Neural Networks (NNs) and Natural Language Processing (NLP) Help Retailers Battle Organized Retail Crime
In 2021, the National Coalition of Law Enforcement and Retailers (C.L.E.A.R.) estimated that organized retail crime resulted in $94 billion in annual losses for retailers, a significant increase from the FBI’s estimate of $30 billion a decade ago. Organized retail crime (ORC) refers to professional shoplifting, cargo theft, retail crime rings and other organized crime occurring in retail environments. These criminals move from store to store and even state to state. Working in teams, ORC groups steal everything from infant formula to tablets. Often, they are stocking up on specified items at the request of an ORC ring leader. Retail Loss Prevention teams are still using traditional investigative techniques to identify, analyze, and link these groups. Advances in Neural Networks and Natural Language Processing can drastically improve the productivity and effectiveness of these teams.
In this article we will discuss:
● Organized Retail Crime Overview
● The Challenge of Building Complex Investigations from Unstructured Data
● How Natural Language Processing (NLP) & Neural Networks (NNs) Exponentially Supercharge ORC Investigations
● How Detective Analytics Leverages NLP & NNs to Battle Organized Retail Crime
Organized Retail Crime Overview
The ORC Crime Cycle
Organized retail crime is much more than mothers shoplifting baby formula in a time of rising inflation. The typical cycle of organized retail crime includes:
Scope of Organized Retail Crime
ORC Spans Multiple Locations and Jurisdictions
ORC crimes span multiple locations and legal jurisdictions. Below is a partial map of one linked crime group of 27 incidents involving three suspects who stole over $7 million in merchandise over a two-year period:
Map of a Single Multi-State ORC Investigation within DetectiveAnalytics
Such geographic diversity makes it very difficult for regional investigators to connect-the-dots. Also, ORC groups like this just don’t target one retail chain, they will target multiple chains in each location.
The Challenge of Building Complex Investigations from Unstructured Incident Data
Inconsistent Incident Reporting
Most retailers use some type of case management or incident reporting system to capture information about retail crimes like shoplifting, burglary, and refund fraud. A major challenge is the inconsistencies in the incident reports that store-level personnel record. Associates may have limited or no training in the best practices to record retail crimes.
The value from basic analytics and machine learning tools fall apart as soon as data becomes inconsistent and unstructured. Most intelligence tools can only offer analytics and ML if the data is in the same exact format with no inconsistencies. But this forced format consistency that allows for basic analytics and ML, comes with a price: Sacrificing the freedom of the associate to report unique aspects of an incident, and omitting all incident data which does not completely conform to the forced format, from all analysis.
The current lack of analytic support due to inconsistent incident data makes it very hard for regional or national loss prevention analysts to identify and analyze ORC groups. Understanding these patterns is critical to the ability to detect, build investigations, and prosecute. The sparsity of collected data poses a major challenge for loss prevention analysts to link crimes together to a specific ORC group.
Difficulty in Harnessing the Treasure-Trove of Information in Incident Narratives, at Scale
Most LP Incident reporting and retail crime intelligence systems require store-associates to report descriptive features into a set of predetermined multiple-choice fields, in order to try and get any insight at all out of the incident reports. Otherwise, the current rudimentary methods of Machine Learning that are being applied for today’s LP teams, just won’t work. This completely misses on, and fails to unlock, the raw power of the incident narrative.
Typically, the narrative field of incident reports contains amazing amounts of detail. Furthermore, they can describe very nuanced and unique aspects of each particular criminal incident that an array of designated fields just can’t replicate. Current LP investigation systems and methods fail at harnessing this information in a scalable manner. It is no surprise that investigation teams must therefore fall-back to relying on the human minds of investigators to both read over and send incident reports and bolo alerts to their colleagues, just to try to make connections manually.
Too Much Data for Investigators to Deal with
Most retail chains have tens-of-thousands of shoplifting incidents per year. To expect an investigator to memorize even 1,000 detailed incident reports, let alone 10,000+ is unrealistic. Furthermore, it is even more unrealistic to expect an investigator to memorize all incidents, and then as they read new incidents, to “rolodex” through their memory to recall similarities between other incidents. Is it worth having investigators manually read through 10,000+ incident reports every year, looking for “needles in the haystack” that can connect-the-dots on investigations?
Is it reasonable, for example, to expect a regional investigator in the NYC area to start reading over thousands of incident reports from the southeastern US just to hope to find a few connections across a large ORC group? Multi-state ORC crews prey on this glaring weakness of current manual investigation methods.
Manually Linking Incidents to Form Investigations
Excel and Word Docs are still the most common tools loss prevention teams use to build investigations and piece them together into ORC groups. Like the detectives of the Dick Tracy era, loss prevention teams manually review incident reports recorded in case management systems and use their intuition to identify potentially linked incidents. The information is recorded into spreadsheets that can be shared with other team members and law enforcement.
A typical loss prevention investigator is responsible for covering dozens or even hundreds of stores in a state or region. A typical medium to large retailer can have 2,000 to 8,000 incidents reported every month. Manually reviewing each incident and seeing if it is possibly linked to past incidents is almost an insurmountable task.
