AVA Predictive Modeling
AVA Predictive Modelling implements state-of-the-art machine learning methods at the AVA Bern
Factsheet
- Schools involved Business School
- Institute(s) Institute for Applied Data Science & Finance
- Research unit(s) Applied AI Research & Solutions
- Strategic thematic field Thematic field "Humane Digital Transformation"
- Funding organisation Others
- Duration (planned) 01.12.2024 - 30.09.2025
- Head of project Prof. Dr. Branka Hadji Misheva
- Project staff Julius Kooistra
- Keywords Public Employment Services, XAI, Machine Learning, Digital Transformation, Clustering, Predictive Analysis
Situation
The Office of Unemployment (AVA) Bern actively supports unemployed citizens with their reintegration into the labour market and financial benefits. It consists of the Public Unemployment Insurance Fund (ALK), which uses insurance benefits to bridge the financial distress caused by unemployment, and the Regional Employment Centres (RAV), which help customers find work with regular discussions and a wide range of offers to enable them to integrate quickly and permanently into the job market. This project explores the potential of using machine learning to enhance AVA’s services by identifying patterns and predicting unemployment duration using proprietary AVA data. By leveraging AI-driven insights, the goal is to support counsellors in decision-making, optimize resource allocation, and ultimately reduce the average duration of unemployment. A more efficient reintegration process benefits both unemployed individuals by restoring confidence and self-sufficiency, and AVA, by reducing financial strain on public funds. Through this collaboration, we aim to contribute to a more effective, data-driven approach to employment services, ensuring sustainable benefits for both individuals and society as a whole.
Course of action
To successfully complete this project, we have developed and integrated a comprehensive data pipeline that enables automated, standardized data transformation of new customer data. This pipeline consists of four key components: preprocessing, clustering, predicting, and explaining. It leverages state-of-the-art methodologies and technologies to optimize AVA’s decision-making processes. • Preprocessing: Cleans the dataset, generates new features, and aggregates redundant information through a structured mapping process. This ensures data consistency and full control over transformations. • Clustering: Groups customers based on mathematical similarity in the transformed dataset, allowing for targeted labour market measures tailored to specific customer profiles. • Predicting: Assesses the probability of long-term unemployment and flags customers as “at risk” if their probability exceeds a predefined threshold. This facilitates more effective resource allocation. • Explaining: Utilizes Shapley values to interpret model decisions, providing counsellors with clear insights into why a customer is classified a certain way. This enhances transparency and trust in AI-driven recommendations. By integrating this AI-powered pipeline, AVA can proactively support unemployed individuals, improve efficiency, and optimize the use of public resources, ensuring a more effective and sustainable approach to labour market reintegration.
Result
Through close collaboration with AVA Bern, we have developed detailed profiles for the identified customer clusters. Using these profiles, AVA has internally built a framework of tailored labour market measures, allowing for the assignment of tailored labour market measures to specific customer profiles. To improve predictive accuracy, we optimized the risk classification model, ensuring that fewer at-risk individuals are mistakenly classified as “not at risk.” This minimizes the likelihood of overlooking those who may require urgent support, ultimately contributing to reducing long-term unemployment. By combining clustering and predictive analysis, we identified distinct risk groups, with some customer segments showing significantly lower and others significantly higher proportions of long-term unemployment. These insights help AVA refine its approach to resource allocation and intervention strategies. Finally, we have identified the key drivers of long-term unemployment at individual, cluster, and population levels. At the individual level, these factors support personalized labour market interventions, while insights at the cluster and population levels contribute to the development of proactive, data-driven labour market policies.
Looking ahead
The established clustering and predictive models provide a solid foundation for more refined and proactive interventions, helping to reduce unemployment duration and optimize resource allocation. This project has led to a fruitful collaboration with AVA Bern, including the Map of Job Offers (MOJO) project linked below.
