Retail security professionals are continually striving to fend off threats. They face a never-ending battle. A solution is implemented only to have offenders circumvent it within days or months. Appriss has joined The Retail Equation and Sysrepublic as one company to bring threat detection and prevention to a new level using the latest methods in artificial intelligence (AI), vast combined data, and expertise from the three organizations.
For years, each company individually pioneered advanced analytical methods and amassed data from both retailers and criminal justice. On any given day, The Retail Equation and Sysrepublic process between 20 and 30 million retail transactions with more than 40 billion transactions in total. Retailers track thousands of known threats using Sysrepublic’s case management solution. Appriss monitors more than 10 million incarcerations per year in more than 2,300 local jails, county jails, and departments of correction from 48 states, providing a rich source of known offenders for retail shoplifting and other types of related crimes. In addition to this data, there are billions of social media posts daily that provide clues to emerging threats.
To use this data, the data sciences team—composed of doctorate-level experts in computer science, machine learning, criminology, and advanced analytics—at the newly combined company is applying the latest advances in AI to bring threat detection and prevention to a new level. In the past ten years, the explosion of data and parallel computing has produced leaps forward in machine learning, deep learning, and other AI methods that are beginning to change the world. The Retail Equation and Sysrepublic are working to bring this technology to retail security.
Combining Big Data and Artificial Intelligence
By combining data sources, machine-learning algorithms look for patterns in the retail and social media data that are associated with known cases of shoplifting, employee fraud, organized retail crime, and so forth. This process forms the backbone of the next generation of return authorization, employee fraud detection, shrink forecasting, and a number of other retail security tools.
Traditionally, each new case of fraud or abuse leads a security team to study the patterns and to look for similar patterns in the data to catch similar cases. Those patterns are usually searched for using an exception-reporting query. This approach will catch cases, but it is not as accurate and timely as using methods based on machine learning, which can analyze thousands of factors simultaneously to root out the factors that are most accurately related to a threat.
With the number of new cases and fraud techniques emerging daily, it would be impossible for traditional methods to keep up. This is where AI methods kick in. While machines will not do our thinking for us anytime soon, there are many advances in this field that improve fraud-detection capabilities vastly. Following are a few of the techniques that provide an overview and glossary of this important field.
Machine learning is an advanced analytics tool that can automatically analyze hundreds or thousands of variables (known as features) about an employee or consumer to find hidden relationships between those features and an outcome or outcomes of interest, such as known cases of shoplifting or employee fraud. The algorithms use advanced mathematical structures to represent these complex relationships without input from the analyst. The Retail Equation and Sysrepublic are applying these algorithms to detect both retail return fraud as well as employee fraud. When the algorithms are fit (or learn) dynamically versus statically, it is referred to as online learning. An online learning routine will take new cases of fraud at some periodicity and learn a new relationship automatically. The “learning” refers to the mathematical process that allows the algorithm to most accurately represent the relationship between the variables.
Deep learning is a concept in the AI world that takes machine learning one step closer to true artificial intelligence. In the preceding machine-learning discussion, the machine-learning algorithm is presented with hundreds or thousands of variables. Typically, a human will have to encode these variables about the employee or consumer. The variables may be things like the number of transactions, the population density surrounding a store, or the number of hours an employee has worked in the past ten days. With deep learning, the algorithm will work off of raw pieces of information, such as transactions, and the algorithm will create the variables automatically in successive layers in what is referred to as a deep network. At each layer, more specific information is created. For example, the first layer may contain all raw transactions, the second layer may contain the average basket amount for an employee, and the third layer may contain the deviation of that average from the other employees in the store. The key to deep learning is that the algorithm figures out these variable constructs automatically by learning what is important for achieving a certain goal like detecting gift-card fraud.
Machine learning and deep learning are subsets of artificial intelligence. AI is a general area of research that encompasses a variety of applications where intelligent machines or algorithms are programmed and created to make automatic decisions or classifications when presented with data. AI can be broadly classified into general AI and narrow AI. General AI includes systems that more resemble a human and can think on their own, while narrow AI is designed to solve specific problems. Most experts agree that general AI is decades away. Narrow AI includes solutions like self-driving cars, facial-recognition technology, or automated fraud-detection systems. Some of the latest advances in this area include a computer that can compete on Jeopardy, play games like chess or Atari, and read medical journals and determine likely paths for treatment of a patient. AI can apply to retail security by (1) automatically learning new types of employee or consumer fraud and abuse, detecting, and preventing that emerging threat; (2) determining shifts in social media that are related to increases in shrink and reported in-store incidents; and (3) forecasting next year’s shrink in real time as new events in and around the store unfold.
One recent success in this area used social media data to predict store shrink. The Retail Equation analyzed a feed from Twitter that included between 20,000 and 100,000 tweets per day. More than 10 million tweets were collected in total. Tweets that occurred near a retail location were converted into metrics that determined tweet sentiment and the presence of key topics of interest using text analytics assigning the tweets mathematical values. The company was able to forecast shrink with similar accuracy found in similar multivariate shrink-prediction models used by retailers. Unlike crime reports, social media data is updated in real time and is far more broadly based on the voice of the shopper.
The Retail Equation and Sysrepublic actively use techniques from AI and are currently increasing the adoption of more advanced techniques from this new area to augment their existing methods. This introduces a new level of threat protection for retailers.