmaceutical Data and Privacy
Today's New York Times has another misguided article on privacy in the medical world. This article seems to be designed to scare Americans into believing that health care privacy is endangered, and that such data is regularly and wantonly traded among companies.
My perspective is different, since I am coming from the side of analyzing data.
First, the pharmaceutical industry is different from virually every other industry in the United States. For the most part, it is illegal for pharmaceutical manufacturers to identify the users of their products. This is based originated with the Health Information Portability and Privacy Act (HIPAA), explained in more detail at this government site.
What is absurd about this situation is that pharmaceutical companies are, in theory, responsible for the health of the millions of people who use their products. To give an example of the dangers, imagine that you have a popular product that causes cardiac damage after several months of use. The cardiac damage, in turn, is sometimes fatal. How does the manufacturer connect the use of the product to death registries? The simple answer. They cannot.
This is not a made-up example. Millions of people used Cox-2 inhibitors, which were on the market until 2004, when Merck voluntarily took Vioxx off the market. This issue here is whether the industry could have known earlier that such dangers lurked in the use of the drug. My contention is that the manufactureres do not have a chance, because they could not do something that virutally every other company can do -- match their customer records to publicly available mortality records.
To be clear about the laws related to drugs. If someone has an adverse reaction while on a drug, then that must be reported to the pharmaceutical company. However, if the adverse reaction is detected a certain amount of time after the patient stops therapy (I believe two weeks), then there is not reporting requirement. Guess what. Cardiac damage caused by Cox-2 inhibitors does not necessarily kill patients right away. Nor is the damage necessarily detected while the patient is still on the therapy.
I have used deidentified records at pharmaceutical clients for various analyses that have ranged from the amusing (anniversary effects in the scripts for ED therapies) to the socially useful (do poor patients have less adherence due to copayments) to the actionable (what messages to give to prescribers). In all cases, we have had to do more work than necessary because of the de-identification requirements, and to make assumptions and work-arounds that may have hurt the analyses. And, contrary to what the New York Times article may lead you to believe, both IMS and Verispan take privacy very seriously. Were I inclined to try to identify particular records, it would be virtually impossible.
Every time a drug is used, there is perhaps an opportunity to learn about its effectiveness and interactions with other therapies. In many cases, these are questions that scientists do not even know to ask, and such exploratory data mining can be critical in establishing hypotheses. Questions such as:
My perspective is different, since I am coming from the side of analyzing data.
First, the pharmaceutical industry is different from virually every other industry in the United States. For the most part, it is illegal for pharmaceutical manufacturers to identify the users of their products. This is based originated with the Health Information Portability and Privacy Act (HIPAA), explained in more detail at this government site.
What is absurd about this situation is that pharmaceutical companies are, in theory, responsible for the health of the millions of people who use their products. To give an example of the dangers, imagine that you have a popular product that causes cardiac damage after several months of use. The cardiac damage, in turn, is sometimes fatal. How does the manufacturer connect the use of the product to death registries? The simple answer. They cannot.
This is not a made-up example. Millions of people used Cox-2 inhibitors, which were on the market until 2004, when Merck voluntarily took Vioxx off the market. This issue here is whether the industry could have known earlier that such dangers lurked in the use of the drug. My contention is that the manufactureres do not have a chance, because they could not do something that virutally every other company can do -- match their customer records to publicly available mortality records.
To be clear about the laws related to drugs. If someone has an adverse reaction while on a drug, then that must be reported to the pharmaceutical company. However, if the adverse reaction is detected a certain amount of time after the patient stops therapy (I believe two weeks), then there is not reporting requirement. Guess what. Cardiac damage caused by Cox-2 inhibitors does not necessarily kill patients right away. Nor is the damage necessarily detected while the patient is still on the therapy.
I have used deidentified records at pharmaceutical clients for various analyses that have ranged from the amusing (anniversary effects in the scripts for ED therapies) to the socially useful (do poor patients have less adherence due to copayments) to the actionable (what messages to give to prescribers). In all cases, we have had to do more work than necessary because of the de-identification requirements, and to make assumptions and work-arounds that may have hurt the analyses. And, contrary to what the New York Times article may lead you to believe, both IMS and Verispan take privacy very seriously. Were I inclined to try to identify particular records, it would be virtually impossible.
