## High Kappa Values are not Necessary for High Quality Corpora

October 2, 2012

I’m not a big fan of kappa statistics, to say the least. I point out several problems with kappa statistics right after the initial studies in this talk on annotation modeling.

I just got back from another talk on annotation where I was ranting again about the uselessness of kappa. In particular, this blog post is an attempt to demonstrate why a high kappa is not necessary. The whole point of building annotation models a la Dawid and Skene (as applied by Snow et al. in their EMNLP paper on gather NLP data with Mechanical Turk) is that you can create a high-reliability corpus without even having high accuracy, much less acceptable kappa values — it’s the same kind of result as using boosting to combine multiple weak learners into a strong learner.

So I came up with some R code to demonstrate why a high kappa is not necessary without even bothering with generative annotation models. Specifically, I’ll show how you can wind up with a high-quality corpus even in the face of low kappa scores.

The key point is that annotator accuracy fully determines the accuracy of the resulting entries in the corpus. Chance adjustment has nothing at all to do with corpus accuracy. That’s what I mean when I say that kappa is not predictive. If I only know the annotator accuracies, I can tell you expected accuracy of entries in the corpus, but if I only know kappa, I can’t tell you anything about the accuracy of the corpus (other than that all else being equal, higher kappa is better; but that’s also true of agreement, so kappa’s not adding anything).

First, the pretty picture (the colors are in honor of my hometown baseball team, the Detroit Tigers, clinching a playoff position).

What you’re looking at is a plot of the kappa value vs. annotator accuracy and category prevalence in a binary classification problem. (It’s only the upper-right corner of a larger diagram that would let accuracy run from 0 to 1 and kappa from 0 to 1. Here’s the whole plot for comparison.

Note that the results are symmetric in both accuracy and prevalence, because very low accuracy leads to good agreement in the same way that very high accuracy does.)

How did I calculate the values? First, I assumed accuracy was the same for both positive and negative categories (usually not the case — most annotators are biased). Prevalence is defined as the fraction of items belonging to category 1 (usually the “positive” category).

Everything else follows from the definitions of kappa, to result in the following definition in R to compute expected kappa from binary classification data with a given prevalence of category 1 answers and a pair of annotators with the same accuracies.

kappa_fun = function(prev,acc) {
agr = acc^2 + (1 - acc)^2;
cat1 = acc * prev + (1 - acc) * (1 - prev);
e_agr = cat1^2 + (1 - cat1)^2;
return((agr - e_agr) / (1 - e_agr));
}


Just as an example, let’s look at prevalence = 0.2 and accuracy = 0.9 with say 1000 examples. The expected contingency table would be

 Cat1 Cat2 Cat1 170 90 Cat2 90 650

and the kappa coefficient would be 0.53, below anyone’s notion of “acceptable”.

The chance of actual agreement is the accuracy squared (both annotators are correct and hence agree) plus one minus the accuracy squared (both annotators are wrong and hence agree — two wrongs make a right for kappa, another of its problems).

The proportion of category 1 responses (say positive responses) is the accuracy times the prevalence (true category is positive, correct response) plus one minus accuracy times one minus prevalence (true category is negative, wrong response).

Next, I calculate expected agreement a la Cohen’s kappa (which is the same as Scott’s pi in this case because the annotators have identical behavior and hence everything’s symmetric), which is just the resulting agreement from voting according to the prevalences. So that’s just the probability of category 1 squared (both annotators respond category 1) and the probability of a category 2 response (1 minus the probability of a category 1 response) squared.

Finally, I return the kappa value itself, which is defined as usual.

Back to the plot. The white border is set at .66, the lower-end threshold established by Krippendorf for somewhat acceptable kappas; the higher-end threshold of acceptable kappas set by Krippendorf was 0.8, and is also indicated on the legend.

In my own experience, there are almost no 90% accurate annotators for natural language data. It’s just too messy. But you need well more than 90% accuracy to get into acceptable kappa range on a binary classification problem. Especially if prevalence is high, because as prevalence goes up, kappa goes down.

I hope this demonstrates why having a high kappa is not necessary.

I should add that Ron Artstein asked me after my talk what I thought would be a good thing to present if not kappa. I said basic agreement is more informative than kappa about how good the final corpus is going to be, but I want to go one step further and suggest you just inspect a contingency table. It’ll tell you not only what the agreement is, but also what each annotator’s bias is relative to the other (evidenced by asymmetric contingency tables).

