How Real-Time Stream Processing Makes Machine Learning More Meaningful

What do you get when you combine machine learning and real-time stream processing? The secret sauce of real-time machine learning … A machine learning powerhouse … The Holy Grail of machine learning … You get the idea.  

To illustrate the potential of machine learning combined with real-time stream processing, let’s look at fraud detection. This application hits close to home because most companies accept some form of digital payment and because of the far-reaching impact of fraud. Indeed, whether digital payments are made online or mobile, or in a physical retail space, the price of slow or reduced fraud detection accuracy is extremely high. For example, financial services companies absorb the penalty of false negative and false positive results through loss of fee revenue, compromised customer relationships, and losses due to fraud. Merchants, meanwhile, may lose sales and incur higher processing fees and fraud losses.

Just how long do these institutions have to detect fraud? Milliseconds. It is a small window of time to gather relevant historical and recent patterns of behavior from the limited amount of data in a transaction — a credit card number, an amount, a date — and to map that data to an existing customer and a merchant. With that context, a fraud detection algorithm calculates the probability of fraud, and the card processing provider decides whether that probability is high enough to grant a stop to the transaction.

But what if the process could be optimized? If the amount of time it takes to gather relevant information and run the fraud detection model could be reduced, additional checks could be applied in the same small window of time available between swiping a card and determining whether a transaction should be approved. 

And this is where Hazelcast enters the equation. With your real-time machine learning models running on the Hazelcast Stream Processing Platform, you can deliver fraud detection at scale thanks to unified data store and stream processing engine capabilities. Compared to building a pipeline piece-by-piece, the unified real-time data platform enables a sophisticated real-time fraud detection data pipeline that meets the most demanding throughput, latency, and performance requirements.

For example, you can store relevant cardholder and merchant data in a highly distributed, low-latency data store, reducing time wasted fetching data from siloed sources, and utilizing the stream processing capability to establish the latest card usage patterns. And because this is in a single platform versus multiple systems from multiple vendors, your ML models can run next to the customer and merchant data necessary to render a decision for the transaction. Thus, you’ve built a sophisticated, scalable, real-time fraud detection pipeline with all of these capabilities (and more) in a single runtime.

With the additional time afforded by optimizing fraud detection execution, the net result is more checks to catch fraud, resulting in better peace of mind for you, the merchant, and the customer — both of whom are still in the middle of a transaction. 

Traditionally, companies architecting state-of-the-art fraud detection pipelines must piece together several technologies. You need a prediction service that accepts fraud prediction requests (compute intensive), retrieves customer and merchant features (low-latency data store – memory intensive), and calculates recent card use patterns (stream processing – compute intensive). All of these are usually running in separate software and infrastructure stacks. You also need the fraud detection ML model (compute), built on Python (compute intensive), running on yet another software and infrastructure stack.

Inevitably, this architecture leads to a complex implementation of technologies and teams from multiple vendors. This fraud detection solution’s maintenance and ongoing operation must factor into long-term maintenance plans and costs. Moreover, piecing these technologies together is not trivial, regardless of whether you plan to run the application on-premise or via a cloud provider. There will invariably be gaps and integration issues among different products. Not to mention, the more hops necessary to communicate with different products and platforms will increase latency, the opposite of what you are going for with real-time fraud detection.

Hazelcast integrates stream processing and a fast data store into a single runtime; not only does this simplify the overall architecture, it lowers TCO by taking advantage of all underlying infrastructure (cores and memory) available to the Hazelcast cluster.

With Hazelcast, a real-time fraud detection pipeline leveraging ML typically consists of 5 stages: 

1. Ingest – pull in transactions from source
2. Enrich – contextualize incoming transaction (for example, customer, merchant, and recent behavior patterns)
3. Transform – features into model inputs
4. Predict – run machine learning model on Hazelcast (Python, ONNX, Pytorch, Tensorflow)
5. Act – apply additional emerging fraud checks, run model explainability, store predictions (and explanations) for future analysis, check model performance, and alert on model performance degradation to signal model re-training needs to data scientists

The applications for the kind of real-time machine learning enabled by Hazelcast are almost endless. Think of all the ways in which healthcare, telecommunications, e-commerce, manufacturing, transportation … you name it … could benefit from the seamless integration of your machine learning models with historical and real-time data that Hazelcast powers.   

For a detailed demonstration of Hazelcast, check out our on-demand talk about fraud detection.

You can also attend and watch our upcoming webinar on July 27 with Finextra on how management of real-time data can be used to combat fraudsters’ sophisticated attack vectors in a time-effective manner. Register today!