Every time a drug is used, there is perhaps an opportunity to learn about its effectiveness and interactions with other therapies. In many cases, these are questions that scientists do not even know to ask, and such exploratory data mining can be critical in establishing hypotheses. Questions such as:
- Are the therapies equally effective, regardless of gender, age, race, and geography?
- Do demographics affect adherence?
- What interactions does a given therapy have with other therapies?
- Does the use of a particular therapy have an effect on mortality?
Monday, July 27, 2009
Time to Event Models, When the Event Is Not Churn
Dear Data Miners,
I am trying to build a churn model to predict WHEN customers will become paying members. Process:
1. Person comes to our web site.
2. They register for free to use the site.
3. If the want to have more access to the site and use more features they pay us.
What are the issues I should consider when I decide to set a cut date. The first step towards censoring the data.
For a classic churn model , we want to know when someone will stop paying us and leave our phone company. We censor those that we don’t know their final status pass our censor point.
I want to know when they will pay us and censor those I don’t know if they will pay us in the future.
Is the cut date choice arbitrary or is there some sampling rule?
Thank you;
Daryl
Daryl,
Your example is a time-to-event model that does not represent churn. There are many such examples in business (and this is something discussed in Data Analysis Using SQL and Excel in a bit of depth).
Think of your situation as two different time-to-event problems:
(1) A person visits the web site, what happens next? Does the person return to the web site or register? This is a time-to-event problem and analysis can provide information on customer registrations, particularly the lag between the initial visit and the registration.
(2) A person registers for free, how long until that person buys something? This can provide insight on paying visitors.
Once you have broken the problem into these pieces, imagining the customer signature is easier. For the first problem, the customer signature is a picture of customers when they initially visit (or for each pre-registration visit, for a time-to-next event problem). The "prediction" columns are the date of the registration (or for time-to-event, the date of the next visit and whether it involves a registration).
The second component is a picture of the customer when they first register, and the prediction columns are when (and whether) the customer every pays for anything. In this case, it is very important to treat this as a time-to-event problem, because older registrations have had more opportunity to pay for something and the analysis needs to take this into account.
As for the censor date, it is the most recent date of the data. So, if you have data through the end of yesterday, then that is the censor date. For instance, for the second component of the analysis, customers who registered before yesterday but never paid would have their outcomes censored (these customers have not paid yet but they may pay in the future).
I am trying to build a churn model to predict WHEN customers will become paying members. Process:
1. Person comes to our web site.
2. They register for free to use the site.
3. If the want to have more access to the site and use more features they pay us.
What are the issues I should consider when I decide to set a cut date. The first step towards censoring the data.
For a classic churn model , we want to know when someone will stop paying us and leave our phone company. We censor those that we don’t know their final status pass our censor point.
I want to know when they will pay us and censor those I don’t know if they will pay us in the future.
Is the cut date choice arbitrary or is there some sampling rule?
Thank you;
Daryl
Daryl,
Your example is a time-to-event model that does not represent churn. There are many such examples in business (and this is something discussed in Data Analysis Using SQL and Excel in a bit of depth).
Think of your situation as two different time-to-event problems:
(1) A person visits the web site, what happens next? Does the person return to the web site or register? This is a time-to-event problem and analysis can provide information on customer registrations, particularly the lag between the initial visit and the registration.
(2) A person registers for free, how long until that person buys something? This can provide insight on paying visitors.
Once you have broken the problem into these pieces, imagining the customer signature is easier. For the first problem, the customer signature is a picture of customers when they initially visit (or for each pre-registration visit, for a time-to-next event problem). The "prediction" columns are the date of the registration (or for time-to-event, the date of the next visit and whether it involves a registration).
The second component is a picture of the customer when they first register, and the prediction columns are when (and whether) the customer every pays for anything. In this case, it is very important to treat this as a time-to-event problem, because older registrations have had more opportunity to pay for something and the analysis needs to take this into account.
As for the censor date, it is the most recent date of the data. So, if you have data through the end of yesterday, then that is the censor date. For instance, for the second component of the analysis, customers who registered before yesterday but never paid would have their outcomes censored (these customers have not paid yet but they may pay in the future).
Friday, June 26, 2009
When Customers Start and End
In texts on credit scoring, some effort almost always goes into defining what is to be considered as a "bad" credit. The Basel framework provides rather a precise definition of what is to be considered a default.
