Documentation / Product / Features / Descriptive & Predictive Modeling


SegmentML is Lytics' take on machine learning for marketing. It is a platform for identifying likely candidates for other audiences. It can be used to answer questions such as:

  1. Which users look like subscribers but aren't?
  2. Which users have purchased, but aren't likely to purchase again?
  3. Which users are still opening email but might not for much longer?

How does SegmentML work?

SegmentML uses existing rich user-profile information as a bed of features to test similarities and dissimilarities with. By using this feature information, SegmentML programmatically finds the answers to hundreds of questions to make a well-informed judgment of an individual.

Here is an example of how a marketer may try something similar by hand:

The goal is to identify users who are likely to buy a travel package for the summer.

  1. People buy travel packages months in advance, so identify users who are visiting in winter and spring.
  2. People who already get promotions are much more likely to buy expensive items.
  3. People who have already bought a package are likely to buy again.

Targeting users that match these three rules will certainly improve the efficacy of a campaign, but why stop at three rules? Are there more factors that can be used to refine this group of users?

This is how SegmentML would do it instead:

The goal is to identify users who are likely to buy a travel package for the summer.

  1. What do people who have bought travel packages in the past "look like" (i.e., what features do they have in common)?
  2. Analyze all the information known for these people who have bought travel packages.
  3. Determine which features are significant and which values of these features are significant (e.g., visiting the website is important, specifically three to five times).
  4. Analyze all the information known about people who have not bought travel packages.
  5. Determine which users share common features with past summer travel package purchasers.
  6. Take the most similar users to use in targeting.

The key difference here is that a marketer may use a handful of criteria using logical heuristics to define a segment of users while SegmentML will look at hundreds of factors including potential non-obvious, impactful criteria to define a segment of users for the same purpose.

How does SegmentML get used?

SegmentML models are defined using a source audience and a target audience. Then scores for each SegmentML model are associated with each user, much like Lytics Scores and Lytics Content Affinity.

Then, the SegmentML scores become available for use in audiences through a special user field (since SegmentML is a custom service, the SegmentML user field is currently unsupported by the audience builder). The user fields operates a lot like Lytics scores, in that there is a fixed domain of possible values and the value threshold can be played with to create an audience of the desired size.