Kako uspeti na Kickstarter-ju?

Novak, Benjamin (author), Zupan, Blaž (mentor)

Platforme za množično financiranje, kot je Kickstarter, v zadnjih letih postajajo vedno bolj priljubljene. Na njih poskušajo razvojne ekipe s kreativnimi projekti pridobiti bodoče kupce in podpornike. Vendar uspeh ni zagotovljen, saj je skoraj dve tretjini predlogov neuspešnih. V nalogi smo iz portala za množično financiranje Kickstarter pridobili podatke o opisu in uspešnosti projektov. Naš cilj je bil zgraditi model, ki bi iz opisa projekta znal napovedati uspešnost kampanje in bi skladno z rezultati v sorodnih delih dosegel točnost napovedi AUC vsaj 0,85. V delu predstavimo našo rešitev in tehnike strojnega učenja, ki smo jih uporabili. Zgrajene modele smo vrednotili s prečnim preverjanjem in na novih projektih. Ugotovili smo, da so pri napovedovanju uspešnosti najpomembnejši število projektov, ki jih je avtor podprl, ciljna vsota, število slik v opisu projekta in število ponujenih nagrad. Na testnih podatkih novih projektov smo dosegli točnost AUC = 0,93.

• Language: Slovenian
• Keywords: Kickstarter, množično financiranje, napovedovanje uspešnosti, strojno učenje, klasifikacija, iskanje značilk
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-84441
• Publication date in RUL: 23.08.2016

Secondary language

• Language: English
• Title: How to succeed on the Kickstarter?

Abstract

Crowdfunding platforms such as Kickstarter are becoming increasingly popular. These platforms are widely used by development teams which are trying to get new buyers and supporters using different creative projects. However, success is not guaranteed since two thirds of the project suggestions fail to achieve their goal. In our thesis, we gathered descriptions and success of different projects on Kickstarter. Our goal was to create a model that could predict success of project compaigns. With this model, we also wanted to reach prediction accuracy AUC = 0,85 that could be compared with the results of other related studies. In the thesis, we present our solution and techniques of machine learning that were used to gather data. These models were later assessed with cross validation and new projects. The results showed that the most important attributes are the number of the projects supported by the author, the goal, the number of pictures in the description of the project and the award number. AUC score accomplished on the test data of the new projects was 0,93.

• Keywords: Kickstarter, crowdfunding, machine learning, classification, feature engineering