Hilfswerk Österreich is one of Austria’s leading charities. Their
mission is to provide support in overcoming health, family, or social
challenges. They provide services all over Austria, ranging from care
for the elderly, child care, to social living and supermarkets.
Within the framework of their 24-hour care services for persons with physical disabilities, Hilfswerk Österreich provide a matching service between patients and live-in care providers. These matchings require intense care and administrative effort; accordingly, early cancellations caused by unsuitable combinations are undesirable. The data4good Hackathon project will examine historical cases, in order to reveal conflicts in ongoing relationships.
A team of very talented data scientists have decided to help solving this challenge. Their goal during the hackathon is to examine historical cases, in order to get insights on which data features are associated with cancellations. Another goal is interpretability of the results: we love algorithms and models, but we also aim at providing some valuable clues to Hilfswerk.
The data come from Hilfswerk's 24-Stunden-Betreuung - Visitenblatt (visits) and Telefonprotokoll (phone calls). The records from the pdf forms resulted in two dataframes containing heterogeneous information about the cases. Each case belongs to one of the two classes (normal or conflict) and has a few to several records in time order. The records include information on service activities that have been done (categorical and text type) as well as conversations with the caregivers and patients.
In other words, the scientists are dealing with a binomial classification problem and aim at predicting contract cancellation based on saved records of normal and conflict cases. Futher challenges:
The team calculated sentiments present in each individual case report. As expected, conflict cases have a lower (unhappier) sentiment than non-conflict cases. Their analysis resulted in several insights:
Calculated based on SentiWS lexicon - a publicly available German-language resource for sentiment analysis.
Calculated based on the tf-idf analysis of the normal and conflict training cases via Quanteda