DeepLIFT (Deep Learning Important FeaTures)

DeepLIFT computes feature importance by comparing the activation of each neuron to a reference activation derived from a neutral or baseline input. This reference represents a 'default' state—such as a zeroed-out transaction or an average user profile—against which deviations are measured.

• The difference-from-reference approach avoids saturation problems that plague gradient-only methods
• For fraud detection, the reference could be a legitimate transaction template, highlighting anomalous deviations
• Enables meaningful attribution even when gradients are near zero (e.g., deep in a ReLU network's inactive region)
• The choice of reference is critical: a well-chosen reference yields explanations aligned with domain expectations

DeepLIFT enforces the Summation-to-Delta axiom, which requires that the sum of all feature contributions equals the difference between the model's output for the actual input and its output for the reference input. This guarantees completeness.

• Mathematically: Σᵢ Cᵢ = F(x) - F(x₀), where Cᵢ is the contribution of feature i
• Ensures no attribution is lost or double-counted during backpropagation
• Critical for regulatory audit trails in fraud detection, where total risk scores must be fully decomposable
• Contrasts with methods like LRP where conservation may not hold by design

DeepLIFT defines a multiplier for each neuron: the ratio of its contribution to the difference-from-reference in its input. These multipliers are backpropagated through the network using chain-rule-like rules specific to each activation function.

• For a linear layer, the multiplier is simply the weight; for ReLU, it depends on whether the neuron is active in both actual and reference inputs
• The Rescale Rule and RevealCancel Rule handle nonlinearities differently, with RevealCancel addressing cases where positive and negative contributions cancel
• This multiplier framework allows efficient, single-pass attribution without requiring multiple forward/backward passes
• Enables real-time explanation generation in real-time fraud scoring pipelines

A key advantage of DeepLIFT over gradient-based methods like Integrated Gradients is its ability to handle saturation: regions where a neuron's output is insensitive to input changes despite the input being highly influential.

• In a sigmoid or tanh neuron near its asymptote, gradients approach zero, masking true feature importance
• DeepLIFT uses the difference-from-reference to bypass this, attributing based on the actual activation delta
• For fraud models with deep architectures, this prevents vanishing attribution signals in saturated layers
• Particularly valuable when explaining decisions on extreme outlier transactions where activations saturate

DeepLIFT offers two distinct rules for propagating contributions through nonlinear activations, each suited to different explanation needs.

• Rescale Rule: Assigns contributions proportionally to the ratio of output delta to input delta; simple and efficient but can mask cancelling effects
• RevealCancel Rule: Separately tracks positive and negative contributions, revealing when features have opposing influences that cancel in the final output
• In fraud detection, RevealCancel can expose cases where a legitimate feature (e.g., known merchant) offsets a suspicious one (e.g., unusual location), providing nuanced reason codes
• The choice of rule impacts the granularity and interpretability of the resulting feature attributions

DeepLIFT provides post-hoc explainability for deep neural networks used in financial fraud detection, decomposing anomaly scores into per-feature contributions that compliance teams can audit.

• Enables generation of adverse action reason codes by ranking feature contributions for a flagged transaction
• Supports algorithmic audit trail requirements by providing deterministic, reproducible attributions for every decision
• Integrates with model governance frameworks by offering mathematically grounded explanations that satisfy regulatory scrutiny
• Compared to SHAP, DeepLIFT is computationally lighter for deep networks, making it suitable for batch explanation of high-volume transaction data