TLDR: PAX-TS is a new, flexible method for explaining why time series forecasting models make their predictions. It works by making small, localized changes to the input data and observing the impact, providing explanations at different levels of detail. This includes understanding the importance of specific time steps, identifying overall temporal patterns, and characterizing how different data channels influence each other in multivariate forecasts. PAX-TS helps users understand complex AI forecasts, answer practical questions, and builds trust in predictive models.
Time series forecasting, a critical tool for predicting everything from stock prices to weather patterns, has seen remarkable advancements thanks to sophisticated models like transformers and large language models. However, these powerful models often operate like ‘black boxes,’ making predictions without revealing the ‘why’ behind them. This lack of transparency poses a significant challenge for users who need to understand and trust these forecasts.
Existing methods for explaining artificial intelligence (XAI), such as LIME, were primarily designed for classification tasks and struggle to provide meaningful insights for time series forecasting, where models predict multiple values over time rather than a single label. This gap leaves practitioners and domain experts without answers to crucial questions like, “Why is the forecast for productivity at 3 PM so low?” or “How can we increase the overall trend of the forecast?”
Introducing PAX-TS: A New Approach to Explainable Time Series Forecasting
A new research paper introduces PAX-TS (Perturbation Analysis eXplanations for Time Series forecasting), a novel, model-agnostic, and post-hoc algorithm designed specifically to address these challenges. PAX-TS offers multi-granular explanations for time series forecasting models and their predictions, making complex AI outputs more understandable. The method is based on localized input perturbations, meaning it makes small, targeted changes to the input data and observes how these changes affect the forecast.
One of the key strengths of PAX-TS is its ability to provide explanations at different levels of detail. From high-level insights into cross-channel correlations in multivariate time series to granular, time-step-level importance explanations, PAX-TS offers a comprehensive view of a model’s behavior. It can even characterize how different data channels influence each other in complex multivariate scenarios.
How PAX-TS Works
At its core, PAX-TS works by systematically perturbing (making small changes to) specific parts of the input time series. These perturbations can be localized to a particular time step, scaled based on summary statistics like mean or variance, or adjusted to reflect changes in the overall trend. After perturbing the input, PAX-TS feeds this altered data to the forecasting model and compares the new forecast to the original one. By analyzing the differences, it calculates a “change ratio” that quantifies how much a specific input change affects a particular aspect of the forecast, such as its maximum value, mean, or the value at a given future time step.
For example, if you want to know which input time steps are most important for predicting the maximum value of a future forecast, PAX-TS can highlight those specific input points and show their positive or negative correlation. This level of detail helps users understand not just what the model predicts, but also which historical data points are driving those predictions.
Uncovering Temporal Patterns and Cross-Channel Relationships
The researchers conducted extensive experiments, evaluating PAX-TS with seven state-of-the-art forecasting algorithms across ten diverse datasets. They found that PAX-TS effectively captures a model’s behavior, revealing noticeable differences in explanations between high-performing and low-performing algorithms on the same datasets.
A particularly insightful aspect of PAX-TS is its ability to visualize temporal dependency patterns. By creating heatmaps that show the correlation between all input time steps and all forecast time steps, PAX-TS identified six distinct classes of patterns. These patterns, such as “Diagonals” (indicating seasonal behavior) or “Bipolar Regions” (strong positive and negative correlations), were found to be strong indicators of forecasting performance. Models exhibiting diagonal or bipolar regional patterns generally achieved the highest accuracy, while those showing a “Last-Timestep” or “Fully Correlated” pattern often performed poorly.
For multivariate time series, PAX-TS excels at illustrating cross-channel correlations. It can generate a graph showing how perturbing one data channel (e.g., indoor temperature) affects the forecasts of other channels (e.g., CO2 concentration). This is invaluable for understanding complex systems where multiple variables interact.
PAX-TS vs. Other Explainability Methods
The paper also compares PAX-TS to other leading explainability algorithms for forecasting, such as ShapTime and TS-MULE. While these methods offer some insights, PAX-TS provides both higher and lower granularity explanations. Unlike its counterparts, PAX-TS can derive importance scores for specific properties of interest (like extrema or summary statistics) and is uniquely suited for analyzing multivariate cross-channel correlations and temporal dependency patterns. Furthermore, PAX-TS is computationally efficient, scaling linearly with the inference time of the underlying forecasting model, making it practical for real-time applications.
Also Read:
- Fairer Explanations for AI: Introducing MC3G for Actionable Recourse
- GraphFNP: A Unified Approach for Trustworthy and Explainable Graph Neural Networks
Practical Applications and Future Outlook
PAX-TS offers a top-down approach to understanding forecasts. Users can start with a broad overview of cross-channel correlations, then zoom into specific index correlations, and finally examine time-step importance for a particular forecast property. This flexibility makes it suitable for a wide range of real-world scenarios, from optimizing irrigation systems in agriculture by explaining soil moisture forecasts to providing detailed CO2 concentration forecasts for facility managers in smart buildings, or even enhancing hospital occupancy predictions.
By making time series forecasting models more transparent, PAX-TS empowers end-users to ask and answer practical questions about their forecasts, fostering greater trust and enabling more informed decision-making. The research paper, PAX-TS: Model-agnostic multi-granular explanations for time series forecasting via localized perturbations, details the algorithmic procedure and empirical evaluations.