In case anyone’s interested, here’s the R code I then used to generate the fancy plot:

pos = 1;
K = 200;
prevalence = rep(NA,(K + 1)^2);
accuracy = rep(NA,(K + 1)^2);
kappa = rep(NA,(K + 1)^2);
for (m in 1:(K + 1)) {
for (n in 1:(K + 1)) {
prevalence[pos] = (m - 1) / K;
accuracy[pos] = (n - 1) / K;
kappa[pos] = kappa_fun(prevalence[pos],accuracy[pos]);
pos = pos + 1;
}
}
library("ggplot2");
df = data.frame(prevalence=prevalence,
accuracy=accuracy,
kappa=kappa);
kappa_plot =
ggplot(df, aes(prevalence,accuracy,fill = kappa)) +
labs(title = "Kappas for Binary Classification\n") +
geom_tile() +
scale_x_continuous(expand=c(0,0),
breaks=c(0,0.25,0.5,0.75,1),
limits =c(0.5,1)) +
scale_y_continuous(expand=c(0,0),
breaks=seq(0,10,0.1),
limits=c(0.85,1)) +
low="orange", mid="white", high="blue",
breaks=c(1,0.8,0.66,0));


## Refactoring in the Zone

September 17, 2012

I remember very clearly when I first started to work as a professional programmer. I was tasked with first designing and then integrating a new semantic interpreter for the grammars in SpeechWorks’s speech recognizer.

I didn’t know my ass from my elbow and pretty much couldn’t get off the ground on my own.

### Get Adopted by a Great Mentor

Luckily, Sasha Caskey pretty much saved my professional programming life by pair programming with me until I “got it” and on a continuing basis after that so I didn’t forget.

One of Sasha’s memorable early lessons involved refactoring or adding new features. After everything was designed (something I’m still more comfortable with than coding), there was the integration. This involved a C implementation of JavaScript interpreting user programs with high-level dialog control and low-level speech integration. I just couldn’t see how the ends could be made to meet.

### Use the Force

Sasha said something along the lines of “use the force”. What he really said is probably more along the lines of “when you’re dealing with good code like this you just have a sense of where everything should go and if you stick to the plan, it usually works”. It sounded awfully reckless to me, but then I was having trouble seeing how version control could work with 20 programmers sharing a code base.

He applied this philosophy on percolating arguments through call chains, propagating return codes, dealing with exceptions (all hacked up with gotos to end-of-function cleanup blocks because we were using straight-up C), and even figuring out what the bounds of a loop should be.

It works. But only once you get to a certain level of expertise where you know what to expect and how things look if they’re “right”. And only if the other people you work with write idiomatic code.

I was reminded of Sasha’s early lessons on two occasions recently.

### Sometimes it Works

First, I just added print statements to Stan’s modeling language. I needed to pass a standard output stream as well as a standard error stream to be the target of code writes. The error stream was already getting propagated. Even though I didn’t write a lot of the code I’m dealing with, the other people I work with are well-trained C++ coders, so everything just works as expected. It was like the current code was sprinkled with bread crumbs I could follow to do what I needed.

The force worked!

### Sometimes it Doesn’t

Second, I was refactoring some student-written Java code and it’s so non-idiomatic I can’t make heads or tails of it. (Think Daily WTF levels of insanity here.) I had no idea what the original programmer intended or how the code was supposed to implement those intentions.

The force completely failed me.

### Chess Memory in Experts and Novices

This all reminds me of some of my favorite psychology experiments ever, by de Groot in the 1950s with followups by Simon and Chase in the 1970s. (I heard about them in Herb Simon’s wonderful cog psych class/seminar at CMU.) It also relates to the seminal short-term memory experiments of George Miller in the 1950s.

The main takeaway is that chess experts have great memories for board positions if and only if the boards make sense tactically. They’re no better than amateurs at remembering random board positions. Even the errors they make in reconstructing board positions also tend to preserve the tactical arrangement even if all the pieces. Herb’s takeaway message was that “memory” is very tied up with expectations and the ability to “chunk” information into bundles. Miller’s experiments and followups showed we could remember as many bundles of information as single random pieces of information (the famous “7 +/- 2” figure for the number of items humans can hold in short term memory).

Here’s a nice survey of the psychology of chess that covers the above experiments and more.

## Dilbert Meets Big Unstructured Data … and Builds a Framework

September 5, 2012

Best Dilbert ever. Or at least the most relevant to this blog:

Dilbert, 5 September 2012

I’ll give you the setup. Dilbert walks into a bar and strikes up a conversation with a woman who asks him what he does for a living. Dilbert replies, “I’m working on a framework to allow construction of large-scale analytical queries on unstructured data.”

I’ll leave the punchline to the strip.

## Using Luke the Lucene Index Browser to develop Search Queries

July 24, 2012

Luke is a GUI tool written in Java that allows you to browse the contents of a Lucene index, examine individual documents, and run queries over the index. Whether you’re developing with PyLucene, Lucene.NET, or Lucene Core, Luke is your friend.

