In deployed machine learning pipelines, “drift” is changes in the model environment that cause the model performance to degrade over time.  Drift might result from data quality changes.  For example, increasing amounts of missing values in the input data.  Or a company might alter the…

A common problem in machine learning is the “rare case” situation. In many classification problems, the class of interest (fraud, purchase by a web visitor, death of a patient) is rare enough that a data sample may not have enough instances to generate useful predictions.…

Epidemiological terms are top of mind now, due to the pandemic. Here are two that often confuse: incidence and prevalence. For example, I encountered the following sentence on a popular medical web site: “Knee meniscal injuries are common with an incidence of 61 cases per…

An often-overlooked basic part of learning new things is vocabulary: if you don’t fully understand the meaning of terms, you are handicapped. Worse, if you think you do understand, but that understanding is wrong, you are deprived of the ability to identify the gap in…

The classic Gambler’s Ruin puzzle has an actuarial parallel:  “Ruin Theory,” the calculations that govern what an insurance company should charge in premiums to reduce the probability of “ruin” for a given insurance line.  “Ruin” means encountering claims that exhaust initial reserves plus accumulated premiums. …

In Natural Language Processing (our course on the subject starts Jan 15), entity extraction is the process of labeling chunks of text as entities (e.g. people or organizations).  Consider this phrase from the blog on close elections linked above:   “the tie was not between Jefferson…

Type I error in statistical analysis is incorrectly rejecting the null hypothesis - being fooled by random chance into thinking something interesting is happening.  The arcane machinery of statistical inference - significance testing and confidence intervals - was erected to avoid Type I error.  Type II error…

The Relative Risk Ratio and Odds Ratio are both used to measure the medical effect of a treatment or variable to which people are exposed. The effect could be beneficial (from a therapy) or harmful (from a hazard).  Risk is the number of those having…

In a randomized experiment, the endpoint or outcome is a formal measure (statistic) of the result of the experiment.  In a randomized clinical trial preparatory to regulatory submission, there is often more than one outcome, due to the time and expense involved in conducting a…

In generalized linear models, a link function maps a nonlinear relationship to a linear one so that a linear model can be fit (and then mapped to the original form).  For example, in logistic regression, we want to find the probability of success:  P(Y =…

Polytomous, applied to variables (usually outcome variables), means multi-category (i.e. more than two categories).  Synonym:  multinomial. 

Parameterized code in computer programs (or visualizations or spreadsheets) is code where the arguments being operated on are defined once as a parameter, at the beginning, so they do not have to be repeatedly explicitly defined in the body of the code.  This allows for…

A decision stump is a decision tree with just one decision, leading to two or more leaves. For example, in this decision stump a borrower score of 0.475 or greater leads to a classification of “loan will default” while a borrower score less than 0.475…

In biostatistics, hazard, or the hazard rate, is the instantaneous rate of an event (death, failure…).  It is the probability of the event occurring in a (vanishingly) small period of time, divided by the amount of time (mathematically it is the limit of this quantity…

In building statistical and machine learning models, regularization is the addition of penalty terms to predictor coefficients to discourage complex models that would otherwise overfit the data.  An example is ridge regression.

In a standard linear regression, a model is fit to a set of data (the training data); the same linear model applies to all the data.  In local regression methods, multiple models are fit to different neighborhoods of the data. A kernel function is used…

In Monte Carlo sampling for simulation problems, random values are generated from a probability distribution deemed appropriate for a given scenario (uniform, poisson, exponential, etc.).  In simple random sampling, each potential random value within the probability distribution has an equal value of being selected. Just…

The art of statistics and data science lies, in part, in taking a real-world problem and converting it into a well-defined quantitative problem amenable to useful solution. At the technical end of things lies regularization. In data science this involves various methods of simplifying models,…

All students of statistics encounter confidence intervals.  Confidence intervals tell you, roughly, the interval within which you can be, say, 95% confident that the true value of some sample statistic lies.  This is not the precise technical definition, but it is how people use the…

Density is a metric that describes how well-connected a network is

Consider the multi-arm bandit problem where each arm has an unknown probability of paying either 0 or 1, and a specified payoff discount factor of x (i.e. for two successive payoffs, the second is valued at x% of the first, where x < 100%).  The Gittens index is [...]

