3 ms·
Dane here. Disclaimer: I can't say I really know much about this, but here is a bit of background info: STAR is acronym for Styrelsen for Arbejdsmarked og Rekr
by recoveringeng 4y ago
Dane here. Disclaimer: I can't say I really know much about this, but here is a bit of background info:
STAR is acronym for Styrelsen for Arbejdsmarked og Rekrutering (English: Agency for Labor Market and Recruitment).
They have a note describing the system (in Danish) at [1]. I copy/pasted a few relevant sections to Google Translate at [2].
TL;DR:
Algorithm is a decision tree based on questionnaires and outcomes of about 152000 cases.
The note states a precision of 70% and a recall of 32%.
There is no mention of a split between modelling and test data.
[1] https://star.dk/media/12514/2020_01_31_beskrivelse_-_profilafklaringsvaerktoej_til_dagpengemodtagere.pdf https://star.dk/media/12514/2020_01_31_beskrivelse_-_profila...
[2] https://translate.google.com/?sl=da&tl=en&text=Modellen%0APrincippet%20bag%20den%20statistiske%20profilafklaring%20er%2C%20at%20man%20ved%20at%20se%20på%20tidligere%20ledig%C2%AD%20hedsforløb%20kan%20få%20en%20ide%20om%2C%20hvem%20i%20en%20gruppe%20af%20nyledige%20dagpengemodtagere%2C%20der%20er%20i%20risiko%20for%20at%20blive%20langtidsledig.%20Den%20statistiske%20vurdering%20i%20profilafklarings%C2%AD%20værktøjet%20forsøger%20således%20at%20forudsige%20om%20den%20enkelte%20borger%20vil%20være%20ledig%20i%20mere%20end%206%20måneder.%20Dette%20gøres%20ved%20at%20sammenligne%20borgeren%20med%20andre%20nyledige%20dag%C2%AD%20pengemodtagere%2C%20der%20historisk%20set%20er%20blevet%20langtidsledige.%20Maskinlæringsalgorit%C2%AD%20men%20danner%20på%20denne%20baggrund%20kombinationer%20af%20faktorer%20der%2C%20i%20gennemsnit%2C%20bedst%20forudser%20om%20borgeren%20kan%20forventes%20at%20blive%20langtidsledig.%0A%0AMere%20præcist%20består%20modellen%20af%20et%20statistisk%20genereret%20beslutningstræ.%20Beslutnings%C2%AD%20træsalgoritmen%20kategoriserer%20i%20én%20af%20to%20grupper%3A%20risiko%20og%20ikke-risiko%20for%20langtidsle%C2%AD%20dighed.%20Algoritmen%20er%20dermed%20trænet%2C%20på%20data%2C%20til%20at%20identificere%20de%20faktorer%2C%20der%20er%20de%20vigtigste%20for%20at%20forudsige%20langtidsledighed.%20Kategoriseringen%20starter%20med%20den%20faktor%20der%20minimerer%20usikkerheden%20ved%20forudsigelsen%20mest.%20I%20kombination%20med%20denne%20faktor%20vælges%20så%20den%20næste%20faktor%2C%20på%20en%20måde%20hvor%20de%20to%20tilsammen%20minime%C2%AD%20rer%20usikkerheden%20mest%2C%20og%20så%20fremdeles.%20Faktorer%20og%20kombinationer%20af%20faktorer%2C%20som%20ikke%20væsentligt%20bidrager%20til%20at%20mindske%20usikkerheden%2C%20medtages%20ikke.%20Algoritmen%20fører%20til%20”grene”%20på%20beslutningstræet.%0A%0AData%0ADen%20statistiske%20profilafklaring%20er%20baseret%20på%20en%20analyse%20af%20påbegyndte%20dagpengefor%C2%AD%20løb%20i%20perioden%20maj%202016%20til%20juni%202018%2C%20hvor%20borgeren%20samtidig%20har%20udfyldt%20forbere%C2%AD%20delsesskemaet%20i%20forbindelse%20med%20forløbets%20start%20(ca.%20152.000%20forløb).%20Data%20består%20af%20udvalgte%20registerdata%20og%20besvarelserne%20på%20forberedelsesskemaet.%20Registerdata%20og%20spørgsmål%20i%20forberedelsesskema%20er%20udvalgt%20med%20udgangspunkt%20i%20den%20eksisterende%20viden%20om%20de%20mekanismer%2C%20der%20har%20betydning%20for%20ledighed.&op=translate https://translate.google.com/?sl=da&tl=en&text=Modellen%0APr...