At Rise of AI 2019 in Berlin, Georg Wittenburg, co-founder and CEO of Inspirient, argues that AI is redefining which data science tasks are hard, by automating large parts of the discipline itself. He describes building, over about four years, 226 classifiers that automate steps from file-format detection through data cleaning, time series analysis and pattern or outlier detection, and says one such automatically generated result was presented to a company board for the first time in September of the prior year. He uses a Deutsche Bahn workshop and an insurance company case to show how automated results change the conversation between data scientists and business stakeholders, and uses the US EPA vehicle emissions dataset to demonstrate how automation surfaces findings analysts would not have thought to search for, illustrating the "streetlight effect" in data analysis. He closes with real client findings, from fraud detection to misidentified target audiences, and in the Q&A predicts the population of data scientists will keep growing even as automation shifts their work toward communicating results rather than producing them.
Key takeaways
- Inspirient built 226 classifiers over about four years to automate steps of the data science process, from file-format detection to pattern and outlier detection.
- An automatically generated data science result was presented to a company board for the first time in September of the year before the talk.
- Gartner estimated roughly 40 percent automation potential in data science by 2020, against data volumes growing about 40 percent per year.
- A feedback loop letting users mark results relevant or not lets the system learn what different audiences, operational staff versus board members, care about.
- Wittenburg predicts the population of data scientists will keep growing, but their work will shift from producing results to communicating and deep-diving on them.
Detailed notes
From a trivial classifier to 226 automated data science steps
Wittenburg opens with a mismatch he noticed: data science can classify handwriting, emotion in text and threats in real time, yet still stumbles on a trivial question such as whether a data file is delimited by tabs, semicolons or commas. He says he built a classifier for that problem about four years before the talk, found it easy, and kept going: classifiers to detect file structure, to determine what a dataset is about (a time series, a P&L item, a location), to decide which analyses apply, and to detect patterns. Over four years his team built 226 classifiers, which he says can automate a decent share of what data scientists normally do, including cleaning, slicing and dicing, time series analysis, and segmentation or outlier detection.
How automated results change the room
He describes a workshop about two and a half years earlier at Deutsche Bahn where data scientists and business stakeholders discussed automatically generated results together, and reports that the automation surfaced gaps in communication: business stakeholders learned things the data scientists had already known for two years but never mentioned, and vice versa. He says that in September of the prior year, an automatically generated result was presented to a company board for the first time. A similar case with an insurance company's analyst team is used to show that a feedback loop, letting users mark results as relevant or not, lets the system learn what different audiences, operational teams versus board members, care about.
The streetlight effect and the EPA emissions example
Wittenburg cites Gartner's estimate of about 40 percent automation potential in data science by 2020, against data volumes growing roughly 40 percent a year, and frames automation as lowering the cost-time curve for getting value from data. He tells a joke about a man searching for a dropped coin under a streetlight rather than where he lost it, naming this the streetlight effect: analysts under time pressure search where it is easiest, not comprehensively. He illustrates this with the US Environmental Protection Agency's vehicle emissions dataset, where an automated correlation analysis of NOx and CO2 surfaces an outlier, a manufacturer not identifiable with certainty from the talk, a finding an analyst would only find by chance without a systematic search.
What clients actually found in their data
He lists findings automated analysis surfaced for clients over three years: a fraud network draining accounts at a top-ten international bank, abnormal spending by external consultants at a government agency, mishandled business loans at a top-three German bank, damage at a German train station, underpaid employees, an IT cloud instance located in the wrong country, and, his favorite, an online business discovering its actual target audience was women, a fact its own investor was startled had never been checked.
Q&A highlights
Asked about industries served, Wittenburg says the approach is not industry-specific but sees most demand from finance, audit, compliance, factory-floor monitoring and, less predictably, logistics. Asked how a customer can trust the system's answers, he compares it to recognizing the right gift for a loved one without being able to define it in advance, and says the system surfaces candidate findings for the user to prioritize rather than testing a single hypothesis. Asked what is actually "AI" in the pipeline, he says some classifiers, like file-format detection, are pure business logic, while others, such as recognizing postcodes, currencies, locations or names within data, use more advanced AI techniques. Asked whether the number of data scientists will fall, he predicts the population will keep growing because of the sheer volume of data, but that their work will shift from tasks that can be automated toward communicating results with business counterparts and pursuing deeper analysis.
Quotes
It's not that they don't have any data, but the problem is to get value out of that data.
Georg Wittenburg
I think that the population of data scientists will actually be growing because there's so much data to analyze.
Georg Wittenburg
Questions answered in this talk
How does AI automate parts of the data science process?
According to Georg Wittenburg at Rise of AI 2019, his company Inspirient built 226 classifiers over four years that automate steps such as detecting file format and structure, cleaning data, running time series analysis, and detecting patterns, segments and outliers.
How much of data science can be automated by AI?
According to Georg Wittenburg at Rise of AI 2019, Gartner estimated an automation potential of about 40 percent in data science by 2020, while a roughly similar percentage growth in data volume each year keeps demand for analysis high.
Will AI reduce the number of data scientist jobs?
According to Georg Wittenburg at Rise of AI 2019, he expects the population of data scientists to keep growing due to the volume of data, but their work to shift away from tasks that can be automated toward communicating findings with business stakeholders and doing deeper analysis.