In an effort to give users a better starting point a few sample attribute building ideas and details are contained within this article. The ideas for these attributes have been sourced from a number of use cases we've handled over the years and can be achieved within the confines of a sample dataset. Your use cases and data may be completely different, the goal here is to get the creative juices flowing.
As a reminder, all Attribute Builder apps use a Data Layer as the source. If you are unsure of what a Data Layer is please refer to the Data Layers documentation or ask your organization's Audiences Solution owner if you do not know what Data Layer to use.
Common e-commerce website attributes that marketers need to be able to build audiences off of center around product-related information, such as brands. This section walks through one method of building brand attributes. The brand-related information is coming from an event table and a product table.
In the screenshot below:
- Recency - 1 day, users are considered for the attribute for one day.
- Frequency - 1, users have interacted with the brand at least one time.
- Attribute Name - lg, using free text code mode and a lower() function on pagename to accommodate upper, lower, mixed case scenarios.
- Attribute Name - sony, using basic mode using condition selections.
Machine Learning Model Attributes
After a machine learning model is built the results are typically in the form of a raw numeric score or labeled categories depending on the type and purpose of the model.
For example, a purchase propensity model would score a visitor from 0 to 1 on their propensity to return to the website and make a purchase. The score output can be included in a data layer providing the ability to build audiences off of. This allows marketers to target the selected audience or for analysts to perform journey analytics. Segmentation models typically apply labels to the derived audiences based on the features they most closely align with (i.e. gaming, appliances, pre-retirees, binge-watchers).
Data scientists and data engineers may provide the model output in one of several methods in order from most granular to least granular:
- Raw scores from 0 to 1 (most granular)
- Scores ranked and placed into percentiles from 1 to 100
- Bucketed and given labels such as High, Medium, and Low (least granular)
In the screenshot below attributes are being created from a purchase propensity model scoring:
- Recency - 14 days, provide the highest score for a visitor in the past 14 days
- Frequency - 1, scored at least once in within the set conditions
- Left Value - max(predictions), attributes are group by identifier, this determines the highest score for a visitor ID.
Online commerce businesses typically have a common set of variables they track to help understand behaviors on the website, such as:
- Product view
- Cart add
- Cart remove
The below screenshot are some examples of how these common metrics can be set up as attributes:
- Recency - 7 days, determine if the visitor did each of the actions during the last seven days.
- Frequency - 1, performed the action at least once in within the set conditions.
- Attribute Name - Cart Add, attributes are group by identifier, this determines if the visitor performed a cart add action during the recency window.
- Attribute Name - Product View, attributes are group by identifier, this determines if the visitor viewed a product during the recency window.
- Attribute Name - Purchase, attributes are group by identifier, this determines if the visitor purchased during the recency window.