Machine Learning (ML) for core insurance systems does not have to be a black box!

dr Ibrahim Halfaoui, FadataFadata

With regard to process optimization, insurance is no longer involved in the use of ML Technologies come round. The integration into core insurance systems will support critical business processes with a high degree of automation. From a cost and security perspective, two points are crucial: the ML process used and the transparency of the model.

from Dr. Ibrahim Halfaoui, Fadata

Dhe machine learning (ML) is gaining importance in almost all industries.

The possible areas of application are diverse: ML can automate subtasks and routine activities as well as complex interactions.”

The insurance industry is already increasingly using ML-based solutions, especially in marketing, sales and support, keywords being chatbots or conversational AI. For example, virtual agents accompany customers with the help of speech recognition during service requests and thus optimize the customer experience. Insurers can also more easily generate new leads based on automated analysis of customer behavior. Last but not least, insurers have the opportunity to analyze customer feedback from support calls using text and speech recognition and thus optimize their marketing strategy.

However, the use of ML by insurance companies will not be limited to sales or marketing issues, the core insurance systems themselves are increasingly coming into focus.”

This involves, for example, the processing of health insurance claims, the transformation of underwriting processes in life insurance or the automation of case processing in property insurance.

Claims management provides a good example of the concrete advantage that ML-based solutions offer. ML technologies support the automated analysis of damage photos or the automated extraction of information from unstructured data such as scanned files. Cognitive applications make it easier to detect fraud attempts and improve the accuracy of claims settlement. On the one hand, process automation and optimization serves to reduce the risk of fraud. On the other hand, it can also be used for the immediate settlement of minor claims. For insurers, the automatic processing of small claims – even in the case of fraud – is profitable because it avoids high process costs caused by human interaction.

Weak supervision is the future

When using ML in the insurance sector, however, two fundamental questions arise. Which procedure is actually used and how can security and transparency be guaranteed?

author dr Ibrahim Halfaoui, Fadata

Ibrahim Halfaoui is an expert, developer and researcher focused on the topic of “Artificial Intelligence” and has more than ten years of experience in the industry. He studied at the Technical University of Munich and did his doctorate on “Deep learning for visual scene understanding in autonomous driving”. During his time at HUAWEI, he led various AI projects in the automotive field with a focus on computer vision and deep learning. He has been with Fadata since 2021 (website) served as a Machine Learning Lead to build Fadata’s AI roadmap and help modernize the final solution.

In principle, a distinction must be made between ML technologies: Unsupervised, Supervised and Reinforcement Learning. In unsupervised learning, an algorithm determines patterns based on the data sets provided. Supervised learning, on the other hand, is based on labeled data sets. This means that target values ​​or categories are already assigned to the data. Finally, in reinforcement learning, an agent interacts with its environment and learns on the basis of rewards or punishments. The method is promising, but highly complex and can only be used to a limited extent in safety-critical areas, since reliable use of the learned rules cannot be guaranteed without additional measures.

Currently, ML frameworks for insurance are mainly based on the supervised learning method and the use of hundreds of thousands of real customer data sets that are representative and cover different use cases. The data is used for training and for simulations with continuous checking of the performance and result quality of the ML model. The disadvantages of this method are the high manual effort and the costs. Mainly the labeling is labor intensive and therefore costly.

A new development is emerging here. Weak supervision or self-supervised learning combine the advantages of unsupervised and supervised learning: high quality results and low costs. This method relies on various techniques to replace manually created labels with automatically generated ones; Examples are Knowledge Distillation, Contrastive Learning or Triplet Loss. Fadata will also follow this path in the future with its INSIS-ML framework, which is used to optimize and accelerate critical core processes of insurance companies.

ML transparency is doable

Even if ML-based procedures offer an insurer numerous advantages, unconditional, uncritical use is not a viable option – if only because of the BaFin specifications and the insurance supervisory requirements for IT (VAIT). They contain strict documentation and control obligations as well as transparency regulations for the data processed in the IT systems. The last VAIT circular states, among other things:

Significant changes to the IT systems as part of IT projects, their impact on the IT structure and IT process organization and the associated IT processes must be evaluated in advance as part of an impact analysis. In doing so, the company has to analyze in particular the effects of the planned changes on the control procedures and the control intensity.” (Source)

How does an ML deployment now correspond to the VAIT requirements? It is clear that an ML solution must not remain a black box. First, when using a commercial ML framework, an insurer needs to know the analysis method used. For transparency and traceability, the trained model and the pattern recognition should be explainable, for example with regard to the correlation of data and models. Equally important are detailed performance metrics of the model, which show, for example, the percentage of correct forecasts.

If possible, an insurance company should also be able to use an ML framework interactively, for example for case-by-case analysis, scenario modeling and simulations. For example, it can check how inputs can influence a predicted output, for example whether the address of an insured person influences the result when detecting fraud.

There is no question that ML will play an increasingly important role in the insurance industry. After all, ML offers far-reaching benefits.”

With the help of ML, offers can be optimized for customers, cases of fraud can be detected, claims accuracy can be improved and ultimately employee productivity can be increased. In addition, ML contributes to an increase in customer satisfaction through faster processing of applications and damage reports. Last but not least, process optimization can generally contribute to reducing costs.

However, ML models require continuous maintenance after commissioning. The result is a considerable effort for a company. This is where the MLOps approach comes into its own, encompassing management of the entire ML lifecycle, from software development to business metrics analysis and operations. Insurers should also consider the MLOps process model as it allows them to successfully deploy and securely monitor ML processes in production environments.dr Ibrahim Halfaoui, Fadata

You can find this article on the Internet at the website:

#Machine #Learning #core #insurance #systems #black #box

Leave a Comment

Your email address will not be published.