Downloads are available from http://code.google.com/p/luke/ You can download the source or just the standalone binary .jar file. Right now we’re using the just-released Luke 4.0.0-ALPHA standalone binary. To run Luke, just fire up the jar file from the command line:

> java -jar lukeall-4.0.0-ALPHA.jar

Once launched, Luke opens with a dialog menu to open an index. We’re using an index called fed-papers.idx, which is an index over the Federalist Papers, a set of 85 op-ed pieces written by James Madison, Alexander Hamilton, and/or John Jay. This is taken from the example in our recently updated Lucene 3 tutorial. To build the index yourself, download the data and accompanying source code and follow the instructions in the tutorial. Each of the 85 papers is treated as a document. The text of the document has been indexed using Lucene’s StandardAnalyzer and is stored in a field named text.

The following is a screenshot of the Luke GUI upon first opening the index. We’ve circled the main controls on the top left corner of the GUI. Luke opens with the Overview tab pane.

All 85 documents have a similar structure. They start with an identifying number, e.g. FEDERALIST No. 1. The opening salutation is: To the People of the State of New York:. Looking at the top terms in the index, we see that there are 85 documents which contain the words {federalist, people, state, york} as well as common function words. The StandardAnalyzer has stop-listed other function words, including {to, of, the, no}. Otherwise, these too would be among the top terms with a document frequency of 85.

### Exploring Document Indexing

The Document tab pane lets you examine individual documents in the index. In the screenshot below we have selected the 47th document in the index, FEDERALIST No. 48. Circled in red are the series of controls used to get to this point. First we chose the Document tab pane, and then used the “Browse by document number” control to choose document 47. The bottom half of the window shows the document fields.

Clicking on the “Reconstruct & Edit” control opens a new pop-up window which allows us to inspect the document contents on a field-by-field basis. Below we show side-by-side screenshots of the text field. On the left we see the raw text as stored; on the right we see the result of tokenization and indexing via the Lucene StandardAnalyzer. Lucene’s StandardAnalyzer includes a StandardTokenizer, StandardFilter, Tokenization happens first. This removes punctuation and assigns each token a position. Then the tokens are lower-cased and stop-listed. LowerCaseFilter and StopFilter.

We have circled the text of the opening salutation: To the People of the State of New York:. It consists of 9 words followed by 1 punctuation symbol. The tokenizer strips out the final colon. The stop-list filter removes the tokens at positions {1, 2, 4, 5, 7}. The text of tokens which have been indexed are displayed as is. Tokens which were stop-listed are missing from the index, so Luke displays the string null followed by the number of consecutive stop-listed tokens.

### Exploring Search

The Search tab pane packs a lot of controls. The annotated screenshot below shows the results of running a search over the index.

In the top right quadrant is an embedded tab pane (tabs circled in red). On the Analysis tab, we have selected the StandardAnalyzer from a pull-down list of Analyzers and selected the default search field from a pull-down list of all fields in the index. Given this, Luke constructs the QueryParser used to parse queries entered into the text box in the upper left quadrant. We entered the words: “Powers of the Judiciary” into this text box, circled in red. Directly below is it the parsed query, also circled in red. Clicking on the “Search” control runs this search over the index. The bottom pane displays the ranked results.

Luke passes the input string from the search text box to the QueryParser which parses it using the specified analyzer and creates a set of search terms, one term per token, over the specified default field. In this example, the StandardAnalyzer tokenizes, lowercases, and stop-lists these words, resulting in two search terms text:powers and text:judiciary. The result of this search pulls back all documents which contain the words powers and/or judiciary.

To drill down on Lucene search, we use the Explain Structure control, circled in red on the following screenshot.

When clicked, this pops up another pane which shows all details of the query constructed from the search expression. The structure of the query used in this example is:

lucene.BooleanQuery
clauses=2, maxClauses=1024
Clause 0: SHOULD
lucene.TermQuery
Term: field='text' text='powers'
Clause 1: SHOULD
lucene.TermQuery
Term: field='text' text='judiciary'

The search expression can be any legal search string according to the Lucene Query Parser Syntax. To search for the phrase powers of the judiciary we need to enclose it in double quotes. But this new search produces no results, which is clearly wrong, since this phrase occurs in the title of FEDERALIST No. 80 (see search results above).

In the query details we see that the input has been parsed into text:"powers judiciary". The Explain Structure feature returns the following:

lucene.PhraseQuery, slop=0
pos: [0,1]
Term 0: field='text' text='powers'
Term 1: field='text' text='judiciary'

The problem here is that the token positions specified don’t allow for the intervening stop-listed words. To correct this, we need to adjust Luke’s default settings on the QueryParser. The screenshot below shows both the controls changed and the query results.