But I have rarely seen the same in predicting cross-sell, up-sell or churn. I do however, remember attending an SPSS conference where churn of pre-paid cards was discussed. Churn, in that case, was defined as a number of consecutive periods where the number of calls fell below a certain level.
In the past, I've used start and end dates of contracts, as well as a simple increase (or decrease) in the number of products that a customer has over time as indicators of what to target.
I'd be really interested in hearing how you define and extract targets, be it in telecom, banking, cards or any other business where you use prediction. For instance, how would you go looking for customers that have churned? Or for that matter, customers where up-sell has been successful?
This may be too simple a question, but if there are standard methods that you use, I'd be really interested in learning about them.
--Ola
Ola,
This is not a simple question at all. Or rather, the simplest questions are often the most illuminating.
The place where I see the biggest issues in defining starts and stops is in survival data mining (obligatory plug for my book Data Analysis Using SQL and Excel, which has two chapters on the subject). For the start date, I try to use (or approximate as closely as possible) the date when two things have occurred: the company has agreed to provide a product or service, and the customer has agreed to pay for it. In the case of post-pay telecoms, this would be the activation date -- and there are similar dates in many other industries, as varied as credit cards, cable subscriptions, and health insurance.
The activation date is often well-defined because the number of active customers gets reported through some system tied to the financial systems. Even so, there are anomalies. I recently completed a project at a large newspaper, and used their service start date as the activation date. Alas, at time, customers with start dates did not necessarily actually receive the paper on the date -- often because the newspaper delivery person could not find the address.
The stop date is even more fraught with complication, because there are a variety of different dates to choose from. For voluntary churn, there is the date the customer requests termination of the service. There is also the date when the service is actually turned off. Which to use? It depends on the application. To count active customers, we want the service cut-off date. To plan for customer retention efforts, we want to know when they call in.
Involuntary churn is also complicated, because there are a series of steps, often called the Dunning Process, which keeps track of customers who do not pay. At what point does a non-paying customer stop? When the service stops? When the bill is written off or settled? At some arbitrary point, such as 60 or 90 days of non-payment? To further confuse the situation, the business may change its rules over time. So, during some periods of time or for some customers, 60 days of non-payment results in service cutoff. For other periods or customers, 90 days might be the rule.
Often, I find multiple time-to-event problems in this scenario. How long does it take a non-paying customer to stop, if ever? How long after customers sign up do they begin?
In your particular case, the contract start date is probably a good place to start. However, the contract end date might or might not be appropriate, since this might not be updated to reflect when a customer actually stops.
--gordon
But I have rarely seen the same in predicting cross-sell, up-sell or churn. I do however, remember attending an SPSS conference where churn of pre-paid cards was discussed. Churn, in that case, was defined as a number of consecutive periods where the number of calls fell below a certain level.
In the past, I've used start and end dates of contracts, as well as a simple increase (or decrease) in the number of products that a customer has over time as indicators of what to target.
I'd be really interested in hearing how you define and extract targets, be it in telecom, banking, cards or any other business where you use prediction. For instance, how would you go looking for customers that have churned? Or for that matter, customers where up-sell has been successful?
This may be too simple a question, but if there are standard methods that you use, I'd be really interested in learning about them.
--Ola
Ola,
This is not a simple question at all. Or rather, the simplest questions are often the most illuminating.
The place where I see the biggest issues in defining starts and stops is in survival data mining (obligatory plug for my book Data Analysis Using SQL and Excel, which has two chapters on the subject). For the start date, I try to use (or approximate as closely as possible) the date when two things have occurred: the company has agreed to provide a product or service, and the customer has agreed to pay for it. In the case of post-pay telecoms, this would be the activation date -- and there are similar dates in many other industries, as varied as credit cards, cable subscriptions, and health insurance.
The activation date is often well-defined because the number of active customers gets reported through some system tied to the financial systems. Even so, there are anomalies. I recently completed a project at a large newspaper, and used their service start date as the activation date. Alas, at time, customers with start dates did not necessarily actually receive the paper on the date -- often because the newspaper delivery person could not find the address.
The stop date is even more fraught with complication, because there are a variety of different dates to choose from. For voluntary churn, there is the date the customer requests termination of the service. There is also the date when the service is actually turned off. Which to use? It depends on the application. To count active customers, we want the service cut-off date. To plan for customer retention efforts, we want to know when they call in.