There are various ways to recommend additional products to an online purchaser, and the most effective ones rely on prior purchase or rating history -

Autoregressive refers to time series forecasting models (AR models) in which the independent variables (predictors) are prior values of the time series itself.

To a statistician, a sample is a collection of observations (cases).  To a machine learner, it’s a single observation.  Modern data science has its origin in several different fields, which leads to potentially confusing  synonyms, like these:

When variables have binary (yes/no) values, a couple of issues come up when measuring distance or similarity between records.  One of them is the "yacht owner" problem.

Rectangular data are the staple of statistical and machine learning models.  Rectangular data are multivariate cross-sectional data (i.e. not time-series or repeated measure) in which each column is a variable (feature), and each row is a case or record.

A dummy variable is a binary (0/1) variable created to indicate whether a case belongs to a particular category.  Typically a dummy variable will be derived from a multi-category variable. For example, an insurance policy might be residential, commercial or automotive, and there would be three dummy variables created:

Curbstoning, to an established auto dealer, is the practice of unlicensed car dealers selling cars from streetside, where the cars may be parked along the curb.  With a pretense of being an individual selling a car on his or her own, and with no fixed…

QUESTION:  The rate of residential insurance fraud is 10% (one out of ten claims is fraudulent).  A consultant has proposed a machine learning system to review claims and classify them as fraud or no-fraud.  The system is 90% effective in detecting the fraudulent claims, but only 80% effective in correctly classifying the non-fraud claims (it mistakenly labels one in five as "fraud").  If the system classifies a claim as fraudulent, what is the probability that it really is fraudulent?

The Receiver Operating Characteristics (ROC) curve is a measure of how well a statistical or machine learning model (or a medical diagnostic procedure) can distinguish between two classes, say 1’s and 0’s.  For example, fraudulent insurance claims (1’s) and non-fraudulent ones (0’s). It plots two quantities:

A prospective study is one that identifies a scientific (usually medical) problem to be studied, specifies a study design protocol (e.g. what you’re measuring, who you’re measuring, how many subjects, etc.), and then gathers data in the future in accordance with the design. The definition…

It is 100 years since R A Fischer introduced the concept of "variance" (in his 1918 paper The Correlation Between Relatives on the Supposition of Mendelian Inheritance).

"Bag" refers to "bootstrap aggregating," repeatedly drawing of bootstrap samples from a dataset and aggregating the results of statistical models applied to the bootstrap samples. (A bootstrap sample is a resample drawn with replacement.)

I used the term in my message about bagging and several people asked for a review of the bootstrap. Put simply, to bootstrap a dataset is to draw a resample from the data, randomly and with replacement.

The field of data science is rife with terminology anomalies, arising from the fact that the field comes from multiple disciplines.

Benford's law describes an expected distribution of the first digit in many naturally-occurring datasets.

Contingency tables are tables of counts of events or things, cross-tabulated by row and column.

The easiest way to think of a spline is to first think of linear regression - a single linear relationship between an outcome variable and various predictor variables. 

To some, NLP = natural language processing, a form of text analytics arising from the field of computational linguistics.

As applied to statistical models - "overfit" means the model is too accurate, and fitting noise, not signal. For example, the complex polynomial curve in the figure fits the data with no error, but you would not want to rely on it to predict accurately for new data:

In statistics, "n" denotes the size of a dataset, typically a sample, in terms of the number of observations or records.

