Improving Labour Inspection Efficiency via Machine Learning

Abstract

Labour inspections are carried out nationwide by governmental agencies in countries that have ratified the International Labour Organization’s Labour Inspection Convention (1947), to enforce decent working conditions and prevent injuries in workplaces. The inspections are conducted by individual inspectors, typically using checklists to survey inspected workplaces for non-compliance to health, environment, and safetyrelated regulations. Carrying out inspections efficiently is becoming increasingly more difficult, as workplaces are becoming more diversified and complex. This thesis therefore investigates the potential for improving the efficiency of labour inspections via machine learning (ML). Current research into this topic is very limited, so we first investigate what kind of data and ML methods that could be used to support labour inspection tasks. The investigation also involves assessing different baselineMLmethods for selectingworkplaces for inspection, and for selecting relevant (predefined) labour inspection checklists. We also assess various feature selection methods to maximize the performance of the baselines. Although the initial results are promising, we found that it was difficult to achieve good prediction accuracy even for the best-performing methods. We also propose ML methods for generating new checklists that could efficiently aid inspectors in identifying working environment violations. One of these methods can be used to generate dynamic checklists, which can be continuously adapted to any new information that surfaces during inspections. Our work also includes proposed explanation approaches to make the dynamic checklists more interpretable for inspectors. We then look further into how such ML-based checklists should be evaluated, by comparing the results from cross-validation performance estimates on existing data to the results from a field study where the checklists are tested in real-world labour inspections. The results of the comparison suggest that the cross-validation performance may not reflect the real-world field performance of the checklists. However, the results from the field study also show that ML-based dynamic checklists significantly increase the number of violations found in the inspections, improving inspection efficiency. The overall results from this Ph.D. suggest a great potential for using ML in labour inspection tasks. Therefore, our work could promote more independent research on the topic.