Involuntary churn is also complicated, because there are a series of steps, often called the Dunning Process, which keeps track of customers who do not pay. At what point does a non-paying customer stop? When the service stops? When the bill is written off or settled? At some arbitrary point, such as 60 or 90 days of non-payment? To further confuse the situation, the business may change its rules over time. So, during some periods of time or for some customers, 60 days of non-payment results in service cutoff. For other periods or customers, 90 days might be the rule.
Often, I find multiple time-to-event problems in this scenario. How long does it take a non-paying customer to stop, if ever? How long after customers sign up do they begin?
In your particular case, the contract start date is probably a good place to start. However, the contract end date might or might not be appropriate, since this might not be updated to reflect when a customer actually stops.
--gordon
Monday, June 8, 2009
Confidence in Logistic Regression Coefficients
I work in the marketing team of a telecom company and I recently encountered an annoying problem with an upsell model. Since the monthly sale rate is less than 1% of our customer base, I used oversampling as you mentioned in your book ‘Mastering data mining’ with data over the last 3 sales months so that I had a ratio of about 15% buyers and 85% non-buyers (sample size of about 20K). Using alpha=5%, I got parameter estimates which were from a business perspective entirely explicable. However, when I then re-estimated the model on the total customer base to obtain the ‘true’ parameter estimates which I will use for my monthly scoring two effects were suddenly insignificant at alpha=5%.
I never encountered this and was wondering what to do with these effects: should I kick them out of the model or not ? I decided to keep them in since they did have some business meaning and concluded that they must have become insignificant since it is only a micro-segment in your entire population.
To your opinion, did I interpret this correctly ? . . .
Many thanks in advance for your advice,
Wendy
Michael responds:
Hi Wendy,
This question has come up on the blog before. The short answer is that with a logistic regression model trained at one concentration of responders, it is a bit tricky to adjust the model to reflect the actual probability of response on the true population. I suggest you look at some papers by Gary King on this topic.
Gordon responds:
Wendy, I am not sure that Prof. King deals directly with your issue, of changing confidence in the coefficients estimates. To be honest, I have never considered this issue. Since you bring it up, though, I am not surprised that it may happen.
My first comment is that the results seem usable, since they are explainable. Sometimes statistical modeling stumbles on relationships in the data that make sense, although they may not be fully statistically significant. Similarly, some relationships may be statistically significant, but have no meaning in the real world. So, use the variables!
Second, if I do a regresson on a set of data, and then duplicate the data (to make it twice as big) and run it again, I'll get the same estimates as on the orignal data. However, the confidence in the coefficients will increase. I suspect that something similar is happening on your data.
If you want to fix that particular problem, then use a tool (such as SAS Enterprise Miner and probably proc logistic) that supports a frequency option on each row. Set the frequency to one for the more common events and to an appropriate value less than one for more common events. I do this as a matter of habit, because it works best for decision trees. You have pointed out that the confidence in the coefficients is also affected by the frequencies, so this is a good habit with regressions as well.
I never encountered this and was wondering what to do with these effects: should I kick them out of the model or not ? I decided to keep them in since they did have some business meaning and concluded that they must have become insignificant since it is only a micro-segment in your entire population.
To your opinion, did I interpret this correctly ? . . .
Many thanks in advance for your advice,
Wendy
 |
Michael responds:
Hi Wendy,
This question has come up on the blog before. The short answer is that with a logistic regression model trained at one concentration of responders, it is a bit tricky to adjust the model to reflect the actual probability of response on the true population. I suggest you look at some papers by Gary King on this topic.
Gordon responds:
Wendy, I am not sure that Prof. King deals directly with your issue, of changing confidence in the coefficients estimates. To be honest, I have never considered this issue. Since you bring it up, though, I am not surprised that it may happen.
My first comment is that the results seem usable, since they are explainable. Sometimes statistical modeling stumbles on relationships in the data that make sense, although they may not be fully statistically significant. Similarly, some relationships may be statistically significant, but have no meaning in the real world. So, use the variables!
Second, if I do a regresson on a set of data, and then duplicate the data (to make it twice as big) and run it again, I'll get the same estimates as on the orignal data. However, the confidence in the coefficients will increase. I suspect that something similar is happening on your data.
