Skip to main content
The Model Explainability section provides transparency into how the model arrives at its predictions. By utilizing SHAP (SHapley Additive exPlanations) values, the platform breaks down global feature importance, granular feature behavior, and individual prediction logic. This ensures that users can trust, validate, and understand the core drivers behind the model’s targeting recommendations.

​
Global Feature Importance

Global Feature Importance
This chart ranks the model’s variables based on their overall impact on predictions across the entire holdout dataset. The bar chart displays the mean absolute SHAP value for each feature. The longer the bar, the more influential that specific variable is when the model calculates a prediction.

​
What this chart tells the user

  • Primary Drivers: The top-ranked features represent the most critical data points for the model. For instance, age can act as an intuitive anchor for decision-making because many users understand how life stage may influence response likelihood.
  • Interpretation by Business Case:
    • In Classification (e.g., DM Propensity): A long bar means this feature heavily dictates whether a user’s probability of responding shifts up or down.
    • In Regression (e.g., DM Expected Amount): A long bar means this feature heavily dictates the predicted donation amount assigned to a donor.
  • The Long Tail of Data: The “Sum of 128 other features” at the bottom highlights that while individual secondary features may each have a smaller impact, their combined contribution can still be meaningful. For example, a model may have no additional named features beyond the top drivers, in which case there is no long-tail contribution; in other cases, many small features together can add a high cumulative impact and materially refine the prediction.

​
Feature Impact and Directional Behavior (Beeswarm)

Feature Impact and Directional Behavior
This chart combines feature importance with directional impact, showing not just which features matter, but how high or low values of those features push the model’s final prediction score. Each dot represents a single individual in the holdout dataset.
  • Vertical Position: Groups the dots by feature (ordered from most important at the top to least important).
  • Horizontal Position (X-axis): Shows the SHAP value. Points to the right of the center line (0.0) indicate a positive impact (increasing the model score); points to the left indicate a negative impact.
  • Color Scale: Represents the actual value of the feature for that individual (Red = High value, Blue = Low value).

​
How to read the colours and direction

  • Positive Correlations (Red on the Right): For features like Age, the red cluster sits entirely on the right side.
    • Classification Example: A high historical response rate (red dots) pushes the score to the right, meaning a higher probability of conversion/outreach success.
    • Regression Example: A high historical response rate (red dots) pushes the score to the right, meaning the model will predict a higher donation probability for this individual.
  • Negative Correlations (Red on the Left / Blue on the Right): For features where blue clusters push to the right, lower raw values provide a stronger lift to the prediction score (e.g., fewer days since the last transaction meaning higher engagement).
  • Density and Outliers: The thickness of the “swarms” indicates how common a specific behavior is across the dataset, helping users spot whether a feature affects the majority of the audience or just a select few outliers.

​
Feature Dependence Analysis

Feature Dependence Analysis
This scatter plot isolates a single key feature to map exactly how its raw values translate into model impact (SHAP values). This view helps users detect non-linear relationships, step-functions, or saturation thresholds in the model’s logic. The x-axis displays the actual raw value of the selected feature (e.g., dm_h_response_rate from 0.0 to 1.0). The y-axis displays the resulting SHAP value (impact on model output). Each dot represents one individual from the holdout dataset.

​
What this chart tells the user

  • Clear Response Thresholds: The distinct “steps” or sudden vertical jumps in the plot (such as around the 0.12 and 0.20 marks) reveal specific inflection points where the model suddenly changes how it treats a user.
  • Dual-Case Interpretation:
    • Classification Context: At an x-value of 0.4, the SHAP value jumps to +0.35, adding a massive boost to the individual’s likelihood to churn or open an email.
    • Regression Context: That same jump at 0.4 translates directly into a baseline increase in the expected lifetime value (LTV) calculated for that profile.
  • Diminishing Returns & Saturation: Users can observe where the curve flattens out. If the SHAP value stops rising after a certain threshold on the x-axis, it indicates that increasing the feature value beyond that point yields no additional predictive lift.

​
Individual Prediction Breakdown (Waterfall)

Individual Prediction Breakdown
The Waterfall chart moves from global patterns to a hyper-local view, detailing exactly why a single specific sample or individual received their final score. It functions as a digital “receipt” for an individual prediction.
  • E[f(x)]E[f(x)] (Baseline): The starting value at the bottom represents the average prediction score across the entire dataset before looking at this individual’s specific data.
  • f(x)f(x) (Final Score): The value at the top represents this specific individual’s final predicted score.
  • Colored Bars:
    • Red Bars (+): Factors that increased this individual’s score above the baseline.
    • Blue Bars (-): Factors that dragged this individual’s score down.

​
What this chart tells the user

  • The Winning/Losing Attributes: It allows users to pinpoint the exact “tipping points” for an individual profile.
    • Classification Application: In this sample, a low historical response rate dragged the score down by -0.14 (reducing their conversion probability), while a long time since the first response (+0.08) pulled the probability back up.
    • Regression Application: In an amount-destination model, the baseline expectation might be 10.Thelowresponserate(−0.14)scalesdownthefinalexpecteddollarvalue,whilepositiveredattributesscaletheexpectedgiftsizebackuptoitsfinalpositionat10. The low response rate (-0.14) scales down the final expected dollar value, while positive red attributes scale the expected gift size back up to its final position at f(x)$.
  • Operational Transparency: This level of granularity is highly valuable for customer success teams or campaign managers who need to understand the precise mix of behavioral history that qualified or disqualified a specific high-value segment.