A classification problem
This is a small Data Science project as part of the Data Science bootcampt neuefische, Fall 2020. The main aim of this project are a throughout Exploratory Data Analysis and the construction and validation of various Machine Learning Classification Models.
In recent years, the range of funding options for projects created by individuals and small companies has expanded considerably. In addition to savings, bank loans, friends & family funding and other traditional options, crowdfunding has become a popular and readily available alternative. Kickstarter, founded in 2009, is one particularly well-known and popular crowdfunding platform. A huge variety of factors contribute to the success or failure of a project. Some of these can be quantified or categorized, which allows for the construction of a machine learning model to attempt to predict whether a campaign will succeed or not.
- SuccessPrediction_presentation.pdf presentation deck as PDF for non-technical stakeholders´
- SuccessPrediction_EDA.ipynb Exploratory Data Analysis notebook incl. Data Cleaning, Feature Engneering and Visualisation (PEP8, technical audience)
- SuccessPrediction_ML.ipynb Machine Learning notebook incl. Models, Visualisation on findings and recommendations (PEP8, technical audience)
- df_model.pkl cleaned data as pickel data file
Synthetic cap of number of occurrences at about 2400 projects in an Subcategory.
Success by Subcategory is very different for the most frequent subcategories
The number of backers has a great correlation to the successrate
This project focuses on the F1-Score as target metric. Evaluated Models and their F1 Score:
-
Random Forrest: 0.81
-
XGBoost: 0.8
-
SVM: 0.69
-
Logistic Regression: 0.78
Feature Importance for Random Forrest
The Random Forrest model performes slightly better as the XGboost model, with an f1-score of 0.81 and slighly better computation time of XGboost. As for the features: The category has the greatest influence on the success of a project. As part of the EDA, the median_avg_pledge_per_backer_in_subcat feature is designed. This feature gives an understanding of how much money the average backer in an sub category is willing to back a project. This has great influence on the success rate of a project.
- Feature Engneering: Average USD-Pledge per Subcategory to evaluate better supported subcategories.
- ROC Curves are another way to illustrate the performance of a specific model. Compared to the F1-Score, ROC Curves contain more information and, hence, generate a deeper understanding of the model's performance. They are especially great for investigating and comparing different models. Besides, ROC curves are a great tool for communicating results. Hence, in future work, ROC curves should generate a deeper understanding of the model's performance.
- The used confusion matrix fulfills its purpose: comparing the predicted and actual successes and failures. The color-coding is a great addition as it shows the severity of various values at a glance. Nevertheless, percentage values and labeling would be a great addition and is part of future work.
- Aiming for the stars: The blurb (a string description) and the photo (mostly different versions of the main image) have been droped in the current analysis earlie on. Splitting the blurb into its fragments could potentially surface a corrolation between success and metaphorical, postive words. In addition, Analysing the images, e.g., for items or for color palets, could bring great insights and be a very steep but rewarding learning curve.
The data was crawled by the company WebRobots ('We provide B2B web crawling and scraping services') directly from the Kickstarter-Website. At the point of writing, crawled data includes data within the timeframe 2014-04-22 to 2020-08-13.