Comment on page
Prospecting Clean Rooms
Build lookalike models to increase campaign addressability
This feature is currently in beta. Please contact your CSM to request access.
They enable the receiving partner to build a lookalike model on their own identity graph, using the private set intersection, computed by the clean room, as the seed.
A prospecting clean room involves two partners. There is a sender, requesting a prospecting model from their DCN partner, and a receiver which will configure and run the model for its partner.
As a sender, you must to select either a source audience or all available data, and request a prospecting model from one or more partner(s). A distinct clean room is created for each selected partner.
An example of prospecting request notification
Upon acceptance, you will be required to select an audience to score (evaluation audience) or use all available data. You will then need to select traits to use in the prospecting model.
Traits selection example
As you select more traits, the Current Prospecting Modelling Success Chance indicator will change from
High. The indicated success chance is based on the coverage of the traits you selected.
Even though selecting all traits usually boosts model performance, exceptions exist. For instance, if your partner's clients all live in Florida but your identity graph covers the entire USA, you might exclude geographic traits. This way, the model focuses on more relevant socio-demographic and attitudinal traits. You can later create an audience that uses the clean room result, and exclude non-Florida residents from it.
An example of configured prospecting clean room
Once the configuration is complete, the model will begin to run. This may take approximately five minutes.
Once you receive a notification confirming the model's completion, two reports will be accessible for you to review:
Although the clean room calculates an exact score for each individual cluster, it does not disclose this information to preserve privacy. Instead, it categorizes clusters into percentiles, with the first percentile representing clusters most similar to your partner's audience. This report aids you in determining the number of percentiles necessary to construct a prospecting audience sizable enough for a campaign, yet still performant. An audience can be directly created from this report.
An example of cluster distribution
Instead, it provides up to ten absolute Top Contributing Features, as determined by the model. These features can represent either very strong positive signals or very strong negative signals. Clusters with these positive features are likely to be in the top percentiles, whereas clusters with negative features will probably fall into the bottom percentiles. The features are displayed in alphabetical order as a mechanism to preserve privacy.
An example of top contributing features report
Because of the absolute nature of Top Contributing Features, you could get all three features
Non-Binaryin the top 10. This means
Genderis a very important trait in the model, as one feature is likely overwhelmingly positive while the other two are negative.
Once the clean room successfully generates results, you will be able to create audiences out of them, either by navigating to the Audience Builder, directly from the Prospecting clean room, or through the Optable CLI. You can build as many simple (e.g.
percentiles 10 to 20) or composite (e.g.
Top 15% AND State=Florida) audiences from them.
An example of prospecting audience building
Prospecting clean rooms integrate with Optable's differential privacy budgeting system, allowing DCN customers to monitor privacy protected clean room activity with its partners, as well as track its rate of collaboration versus a system assigned and replenished privacy budget. With the built in notifications, a DCN customer can better assess and decide its level of risk tolerance for re-identification resulting from clean rooms, on a partner by partner basis.
In addition, multiple mechanisms are built in the prospecting clean room to ensure privacy.
- Clean room will refuse to output results if the intersection between the two audiences is too low
- Clean room will refuse to output results if the match rate is too high
- Clean room will refuse to output results if the count of scored clusters (containing at least one feature from the model) is too low
- Clean room does not evaluate features that are too rare
- Only the top 10 contributing features are exposed in reports
- Top contributing features are ordered alphabetically and not in order of importance
- Top contributing features does not tell whether it is very positive or negative
- Feature importance scores are not exposed
- Cluster similarity scores are not exposed
- Clusters are grouped in percentiles
- Most dissimilar clusters are not included in percentiles grouping
- Bottom percentiles are grouped together in percentile ranges
- The intersection of the two audiences are not removed from the percentiles
Last modified 5mo ago