Lyons Davidson Solicitors and UWE Bristol are nearing completion of their 2nd Knowledge Transfer Partnership (KTP), focussing on emerging technology to develop an AI/Machine Learning system which will streamline their business.
The provision of legal services is in a period of rapid transition and the project is responding to a business challenge created by new legislation that will be critical in future proofing Lyons Davidson.
The company embraces innovation and technology, to deliver efficient and effective legal solutions with a client-centric approach, and their goal is to work toward an automated claims process putting client experience at its core.
The projects’ emphasis on fusing Machine Learning and other learning paradigms to create useful predictive models is complementary to the previous, successfully completed KTP, which embedded expertise in a different form of AI.
The project aims to create a new AI/Machine Learning system, which will predict likely settlement values and help to advise customers on compensation offers. The ability to automate settlement of claims will improve and shorten the decision-making process, giving Lyons Davidson a competitive edge.
UWE Bristol Academics working in this partnership are Jim Smith, Professor of Interactive Artificial Intelligence and Deputy Director of the Computer Science Research Centre, and Dr Mehmet Aydin, Senior Lecturer in Computer Science.
Professor Jim Smith tells us that “the project is addressing how to build and deploy explainable AI systems that can learn predictive models from historical records, such as likely settlement values for a range of insurance claims. A key part of the project is to make these systems interactive, so that company staff can query and recommend changes, so that the models can rapidly adapt in response to human expertise.”
The KTP office team at UWE caught up with Jan Klimaszewski, who is working on the project as associate and Software Engineer in Artificial Intelligence.
What type of AI / Machine Learning is being used within this Knowledge Transfer Partnership project?
We use a Pieciewiese linear regression tree (PL Tree). It is made up of two models a decision tree which provides a set of rules and linear regression) but is also a model in itself. It is a set of rules where the output is a mathematical equation. In general, every decision tree regressor is a set of rules with an output as a mean of some values, but for a PL tree the value is a mathematical equation.
We use active learning that includes human involvement, Case Handlers provide data over time and will be able to flag incorrect predictions as they arise. The model will then adapt a pattern, noticing that something has changed and will try to correct itself to make the right prediction.
How does this model work?
The architecture of the model includes input of parameters including injury types, accident types, general damages (tariff injuries, tariff and non-tariff costs) and special damages. It considers prognosis, length of injury and tariff type.
When a Case Handler uses the model, they input the case details using drop down menus and text boxes. They then click predict and the model will return a mathematical equation. It works out each section of the claim, so there could be a final calculation made up of several equations depending on the rule set.
How has AI / Machine Learning been developed within this project?
We analysed data using various statistical methods. After this we tested against various Machine Learning models; Random Forest and a simple decision tree. We used Machine Learning techniques such as cross validation to evaluate the model and to find the best possible parameters and performance. We used a technique called Gridsearch which allowed us to find the best possible set up and to see if the model had improved or not. These steps were repeated until we had a model that was not over or under fitting.
It is essential to have low error rates in the decision tree, so we used 4 different types of metrics to assess the model (mean absolute error, mean squared error, root mean squared error and r squared error).
We decided to use the PL tree because simple trees and Random forest were too simplistic and performing with low accuracy.
The team got to the point where we had everything, but the model was stopping us progressing further. Research was carried out on other Machine Learning models, one of which used genetic trees, but this was too complex for the solution. We decided to use a PL Tree which provides us with everything required.
The entire journey was about training, testing, comparing models and trying different approaches with data. The team are currently focussing on data preprocessing. Data preprocessing refers to the technique of preparing raw data to make it suitable for building and training Machine Learning models. – Right now we are in the testing phase
What will this mean for Lyons Davidson going forward?
This will make the the negotiation process more efficient. The system will provide guidance to the case handler and allow offers and counteroffers to be made quicker and more consistently. It will also provide a framework to provide more automation within the offer process and help reduce touchpoints on the claim for case handlers.
Is there anything else that you would like to tell us about the project?
The relationship between the business and the academic team has strengthened the project and made it possible to execute.
In summary, the first stage of this knowledge transfer partnership has established a software development framework, creating data tools to present the initial idea of the project. It was then ensured that data is pre-processed without concerns and Case Handlers can generate their own unique rules as needed. A generic re-useable ‘data-dictionary’ framework was created for storing meta-data that allows users to explore and edit.Ongoing knowledge exchange with case handlers will ensure that the data is utilised correctly, and that the input process is designed in the most efficient manner.
A functional model has been created, capable of predicting settlement value ranges. The overall aim is to accurately predict insurer behaviour and achieve the best result for clients in the quickest time.
