What This Chart Shows
This chart breaks down each channel's attributed lift into two components: the Physics Model (interpretable, equation-based attribution) and the Model Residual (AI-learned patterns that capture effects the physics model doesn't explain). This helps you understand how much of each channel's attribution comes from well-understood, explainable relationships versus patterns learned by the machine learning component. Channels with high Physics contribution are more interpretable; channels relying heavily on the Model Residual may warrant closer inspection or additional validation.
Key Questions This Chart Helps Answer
- How much of each channel's attribution comes from the interpretable physics model vs the AI residual?
- Which channels have the most explainable attribution?
- How does the physics vs AI split vary across my top-performing channels?
Axes, Metrics, and Units
Element | Description |
|---|---|
| X-axis | Channel / Series: Top 15 channels by absolute total contribution. |
| Y-axis | Lift Contribution (KPI Units): Attributed lift for each channel, split by model track. |
| Blue bars | Physics Model: Attribution from the interpretable, equation-based model. |
| Purple bars | Model Residual: Attribution from the AI-learned residual, distributed proportionally across channels by their physics contribution. |
| Bar mode | Stacked (default) or grouped, depending on plot type setting. |
| Hover | Shows series name, Physics contribution, and Model Residual contribution. |
Model Tracks Explained:
Track | Description |
|---|---|
| Physics Model | The interpretable, equation-based component of the model. Captures known causal relationships (e.g., spend → response, lag structures). |
| Model Residual | The AI-learned component that captures patterns not explained by the physics model. Distributed to channels proportionally by their absolute physics contribution. |
All values are modeled outputs in KPI units (e.g., revenue, conversions).
Control Options Reference
Control | Meaning |
|---|---|
| Plot type | Stacked (default) or Grouped bar layout. |
How to Interpret the Results
- Stacked view: Total bar height = Physics + Model Residual = Total attributed lift for that channel. The split shows composition.
- High Physics (blue) proportion: The channel's attribution is mostly from the interpretable model—higher confidence in the explanation.
- High Model Residual (purple) proportion: The channel relies more on AI-learned patterns. The attribution may be accurate, but the "why" is less clear.
- Compare across channels: Some channels may have balanced splits; others may lean heavily one way. This reflects the nature of the data and causal relationships.
- Negative values: Either track can be negative (e.g., if the physics model predicts a negative effect or the residual corrects downward). Stacking shows the net effect.
Practical Applications for Marketers
Application | How to use this chart |
|---|---|
| Interpretability assessment | Prioritize channels with high Physics contribution when you need explainable results for stakeholders. |
| Validation targeting | Channels with high Model Residual may benefit from additional testing or holdout experiments to validate the AI-learned effects. |
| Model trust calibration | Understand which channels are driven by known causal mechanisms vs learned patterns when making high-stakes decisions. |
| Diagnostic review | A sudden shift in the Physics/Residual split for a channel may indicate data changes or model updates worth investigating. |
| Stakeholder communication | Use this chart to explain that attribution comes from a rigorous two-track system, not a black-box model. |
Common Mistakes & Misinterpretations
Mistake | Why it is a problem | How to avoid |
|---|---|---|
| Assuming Model Residual = wrong | The AI residual captures real effects that the physics model doesn't fully explain. It's not an error term—it's additional signal. | Treat Model Residual as "learned effects," not noise or error. |
| Ignoring channels with high Model Residual | These channels may still be valuable; they just have less interpretable attribution. Ignoring them could mean missing important performance. | Investigate rather than dismiss; consider validation experiments. |
| Expecting the split to be consistent over time | The Physics/Residual balance can shift with data changes, model updates, or market dynamics. | Monitor trends over time; don't assume a static split. |
| Treating Physics as "certain" | Physics model attribution is more interpretable, but not guaranteed to be correct. It's based on assumed causal structures. | Use Physics as higher-confidence, not infallible. |
| Comparing channels with very different scales | A channel with 90% Physics may have much smaller total lift than one with 50% Physics. The percentage doesn't tell the whole story. | Always consider absolute lift alongside the percentage split. |
Caveats & Considerations
- Residual distribution: The Model Residual is a daily aggregate distributed across channels proportionally by their absolute physics contribution. This is an allocation, not a direct per-channel measurement.
- Top 15 channels: Only the top 15 channels by absolute total contribution are shown. Smaller channels are not displayed.
- Negative contributions: Both Physics and Model Residual can be negative. A channel with positive Physics but negative Residual shows the AI model correcting downward.
- Stacked vs grouped: Stacked view (default) shows total lift as bar height with composition. Grouped view shows the two tracks side-by-side for easier comparison of each component's magnitude.
- Model version dependence: The Physics/Residual split depends on the model configuration. Changes to the physics equations or AI training can shift the balance.