If you want to fix that particular problem, then use a tool (such as SAS Enterprise Miner and probably proc logistic) that supports a frequency option on each row. Set the frequency to one for the more common events and to an appropriate value less than one for more common events. I do this as a matter of habit, because it works best for decision trees. You have pointed out that the confidence in the coefficients is also affected by the frequencies, so this is a good habit with regressions as well.
Sunday, May 10, 2009
Not Enough Data
An article in yesterday's New York Times reminded me of examples of "bad" examples of data mining. By bad examples, I mean that spurious correlations are given credence -- enough credence to make it into a well-reputed national newspaper.
The article, entitled "Eat Quickly, for the Economy's State" is about a leisure time report from the OECD that shows a correlation between the following two variables:
The highlight is a few charts that shows that countries such as Mexico, Canada, and the United States have the lowest time spent eating (under 75 minutes per day) versus countries such as New Zealand, France, and Japan (over 110 minutes per day). The first group of countries have higher growth rates, both in 2008 and for the past few years.
My first problem with the analysis is one of granularity. Leisure time is measured per person, but GNP is measured over everyone. One big component of GNP growth is population growth, and different countries have very different patterns of population growth. The correct measure would be per capital GNP. Taking this into account would dampen the GNP growth figures for growing countries such as Mexico and the United States, and increase the GNP growth figures for lesser growing (or shrinking countries) such as Italy, Germany, and Japan.
Also, the countries where people eat more leisurely have other characteristics in common. In particular, they tend to have older populations and lower (or even negative) rates of population growth. One wonders if speed eating is a characteristic of younger people and leisurely eating is a characteristic of older people.
The biggest problem, though, is that this is, in all likelihood, a spurious correlation. One of the original definitions of data mining, which may still be used in the ecoonomics and political world, is a negative one: data mining is looking for data to support a conclusion. The OECD surveys were done in 17 different countries. The specific result in the NYT article is "Counties in which people eat and drink less than 100 minutes per day grow 0.9% faster -- on average -- than countries in which people each and drink more than 100 minutes per day".
In other words, the 17 countries were divided into two groups, and the growth rates were then measured for each group. Let's look at this in more detail.
How many ways are there to divide 17 countries into 2 groups? The answer is 2^17 = 131,072 different ways (any particular country could be in either group). So, if we had 131,072 yes-or-no survey questions, then would would expect any combination to arise, including the combinations where all the high growth countries are in one group and all the low growth countries in the other. (I admit the exact figure is a bit more than 131,072 but that is unimportant to illustrate my point.)
The situation actually gets worse. The results are not yes-or-no; they are numeric measurements which are then used to split the countries into two groups. The splits could be at any value of the measure. So, any given measurement results in 17-1=16 different possible splits (the first group having the country with the lowest measurement, with the two lowest, and so on). Now we only need about 8,192 uncorrelated measurements to get all possibilities.
However, we do not need all possibilities. A glance at the NYT article shows that the country with the worst 2008 growth is Poland, yet it is in the fast-eating group. And Spain -- in the slow eating group -- is the third fastest growing economy (okay, its GNP actually shrank but less than most others). So, we only need an approximation of a split, where the two groups look different. And then, voila! we get a news article.
The problem is that the OECD was able to measure dozens or hundreds of different things in their survey. My guess is that measures such as "weekly hours of work in main job," "time spent retired," and "time spent sleeping" -- just a few of the many possibilities -- did not result in interesting splits. Eventually, though, a measure such as "time spent eating and drinking" results in a split where the different groups look "statistically significant" but they probably are not. If the measure is interesting enough, then it can become an article in the New York Times.
This is probably a problem with statistical significance. The challenge is that a p-value of 0.01 means that something has only a 1% chance of happening at random. However, if we look at 100 different measures, then there is a really, really good chance that one of them will have a p-value of 0.01 or less. By the way, there is a statistical adjustment called the Bonferroni correction to take this into account (this as well as others are described in the Wikipeida).
Fortunately, neither the OECD nor the New York Times talk about this discovery as an example of data mining. It is just poor data analysis, but poor data analysis that can re-enforce lessons in good data analysis. Lately, I have been noticing more examples of articles such as this, where researchers -- or perhaps just journalists -- extrapolate from very small samples to make unsupported conclusions. These are particularly grating when they appear in respected newspapers, magazines, and journals.
