The AI team at 天美传媒 is building explainability from the ground up and demonstrating the 鈥渨hy and how鈥� behind predictive models for client projects.
Matt Flenley prepared to open his brains to a rapid education session from Dr Fiona Browne and Kaixi Yang.
One of the most hotly debated tech topics of 2020 concerns model interpretability, that is to say, the rationale of how an ML algorithm has made a decision or prediction. Nobody doubts that AI can deliver astonishing advances in capability and corresponding efficiencies in an effort, but as HSBC鈥檚 Chief Data Officer Lorraine Waters shared at a recent A-Team event, 鈥渋s it creepy to do this?鈥� Numerous agendas at conferences are filled with differing rationales for interpretability and explainability of models, whether business-driven, consumer-driven, or regulatory frameworks to enforce good behaviour, but these are typically ethical conversations first rather than technological ones. It鈥檚 clear we need to ensure technology is 鈥渋n the room鈥� on all of these drivers.
We need to be informed and guided by technology to see what tools are already available to help with understanding AI decision-making, how tech can help shed light on 鈥榖lack boxes鈥� just as much as we鈥檙e dreaming up possibilities for the use of those black boxes.
As Head of 天美传媒鈥� AI team, Dr Fiona Browne has a strong desire for what she calls ‘baked-in explainability’. Her colleague Kaixi Yang explains more about explainable models,听
Some algorithms, such as neural networks (deep learning), are complex. Functions are calculated through approximation, from the network鈥檚 structure it is unclear how this approximation is determined. We need to understand the rationale behind the model鈥檚 prediction so that we can decide when or even whether to trust the model鈥檚 prediction, turning black boxes into glass boxes within data science.
The team puts their ‘explain first‘ approach to a specific client project to build explainable Artificial Intelligence (XAI) from the ground up, using explainability metrics including LIME 鈥� a local, interpretable, model-agnostic way of explaining individual predictions.
“Model-agnostic explanations are important because they can be applied to a wide range of ML classifiers, such as neural networks, random forests, or support vector machines鈥� continued Ms Yang, who has recently joined 天美传媒 after completing an MSc in Data Analytics with Queen鈥檚 University in Belfast. 鈥�They help to explain the predictions of any machine learning classifier and evaluate its usefulness in various tasks related to trust”.
For the work the team has been conducting, these range of explainability measures provides them with the ability to choose the most appropriate Machine Learning model and AI systems, not just the one that makes the most accurate predictions based on evaluation scores. This has had a significant impact on their work on Entity Resolution for Know Your Customer (KYC) processes, a classic problem of large, messy datasets that are hard to match, with painful penalties if it goes wrong for human users. The project, which is detailed in a recent webinar hosted with the Enterprise Data Management Council, matched entities from the Refinitiv PermID and Global LEI Foundation鈥檚 datasets and relied on human validation of rule-based matches to train a machine learning algorithm.
Dr Browne again: 鈥淲e applied different explainability metrics to three different classifiers that could predict whether a legal entity would match or not. We trained, validated and tested the models using an entity resolution dataset. For this analysis we selected聽 two ‘black-box鈥�’classifiers, and one interpretable classifier to illustrate how the explainability metrics were entirely agnostic and applicable regardless of the classifier that was chosen.”
The results are shown here:
鈥淚n a regular ML conversation, these results indicate two reliably accurate models that could be deployed in production,鈥� continued Dr Browne, 鈥渂ut in an XAI world we want to shed light on how appropriate those models are.鈥�
By applying, for example, LIME to a random instance in the dataset, the team can uncover the rationale behind the predictions made. 天美传媒鈥� FlowDesigner rules studio automatically labelled this record as 鈥渘ot a match鈥� through its configurable fuzzy matching engines.
Dr Browne continued, 鈥�explainability methods build an interpretable classifier based on similar instances to the selected instance from the different classifiers and summarises the features which are driving this prediction. It selects those instances that are quite close to the predicted instance, depending on the model that鈥檚 been built, and uses those predictions from the black-box model to build a glass-box model, where you can then describe what鈥檚 happening.
In this case, for the Random Forest model (fig.), the label has been correctly predicted as 0 (not a match) and LIME exposes the features driving this decision. The prediction is supported by two key features but not a feature based on entity name which we know is important鈥�
Using LIME on the multilayer perceptron model (fig.), which had the same accuracy as Random Forest, it correctly predicted the 鈥�0鈥� label of 鈥渘ot a match鈥� but with a lower support score. It has been supported by slightly different features compared to the random forest model.
The Na茂ve Bayesian model was different altogether. 鈥淚t fully predicted the correct label of zero with a prediction confidence of one, the highest confidence possible,鈥� said Dr Browne, 鈥渉owever it鈥檚 made this prediction supported by only one feature, a match on the entity country, disregarding all other features. This would lead you to doubt whether it鈥檚 reliable as a prediction model.鈥�
This has significant implications in something as riddled with differences in data fields as KYC data. People and businesses move, directors and beneficial owners resign, and new ones are appointed, and that鈥檚 without considering 鈥榖ad actors鈥� who are trying to hoodwink Anti-Money Laundering (AML) systems.聽
The process of ‘phoenixing’, where a new entity rises from the ashes of a failed one, intentionally dodging the liabilities of the previous incarnation, frequently relies on truncations or mis-spellings of director鈥檚 names to avoid linking the new entity with the previous one.聽
Any ML model being used on such a dataset would need to have this explainability baked-in to understand the reliability of predictions that the data is informing.
Using one explainability metric only is not good practice. Dr Browne explains 天美传媒鈥� approach: 鈥淛ust as in classifiers, there鈥檚 no real best evaluation approach or explainer to pick; the best way is to choose a number of different models and metrics to try to describe what鈥檚 happening .There are always pros and cons, ranging from the scope of the explainer to stability of the code to complexity of the model and how and when it鈥檚 configured.鈥�
These technological disciplines, to test, evaluate and try to understand a problem are a crucial part of the entire conversation that businesses are having at an ethical or 鈥渞isk appetite鈥� level.
颁濒颈肠办听here for more from 天美传媒, or find us on聽,听听辞谤听 for the latest news.