Sequence Influence Function

The core mechanism relies on a first-order Taylor expansion of the model's loss function. By computing the gradient of the loss with respect to the model parameters for both a training sequence and a test sequence, the method estimates the change in the test loss if that training sequence were infinitesimally up-weighted or removed. This avoids the prohibitive cost of full leave-one-out retraining.

The calculation requires the inverse Hessian-vector product (iHVP) of the model's loss. The Hessian matrix captures the local curvature of the loss landscape and quantifies how resistant the model's parameters are to change. Key implementation details include:

• LiSSA Algorithm: A stochastic estimator that avoids explicitly computing or inverting the full Hessian matrix.
• Conjugate Gradients: An alternative iterative solver for approximating the iHVP with higher precision for convex objectives.

The influence score can be positive or negative. A high positive influence score indicates that removing the training sequence would decrease the test loss, meaning the sequence was harmful to the model's performance on that specific test example. Conversely, a negative score identifies a helpful training sequence whose removal would increase test loss. This enables precise data debugging.

When the training sequence and the test sequence are the same, the calculation yields a self-influence score. Sequences with disproportionately high self-influence are often outliers, mislabeled examples, or duplicates that dominate the model's local parameter updates. This property makes influence functions a powerful tool for automated data quality auditing without manual inspection.

Exact influence calculation scales quadratically with the number of training examples, making it intractable for modern deep learning datasets. Practical implementations rely on:

• Stochastic gradient approximations to reduce the computational graph.
• Layer-specific influence that restricts the Hessian calculation to the final layers of a sequence model.
• K-FAC (Kronecker-Factored Approximate Curvature) to efficiently approximate the Fisher information matrix as a block-diagonal structure.

Unlike feature attribution methods that score individual time steps, the Sequence Influence Function treats an entire training trajectory as an atomic unit. This is critical for applications where the context of the full sequence matters, such as identifying a specific patient's longitudinal health record that biases a diagnostic model or a financial time series that causes a trading algorithm to overfit to a historical regime.

Adapts Shapley values to the time domain, fairly distributing a prediction's outcome among individual time steps. It treats each time step as a cooperative game player, computing the average marginal contribution of a step across all possible step subsets. This provides a theoretically grounded, consistent attribution that satisfies axioms like local accuracy and consistency, making it a gold standard for high-stakes auditing in finance.

A direct perturbation method that systematically removes or masks individual time steps from a sequence to measure the resulting change in the model's output. By observing the prediction delta, engineers rank steps by importance. While computationally straightforward, it can miss interaction effects where two steps jointly influence a prediction. Often used as a baseline to validate more sophisticated attribution techniques.

Generates a minimally altered alternative time series that would change the model's forecast to a desired outcome. Unlike influence functions that look backward to training data, counterfactuals look forward to answer 'what if' questions. For example: 'What minimal change in Q2 spending would have flipped the bankruptcy prediction?' Critical for recourse generation and regulatory compliance.

Moves beyond correlation to identify the true causal drivers of a forecast using structural causal models or intervention analysis. While influence functions identify statistical leverage, causal attribution answers whether a training sequence caused the model to learn a specific pattern. Techniques like Granger causality and PCMCI algorithms are employed to validate that attributions reflect genuine mechanistic relationships.

A quantitative evaluation score that measures how accurately an explanation reflects the model's true reasoning. It tests whether highly attributed time steps are also the most sensitive to perturbation. A faithful explanation will show a strong correlation between attributed importance and actual model behavior change when those steps are altered. Essential for comparing and validating different explainability methods.

A diagnostic process that visualizes and quantifies the activations of forget, input, and output gates in an LSTM network. By tracking when the forget gate erases memory or the input gate writes new information, engineers can explain how the model manages long-range dependencies. This mechanistic view complements influence functions by revealing how a training sequence's signal propagates through the architecture.