Adopter-Prediction-Challenge

Imbalanced class classfication challenge on classifying likelihood adopters for a music streaming service.

Advisor - Dr. Mochen Yang (mochenyang.github.io)

Problem Statement

Website XYZ, a music-listening social networking website, follows the “freemium” business model. The website offers basic services for free, and provides a number of additional premium capabilities for a monthly subscription fee. We are interested in predicting which people would be likely to convert from free users to premium subscribers in the next 6 month period, if they are targeted by our promotional campaign.

Task

The task is to build the best predictive model for the next marketing campaign, i.e., for predicting likely adopters (that is, which current non-subscribers are likely to respond to the marketing campaign and sign up for the premium service within 6 months after the campaign).

Results

Challenge completed with first position, metrics:

Metric	Value
F1 score	0.15028
Recall	0.24026
Precision	0.10934
Accuracy	0.95170

Leaderboard, with number of submissions:

Solution Approach

approach.pdf

Dataset

The labeled dataset contains 86,682 records, each record representing a different user of the XYZ website who was targeted in the previous marketing campaign. Each record is described with 25 attributes. Here is a brief description of the attributes (attribute name: type - description):

• adopter: binominal (0 or 1) - whether a user became a subscriber within the 6 month period after the marketing campaign (this is the outcome variable)
• user_id: integer - whether a user became a subscriber within the 6 month period after the marketing campaign (this is the outcome variable)
• age: integer - age in years
• friend_cnt: integer - numbers of friends that the current user has
• avg_friend_age: real - average age of friends (in years)
• avg_friend_male: real (between 0 and 1) - percentage of males among friends
• friend_country_cnt: integer - number of different countries among friends of the current user
• subscriber_friend_cnt: integer - number of friends who are subscribers of the premium service
• songsListened: integer - total number of tracks this user listened (or reported as listened)
• lovedTracks: integer - total number of different songs that the user “liked”
• posts: integer - number of forum or discussion board posts made by the user
• playlists: integer - number of playlists created by the user
• shouts: integer - number of wall posts received by the user
• good_country: boolean - country type of the user: 0 – countries where free usage is more limited, 1 – less limited
• tenure: integer - number of months since the user has registered on the website
• delta_<attrname>: integer - Such attributes refer not to the overall number, but the change to the corresponding number over the 3-month period before the marketing campaign.

Attempted algorithms

• k-NN
• Decision Tree (CART)
• Logistic Regression
• Support Vector Machine
• Random Forest
• Extra Trees Classifier
• LDA Classifier https://scikit-learn.org/stable/modules
• Gradient Boosting
• XGBoost - github.com/dmlc/xgboost
• LightGBM - github.com/microsoft/LightGBM
• Ensembles of above
• Neural Networks