supertype logo dark (wide)
supertype logo dark (wide)
opera-white

Audience Segmentation & lookalike creation with confidence

Supertype helps the AdColony publisher team create better audience targeting tools by layering segmentation and predictive clustering on top of raw device-level fields.

Challenges

Demographic Inference from Non-PII (non personally identifiable information) Data

Opera's video ad platform, AdColony, provides advertisers a cost-effective distribution channel to reach audience with their message. While there are 256 raw device-level fields associated with each impressions served, none of them are demographic data. The advertiser however, would greatly benefit from being able to infer basic demographic profiles (gender, age groups) from these device-level fields and historical activity.

  • Separating 'signals' from 'noise' in a high-dimensional data operating as a SSP (supply-side platform)
  • Enriching these signals so they can be used for high-value activities such as interest-based or demography-based targeting

Advertising Relevance through audience matching

Without the ability to define and present a clear set of audience parameters, advertisers on the AdColony / Opera platform rely on vague, broad data (broadband network, carrier name, device model etc) in their audience targeting, which isn't ideal for most performance-based marketing campaigns.

Consequently, advertisers seek alternative advertising avenues where better audience matching means are available, causing a drop in client retention and acquisition.

Challenges in Numbers

201
raw device fields
6.9 million
rows of raw data each day

Approach

  • Machine Learning Model

    Given 10 million raw data, containing no demographic information, Supertype's data scientists trained a machine learning model that could infer properties signalling a device user's demographic attributes (e.g genders) at an accuracy of 99.57% in a binary classification, based on validated ground truths

  • Dimensionality Reduction

    Supertype's model prototyping process also aims to provide explainability, using multiple approaches to dimensionality reduction such that a 3-dimension visualization can be had on how to arrive at complex, non-linear decision boundaries within the raw data

Interactive Demonstration

Pan, zoom and rotate along the axis of the following visualization. This is a demonstration of a dimensionality reduction algorithm that seeks to infer the gender of a device owner from raw, non-PII data by shrinking 256 dimensions into 3 axes.

Detailed Methodology

Learn from Gerald, who detailed his machine learning approach to dimensionality reduction -- while being on the data science consultancy team working with AdColony.

Read Article

Deliverables & Results

Qualifying the value of Supertype’s data science & engineering work

More Desirable to Advertisers

Ability to direct a campaign's messaging towards a key a demographic segment is a top priority of many advertisers' campaign objective

Improve bottom line with lookalike audiences

Opera's AdColony team can increase revenue opportunities by offering groups of audience that are behaviorally- and characteristically- similar to an advertiser post the initial campaign, modeled from the custom-trained machine learning algorithms

Better product offerings with greater ROI

By producing non-linear decision boundaries that classify device into distinct groups with an accuracy of 99.57% on out-of-sample tests, Supertype laid the foundation for new advertising products with an added layer of audience targeting abstraction.

Contact Us

Supertype_adcolony_opera

Want more case study?

This case study, along with others in our archive is also available as a PDF. To schedule a call with us and learn more about our past projects, leave your contact method below.

Your work email, so we can reach you. We respect your privacy and do not engage in any spamming activities.

Inquiries

s@supertype.ai

Back to homepage

More content from Supertype