First we go to the QueryParser tab in the set of search controls in the top right quadrant, then we click the checkbox labeled Enable positionIncrements. Now the parsed query looks like this: text:"powers ? ? judiciary", which translates to the following programmatic query:

lucene.PhraseQuery, slop=0
pos: [0,3]
Term 0: field='text' text='powers'
Term 1: field='text' text='judiciary'

Finally, we select the single search result, and click on the Explain control (circled in red). This pops up another window which explains the query score (outlined in red). Here is the text of the explanation:

0.0847  weight(text:"powers ? ? judiciary" in 79) [DefaultSimilarity], result of:
0.0847  fieldWeight in 79, product of:
1.0000  tf(freq=1.0), with freq of:
1.0000  phraseFreq=1.0
3.6129  idf(), sum of:
1.3483  idf(docFreq=59, maxDocs=85)
2.2646  idf(docFreq=23, maxDocs=85)
0.0234  fieldNorm(doc=79)

### Using the Lucene XML Query Parser

The Lucene API contains many kinds of queries beyond those generated by the QueryParser. You can use Luke to develop these queries as well, via the Lucene XML Query Parser. It is almost impossible to find any on-line documentation on this most excellent contribution to Lucene by Mark Harwood. Luckily, it is distributed with Lucene source tgz files. For Lucene 3.6 the path to this documentation is: lucene-3.6.0/contrib/xml-query-parser/docs/index.html.

Recast in as a Lucene XML Query, our original search term:powers term:judiciary becomes:

 <BooleanQuery fieldName="text">
<Clause occurs="should">
<TermQuery>powers</TermQuery>
</Clause>
<Clause occurs="should">
<TermQuery>judiciary</TermQuery>
</Clause>
</BooleanQuery>

The screenshot below shows the results of pasting the above XML into the search expression textbox with the QueryParser control Use XML Query Parser turned on. (Note: currently this works in Luke 3.5 but not in Luke 4.0-ALPHA. We hate the bleeding edge).

The XML rewrites to exactly the same query as our first query.

Big deal, you might be saying right about now. That’s a whole lotta chopping for not much kindling. Au contraire! One of the reasons for developing the XML Query Syntax is this:

To bridge the growing gap between Lucene query/filtering functionality and the set of functionality accessible through the standard Lucene QueryParser syntax.

For example, a while back, Mark Miller blogged about Lucene SpanQueries. This is similar to but not the same as PhraseQuery. We can use Luke to compare and contrast the difference between the two. Here’s our earlier phrase query Powers of the Judiciary recast as a span query.

  <SpanNear slop="3" inOrder="true" fieldName="text">
<SpanTerm>powers</SpanTerm>
<SpanTerm>judiciary</SpanTerm>
</SpanNear>

The screenshot below shows the results of running this query.

FEDERALIST No. 48 scores slightly better on this query than does FEDERALIST No. 80. Understanding why is left as an exercise to the reader.

## MacBook Pro 15″ Retina Display Awesomeness

July 14, 2012

I just received my new MacBook Pro 15″ with the Retina display.

First, I have to mention how blown away I was that Apple has a feature (the “Migration Assistant“) that lets you clone your last computer. An hour or two after setting up, the new MacBook Pro had all the software, data, and settings (well, almost all) from my previous computer, a MacBook Air. All done over my home wireless network (though our sysadmin here at Columbia strongly recommended a wired connection, my Air doesn’t have a port and I didn’t have a dongle).

Yes, text is just as beautiful as on the iPad3. So are photos and images. Everything else I use is looking awfully pixelated in comparison (such as this blog post I’m typing into Safari on my 27″ iMac).

The biggest downside is that it’s big (15″ diagonal screen vs. 13″ on my MacBook Air) and heavy (4.5 lbs vs. 2.9 lbs for the Air). Though big isn’t so bad — the 15” screen seems luxurious after the Air’s rather cramped confines. Some software’s not up to the display, so the text looks really bad on the new MacBook Pro. Firefox and Thunderbird, for instance, look terrible. Overall, it’s just not as nice to handle as the Air. (Not to mention Columbia slapping the ugliest anti-theft stickers ever on it. Now I look like both a hipster clone and a corporate drone at the same time.) The magsafe cable has a very strong magnet compared to the Air’s and sticks out a bit more. And to add insult to injury, they’re not interchangeable, so we had to throw more money toward Cupertino.

