Input and output in Attribution AI
The following document outlines the different input and outputs utilized in Attribution AI.
Attribution AI input data
Attribution AI works by analyzing the following datasets to calculate algorithmic scores:
- Adobe Analytics datasets using the Analytics source connector
- Experience Event (EE) datasets in general from Adobe Experience Platform schema
- Consumer Experience Event (CEE) datasets
You can now add multiple datasets from different sources based on the identity map (field) if each of the datasets shares the same identity type (namespace) such as an ECID. After you select an identity and a namespace, ID Column completeness metrics appear which indicate the volume of data being stitched. To learn more about adding multiple datasets, visit the Attribution AI user guide.
The channel information is not always mapped by default. In some cases, if the mediaChannel (field) is blank, you would not be able to “continue” until you map a field to mediaChannel as it is a required column. If the channel is detected in the dataset, it is mapped to mediaChannel by default. The other columns such as media type and media action are still optional.
After you map the channel field, continue to the ‘Define events’ step where you can select the conversion events, touchpoint events, and choose specific fields from individual datasets.
For more details on setting up the Consumer Experience Event (CEE) schema, please refer to the Intelligent Services data preparation guide. For more information on mapping Adobe Analytics data, visit the Analytics field mappings documentation.
Not all the columns in the Consumer Experience Event (CEE) schema are mandatory for Attribution AI.
You can configure the touch points using any fields recommended below in the schema or selected dataset.
Typically, attribution is run on conversion columns such as order, purchases, and checkouts under “commerce”. The columns for “channel” and “marketing” are used to define touchpoints for Attribution AI (for example, channel._type = 'https://ns.adobe.com/xdm/channel-types/email'
). For optimal results and insights, it is highly recommended that you include as many conversion and touchpoint columns as possible. Additionally, you are not limited to just the above columns. You can include any other recommended or custom columns as a conversion or touchpoint definition.
Experience event (EE) Datasets do not need to explicitly have Channel and Marketing mixins as long as the channel or campaign information relevant to configure a touchpoint is present in one of mixin or pass through fields.
channel.typeAtSource
(for example, channel.typeAtSource = 'email'
).Historical data data-requirements
- You need to provide at least 3 months (90 days) of data to run a good model.
- You need at least 1000 conversions.
Attribution AI requires historical data as input for model training. The data duration required is mainly determined by two key factors: training window and look-back window. Input with shorter training windows are more sensitive to recent trends, while longer training windows help produce more stable and accurate models. It’s important to model the objective with historical data that best represents your business goals.
The training window configuration filters conversion events set to be included for model training based on occurrence time. Currently, the minimum training window is 1 quarter (90 days). The lookback window provides a time frame indicating how many days prior to the conversion event touchpoints related to this conversion event should be included. These two concepts together determine the amount of input data (measured by days) that is required for an application.
By default, Attribution AI defines the training window as the most recent 2 quarters (6 months) and lookback window as 56 days. In other words, the model will take into consideration all of the defined conversion event(s) that have occurred in the past 2 quarters and look for all the touchpoints that have occurred within 56 days prior to the associated conversion event(s).
Formula:
Minimum length of data required = training window + lookback window
Example:
- You want to attribute conversion events that have happened within the last 90 days (3 months) and track all the touchpoints that have happened within 4 weeks prior the conversion event. The input data duration should span over the past 90 days + 28 days (4 weeks). The training window is 90 days and the lookback window is 28 days totaling 118 days.
Attribution AI output data
Attribution AI outputs the following:
Example output schema:
Raw granular scores raw-granular-scores
Attribution AI outputs attribution scores in the most granular level possible so that you can slice and dice the scores by any score column. To view these scores in the UI, read the section on viewing raw score paths. To download the scores using the API visit the downloading scores in Attribution AI document.
- The reporting column is included in the configuration page either as part of touchpoint or conversion definition configuration.
- The reporting column is included in additional score dataset columns.
