user scoring for fun and profit


predictive modeling in a social shopping service


Wrapp

social shopping

For our users

  • free and paid digital gift cards
  • send to friends or claim for yourself
  • everything in a convenient mobile app wallet


For our partners
  • viral spread of offers through social media
  • advanced targeting 
  • advanced real-time analytics

wrapp in numbers

1.7 M users
18M gifts sent
160 M friends of users
200+ merchant partners



Campaigns

Last week "Kanelbullens dag" 100k+ gift cards sent in Sweden

why predictive modeling at wrapp?

  • Fraud
  • Recommendations
  • LTV
  • Campaign budgeting

    Fraud

    • Many types of fraud
      • Gift card trading
      • Gift card stacking
      • Gift card hoarding 
      • Credit card fraud
    • Predictors
      • social network activity
      • earlier wrapp activity,
      • demographics
      •  geolocation
      • credit card data
    • Output
      • Fraud risk class

    Recommendations

    Recommend campaigns to users and users for campaigns

    • predictors
      • social activity
      • demographics
    • output
      • your favourite campaign
      • best recipient for a campaign

    order value

    • predictors
      • social activity
      • demographics
      • campaign metadata
    • output
      • predicted order value for a user

    Campaign budgeting

    Not all gift cards are redeemed, merchants want to know how many they have to pay for
  • Campaign metadata
  • Seasonal effects
  • Fraud

    In Wrapp we are facing at least two types of fraud


    • Credit card fraud
    • Account fraud

    building models

    Most of our models are quite simple linear models

    R = M x p
    R =response
    M= model
    p = predictor vector

    Building a good model basically entails setting the weights in M appropriately. This can be done with a few methods:

    • Heuristics (aka. pure guesswork)
    • Logistic regression
    • Machine learning methods

    quantified Heuristics

    Can actually be very good, especially if you  quantify them so they are applied equally. Especially useful for less tangible response variables

    Basic methodology
    1. list everything you can think of that  
      could impact your response
    2. Assign weights to each manually
    3. Test, test, test

    We are using "quantified heuristics" for our fraud filters and our recommendation engine, however for the latter we are collecting statistics to help improve our weighting. 

    linear regression

    Basic methodology
    1. Collect data
    2. Run your favorite regression model (glm)
    3. Profit



    case study: predicting AOV

    who are the misers?


    hypothesis

    a small percentage of users are doing a majority of our "low value transactions"

    1) look at the data



    Dataset: 200k transactions

    aov histogram
    campaign histogram

    2) build a model



    glm(transactions)

    3)  Profit? 


    prediction is just step 1
    what if we change the rules?
    what if we could collect better data?

    future: where do we want to go

    Better predictive modelling requires intervention, not just passive data collection

    Examples: automatically change recommenders
    measure response, apply change them again 




    STAY TUNED FOR A MORE DATA-DRIVEN SHOPPING EXPERIENCE

    user scoring for fun and profit

    By cortex

    user scoring for fun and profit

    • 1,671
    Loading comments...

    More from cortex