I’d say the price is a downside (mine came out to about $2700 before Columbia discounts, including AppleCare). Even if I were buying this myself it’d be worth it, because I’ll average at least 20 hours/week use for two or more years. Additional upsides are 16GB of memory and four cores. With that, it runs the Stan C++ unit tests in under 3 minutes (it takes around 12 minutes on the Air and the Air starts buzzing like an angry fly). The HDMI port saves a dongle, but then the change to Thunderbolt meant buying another one. I don’t know that I’ll get much use of out of USB 3.0 (the iPad 3 is only USB 2.0). I also get 256GB of SSD, though I never filled the 128GB I had on the Air. The ethernet port and HDMI port are handy — two less dongles compared to the MacBook Air if you need either of these ports. I haven’t heard the fan. I’ve heard about it — it’s asymmetrical, which according to my signal processing geek friends, reduces the noise tremendously. It’s either super quiet or the machine’s so powerful the Stan unit tests don’t stress it out. ## grammar why ! matters July 11, 2012 For all those of you who might have read things like this, this, or this, I want to explain why the answer is “yes”, spelling and grammar matter. ### Language is a Tool Language is a tool used for many purposes. If your goal is to entertain, there are different conventions. Singers like Bob Dylan can be highly entertaining while remaining nearly incomprehensible. If your goal is to connect to friends or loved ones, yet other conventions come into play. Sometimes language is used for multiple things at once. ### Language is a Convention Language is a matter of convention. We simply cannot write or say whatever we want to however we want to and be understood. If your purpose is communication, it behooves you to make your message clear. There are exceptions to this, too. I might be trying to communicate how worldly I am by using French or Italian food terms or pronunciations instead of English, even knowing the audience won’t understand them. Communicating means using shared conventions. ### Word Order For instance, consider word order. Consider the following “understood be and to want we however to want we whatever say or write cannot simply we”. You’ve seen that sentence above, only in reverse. In reverse, it’s pretty much impossible to understand. Even in the CBS piece by Steve Tobak, the author mocks bad grammar with “me want food”. Well, that has a subject, verb and object, in perfect English order, which is why it’s so easy to understand. It even has the tense of “want” and the number and lack of determiner for “food” right. The only mistake is the object/subject distinction in “me” vs. “I”! ### Spelling? Tobak goes on to quote a comment, “I jus read your article; ___. Very interesting!” What’s wrong with bad spelling? It’s unpleasant because it slows us down as readers. If it gets bad enough, it can block understanding. I had no problem detangling the last example, but how about “I js rd y ar — int!!!!!!!”? Spelling used to be even more chaotic in English. It’s better in some other languages. ### Disclaimers I’m all for telegraphic speech. It works best in shared contexts. It’s a little harder with a bare Tweet. Language is incredibly tied up with context. Enough world knowledge can get you by, too. I might be able to refer to a TV show by “ST:TNG”, but my mom would have no idea what I was talking about. For some purposes, precision and clarity matter much less. Consider drafting legislation vs. planning to meet at a restaurant vs. saying hello. Telegraphic speech can be very precise. Doctors’ notes to each other are a prime example. You don’t need a verb if everyone knows there’s only one thing to do with a device or a noun if there’s only one device to use. Saying language is conventional and conventions should be followed is a subtly different stance from traditional linguistic prescriptivism. Languages change. If they didn’t, English wouldn’t even exist. I’m not railing against split infinitives, dangled prepositions, a complete failure to understand “who”/”whom” or even “I”/”me”, abandoning adverbial morphology, using “ain’t”, pronouncing “ask” like “axe”, etc. etc. I think these all have a good chance of achieving “proper” English status one day. ## Lucene Tutorial updated for Lucene 3.6 July 5, 2012 [Update: 8 Mar 2014. I’ve just written a quick introduction to Lucene 4: The contents of this introduction are excerpted from the Lucene chapter in my new book: This chapter covers search, indexing, and how to use Lucene for simple text classification tasks. A bonus feature is a quick reference guide to Lucene’s search query syntax.] The current Apache Lucene Java version is 3.6, released in April of 2012. We’ve updated the Lucene 3 tutorial and the accompanying source code to bring it in line with the current API so that it doesn’t use any deprecated methods and my, there are a lot of them. Bob blogged about this tutorial back in February 2011, shortly after Lucene Java rolled over to version 3.0. Like other 3.x minor releases, Lucene 3.6 introduces performance enhancements, bug fixes, new analyzers, and changes that bring the Lucene API in line with Solr. In addition, Lucene 3.6 anticipates Lucene 4, billed as “the next major backwards-incompatible release.” ### Significant changes since version 3.0 • IndexReader delete methods are deprecated and will be removed entirely in Lucene 4. All deletes and updates are done via an IndexWriter. • There is a single IndexWriter constructor that takes two arguments: the index directory and an IndexWriterConfig object. The latter was introduced in Lucene 3.1. It holds configuration information