Machine Learning

Explainable Predictions

The Explainable Predictions feature explains the Machine Learning predictions for each customer profile in 360.

By leveraging this feature, marketers can understand the input factors that contribute towards the machine learning prediction made for a given customer profile. Therefore, marketers can understand all the factors that affect an individual’s prediction, and design campaigns that take all such factors into account for individuals with similar demographics.

Marketers can leverage this mechanism for building campaigns where the contributing factors for individuals are used to target specific demographic behavior, which are otherwise opaque to traditional analytics and reporting.

Description and usage

To access this feature:

  1. Sign in to the CDP user interface.

  2. Click 360 Profiles.

  3. Click the Machine Learning tab.

  4. Click Explain Prediction.

    The system displays a window that explains the positive and negative contribution of the input features to the machine learning prediction of the individual. You can compare these values relative to the baseline prediction or the prediction that does not account for the input factors, and uses the model’s global parameters. You can view the explanation for an individual’s Likelihood to Buy prediction. Hovering on each respective bar shows the relevant explanation for each factor.

    Note

    The Explain Prediction option is available for all supervised models such as Likelihood To Buy, Predictive Lifetime Value, and Likelihood To Pay Full Price. This option is unavailable for unsupervised models such as Behavior Clusters, Product Clustering, and Fuzzy Clustering.

In the preceding example, the profile of Elysee Maunton has a predicted Likelihood to Buy score of 2 - High. The positively contributing factors in the graph indicate that:

  • Elysee is 22.95% more Likely to Buy than an average customer because of Total Transaction Counts over All time.

  • Elysee is 1.02% more Likely to Buy than an average customer because of Number of Purchased Products over All Time.

Total Transaction Counts over All time and Number of Purchased Products over All Time contributed positively towards boosting the score from average to 2-High.

Insight: This indicates that the user made above-average transactions in the past in terms of the number of transactions and the number of products purchased above an average customer’s engagement levels.

Based on the preceding example and negatively contributing factors, the graph indicates that:

  • Elysee is 8.61% less Likely to Buy than an average customer because of Number of days since first transaction.

  • Elysee is 3.89% less Likely to Buy than an average customer because of Number of transactions in last 30 Days.

Number of days since first transaction and Number of transactions in last 30 Days contributed towards reducing the Likelihood to Buy score.

Insight: The user has above-average engagement with the brand, outside the 30-day time window. However, the recent engagement is on the decline, leading to a slight drop in the Likelihood to Buy score.

Conclusion: Elysee is an ideal candidate for a discount or bundle campaign. Running a time-bound campaign on such a user can maximize the conversion.

Suggestions and tips

To gain accurate and aggregate view of demographic behavior, marketers can:

  • Filter the 360 profiles by a demographic criteria like City.

  • Aggregate the contributing factors from a varied selection of profiles to arrive at a conclusion and campaign hypothesis.

For example, for a city Tacoma, it is useful to look at the details of multiple profiles to understand how the Likelihood to Buy contributing factors vary. In the preceding screenshots, the contributing factors vary for:

  • Elysee with a 2 - High Likelihood to Buy score.

  • Kingsy Beveredge with a 4 - Medium Likelihood to Buy score.

  • Gannon Lourens with a 7 - Low Likelihood to Buy score.

The three profiles indicate that:

  • Elysee Maunton’s 2 - High Likelihood to Buy score is bolstered by higher Total Transactions Count and lower Number of Days since first transaction.

  • Kingsley Bevergede’s 4 - Medium Likelihood to Buy (lower score compared to Elysee Maunton) was brought down by the Email send since last time score, suggesting that the user did not engage on that channel through campaigns, and therefore, lookalikes can be a target for a “Revive” campaign.

  • Gannon Loure’s 7 - Low Likelihood to Buy score is slightly bolstered by Transaction date since first transaction (suggesting recent purchase). This suggests that the customer initially made a few purchases but engagement is low subsequently. This can be an ideal case for a discount and re-engage campaign.

Feature dictionary

It is important to understand the input features that feed a machine learning model. To surface the input parameters and their human readable definitions in CDP, you can leverage the Feature Dictionary feature available in all the machine learning models. These are available in the individual dashboards for all machine learning models.

For example, in the Likelihood to Buy ML dashboard, you can scroll down to the Explainability section to view the Feature Dictionary tile. This contains the input features for the Likelihood to Buy model along with their human readable descriptions.

For the Likelihood to Convert model, the Feature Dictionary tile appears as follows:

You can access the same dashboards from 360 Profiles. To view the dashboard, you must click the i button adjacent to the Input Features label.

Did not find what you were looking for?

If this content did not answer your questions, try searching or contacting our support team for further assistance.

Acquia Help

Filter by product:

Machine Learning common questions