Semantic Identification: the New Approach to Look-Alike

Our E-Contenta team works with content and does its best to develop technologies that will smoothen and upgrade this process. Actually, look-alike strategy has long been effectively used by targeting experts. But is there anything specific to native advertising?

Look-alike involves targeting an audience similar to website visitors who performed a targeted action: they looked through several pages, made a request or placed an order. The goal of this targeting is to find a new larger audience with high conversion potential based on known user profile parameters, which are the core of the target audience.

Usually, building a look-alike model refers to detecting similarities between user profiles. At the same time, the profile contains both reliable (such as gender or age obtained by the advertiser within the framework of loyalty programs or available from external DMP systems) and computed properties based on predictive models.

The first properties are really few, so look-alike cannot stick to them to significantly expand the audience. Although people seem to understand the second type, certain unreliability and loss of information about the user lies behind it. For example, a user who has several times watched the news about the drinking water quality problems may be marked as “eco-conscious” and “interested in water filters”, but the entire context of the pages he visited would be lost. Such losses are accepted by classical modeling because there is no computation of similarities between profiles with the complex structure.

E-Contenta has developed a new approach to identifying semantic relations of the pages visited by a user and building a look-alike audience. It is based directly on the initial data, without loss of information at the stage of computed properties formation.

In the new model, user data is presented as a graph containing all information about the content this user interacted with. A graph is a network showing objects and connections between them.

The major difficulty of using this new model lies in finding the proximity of users, since it is necessary to determine the degree of similarity of the corresponding graphs. Successful solution of this problem allows us to better analyze the user and determine whether he will enter the look-alike audience, even if there is little information about him (for example, it is known that he has visited only 10 websites) and find 35% more relevant audience compared to the classical algorithm.

Such innovation makes E-Contenta campaigns even more effective, since the platform will be able to work with the entire array of source data and more accurately determine the core of user interests. If you want to test look-alike targeting within your native campaign, follow the link and submit a request!