that was previously specified directly as additional arguments to the constructor. • IndexWriter optimize methods are deprecated. The merge method(s) supply this functionality. ### Building the Source The ant build file is in the file src/applucene/build.xml and should be run from that directory. The book’s distribution is organized this way so that each chapter’s demo code is roughly standalone, but they are able to share libs. There are some minor dependencies on LingPipe in the example (jar included), but those are just for I/O and could be easily removed or replicated. As an added bonus, the source code now includes the data used in the examples throughout the tutorial, the venerable Federalist Papers from Project Gutenberg. ## More on the Terminology of Averages, Means, and Expectations June 21, 2012 I missed Cosma Shalizi’s comment on my first post on averages versus means. Rather than write a blog-length reply, I’m pulling it out into its own little lexicographic excursion. To borrow stylistically from Cosma’s blog, I’ll warn you, the reader, that this post is more linguistics than statistics. ### Three Concepts and Terminology Presumably everyone’s clear on the distinctions among the three concepts, 1. [arithmetic] sample mean, 2. the mean of a distribution, and 3. the expectation of a random variable. The relations among these concepts is very rich, which is what I conjecture is causing their conflation. Let’s set aside the discussion of “average”, as it’s less precise terminologically. But even the very precision of the term “average” is debatable! The Random House Dictionary lists the meaning of “average” in statistics as the arithmetic mean. Wikipedia, on the other hand, agrees with Cosma, and lists a number of statistics of centrality (sample mean, sample median, etc.) as being candidates for the meaning of “average”. ### Distinguishing Means and Sample Means Getting to the references Cosma asked for, all of my intro textbooks (Ash, Feller, Degroot and Schervish, Larsen and Marx) distinguished sense (1) from senses (2) and (3). Even the Wikipedia entry for "mean" leads off with In statistics, mean has two related meanings: the arithmetic mean (and is distinguished from the geometric mean or harmonic mean). the expected value of a random variable, which is also called the population mean. Unfortunately, the invocation of the population mean here is problematic. Random variables aren’t intrinsically related to populations in any way (at least under the Bayesian conception of what can be modeled as random). Populations can be characterized by a set of (conditionally) independent and identically distributed (i.i.d.) random variables, each corresponding to a measureable quantity of a member of the population. And of course, averages of random variables are themselves random variables. This reminds me to the common typological mistake of talking about “sampling from a random variable” (follow the link for Google hits for the phrase). ### Population Means and Empirical Distributions The Wikipedia introduces a fourth concept, population mean, which is just the arithmetic mean of a given population. This is related to the point Cosma brought up in his comment that you can think of a sample mean as the mean of a distribution with the same distribution as the empirically observed distribution. For instance, if you observe three heads and a tail in three coin flips, you create a discrete random variable $X$ with $p_X(1) = 0.75$ and $p_X(0) = 0.25$, then the average number of heads is equal to the expectation of $X$ or the mean of $p_X$. ### Conflating Means and Expectations I was surprised that like the Wikipedia, almost all the sources I consulted explicitly conflated senses (2) and (3). Feller’s 1950 Introduction to Probability Theory and Applications, Vol 1 says the following on page 221. The terms mean, average, and mathematical expectation are synonymous. We also speak of the mean of a distribution instead of referring to a corresponding random variable. The second sentence is telling. Distributions have means independently of whether we’re talking about a random variable or not. If one forbids talk of distributions as first-class objects with their own existence free of random variables, one might argue that concepts (2) and (3) should always be conflated. ### Metonomy and Lexical Semantic Coercion I think the short story about what’s going on in conflating (2) and (3) is metonymy. For example, I can use “New York” to refer to the New York Yankees or the city government, but no one will understand you if you try to use “New York Yankees” to refer to the city or the government. I’m taking one aspect of the team, namely its location, and using that to refer to the whole. This can happen implicitly with other kinds of associations. I can talk about the “mean” of a random variable $X$ by implicitly invoking its probability function $p_X(x)$. I can also talk about the expectation of a distribution by implicitly invoking the appropriate random variable. Sometimes authors try to sidestep random variable notation by writing $\mathbb{E}_{\pi(x)}[x]$, which to my type-sensitive mind appears ill-formed; what they really mean to write is $\mathbb{E}[X]$ where $p_X(x) = \pi(x)$. I found it painfully difficult to learn statistics because of this sloppiness with respect to the types of things, especially among less careful authors. Bayesians, myself now included, often write $x$ for both a random variable and a bound variable; see Gelman et al.’s Bayesian Data Analysis, for a typical example of this style. Settling down into my linguistic armchair, I’ll conclude by noting that it seems to me more felicitous to say