The exploding use of lookalike modeling within digital in the last decade is tied to the evolution of machine learning’s ability to drive ROI across the funnel. Yet there’s still a lack of clarity on the best lookalike modeling trends to use and where to use these approaches. Additionally, as most of the attention goes to the algorithms that power machine learning, very little consideration is given to the actual inputs (data) that fuel lookalike modeling trends in the first place. Given the complexity of lookalike modeling, this pattern is concerning.
Lookalike modeling is both an art and a science.
The inputs (and how those inputs are combined) influence the eventual predictive power of a lookalike model more than any machine-learning algorithm.
As complex lookalike algorithms continue to become readily available, data becomes the ultimate differentiator.
Having predictive and proprietary data that is relevant to your business is not only a significant competitive edge. Moreover, it increases the predictive power of lookalike models, irrespective of the application of the model (customer retention, personalization, acquisition, or other CRM activities).
For example, testing the 3,500 proprietary data signals contained within Zeta’s Data Cloud indicated a huge influence on the predictive power of Zeta’s lookalike modeling for both CRM and acquisition. (In other words, using the right kind of data in your lookalike models is both a real differentiator and an advantage.)
So, with all of that said, here are 3 Next-Generation Lookalike Modeling Trends for Marketing and Advertising…
Trend #1 – Lookalike models transition from supervised to unsupervised learning approaches
Lookalike models are used to characterize the dominant traits of a specific set of records (be they audience segments, campaign convertors, or something else entirely), especially against another set of records opposite in nature.
For instance, lookalikes of a set of convertors are built by mathematically teasing out the dominant traits of said convertors and then comparing them to a set of non-convertors.
The algorithms that power this comparison-based lookalike model are known as “supervised learning algorithms”, and they do a great job of describing “labeled” audiences.
In the example above, the labels indicate whether a record is a “convertor” or “not a convertor.”
But it is in situations where labeled data is not readily available that true marketing and advertising ROI is achieved. For example…
- Why does a customer make multiple purchases from you, and what are the traits of those customers?
- What insights are you missing from your campaign because you’re looking in the wrong place?
Answering these questions (and addressing other situations where labeled data is not readily available) requires incorporating “unsupervised learning algorithms” into lookalike modeling.
Unsupervised algorithms don’t require labeled data and they don’t need to be “told” what to find. They can analyze the available data to find hidden patterns within campaigns, audiences, and more.
The patterns discovered are a goldmine of actionable opportunity for marketers and advertisers which is why the shift from supervised to unsupervised learning algorithms is happening slowly but surely.
Trend #2 – Keeping humans in the loop to ensure modeling accuracy
The algorithms used to power lookalike models regularly provide an accurate approximation of what’s happening in the real-world.
However, they’re not perfect.
When allowed to run loose without any human oversight, they can create inaccurate models.
For example, given a set of inputs, an algorithm designed to identify lookalikes of a prospect who signs up for a credit card can over-index on traits that aren’t compliant with fair credit lending practices.
As a result, marketers and advertisers might fail to reach key underserved populations. To ensure inaccuracies like this don’t happen in their lookalike models, marketers and advertisers are increasingly keeping people in the loop.
By allowing human eyes to review the initial findings of a model and approve (or reject) its use for downstream purposes, marketers and advertisers can mitigate their exposure to risk from the deployment of an inaccurate model.
Trend #3 – Data distribution using sub-models will power lookalike modeling
“Data privacy” is a buzzword in marketing and advertising these days—it commands the attention of consumers, regulators, and agencies alike.
As a result, no marketing or advertising professional wants to obtain insights about customers or prospects that aren’t privacy compliant. This reality casts doubt on the future of data appending by way of sharing owned, first-party data with external, third-party data vendors that provide enrichment services.
Going forward, lookalike models will be built and distributed using multiple sub-models which will determine where the data resides.
The outputs from these sub-models will come together in a secure, cleanroom location where they’ll be collated so a primary modeling decision can be made.
This will be a departure from the current state of lookalike modeling where all the data required to build a model is pooled in one location.
Wrapping up lookalike modeling trends for marketing and advertising…
In the years ahead, lookalike modeling will continue to embrace more sophisticated, accurate, and distributed techniques that increase ROI.
Marketers and advertisers would be wise to stay ahead of these emerging lookalike modeling trends so they can ensure they’re always obtaining the best business outcomes possible for the brands they represent.