Analiza časovnih in prostorskih podatkov pri osebnih zavarovanjih

MOŽEK, DAMIR (author), Kononenko, Igor (mentor)

V tem magistrskem delu predstavljamo razvoj metodologije za analizo zavarovalniških podatkov. Ker so zavarovalniški podatki predstavljeni časovno in prostorsko, je za razvoj metodologije potreben poseben pristop. V ta namen uporabimo prijeme prostorsko-časovnega podatkovnega rudarjenja, ki nam omogočajo ustrezno obravnavo časovnih in prostorskih atributov. Pri analizi podatkov se omejimo na podatke osebnih zavarovanj. V obravnavo zajamemo podatke o prijavah škodnih dogodkov na področju nezgod in bolezni. Zavarovalniške podatke navežemo na podatke o vremenskih razmerah v dnevu prijave odškodninskega zahtevka. S to navezavo želimo izkoristiti tezo o vplivu vremena na pojav nezgod. Nadejamo se, da bomo na ta način lažje napovedovali število odškodninskih zahtevkov in povprečno višino izplačanega odškodninskega zahtevka. Najprej se lotimo reševanja klasifikacijskega problema. Uporabimo nekaj osnovnih klasifikacijskih algoritmov, vendar se stopnja uspešnosti napovedovanja pri vseh algoritmih izkaže za izredno nizko. Ker so problemi po naravi regresijski, se preizkusimo še v reševanju regresijskega problema. Tudi regresijski algoritmi ne dajo dosti boljših rezultatov. Preverimo ustreznost časovnega okna učne množice. Dobimo potrditev, da je časovno okno glede na podatke izbrano ustrezno. Nadalje preverimo, če instance obravnavajo ustrezno časovno skalo. Pridemo do sklepa, da je časovna skala izbrana ustrezno. Poskusimo lokalizirati problem tako, da podatke razdelimo. Pri različnih primerih razreza Slovenije pridemo do ugotovitve, da je ocena napovedi za vsako izmed lokalnih območij slabša od napovedi za celotno Slovenijo. Postavljena metodologija se izkaže za delno uporabno. Uporabimo jo lahko za napovedovanje števila nezgod. S pomočjo diagnostike dobimo potrditev, da je za neuspeh kriva majhnost množice obravnavanih dogodkov. Podamo predloge glede možnosti izboljšav. Nadejamo se, da se uporabnost postavljene metodologije pokaže v prihodnosti. Za potrebe magistrskega dela pripravimo še zemljevide za spremljanje nezgod in bolezni v Sloveniji skozi čas. Na podlagi zemljevidov ugotavljamo trende gibanja za prihodnost. Identificiramo najpomembnejše dejavnike, odgovorne za nastanek nezgod. Na koncu naredimo še analizo vpliva nadmorske višine na pojav nezgod.

• Language: Slovenian
• Keywords: prostorsko-časovno podatkovno rudarjenje, strojno učenje, osebna zavarovanja, nezgode, vreme
• Work type: Master's thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-85519
• Publication date in RUL: 15.09.2016

Secondary language

• Language: English
• Title: Analysing of temporal and spatial data in life insurance

Abstract

In this master’s thesis we present the development of the methodology for the analysis of insurance data. Due to the fact that insurance data are presented temporally and spatially, a special approach is necessary for the development of the methodology. For this purpose we use approaches of spatio-temporal data mining, which enable us the appropriate treatment of temporal and spatial attributes. When analysing data we limit ourselves on the data of personal insurance. Into the treatment we capture data on reports of the loss events in the field of accidents and diseases. Insurance data are linked to the data on weather conditions on the day of the report of the claim of compensation. By this linkage we wish to use the thesis on the influence of the weather on the accidents. We hope that in this way we shall predict the number of the claims of compensation and the average amount of the disbursed claims of compensation more easily. Firstly, we deal with the solving of the classification problem. We use some of the basic classification algorithms, but the level of successfulness of predicting in case of all algorithms proves to be extremely low. Due to the fact that the problems are by nature regression, we try to solve the regression problem too. Even regression algorithms do not offer much better results. We check the adequacy of training set time window. We get the confirmation that the time windows is selected appropriately with respect to the data. Furthermore, we check if the instances deal with the appropriate time scale. We come to the conclusion that the time scale is selected appropriately. We try to localize the problem: we divide the data. In different cases of the cut of Slovenia we come to the ascertainment that the estimate of the prediction for each of the local areas is worse than for the entire Slovenia. The set methodology proves to be partially useful. It can be used for predicting the number of accidents. By means of diagnostics we receive the confirmation that the failure is due to the smallness of the multitude of treated events. We give proposals regarding the possibility of improvements. We hope that the usefulness of the set methodology will become evident in the future. For the needs of the master’s thesis we also prepare the maps for following the accidents and diseases in Slovenia through time. On the basis of the maps we ascertain the trends for the future. We identify the most important factors, responsible for the emergence of the accidents. At the end we perform the analysis of the influence of the altitude for the emergence of the accidents.

• Keywords: spatio-temporal data mining, machine learning, personal insurance, accidents, weather