the expectation of a random variable is the mean of its distribution. than to say the mean of a random variable is the expectation of its distribution. ## 0/1 Loss Meaningless for Predicting Rare Events such as Exploding Manholes June 14, 2012 [Update: 19 June 2012: Becky just wrote me to clarify which tools they were using for what (quoted with permission, of course — thanks, Becky): … we aren’t using BART to rank structures, we use an independently learned ranked list to bin the structures before we apply BART. We use BART to do a treatment analysis where the y values represent whether there was an event, then we compute the role that the treatment variable plays in the prediction. Here’s a journal paper that describes our initial ranking method http://www.springerlink.com/content/3034h0j334211484/ and the pre-publication version http://www1.ccls.columbia.edu/%7Ebeck/pubs/ConedPaperRevision-v5.pdf The algorithm for doing the ranking was modified a few years ago, and now Cynthia is taking a new approach that uses survival analysis.] ### Rare Events Let’s suppose you’re building a model to predict rare events, like manhole explosions in the Con-Ed system in New York (this is the real case at hand — see below for more info). For a different example, consider modeling the probability of a driver getting into a traffic accident in the next week. The problem with both of these situations is that even with all the predictors in hand (last maintenance, number of cables, voltages, etc. in the Con-Ed case; driving record, miles driven, etc. in the driving case), the estimated probability for any given manhole exploding (any person getting into an accident next week) is less than 50%. ### The Problem with 0/1 Loss A typical approach in machine learning in general, and particularly in NLP, is to use 0/1 loss. This forces the system to make a simple yea/nay (aka 0/1) prediction for every manhole about whether it will explode in the next year or not. Then we compare those predictions to reality, assigning a loss of 1 if you predict the wrong outcome and 0 if you predict correctly, then summing these losses over all manholes. The way to minimize expected loss is to predict 1 if the probability estimate of failure is greater than 0.5 and 0 otherwise. If all of the probability estimates are below 0.5, all predictions are 0 (no explosion) for every manhole. Consequently, the loss is always the number of explosions. Unfortunately, this is the best you can do if your loss is 0/1 and you have to make 0/1 predictions. So we’ve minimized 0/1 loss and in so doing created a useless 0/1 classifier. ### A Hacked Threshold? There’s something fishy about a classifier that returns all 0 predictions. Maybe we can adjust the threshold for predicting explosions below 0.5. Equivalently, for 0/1 classification purposes, we could rescale the probability estimates. Sure, it gives us some predicted explosions, but the result is a non-optimal 0/1 classifier. The reason it’s non-optimal in 0/1-loss terms is that each prediction of an explosion is likely to be wrong, but in aggregate some of them will be right. ### It’s not a 0/1 Classification Problem The problem in 0/1 classification arises from converting estimates of explosion of less than 50% per manhole to 0/1 predictions minimizing expected loss. Suppose our probability estimates are close, at least in the sense that for any given manhole there’s only a very small chance it’ll explode no matter what its features are. Some manholes do explode and the all-0 predictions are wrong for every exploding manhole. What Con-Ed really cares about is finding the most at-risk properties in its network and supplying them maintenance (as well as understanding what the risk factors are). This is a very different problem. ### A Better Idea Take the probabilities seriously. If your model predicts a 10% chance of explosion for each of 100 manholes, you expect to see 10 explosions. You just don’t know which of the 100 manholes they’ll be. You can measure these marginal predictions (number of predicted explosions) to gauge how accurate your model’s probability estimates are. We’d really like a general evaluation that will measure how good our probability estimates are, not how good our 0/1 predictions are. Log loss does just that. Suppose you have $N$ outcomes $y_1,\ldots,y_N$ with corresoponding predictors (aka features), $x_1,\ldots,x_N$, and your model has parameter $\theta$. The log loss for parameter (point) estimate $\hat{\theta}$ is ${\mathcal L}(\hat{\theta}) - \sum_{n=1}^N \, \log \, p(y_n|\hat{\theta};x_n)$ That is, it’s the negative log probability (the negative turns gain into loss) of the actual outcomes given your model; the summation is called the log likelihood when viewed as a function of $\theta$, so log loss is really just the negative log likelihood. This is what you want to optimize if you don’t know anything else. And it’s exactly what most probabilistic estimators optimize for classifiers (e.g., logistic regression, BART [see below]). ### Decision Theory The right thing to do for the Con-Ed case is to break out some decision theory. We can assign weights to various prediction/outcome pairs (true positive, false positive, true negative, false negative), and then try to optimize weights. If there’s a huge penalty for a false negative (saying there won’t be an explosion when there is), then you are best served by acting on low-probability information, such as servicing even low-probability manholes. For example, if there is a$100 cost for a manhole blowing up and it costs \$1 to service a manhole so it doesn’t blow up, then even a 1% chance of blowing up is enough to send out the service team.

