Lookalike Model Builder
The Lookalike Model Builder provides an interface for marketers to quickly build custom machine learning models and Predictive Audiences based on customer data. In the Lytics UI, the Model Builder is located under the Lookalike Models tab within the Laboratory section, which will serve as a hub for marketing teams to get hands on with data science.
To get started, click Create New Model at the top right.
This opens up the lookalike model builder.
For most use cases, building a model by setting the basic configuration parameters is sufficient. The only required parameters are the selection of a source and target audience, which are very important for building a usable model.
If you select audiences that are too dissimilar, the model may be unable to find lookalikes in the source audience. Learn more about selecting the right audiences for your use case. The size of each audience is also important.
If your selected audience exceeds the maximum size, you can add filters to refine it. For example, if the source audience is “Unknown users” you could add a filter for “Active in the last 30 days” to ensure you aren’t targeting unknown users with stale cookie identifiers.
The basic model parameters are defined below. Scroll to the right to see examples.
|Source Audience required||Select an existing audience as the source to find lookalikes from.|
|Target Audience required||Select an existing audience as the target (users you want to find more of).|
|Custom Model Name optional||If no custom model name is provided, the default name will be |
|Auto Tune optional||Use an automated "intelligent" feature selection process and make a best attempt at building the healthiest model.||Checked|
|Model Training Only optional||Build a model without scoring users. Useful for testing and debugging purposes.||Checked|
Even if you select Auto Tune, you can still specify advanced configuration options, like sample size. See the full list below.
For additional model configuration, select the Advanced Options.
For manually built models (without Auto Tune), one or more features must be selected for the model build, such as
|Use Scores optional||Leverage Lytics Behavioral Scores as features for the model.||Checked|
|Use Content optional||Leverage Lytics Content Affinity as features for the model.||Checked|
|Additional Fields optional||Select fields in the user schema as features for the model.|
|Continuously Re-train optional||Retrains the model every week with a new training sample.||Unchecked|
|Sample Size optional||The sample size of users for the model training set. Allowable range: 100 - 50000 users.|
Create Predictive Audiences
Once a Lookalike Model is built and users are scored (make sure the Model Training Only option is unchecked), you can create Predictive Audiences with different prediction decision thresholds for the model. Learn more about creating audiences with custom rules.
From the Lookalike Models list view, click the model you'd like to use to build a Predictive Audience. Then find the Model Usage section and click Create Predictive Audience.
This opens the Audience Builder. All Lookalike Models are keys under the user field
segment_prediction. The values are the model prediction.
By lowering the prediction decision threshold, you can expand the reach (i.e. audience size) for targeting while sacrificing some accuracy. Selecting the appropriate threshold for targeting customers allows you to balance accuracy with reach based on your marketing campaign goals.
Using Lookalike Model Percentiles
Another option to build a Predictive Audience is by using the
Lookalike Model Percentiles field. Similar to the
segment_prediction field, the Lookalike Models are keys for the
Lookalike Model Percentiles field.
The percentile for a model represents the value at which a percentage of the predictions fall below. For example, the 80th percentile represents the prediction score at which 80% of all other scores fall below. This allows for an easy, systematic way to account for the shape of a model's prediction distribution, as it can sometimes be hard to determine who the best users are for a model if the distriburion is skewed is any direction. Continuing the example, by using percentiles, we can take users with a
Lookalike Model Percentiles value of at least 80 to give us the top 20% of users, without having to figure out what prediction score is considered "good".