How Natural Language Processing (NLP) & Neural Networks (NNs) Exponentially Supercharge ORC Investigations
As we have seen, Loss Prevention teams face a herculean task in identifying, analyzing, and prosecuting ORC groups. Figuring out how to identify how individual incidents from dozens or hundreds of stores are linked to an ORC group is almost an impossible task.While an LP analyst’s intuition helps, it can go only so far.
The complexity of crime-linking is beyond mind-boggling. A typical national retailer could record 4,000 incidents in a single month - 48,000 in one year. The mathematics of crime-linking shows that there are a possible 1,151,976,000 (1billion+) connections between these 48,000 incidents ((n2-2)/2). While not all of these incidents are linked to ORC groups, a non-trivial number are. AI technologies like Natural Language Processing and Neural Networks can help retailers conquer these complexities.
How Neural Networks and Natural Language Processing Can Analyze Loss Prevention Incidents
The raw material Loss Prevention analysts start with are incident reports filed by store personnel. These are usually recorded in case management or incident management systems. These reports usually include some specific fields (incident number, date, time, SKUs of stolen items, etc.) and a general narrative text field that contains a description of the incident, like the following:
“Today, a male suspect came in. He was very suspicious, and did something odd where he went up to an associate and talked about how his day was going, before putting items in his jacket and leaving. This is very similar to another incident I had heard about at the nearby store.“
Notice above, that this incident report narrative describes a peculiar activity that would not be captured in designated multi-choice fields. (The unique way this shoplifter likes to strike up a friendly conversation with an associate before stealing.) But this is precisely where NLP and NNs can shine.
Natural Language Processing (NLP) is a branch of Artificial Intelligence (AI) that enables machines to understand the human language. Its goal is to build systems that can make sense of text and automatically perform tasks like translation, spell check, or topic classification. NLP can extract and standardize entities from case management records. NLP can recognize many things from incident reports, including: suspects, vehicles, MOs, peculiar activities, family members working together, patterns in time and location strikes, among many others.
NLP can further help understand the extracted entities. For example, one incident report could contain the name Robert Smith and another Bob Smith. NLP algorithms could conclude that the two names may most likely reference the same person. It could also determine that Robert Smith is an employee and not a suspect.
The use of NLP makes it possible to extract, with a high degree of accuracy, relevant facts from incident reports. Combined with Neural Networks, makes investigations, at-scale, a breeze.
How NLP AND NNs Can Detect How Incidents Are Linked to ORC Groups
Machine learning is the process of applying algorithms that teach machines how to automatically learn and improve from experience without being explicitly programmed. NLP & NNs are specific types of Machine Learning that are perfect for automating loss prevention investigations at scale. For ORC incidents, algorithms can be applied to the extracted facts to determine whether or not two incidents are linked and the work of an organized crime group. Using computational power, it is possible to determine which of the millions of possible linkages among a retailer’s incident reports are the work of the same crime group. What would take an analyst weeks to accomplish manually can be done by computer in a matter of minutes, with a significantly higher degree of accuracy.
How Detective Analytics Leverages NLP & NNs to Battle Organized Retail Crime
Detective Analytics has built a robust Artificial Intelligence Powered solution to help retailers battle organized retail crime. DA has four core functions:
Ingest Incident Data
Detective Analytics can accept outputs from any incident/case management system. DA connects to the data and then automatically translates it into a common data format, which is used for further analysis. DA offers direct integration with many popular incident management systems. If a direct integration is not available, standard reports can be exported and loaded into the DA database in a variety of methods, including STFP.
Analyze & Link Incidents
Next, DA applies its natural language processing and machine learning algorithms to extract facts from the incident data and create linked groups. It is important to note that this happens automatically every time an incident is added to the database. What would normally take weeks for a loss prevention analyst to do manually using tools like Excel happens in mere minutes. This time savings can be reinvested into more value-added activities for the LP team.
Visualizations
Detective Analytics provides dozens of ways to visualize incidents and investigations. Some sample dashboards include:
Investigation Link Analysis Charts and Maps
Incident Heatmaps
Incident Time Charts
Drill Down Reports
Users can drill down into the details of each item in the visualizations:
Pareto Charts
Value Added Functions
In addition to the visualizations, Detective Analytics provides a number of value-added functions including:
Linked Group Investigation Law Enforcement Reports
Detective Analytics generates comprehensive reports that summarize incidents, suspects, license plates, etc for law enforcement. These are live clickable reports that authorized law enforcement personnel can access to support their investigations and prosecutions.
Predictive Policing
Detective Analytics provides predictions about where a linked organized crime group will strike next using historical data and predictive analytics.
Labor Optimization
Using the predictive policing and other analyses, DA offers recommendations on how LP labor and guards can be deployed to mitigate anticipated ORC activity.
Appeasement Fraud
All of the tools used to analyze ORC incidents can also be used to battle appeasement fraud. Specific accounts can be flagged after the fact to be shut down to prevent future frauds. DA can even be integrated into your claims payment system to deny fraudulent activity in real time.
Recap
Organized Retail Crime has reached epidemic levels. In 2021 ORC cost U.S. retailers an estimated $94 billion in losses. Retailers face several challenges in battling organized retail crime groups, with the analysis of internal crime incident reports being one of the largest. Artificial Intelligence technologies like natural language processing and machine learning are potential solutions to these challenges. Detective Analytics offers a unique artificial intelligence solution for retail loss prevention teams that is proven to work at scale in most any environment.
