launch hard or go home! 0.005predicting the success of kickstarter projects · 2019-12-14 ·...

1
Launch Hard or Go Home! Predicting the Success of Kickstarter Projects Vincent Etter, Matthias Grossglauser and Patrick Thiran 1. What is Kickstarter? I Crowdfunding website launched in 2009 I People can create a page to raise money for a project: I They have to decide on a funding goal and a campaign duration I People pledge money to the project in exchange for various rewards I Backers are charged only if the funding goal is reached at the end of the campaign I As of June 2013: I More than 42 000 projects funded I $ 555 million raised I 4.1 million of backers I Only half of the projects reach their goal: can we predict which? 2. Project Example: The Veronica Mars Movie Project I Project to create a movie sequel of a famous TV show I Campaign lasted from March 13 th to April 13 th 2013 I Ambitious funding goal: $ 2 million I Huge success: 91 585 backers, $ 5 702 153 raised Mar 16 2013 Mar 19 2013 Mar 22 2013 Mar 25 2013 Mar 28 2013 Mar 31 2013 Apr 03 2013 Apr 06 2013 Apr 09 2013 Apr 12 2013 0 1 2 3 4 5 6 Money (million $) Project goal Money pledged Number of backers 0 16 33 50 66 83 100 Backers (thousands) 3. Our Dataset I Web crawler started in September 2012 I Automatically discovers new projects on the Recently launched page I Regularly checks the status of live projects: I Number of backers I Money pledged I Monitor Twitter in parallel to record mentions of Kickstarter I Preprocessing: I Time is normalized to [0, 1] I Pledged money is normalized with respect to the project’s goal 4. Dataset Summary Projects Backers Pledges Tweets 16 042 1 309 295 2 265 156 738 176 I Average project statistics: Successful Failed Total Number 7739 8303 16042 Proportion 48.24% 51.76% 100% Goal ($) 9595 34 693 22 585 Duration (days) 30.89 33.50 32.24 Number of backers 262 25 139 Final amount 216.60% 11.40% 110.39% Number of tweets 73 20 46 I Pledged money for successful/failed projects relative to their goal: 5. Time-series Predictors I Predict success of a project based on its pledged money over time I Use partial information: from time 0 to time t, t 1 (trajectory) I k-Nearest Neighbors I Find the k projects that have the closest trajectories I Predict success if the majority of them are successful, failure otherwise I Markov Chain I Discretize the (time, money) space into a N T × N M grid I Consider the pledged money M (n) at each time step n as a random variable I Learn transition probabilities P (n) [0, 1] N M ×N M for each n ∈{1,...,N T }: P(M (n + 1) = m n+1 | M (n)= m n ,n)= P m n ,m n+1 (n). Transition probabilities Estimated success probabilities 6. Social Predictors I Instead of using pledged money, use social features at time t I Train a SVM classifier using project goal/duration and social features I Tweets I Number of tweets/replies/retweets I Number of unique users that tweeted I Estimated number of backers (using tweet’s text, e.g. “I just backed project X”) I Co-backers graph I Build a graph where vertices are projects I Edges between projects represent common backers I Extract several features from this graph: I Number of projects with common backers I Number/proportion of successful projects with common backers I Number of backers, number/proportion of first-time backers (i.e. with only one backed project) Users backing projects 2 1 Corresponding co-backers graph 7. Results I Can we use social predictors to improve by the time-series ones? I Train a SVM to combine the four individual predictions into one I Very useful at the start: first combined prediction 3.6% more accurate Laboratory for Communications and Applications Swiss Federal Institute of Technology in Lausanne (EPFL)

Upload: others

Post on 22-Jun-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Launch Hard or Go Home! 0.005Predicting the Success of Kickstarter Projects · 2019-12-14 · Launch Hard or Go Home! Predicting the Success of Kickstarter Projects Vincent Etter,

Launch Hard or Go Home!Predicting the Success of Kickstarter Projects

Vincent Etter, Matthias Grossglauser and Patrick Thiran

1. What is Kickstarter?

I Crowdfunding website launched in 2009I People can create a page to raise money for a project:

I They have to decide on a funding goal and a campaign durationI People pledge money to the project in exchange for various rewardsI Backers are charged only if the funding goal is reached at the end of the campaign

I As of June 2013:I More than 42 000 projects fundedI $ 555 million raisedI 4.1 million of backers

I Only half of the projects reach their goal: can we predict which?

2. Project Example: �The Veronica Mars Movie Project �

I Project to create a movie sequel of a famous TV showI Campaign lasted from March 13th to April 13th 2013I Ambitious funding goal: $ 2 millionI Huge success: 91 585 backers, $ 5 702 153 raised

Mar 16 2013 Mar 19 2013 Mar 22 2013 Mar 25 2013 Mar 28 2013 Mar 31 2013 Apr 03 2013 Apr 06 2013 Apr 09 2013 Apr 12 20130

1

2

3

4

5

6

Mon

ey(m

illio

n$)

Project goalMoney pledgedNumber of backers

0

16

33

50

66

83

100

Bac

kers

(tho

usan

ds)

3. Our Dataset

I Web crawler started in September 2012I Automatically discovers new projects on the Recently launched pageI Regularly checks the status of live projects:

I Number of backersI Money pledged

I Monitor Twitter in parallel to record mentions of KickstarterI Preprocessing:

I Time is normalized to [0, 1]I Pledged money is normalized with respect to the project’s goal

4. Dataset Summary

Projects Backers Pledges Tweets16 042 1 309 295 2 265 156 738 176

I Average project statistics:

Successful Failed TotalNumber 7739 8303 16042

Proportion 48.24% 51.76% 100%Goal ($) 9595 34 693 22 585

Duration (days) 30.89 33.50 32.24Number of backers 262 25 139

Final amount 216.60% 11.40% 110.39%Number of tweets 73 20 46

I Pledged money for successful/failed projects relative to their goal:

5. Time-series Predictors

I Predict success of a project based on its pledged money over timeI Use partial information: from time 0 to time t, t ≤ 1 (trajectory)

I k-Nearest NeighborsI Find the k projects that have the closest trajectoriesI Predict success if the majority of them are successful, failure otherwise

I Markov ChainI Discretize the (time, money) space into a NT ×NM gridI Consider the pledged money M(n) at each time step n as a random variableI Learn transition probabilities P (n) ∈ [0, 1]NM×NM for each n ∈ {1, . . . , NT}:

P(M(n + 1) = mn+1 |M(n) = mn, n) = Pmn,mn+1(n).

Transition probabilities Estimated success probabilities

6. Social Predictors

I Instead of using pledged money, use social features at time t

I Train a SVM classifier using project goal/duration and social features

I TweetsI Number of tweets/replies/retweetsI Number of unique users that tweetedI Estimated number of backers (using tweet’s text, e.g. “I just backed project X”)

I Co-backers graphI Build a graph where vertices are projectsI Edges between projects represent common backersI Extract several features from this graph:

I Number of projects with common backersI Number/proportion of successful projects with common backersI Number of backers, number/proportion of first-time backers (i.e. with only one backed project)

Users backing projects

2 1

Corresponding co-backers graph

7. Results

I Can we use social predictors to improve by the time-series ones?I Train a SVM to combine the four individual predictions into one

I Very useful at the start: first combined prediction 3.6% more accurate

Laboratory for Communications and Applications Swiss Federal Institute of Technology in Lausanne (EPFL)