A corpus is a body of documents to be used in a text mining task.  Some corpuses are standard public collections of documents that are commonly used to benchmark and tune new text mining algorithms.  More typically, the corpus is a body of documents for…

Causal modeling is aimed at advancing reasonable hypotheses about underlying causal relationships between the dependent and independent variables. Consider for example a simple linear model: y = a0 + a1 x1 + a2 x2 + e where y is the dependent variable, x1 and x2…

In an experiment, an arm is a treatment protocol - for example, drug A, or placebo.   In medical trials, an arm corresponds to a patient group receiving a specified therapy.  The term is also relevant for bandit algorithms for web testing, where an arm consists…

A sparse matrix typically refers to a very large matrix of variables (features) and records (cases) in which most cells are empty or 0-valued.  An example might be a binary matrix used to power web searches - columns representing search terms and rows representing searches,…

We continue our effort to shed light on potentially confusing usage of terms in the different data science communities. In statistics, a sample is a collection of observations or records.  It is often, but not always, randomly drawn.  In matrix form, the rows are records…

With this entry, we inaugurate a new effort to shed light on potentially confusing usage of terms in the different data science communities. In statistics and machine learning, normalization of variables means to subtract the mean and divide by the standard deviation.  When there are…

HDFS is the Hadoop Distributed File System.  It is designed to accommodate parallel processing on clusters of commodity hardware, and to be fault tolerant.

The Kruskal-Wallis test is a nonparametric test for finding if three or more independent samples come from populations having the same distribution. It is a nonparametric version of ANOVA.

A "discovery" is a hypothesis test that yields a statistically significant result. The false discovery rate is the proportion of discoveries that are, in reality, not significant (a Type-I error). The true false discovery rate is not known, since the true state of nature is not known (if it were, there would be no need for statistical inference).

The 2006 Netflix Contest has come to convey the idea of crowdsourced predictive modeling, in which a dataset and a prediction challenge are made publicly available.  Individuals and teams then compete to develop the best performing model.

This week's word is actually a letter.  R is a statistical computing and programming language and program, a derivative of the commercial S-PLUS program, which, in turn, was an offshoot of S from Bell Labs.

In time series forecasting, a moving average is a smoothing method in which the forecast for time t is the average value for the w periods ending with time t-1.

In regression models, an interaction term captures the joint effect of two variables that is not captured in the modeling of the two terms individually.

A naive forecast or prediction is one that is extremely simple and does not rely on a statistical model (or can be expressed as a very basic form of a model).

In discrete response models, overdispersion occurs when there is more correlation in the data than is allowed by the assumptions that the model makes.

In a classification model, the confusion matrix shows the counts of correct and erroneous classifications.  In a binary classification problem, the matrix consists of 4 cells.

The predictors in a predictive model are sometimes given different terms by different disciplines.  Traditional statisticians think in terms of variables.

Structured data is data that is in a form that can be used to develop statistical or machine learning models (typically a matrix where rows are records and columns are variables or features).

Censoring in time-series data occurs when some event causes subjects to cease producing data for reasons beyond the control of the investigator, or for reasons external to the issue being studied.

Predictive modeling is the process of using a statistical or machine learning model to predict the value of a target variable (e.g. default or no-default) on the basis of a series of predictor variables (e.g. income, house value, outstanding debt, etc.).

Goodness-of-fit measures the difference between an observed frequency distribution and a theoretical probability distribution which

An adjacency matrix describes the relationships in a network. Nodes are listed in the top..

In a test of significance (also called a hypothesis test), Type I error is the error of rejecting the null hypothesis when it is true -- of saying an effect or event is statistically significant when it is not.

Data partitioning in data mining is the division of the whole data available into two or three non-overlapping sets: the training set (used to fit the model), the validation set (used to compared models), and the test set (used to predict performance on new data).

Longitudinal data records multiple observations over time for a set of individuals or units. A typical..

Cross-sectional data refer to observations of many different individuals (subjects, objects) at a given time, each observation belonging to a different individual.  A simple...

CHAID stands for Chi-squared Automatic Interaction Detector. It is a method for building classification trees and regression trees from a training sample comprising already-classified objects.

Also called the training sample, training set, calibration sample.  The context is predictive modeling (also called supervised data mining) -  where you have data with multiple predictor variables and a single known outcome or target variable.

The centroid is a measure of center in multi-dimensional space.

Promoting better understanding of statistics throughout the world.

New Editor of Journal of Statistics Education

Read Peter's Letter to the Editor in Saturday's Washington Post.

Last year's popular story out of the Predictive Analytics World conference series was Andrew Pole's presentation of Target's methodology for predicting which customers were pregnant.

Facebook began trading around 11:30 this morning, and I spent 8 minutes