Decision Mechanics

Insight. Applied.

Twitter’s anomaly detection package for R

By

twitter time series from Christmas Eve 2014

Twitter have an interest in detecting anomalies in their service. Anomalies could be down to user engagement, spamming or technical issues. Regardless of the reasons, it’s something they want to know about when it happens.

To aid detection of anomalies in their time series data they have developed, and open-sourced, an anomaly detection package for R. Their algorithm is based on the Generalized ESD Test and can detect both global and local anomalies. The package is also capable of detecting anomalies when seasonal and trend factors are present in the time series data.

Visual p-hacking

By

It’s that time of year when we start getting the “best x of 2015″ posts. Nathan Yau of FlowingData just published his list of the best visualization projects. Yau reckons that this was the year of using visualization to teach about data and statistics.

My favorite is “Science Isn’t Broken” by Christie Aschwanden of FiveThirtyEight. It’s a visual interactive demonstration of how you can shape the results of a study through p-hacking (using US political parties and the economy as the example). The choices you make as to variables, data, methods, etc. have a significant impact on the outcomes of an analysis. Quantification and statistical significance are no guarantee of quality.

Subjectivity in data science

By

An article recently published in Nature reinforces the fact that the real challenge in data science is not mastery of the technical tools, but the ability to understand and define the problem.

Researchers posed the question of whether the color of a soccer player’s skin is a factor in how many red cards (serious reprimands) he receives. Seems like a pretty straightforward analysis.

The authors sent a data set to 29 research teams and asked them to answer the question. 20 research teams found there was a difference. Nine said there was no relationship between skin color and number of red cards received. Of the 20 who found a difference, two reported that dark-skinned players were less likely to receive red cards.

So, what explains this wide variation in results? Bias? Incompetence? No. It’s down to things like choice of what data is important, selection of analysis methods, etc.

If you have the resources, it seems like the best thing to do is crowd-source your answers. Have multiple researchers do the analysis, compare/contrast the results, share the insights across the teams, redo the analyses and then accept the majority answer. Of course, if you can only afford to employ one team, you need to be aware that data science isn’t an exact science…

Analyzing an Isle of Man TT legend

By

The BBC has an article on using a sensor array to determine what makes 23-time Isle of Man TT winner John McGuinness so quick.

Motorcycle riders have been tackling the 38-mile street circuit for over a hundred years. As it’s run on (closed) public roads, it’s an incredibly dangerous race. Riders average 212kph (132mpg) round the course—often coming within inches of stone walls and buildings.

John McGuinness is one of the most successful riders ever to tackle the course and data analysis company EMC decided to find out why he’s so quick. They fitted 50 sensors to him and the bike—collecting data on speed, acceleration, lean angle, throttle use, braking, body position, heart rate, etc.

One key finding is that he’s as cool as a cucumber, recording a heart rate of a mere 120 beats per minute when travelling at over 300kph on country roads. That lets him conserve energy and concentrate more than less skilled riders.

Applying machine learning techniques to the data revealed that only 14 variables were influential in his performance. McGuinness was just a little bit better in most areas than other riders, but each of these small advantages added up to a significant difference in lap times.