JBS Solutions values its partnerships with major cloud solution providers and their platforms— including Amazon. 90% of our engineers have extensive experience building and architecting AWS cloud-based solutions, many for more than 10 years. So, as we discuss opportunities where retailers can leverage machine learning (ML), we’ll highlight some of the AWS tools that make getting started even easier.
In a connected world, detecting fraudulent transactions is no longer a matter of requesting a customer’s ID at the checkout. Nor do the fraud detection systems that card-issuing banks have in place always catch everything. Cleaning and reconciling fraudulent transactions is time-consuming and expensive, which is why retailers need a solution of their own.
Traditional fraud detection mechanisms are usually rules-based. They are effective to a degree, but they present retailers (and their customers) with a few serious issues.
- False positives: Often, rules will block genuine transactions by mistake
- Fixed outcomes: Thresholds for rules can change over time, which can invalidate them.
- Maintenance: Rules (and the rule engine) can be challenging to manage and scale.
For ML, fraud detection is a “classification” problem, similar to spam detectors or loan default prediction. It can classify a transaction as legitimate or fraudulent based on amount, merchant, location, time, and other factors. ML can complement rules-based fraud detection to provide a more holistic fraud detection strategy.
- Good first Ml solution for retailers
- Speed and scale
- Less expensive than more traditional rules-based systems
- More accurate/effective
In a nutshell, buyer segmentation divides customers into smaller groups based on shared characteristics, so the organization can better tailor and target its marketing efforts to the appropriate segments.
Traditional segmentation uses demographics and other characteristics like age, gender, location, prior product purchases, and marital status. The results are useful, but this approach presents retailers with several problems:
- categories are too broad/fixed: Not every 40-year-old likes the same thing.
- Easily outdated: People move, they marry (or divorce), and their tastes change.
- Not dynamic: They can’t take in-the-moment data (e.g., clickstream data) into account to maximize current relevance.
The key to optimal buyer segmentation is to leverage customer behavioral data in addition to demographics and other characteristics. This data might include RMF (Recency, Frequency, and Monetary) data about a customer’s purchases, as well as real-time data both online (clickstream) or in the store (video and location). ML algorithms can divide buyers into groups that otherwise would be incredibly difficult or impossible to find via traditional means.
- Good first Ml solution for retailers
- Model is less complex than typical supervised learning models
- Automatic and dynamic behavioral connections between customers (clusters)
- Connections/groupings are always up-to-date, even in-the-moment
Also, with tools like AWS SageMaker, retailers get a wide variety of ML and deep learning algorithms that they can use for buyer segmentation.
- Fully managed service to build, train, and deploy machine learning models
- Out-of-the-box tools for creating and managing training datasets
- Automatic hyper-parameter tuning
- Automatic model monitoring to determine effectiveness
A good recommendation solution can make finding relevant products and content more efficient, save customers time, and result in more sales. According to a McKinsey study, 35% of Amazon sales come from ML-driven product recommendations. An ML-driven content recommendations account for 80% of what Netflix customers watch. The better the recommendations, the more (and faster) customers will buy.
Product recommendation solutions use a class of algorithms and techniques that suggest the relevancy of items to an individual customer. The algorithms determine relevance based primarily on historical data, such as items searched for, browsed, selected, or bought. The presentation layer can then use relevancy ranking to determine what product suggestions to display. The goal is to make it more likely for the shopper to take the next step—whether it is to watch another show or add the item to their cart.
There are two broad categories of algorithms for ML product recommendation solutions:
- collaborative filtering models are based on past interactions between customers and target items. These algorithms work on the idea that historical data alone should be enough to make recommendations.
- content-based systems use historical data, but they may include other detailed information about the customer (location, demographics, behaviors) and your product catalog (geographic availability, product metadata, rankings, and return rates). The idea is that the better you know your customers and your products, the better recommendation you can make. Tastes change.
Each of these has its benefits and drawbacks. The former requires fewer data points about a customer, but its rankings lack the quality and personalization of content-based solutions. The latter, of course, requires a much more sophisticated data pipeline.
While implementing a product recommendation solution is probably not the first ML project you should tackle, the benefits of a quality solution are apparent. (Amazon and Netflix—not to mention their customers— would agree.)
Inventory/supply chain optimization
Inventory and supply chain management has never been a simple problem, as it involves so many moving parts in a retail organization. Retailers have gotten pretty good at forecasting, but without a real supply chain optimization solution in place, it is an educated guessing game. An ML solution can take a lot of the guesswork off a retailer’s plate.
While a comprehensive inventory and supply chain optimization solution is one of the more complex ML initiatives a retailer can begin, it can also be one of the most rewarding. A McKinsey study claims ML-enabled supply chain management can reduce forecasting errors up to 50%, help scale back inventory up to 50%, and reduce lost sales by up to 65%.
ML solutions for retail inventory and supply chain can analyze and leverage customer behavioral data to:
- More accurately predict near-future demand for optimal inventory levels
- Better understand how price changes impact sales
- Minimize ideal stock and optimize inventory storage
- Predict the effects of cannibalization, halo effect, and deferred demand due to promotion of a particular item
- Incorporate more factors than just sales data to deliver much more accurate results
- Use third-party data in their ML solution (e.g., using weather data to forecast when you will need more products for family outings and cookouts)
- Leverage the power of complex algorithms, including
- Random forests
- Deep learning (neural nets)
Despite the tremendous benefits, however, inventory and supply chain optimization is not the best candidate for a retailer just starting with ML— unless it addresses only a small component of the overall system.
Other retail applications for ML
The use cases above are only four of many possible applications of ML to optimize retail sales, marketing, and operations. Others include:
- Pricing optimizations
- Optimized search results
- Predictions and decisions
- Image recognition and computer vision
For even more great information, download on whitepaper Getting Started with Machine Learning In Retail.