Featured Product
This Week in Quality Digest Live
Statistics Features
Donald J. Wheeler
Tracking the process while making different products
Harish Jose
How to generate an OC curve based on sample size and number of rejects
Donald J. Wheeler
Do you know what really happens in phase two?
Jody Muelaner
Keeping an eye on the big picture
Bill Snyder
How we measure innovation,and how innovation drives productivity and affects inequality

More Features

Statistics News
Ability to subscribe with single-user minimum, floating license, and no long-term commitment
A guide for practitioners and managers
Gain visibility into real-time quality data to improve manufacturing process efficiency, quality, and profits
Tool for nonstatisticians automatically generates models that glean insights from complex data sets
Version 3.1 increases flexibility and ease of use with expanded data formatting features
Provides accurate visual representations of the plan-do-study-act cycle
SQCpack and GAGEpack offer a comprehensive approach to improving product quality and consistency
Ask questions, exchange ideas and best practices, share product tips, discuss challenges in quality improvement initiatives
Strategic investment positions EtQ to accelerate innovation efforts and growth strategy

More News

Rob Matheson

Statistics

Democratizing Data Science

Tool for nonstatisticians automatically generates models that glean insights from complex data sets

Published: Wednesday, February 13, 2019 - 12:00

Democratizing data science is the notion that anyone, with little to no expertise, can do data science if provided ample data and user-friendly analytics tools. MIT researchers are hoping to support that idea with a new tool for nonstatisticians that automatically generates models for analyzing raw data. The tool ingests data sets and generates sophisticated statistical models typically used by experts to analyze, interpret, and predict underlying patterns in data.

The tool currently lives on Jupyter Notebook, an open-source web framework that allows users to run programs interactively in their browsers. Users need only write a few lines of code to uncover insights into, for instance, financial trends, air travel, voting patterns, the spread of disease, and other trends.

In a paper presented at this week’s ACM SIGPLAN Symposium on Principles of Programming Languages, the researchers show their tool can accurately extract patterns and make predictions from real-world data sets, and even outperform manually constructed models in certain data-analytics tasks.

“The high-level goal is making data science accessible to people who are not experts in statistics,” says first author Feras Saad, a Ph.D. student in the Department of Electrical Engineering and Computer Science (EECS). “People have a lot of data sets that are sitting around, and our goal is to build systems that let people automatically get models they can use to ask questions about that data.”

Ultimately, the tool addresses a bottleneck in the data science field, says co-author Vikash Mansinghka, a researcher in the Department of Brain and Cognitive Sciences (BCS) who runs the Probabilistic Computing Project. “There is a widely recognized shortage of people who understand how to model data well,” he says. “This is a problem in governments, the nonprofit sector, and places where people can’t afford data scientists.”

The paper’s other co-authors are Marco Cusumano-Towner, an EECS PhD student; Ulrich Schaechtle, a BCS postdoc with the Probabilistic Computing Project; and Martin Rinard, an EECS professor and researcher in the Computer Science and Artificial Intelligence Laboratory.

Bayesian modeling

The work uses Bayesian modeling, a statistics method that continuously updates the probability of a variable as more information about that variable becomes available. For instance, statistician and writer Nate Silver uses Bayesian-based models for his popular website FiveThirtyEight. Leading up to a presidential election, the site’s models make an initial prediction that one of the candidates will win, based on various polls and other economic and demographic data. This prediction is the variable. On Election Day, the model uses that information, and weighs incoming votes and other data, to continuously update that probability of a candidate’s potential of winning.

More generally, Bayesian models can be used to “forecast”—predict an unknown value in the dataset—and to uncover patterns in data and relationships between variables. In their work, the researchers focused on two types of data sets: time-series, a sequence of data points in chronological order; and tabular data, where each row represents an entity of interest and each column represents an attribute.

Time-series data sets can be used to predict, say, airline traffic in the coming months or years. A probabilistic model crunches scores of historical traffic data and produces a time-series chart with future traffic patterns plotted along the line. The model may also uncover periodic fluctuations correlated with other variables, such as time of year.

On the other hand, a tabular dataset used for, say, sociological research, may contain hundreds to millions of rows, each representing an individual person, with variables characterizing occupation, salary, home location, and answers to survey questions. Probabilistic models could be used to fill in missing variables, such as predicting someone’s salary based on occupation and location, or to identify variables that inform one another, such as finding that a person’s age and occupation are predictive of their salary.

Statisticians view Bayesian modeling as a gold standard for constructing models from data. But Bayesian modeling is notoriously time-consuming and challenging. Statisticians first take an educated guess at the necessary model structure and parameters, relying on their general knowledge of the problem and the data. Using a statistical programming environment, such as R, a statistician then builds models, fits parameters, checks results, and repeats the process until they strike an appropriate performance tradeoff that weighs the model’s complexity and model quality.

The researchers’ tool automates a key part of this process. “We’re giving a software system a job you’d have a junior statistician or data scientist do,” Mansinghka says. “The software can answer questions automatically from the data—forecasting predictions or telling you what the structure is—and it can do so rigorously, reporting quantitative measures of uncertainty. This level of automation and rigor is important if we’re trying to make data science more accessible.”

Bayesian synthesis

With the new approach, users write a line of code detailing the raw data’s location. The tool loads that data and creates multiple probabilistic programs that each represent a Bayesian model of the data. All these automatically generated models are written in domain-specific probabilistic programming languages—coding languages developed for specific applications—that are optimized for representing Bayesian models for a specific type of data.

The tool works using a modified version of a technique called “program synthesis,” which automatically creates computer programs given data and a language to work within. The technique is basically computer programming in reverse: Given a set of input-output examples, program synthesis works its way backward, filling in the blanks to construct an algorithm that produces the example outputs based on the example inputs.

The approach is different from ordinary program synthesis in two ways. First, the tool synthesizes probabilistic programs that represent Bayesian models for data, whereas traditional methods produce programs that do not model data at all. Second, the tool synthesizes multiple programs simultaneously, while traditional methods produce only one at a time. Users can pick and choose which models best fit their application.

“When the system makes a model, it spits out a piece of code written in one of these domain-specific probabilistic programming languages... that people can understand and interpret,” Mansinghka says. “For example, users can check if a time series dataset like airline traffic volume has seasonal variation just by reading the code—unlike with black-box machine learning and statistics methods, where users have to trust a model’s predictions but can’t read it to understand its structure.”

Probabilistic programming is an emerging field at the intersection of programming languages, artificial intelligence, and statistics. This year, MIT hosted the first International Conference on Probabilistic Programming, which had more than 200 attendees, including leading industry players in probabilistic programming such as Microsoft, Uber, and Google.

“My team at Google AI builds probabilistic programming tools on top of TensorFlow. Probabilistic programming is an important area for Google, and time series modeling is a promising application area, with many use cases at Google and for Google’s users,” says Ryan M. Rifkin, a Google researcher who was not involved in the research. The researchers’ paper “shows how to apply probabilistic programming to solve this important problem—and reduces the effort needed to get started, by showing how the probabilistic programs can be synthesized from data, rather than written by people.”

Discuss

About The Author

Rob Matheson’s picture

Rob Matheson

Rob Matheson is a writer, editor, and web producer at MIT News, as well as a contributing writer at MIT Technology Review of the Massachusetts Institute of Technology in Cambridge, Massachusetts.