The following table outlines the schema fields in the raw scores example output:
Example: 2020-06-09T00:01:51.000Z
Example: “Order”, “Purchase”, “Visit”
Example: 575525617716-0-edc2ed37-1aab-4750-a820-1c2b3844b8c4
Example: 4461-edc2ed37-1aab-4750-a820-1c2b3844b8c4
Example: _atsdsnrmmsv2
Example: Attribution AI Scores - Model Name__2020
Example: ORDER_US
Example: Order, Lead, Visit
Example: Adobe Analytics
Example: Adobe.com
Example: Order
placeContext.geo.countryCode
.Example: US
Example: 99.9
Example: RX 1080 ti
Example: Gpus
Example: 1 1080 ti
Example: 2020-06-09T00:01:51.000Z
Example: MJ-03-XS-Black
Example: 2020-06-09T00:01:51.000Z
Example: city: San Jose
id
and namespace
.Example: 17348762725408656344688320891369597404
Example: aaid
Example: PAID_SEARCH_CLICK
Viewing raw score paths (UI) raw-score-path
You can view the path to your raw scores in the UI. Start by selecting Schemas in the Platform UI then search for and select your attribution AI scores schema from within the Browse tab.
Next, select a field within the Structure window of the UI, the Field properties tab opens. Within Field properties is the path field that maps to your raw scores.
Aggregated attribution scores aggregated-attribution-scores
Aggregated scores can be downloaded in CSV format from the Platform UI if the date range is less than 30 days.
Attribution AI supports two categories of attribution scores, algorithmic and rule-based scores.
Attribution AI produces two different types of algorithmic scores, incremental and influenced. An influenced score is the fraction of the conversion that each marketing touchpoint is responsible for. An incremental score is the amount of marginal impact directly caused by the marketing touchpoint. The main difference between the incremental score and the influenced score is that the incremental score takes the baseline effect into account. It does not assume that a conversion is caused purely by the preceding marketing touchpoints.
Here is a quick look at an Attribution AI schema output example from the Adobe Experience Platform UI:
See the table below for more details about each of these attribution scores:
Raw Score reference (attribution scores)
The table below maps the attribution scores to the raw scores. If you wish to download your raw scores, visit the downloading scores in Attribution AI documentation.
Aggregated Scores aggregated-scores
Aggregated scores can be downloaded in CSV format from the Platform UI if the date range is less than 30 days. See the table below for more details about each of these aggregate columns.
Example: 2016-05-02
Example: 2017-04-21
Example: ORDER_AMER
Example: ORDER
Example: PAID_SEARCH_CLICK
Example: CC
Example: gpus, laptops
Example: US
Example: Paid Conversion
Example: PAID, OWNED
channel._type
property that is used to provide a rough classification of channels with similar properties in Consumer Experience Event XDM.Example: SEARCH
mediaAction
property is used to provide a type of experience event media action.Example: CLICK
Example: COMMERCIAL
Example: Thanksgiving Sale
Raw Score reference (aggregated)
The table below maps the aggregated scores to the raw scores. If you wish to download your raw scores, visit the downloading scores in Attribution AI documentation. To view the raw score paths from within the UI, visit the section on viewing raw score paths within this document.
- Attribution AI uses only updated data for further training and scoring. Likewise, when you request to delete data, Customer AI refrains from using the deleted data.
- Attribution AI leverages Platform datasets. To support consumer rights requests a brand may receive, brands should use Platform Privacy Service to submit consumer requests of access and delete to remove their data across the data lake, Identity Service, and Real-Time Customer Profile.
- All datasets we use for input/output of models will follow Platform guidelines. Platform Data Encryption applies for data at-rest and in-transit. See the documentation to learn more about data encryption
Next steps next-steps
Once you have prepared your data and have all your credentials and schemas in place, start by following the Attribution AI user guide. This guide walks you through creating an instance for Attribution AI.