Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
SHAP for Time Series (Temporal SHAP)
Adaptation of the SHAP (SHapley Additive exPlanations) method that takes into account the temporal dependency of observations to assign importance to each time step in the final prediction.
LSTM-Attribution
Interpretability technique specific to LSTM-type recurrent neural networks, which quantifies the contribution of each hidden state or memory cell to the model's output.
Guided Temporal Perturbation
Interpretability approach that systematically modifies segments of the time series to observe the impact on prediction, using heuristics to target the most influential periods.
Sequential Saliency Map
Generation of saliency maps adapted to sequential data, where each point in the sequence receives an importance score based on the gradient of the output with respect to the input at that specific time.
Temporal Integrated Gradients
Extension of the Integrated Gradients method that integrates gradients along a path in the time series space, often starting from a baseline sequence (e.g., zeros or an average sequence).
Temporal Counterfactual Explanation
Generation of a minimal alternative time sequence that would have led to a different prediction, allowing understanding of the critical conditions for the model's decision.
Functional ANOVA Decomposition
Statistical method that decomposes the prediction function of a temporal model into main effects (individual periods) and interaction effects (combined periods) to quantify their influence.
Wavelet-based Interpretability
Use of the wavelet transform to decompose the time series into different frequencies and locate the patterns that most influence the model's prediction.
Temporal Association Rules
Extraction of rules of the type 'if pattern A occurs at time t, then prediction B' to explain the model's behavior in terms of understandable temporal patterns.
LIME for Time Series (Time-LIME)
Adaptation of LIME (Local Interpretable Model-agnostic Explanations) that samples segments of the time series to create a local linear model explaining the prediction at a given point.
Temporal Influence Profile
Graphical representation of the impact of each past time step on the current prediction, revealing the model's relevant memory or horizon for a specific task.
Causal Sensitivity Analysis
Evaluation of the model's sensitivity to causal interventions on the time series, distinguishing correlation from causation for a more robust interpretation.
Temporal Prototype Explanation
Method that identifies prototype temporal sequences (most representative) of a prediction class and explains a new prediction by its similarity to these prototypes.
Temporal Error Decomposition
Technique that dissociates the model's prediction error into components linked to specific phases of the time series (e.g., noise, trend, seasonality) to target weaknesses.
Temporal Surrogate Model Interpretation
Training a simple and interpretable model (e.g., ARIMA, linear regression) to locally approximate the behavior of a complex model (e.g., neural network) on a given time window.
RNN Hidden State Visualization
Set of techniques (e.g., PCA, t-SNE) applied to the hidden state vectors of RNNs to visualize the model's internal dynamics and identify phases of learning temporal patterns.