Follow

Lookalike Modeling

The Cxense DMP now supports lookalike modeling. You enable lookalike modeling for an existing segment, setting a target size for the new lookalike segment. The Cxense DMP will automatically use machine learning to classify new users as lookalikes.

For instance, you may have a small segment consisting of female users based on data from you CRM system. You enable lookalike modeling, and set the target size to 50%. A training model is generated based on the original segment, and the 50% of new users who are most similar to the original users will be classified into the lookalike segment.

A new lookalike segment is automatically generated for you, and will appear in the Cxense DMP segment list. You can use this segment directly, but lookalikes are also available as a segment criteria that can be used in multiple segments.

When the lookalike model has been trained, you will start getting statistics about the lookalike model. The lift chart displays how much better we can expect the lookalike model to predict a segment member, compared to random selection. The vertical line represents the currently defined target size.

In the chart above, we find the lift for the lookalike model as the gray line, and we see that for the current target size (5%) the lift is about 10. A lift score of 10 suggests that the lookalike model is 10 times more accurate in predicting a segment member than random guess.

The lift score, and therefore the precision, will typically be higher, the lower the target size is set. It is important to understand that there is a trade-off between the target proportion and the precision we can expect from the model, and we should therefore be careful not to set the target size too high. The higher the target size is set, the less benefit we have from the lookalike model.

A lookalike model is trained based on a set of attributes derived from the user profiles of users inside, and for the users outside the segment. After training a lookalike model, we may inspect the most significant attributes found inside the lookalike model.

In the screenshot above we see a list of attributes which are more frequently found within the segment population (Positive Signals), and attributes which are more frequently found outside the segment population (Negative signals).

For more information, please refer to the lookalike modeling documentation:

Lookalike modeling introduction

We hope you find the lookalike modeling features useful, and we are looking forward to your feedback.

 

 

 

 

 

 

 

Have more questions? Submit a request

Comments

Powered by Zendesk