Brief: https://www.hackerearth.com/challenge/competitive/machine-learning-challenge-2/
Kickstarter is a community of more than 10 million people comprising of creative, tech enthusiasts who help in bringing creative project to life. Till now, more than $3 billion dollars have been contributed by the members in fuelling creative projects. The projects can be literally anything – a device, a game, an app, a film etc.
Kickstarter works on all or nothing basis i.e if a project doesn’t meet it goal, the project owner gets nothing. For example: if a projects’s goal is $500. Even if it gets funded till $499, the project won’t be a success.
Recently, kickstarter released its public data repository to allow researchers and enthusiasts like us to help them solve a problem. Will a project get fully funded ?
In this challenge, you have to predict if a project will get successfully funded or not.
There are three files given to download: train.csv, test.csv and sample_submission.csv The train data consists of sample projects from the May 2009 to May 2015. The test data consists of projects from June 2015 to March 2017.
| Variable | Description |
|---|---|
| project_id | unique id of project |
| name | name of the project |
| desc | description of project |
| goal | the goal (amount) required for the project |
| keywords | keywords which describe project |
| disable communication | whether the project authors has disabled communication option with people donating to the project |
| country | country of project author |
| currency | currency in which goal (amount) is required |
| deadline | till this date the goal must be achieved (in unix timeformat) |
| state_changed_at | at this time the project status changed. Status could be successful, failed, suspended, cancelled etc. (in unix timeformat) |
| created_at | at this time the project was posted on the website(in unix timeformat) |
| launched_at | at this time the project went live on the website(in unix timeformat) |
| backers_count | no. of people who backed the project |
| final_status | whether the project got successfully funded (target variable – 1,0) |
| Variable | Description |
|---|---|
| goal_usd | goal converted to USD using historic exchange rate |
| keywords_len | length of keywords |
| keywords_count | number of keywords |
| name_len | length of project name |
| name_count | number of words in project name |
| desc_len | length of description |
| desc_count | number of words in description |
| name_capitals | number of uppercase letters in project name |
| name_digits | number of digits in the project name |
| name_digits_any | boolean indicating whether there are any digits in project name |
| desc_digits | number of digits in description |
| launched_at_weekday | day of week when the campaign was launched |
| deadline_weekday | day of week of campaign deadline |
| campaign_len | campaign length |
After experimenting with several algorithms (Naive Bayes, SVM) and selecting optimal features & parameters using KBest and GridSearchCV respectively, I settled on AdaBoost with thw following features:
['campaign_len'
,'keywords_count'
,'keywords_len'
,'name_count'
,'name_capitals'
,'desc_digits'
,'name_digits_any'
]
This led to a prediction score of 0.67484 on the test version of the dataset.
Seasonality is something that could be explored further. I found that the day of the week when the campaign starts or finishes play no role in the outcome, but longer-term seasonal fluctuations are likely. Could there be a downturn in funding around summer? or an upturn around the time when people get bonuses?