From Input to Insight: Detecting Tone Through Machine Learning
/
FROM THE RESOURCE GUIDE
Jeremy R. Clopton, CFE, CPA, ACDA, CIDA
Director, Big Data & Analytics, Digital Forensics, BKD, LLP and
H. Bryan Callahan, CFE, CPA/CFF, CVA
Director, Forensics & Valuation Services, BKD, LLP
Big data. Analytics. Machine learning. Artificial intelligence. These topics, and many others, are being used with regularity in all aspects of business — from marketing and operations to recruiting and retention. They should also be topics used regularly when discussing fraud examination techniques. According to the ACFE's 2016 Report to the Nations, proactive data monitoring and analysis were associated with the highest reduction in both median loss and median duration compared to all other anti-fraud controls.
In a recent case, analytics and machine learning were applied to the analysis of a variety of textual data sources. Much of what occurred in the scheme was off-book — not recorded in the company’s financial statements. Without transactional data to rely on, our examiners leveraged other data to use in the investigation and provided information to enhance the interview process.
A large company became aware of a potential theft scheme involving the IT director and some of his direct reports. The allegations were brought to the company’s attention by a whistleblower who had previously been terminated. The individuals involved in the scheme were taking old IT equipment that still had value and selling it on eBay. The user ID used to sell the equipment was in the company’s name, though it was not in the company’s control. Rather, the IT director linked his personal PayPal account to the “company” eBay account. All payments that came through the account deposited directly to his personal account, never remitting funds to the company.
Typically, transactional-based analytics would have been the starting point. However, without transactions to analyze, examiners turned to email and instant messages of the IT department personnel. The first approach — keyword searching — did not net much in the way of direct evidence.
The second approach — tone detection — identified a number of instant messages between the IT director and a supervisor which had a conspiratorial tone (other common tones in examinations include nervous, evasive, anxious and intimate). The topic of those communications was the eBay scheme.
In addition, tone detection also identified a couple of emails between the IT director and some female colleagues that may have been a little too “friendly” for normal professional relationships. While these were not used in this investigation, these types of results can be useful in lawsuits and investigations involving sexual harassment.
Armed with text messages, examiners interviewed the IT director who confessed to the scheme. Both machine learning — in this case specifically tone detection — and traditional analytics using keyword searching were used to successfully uncover the scheme at hand.
These topics and more are covered in the ACFE’s 2-day course, Using Data Analytics to Detect Fraud. Working through the data analysis process and assessing case studies from a data perspective, the course will help you:
- Focus on the analytics process to successfully apply analytics in your examinations.
- Learn the fundamental data analysis techniques and how to perform them in a variety of software solutions.
- Learn about advanced analytics techniques, including text analytics, visual analytics and predictive modeling.
- Strategize how to apply analytics in specific fraud schemes and develop a framework for
- that application.
Transactions, communications, technology and other assets continue to generate more data every day. The use of analytics, machine learning, artificial intelligence and other advanced analytics methods will help anti-fraud professionals evolve their methods to keep up with the complex occupational fraud landscape.
You can read more about this course and more events and seminars in our latest Resource Guide.