Demographic prediction for not registered users at leading Danish media.

Client

Ekstra Bladet

Services Used

Data Engineering

Data Science

Data Modelling

Created with Sketch.

People in the company act on data insights within 1-2 minutes and that wouldn’t be possible without the infrastructure Flowtale built for us.

Lars Gundersen, Head of Data & Insights, Ekstra Bladet
  • Challenge

  • Solution

  • Results

Challenge

Ekstra Bladet, being the 4th most visited website in Denmark after Google, Facebook and Youtube, has some serious big data sets for their user base, split between registered and unregistered users. Registered users had both behaviour and demographics, while unregistered had only behaviour. Ekstra Bladet sought to also predict demography for non-registered users. It was agreed by making a model of the registered users, demography could be accurately predicted for the non-registered users through their behaviour.

Solution

A model was created to assess potential explanatory variables in terms of their ability to predict age and gender. Furthermore, its purpose was to conduct preliminary training of key machine learning models in order to assess their potential predictive performance with respect to age and gender (separately). Gender, in this example, being a behavioural pattern that can fit more sexes.

Predictive Models

By utilising machine learning algorithms, two predictive models were created; (i) a classification model to define age and (ii) a binary classification model to define gender.

Performance Assessment

The model was developed for maximum accuracy, which is defined as the percentage of correct classifications. Model families utilised; (i) gradient boosting, (ii) Random forest, and (iii)Support vector machine. Each family was considered for both age and gender. The model families were trained and benchmarked using the caret package.

Solid Process

The process of the task was separated into three overall steps; (i) Data research - collection and structuring of data for modelling and prediction purposes, (ii) Model building - building the appropriate model for subsequent preaching, and (iii) Prediction - prediction of gender and age.

Results

Diving into the userbase of Ekstra Bladet into gender and age is non-disclosable according to Flowtale’s Secrecy and Protection policies. However, the model accuracies were as following:

92%

“Man” prediction accuracy

48%

“Woman” prediction accuracy

83%

Age prediction, e.g. >44 years