We haven’t changed the model’s probability estimates at all, just how we act on them.

In Bayesian decision theory, you choose actions to minimize expected loss conditioned on the data (i.e., optimize expected outcomes based on the posterior predictions of the model).

### Ranking-Based Evaluations

Suppose we sort the list of manholes in decreasing order of estimated probability of explosion. We can line this up with the actual outcomes. Good system performance is reflected in having the actual explosions ranked high on the list.

Information retrieval supplies a number of metrics for this kind of ranking. The thing I like to see for this kind of application is a precision-recall curve. I’m not a big fan of single-number evaluations like mean average precision, though precision-at-N makes sense in some cases, such as if Con-Ed had a fixed maintenance budget and wanted to know how many potentially exploding manholes it could service.

There’s a long description of these kind evaluations in

Just remember there’s noise in this received curves and that picking an optimal point on them is unlikely to produce such good behavior on held-out data.

With good probability estimates for the events you will get good rankings (there’s a ton of theory around this I’ve never studied).

### About the Exploding Manholes Project

I’ve been hanging out at Columbia’s Center for Computational Learning Systems (CCLS) talking to Becky Passonneau, Haimanti Dutta, Ashish Tomar, and crew about their Con-Ed project of predicting certain kinds of events like exploding manholes. They built a non-parametric regression model using Bayesian additive regression trees with a fair amount of data and many features as predictors.

I just wrote a blog post on Andrew Gelman’s blog that’s related to issues they were having with diagnosing convergence:

But the real problem is that all the predictions are below 0.5 for manholes exploding and the like. So simple 0/1 loss just fails. I thought the histograms of residuals looked fishy until it dawned on me that it actually makes sense for all the predictions to be below 0.5 in this situation.

### Moral of the Story

0/1 loss is not your real friend. Decision theory is.

This whole discussion reminds me of the lottery “paradox”. Each ticket holder is very unlikely to win a lottery, but one of them will win. The “paradox” arises from the inconsistency of the conjunction of beliefs that each person will lose and the belief that someone will win.

Oh, no! Henry Kyburg died in 2007. He was a great guy and decades ahead of his time. He was one of my department’s faculty review board members when I was at CMU. I have a paper in a book he edited from the 80s when we were both working on default logics.

## Computing Autocorrelations and Autocovariances with Fast Fourier Transforms (using Kiss FFT and Eigen)

June 8, 2012

[Update 8 August 2012: We found that for KissFFT if the size of the input is not a power of 2, 3, and 5, then things really slow down. So now we’re padding the input to the next power of 2.]

[Update 6 July 2012: It turns out there’s a slight correction needed to what I originally wrote. The correction is described on this page:

I’m fixing the presentation below to reflect the correction. The change is also reflected in the updated Stan code using Eigen, but not the updated Kiss FFT code.]

Suppose you need to compute all the sample autocorrelations for a sequence of observations

x = x[0],...,x[N-1]

The most efficient way to do this is with the discrete fast fourier transform (FFT) and its inverse; it’s ${\mathcal O}(N \log N)$ versus ${\mathcal O}(N^2)$ for the naive approach. That much I knew. I had both experience with Fourier transforms from my speech reco days (think spectograms) and an understanding of the basic functional analysis principles. I didn’t know how to code it given an FFT library. The web turned out to be not much help — the explanations I found were all over my head complex-analysis-wise and I couldn’t find simple code examples.

Matt Hoffman graciously volunteered to give me a tutorial and wrote an initial prototype. It turns out to be really really simple once you know which way ’round the parts all go.

### Autocorrelations via FFT

Conceptually, the input N-vector x is the time vector and the autocorrelations will be the frequency vector. Here’s the algorithm:

1. create a centered version of x by setting x_cent = x / mean(x);
2. pad x_cent at the end with entries of value 0 to get a new vector of length L = 2^ceil(log2(N));
3. run x_pad through a forward discrete fast fourier transform to get an L-vector z of complex values;
4. replace the entries in z with their norms (the norm of a complex number is the real number resulting of summing the squared real component and squared imaginary component).
5. run z through the inverse discrete FFT to produce an L-vector acov of (unadjusted) autocovariances;
6. trim acov to size N;
7. create a L-vector named mask consisting of N entries with value 1 followed by L-N entries with value 0;
8. compute the forward FFT of mask and put the result in the L-vector adj
9. to get adjusted autocovariance estimates, divide each entry acov[n] by norm(adj[n]), where norm is the complex norm defined above; and
10. to get autocorrelations, set acorr[n] = acov[n] / acov[0] (acov[0], the autocovariance at lag 0, is just the variance).

The autocorrelation and autocovariance N-vectors are returned as acorn and acov respectively.

It’s really fast to do all of them in practice, not just in theory.

Depending on the FFT function you use, you may need to normalize the output (see the code sample below for Stan). Run a test case and make sure that you get the right ratios of values out in the end, then you can figure out what the scaling needs to be.

### Eigen and Kiss FFT

For Stan, we started out with a direct implementation based on Kiss FFT.

• Stan’s original Kiss FFT-based source code (C/C++) [Warning: this function does not have the correction applied; see the current Stan code linked below for an example]

At Ben Goodrich’s suggestion, I reimplemented using the Eigen FFT C++ wrapper for Kiss FFT. Here’s what the Eigen-based version looks like:

As you can see from this contrast, nice C++ library design can make for very simple work on the front end.

Hat’s off to Matt for the tutorial, Thibauld Nion for the nice blog post on the mask-based correction, Kiss FFT for the C implementation, and Eigen for the C++ wrapper.