Why explainability is key to success in machine learning
Source – https://thepaypers.com/
Sean Nierat from PayPal has explained to The Paypers why there is a need of systems that not only make accurate predictions, but also that explain why they’ve arrived at a particular answer
Machine learning (ML) is part of a burgeoning AI industry that could soon become a multitrillion-dollar opportunity for global businesses. It is being used by PayPal, and others to help pioneer advanced data-driven fraud prevention by enhancing human intelligence with a 360-degree view of each customer. Yet, as ML becomes ubiquitous, it’s increasingly being argued that we not only need systems to make accurate predictions but also ones that explain why they’ve arrived at a particular answer.
Tackling bias with transparency
We’ve come a long way from the old days of fraud prevention. It’s undeniable that bad actors are getting smarter, with huge volumes of readily accessible customer data at their disposal and a wealth of tools bought on the dark web. Sophisticated fraud built on these foundations demands an equally sophisticated response. That’s why PayPal uses advanced ML to continually optimize the complex rules written by our client’s in-house fraud and data science teams, and to apply these rules to large datasets in order to spot patterns that humans may miss.
The problem with such systems is that they’re only as good as the data they’re trained on. Increasingly, organizations are concerned about unconscious bias emanating from this data, and the algorithms designed to interpret it. With ML used today in everything from mortgage application approvals to police facial recognition systems, there are important questions to answer – especially in a new era of intense regulatory scrutiny.
Clear box vs black box
This is where clear box ML or ‘explainable AI’ (XAI) approaches come into their own. Black box models like artificial neural networks (ANNs) or deep learning operate so that even the humans that designed them don’t know how decisions are made. However, with XAI, businesses gain vital insight into the whole process, from data collection to decision making.
This additional clarity and transparency offers multiple benefits including:
- improves business confidence in an XAI-powered prediction/ outcome;
- enhances the ability to control and manage algorithms in line with business objectives;
- increases accountability, as systems can be audited;
- improves regulatory compliance efforts;
- enables teams to identify new fraud patterns faster.
A new approach
PayPal’s enterprise Fraud Protection offerings champion clear box, advanced ML through our use of explainability methods like LIME, Shapley, and RL-LIM. Our prediction engine delivers an interpretability plot for every single event, helping to drive customer confidence in the results and continued ongoing improvements.
Our platform is purpose-built to handle both the complex fraud challenges businesses face today and to make the necessary adjustments to help address those of tomorrow. With PayPal, businesses can take a dynamic approach to fraud – streamlining the experience for good customers and adding protection layers when necessary.
PayPal leverages fraud and risk knowledge from its 2-Sided-Network of over 330 million customers and 25 million merchants transacting 12 billion times a year as well as integrated third-party feeds to enable the processing and correlation of vast amounts of heterogeneous data to help deliver actionable business intelligence.
- a purpose-built data lake stores structured and unstructured data from various sources;
- powerful Device Recon analyses hundreds of mobile and desktop device characteristics and behaviours, and applies machine learning models for risk scoring and clustering;
- easy-to-update rules and machine learning algorithms help businesses adapt to changing fraud schemes;
- robust link analysis and data visualization help enable businesses to proactively uncover anomalous patterns indicative of fraud;
- real-time complex authentication helps differentiate trusted from suspicious users.