Data mining is not about finding spurious correlations and claiming some great discovery. It is about extracting valuable information from large quantities of data, information that is stable and useful. Smaller amounts of data often contain many correlations. Often, these correlations are going to be spurious. And without further testing, or at least a mechanism to explain the correlation, the results should not be mentioned at all.
The article, entitled "Eat Quickly, for the Economy's State" is about a leisure time report from the OECD that shows a correlation between the following two variables:
- Change in real GNP in 2008; and,
- Amount of time people spend eating and drinking in a given day.
The highlight is a few charts that shows that countries such as Mexico, Canada, and the United States have the lowest time spent eating (under 75 minutes per day) versus countries such as New Zealand, France, and Japan (over 110 minutes per day). The first group of countries have higher growth rates, both in 2008 and for the past few years.
My first problem with the analysis is one of granularity. Leisure time is measured per person, but GNP is measured over everyone. One big component of GNP growth is population growth, and different countries have very different patterns of population growth. The correct measure would be per capital GNP. Taking this into account would dampen the GNP growth figures for growing countries such as Mexico and the United States, and increase the GNP growth figures for lesser growing (or shrinking countries) such as Italy, Germany, and Japan.
Also, the countries where people eat more leisurely have other characteristics in common. In particular, they tend to have older populations and lower (or even negative) rates of population growth. One wonders if speed eating is a characteristic of younger people and leisurely eating is a characteristic of older people.
The biggest problem, though, is that this is, in all likelihood, a spurious correlation. One of the original definitions of data mining, which may still be used in the ecoonomics and political world, is a negative one: data mining is looking for data to support a conclusion. The OECD surveys were done in 17 different countries. The specific result in the NYT article is "Counties in which people eat and drink less than 100 minutes per day grow 0.9% faster -- on average -- than countries in which people each and drink more than 100 minutes per day".
In other words, the 17 countries were divided into two groups, and the growth rates were then measured for each group. Let's look at this in more detail.
How many ways are there to divide 17 countries into 2 groups? The answer is 2^17 = 131,072 different ways (any particular country could be in either group). So, if we had 131,072 yes-or-no survey questions, then would would expect any combination to arise, including the combinations where all the high growth countries are in one group and all the low growth countries in the other. (I admit the exact figure is a bit more than 131,072 but that is unimportant to illustrate my point.)
The situation actually gets worse. The results are not yes-or-no; they are numeric measurements which are then used to split the countries into two groups. The splits could be at any value of the measure. So, any given measurement results in 17-1=16 different possible splits (the first group having the country with the lowest measurement, with the two lowest, and so on). Now we only need about 8,192 uncorrelated measurements to get all possibilities.
However, we do not need all possibilities. A glance at the NYT article shows that the country with the worst 2008 growth is Poland, yet it is in the fast-eating group. And Spain -- in the slow eating group -- is the third fastest growing economy (okay, its GNP actually shrank but less than most others). So, we only need an approximation of a split, where the two groups look different. And then, voila! we get a news article.
The problem is that the OECD was able to measure dozens or hundreds of different things in their survey. My guess is that measures such as "weekly hours of work in main job," "time spent retired," and "time spent sleeping" -- just a few of the many possibilities -- did not result in interesting splits. Eventually, though, a measure such as "time spent eating and drinking" results in a split where the different groups look "statistically significant" but they probably are not. If the measure is interesting enough, then it can become an article in the New York Times.
This is probably a problem with statistical significance. The challenge is that a p-value of 0.01 means that something has only a 1% chance of happening at random. However, if we look at 100 different measures, then there is a really, really good chance that one of them will have a p-value of 0.01 or less. By the way, there is a statistical adjustment called the Bonferroni correction to take this into account (this as well as others are described in the Wikipeida).
Fortunately, neither the OECD nor the New York Times talk about this discovery as an example of data mining. It is just poor data analysis, but poor data analysis that can re-enforce lessons in good data analysis. Lately, I have been noticing more examples of articles such as this, where researchers -- or perhaps just journalists -- extrapolate from very small samples to make unsupported conclusions. These are particularly grating when they appear in respected newspapers, magazines, and journals.
Data mining is not about finding spurious correlations and claiming some great discovery. It is about extracting valuable information from large quantities of data, information that is stable and useful. Smaller amounts of data often contain many correlations. Often, these correlations are going to be spurious. And without further testing, or at least a mechanism to explain the correlation, the results should not be mentioned at all.
Subscribe to:
Posts (Atom)
Posts
