Analiza tveganja za samomor z uporabo globokih nevronskih mrež

Hudobivnik, Rok (author), Bratko, Ivan (mentor)

Cilj diplomske naloge je na podlagi bioloških podatkov o ljudeh, ki so storili samomor, oz. ljudeh, ki ga niso, naučiti nevronsko mrežo ločevati med tema dvema skupinama. S tem bi lahko v nadaljevanju potencialno razvili način vnaprejšnjega preprečevanja samomorov. Rezultat analize podatkov je pokazal, da nevronska mreža ločuje med skupinama veliko bolj natančno, kot slepo ugibanje za ta primer, s povprečno točnostjo 71,4 % in standardnim odklonom 2,33 %. Tekom pisanja diplomskega dela sem reševal predvsem dva problema, prvi izmed dveh je bil problem manjkajočih vrednosti, ki se je izkazal za glavnega krivca pri omejitvah klasifikacijske točnosti podatkov. Drugi problem je bil problem iskanja prave konfiguracije nevronske mreže, ki bi pri določenem vhodu vrnila najboljšo možno klasifikacijsko točnost. Rezultati in zaključki diplomskega dela se skladajo s predhodnimi analizami teh podatkov.

• Language: Slovenian
• Keywords: Globoko učenje, nevronske mreže, samomor.
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-95107
• Publication date in RUL: 14.09.2017

Secondary language

• Language: English
• Title: Suicide risk analysis using deep neural networks

Abstract

The goal of this thesis was to train a neural network to classify between two groups of people: those who have or have not committed suicide, based on the received biological data set. With the analysis of this data set further research could be performed with a goal of pre-emptive suicide prevention. In the experiments in this thesis, I achieved the average classification accuracy of 71,4 % and standard deviation of 2,33 %. The thesis deals with two distinct problems, first with the problem of missing values, that in the end proved to be the deciding factor for the limitations of the classification accuracy. Second, the problem of finding the optimal configuration of the neural network for this data set. The results and conclusions of this thesis are generally in agreement with other research done on this particular data set.

• Keywords: Deep learning, neural nets, suicide.