PatchDecomp: Interpretable Patch-Based Time Series Forecasting

Imagine you are trying to predict the weather for next week. You have a super-smart computer model that looks at temperature, wind speed, and humidity from the last few days to make its guess.

The Problem:
The computer gives you a perfect prediction, but when you ask, "Why did you think it would rain on Tuesday?" it just says, "Because I said so." It's like a "black box." In the real world (like managing a factory or a power grid), people don't trust predictions they can't understand. They need to know why the model made that decision.

The Solution: PatchDecomp
The authors of this paper created a new method called PatchDecomp. Think of it as a "Time-Traveling Detective" that doesn't just give you an answer, but shows you the evidence.

Here is how it works, using simple analogies:

1. The "Chunking" Strategy (Patching)

Most old models look at time series data (like stock prices or electricity usage) one second at a time. It's like trying to understand a movie by looking at a single frame every second. It's too fast and misses the bigger picture.

PatchDecomp takes a different approach. It chops the timeline into chunks (which they call "patches").

Analogy: Imagine reading a book. Instead of analyzing every single letter, you read it in sentences or paragraphs.
Why it helps: A "patch" might be "the last 24 hours of electricity usage." The model treats that whole day as one unit of information, which makes it easier to spot patterns (like "it's always high on Mondays").

2. The "Recipe" Approach (Decomposition)

This is the magic trick. When the model makes a prediction, it doesn't just spit out a number. It breaks that number down into a recipe.

The Analogy: Imagine you bake a cake and it tastes amazing.
- Old Models: "Here is the cake. It's delicious." (No explanation).
- PatchDecomp: "Here is the cake. It is made of 40% flour from yesterday, 30% sugar from this morning, and 30% eggs from next week's forecast."

PatchDecomp calculates exactly how much each "chunk" of data contributed to the final prediction.

Did the past data (yesterday's temperature) push the prediction up?
Did the future data (a known holiday next week) push it down?
Did an external factor (like a sudden windstorm forecast) change the result?

It draws a chart showing these contributions as colored areas, so you can literally see which piece of the puzzle mattered most.

3. Why This Matters (The "Trust" Factor)

The paper tested this on real-world data, like electricity prices and traffic flow.

Accuracy: It was just as good at predicting the future as the most complex, "black box" models currently in use.
Interpretability: This is the big win. Because it breaks the prediction down by "patches," a human operator can look at the chart and say, "Ah, I see! The model predicted high prices because the system load (demand) is expected to spike tomorrow, not because of some mysterious glitch."

The Big Picture

Think of PatchDecomp as a translator between the complex language of AI and the human language of logic.

Before: "The AI predicts a spike." (User: Confused and suspicious.)
After: "The AI predicts a spike because the 'Tuesday Morning' patch of data and the 'Wind Forecast' patch both strongly suggest it." (User: Understands and trusts the decision.)

In short, PatchDecomp proves you don't have to sacrifice accuracy to get clarity. It lets the AI show its work, just like a good math teacher, so we can trust the answers it gives us.

1. Problem Statement

Time series forecasting (TSF) is critical for domains like manufacturing, logistics, and energy management. While deep learning models (e.g., Transformers, MLPs) have achieved high accuracy, they are often treated as "black boxes," lacking interpretability. This opacity hinders trust and deployment in safety-critical real-world applications.

Existing interpretability methods face specific limitations:

Post-hoc methods (e.g., LIME, SHAP) are computationally expensive and approximate.
Inherently interpretable models often rely on visualizing attention weights or linear matrices. However, these do not clearly explain how specific subsequences (patches) of input variables (including exogenous variables) contribute to the final prediction.
There is a lack of methods that can rigorously decompose the contribution of input patches to output predictions while maintaining state-of-the-art accuracy.

2. Methodology: PatchDecomp

PatchDecomp is a neural network-based TSF model designed to be inherently interpretable by decomposing predictions based on input patches. It handles target variables, historical exogenous variables, future exogenous variables (e.g., weather forecasts), and static variables.

Core Architecture

The model consists of two main components:

A. Patch Encoder

Normalization & Patching: Input time series are standardized using Reversible Instance Normalization (RevIN). The series is divided into fixed-length subsequences (patches) of length $P$ .
Embedding:
- Patch Embedding: Each patch is linearly transformed into a latent vector.
- Positional Encoding: Temporal information is added via linear transformation or embedding.
- Static Variable Embedding: Static variables (e.g., product ID) are embedded and added to the latent vectors.
Encoding: A residual block (combining MLP and skip connections) processes these vectors $N_{enc}$ times to produce latent representations ( $z_{src}$ ) for input patches and ( $z_{tgt}$ ) for output patches.

B. Patch Decoder

Multi-Head Attention: The decoder uses a multi-head attention mechanism where:
- Query: Representations of output patches ( $z_{tgt}$ ).
- Key & Value: Representations of input patches ( $z_{src}$ ).
Decomposable Attention Mechanism: Unlike standard attention which aggregates values, PatchDecomp modifies the computation to preserve contribution information. It computes the element-wise product of the attention weights and the value tensors, then sums along the patch dimension.
Bias Addition: A learnable bias vector is added to the result to account for patch-specific contributions.
Output Generation: The final latent vector is linearly transformed and denormalized to produce the forecast.

Interpretability Mechanism

The key innovation is that the model explicitly calculates the contribution of each input patch to the final prediction.

The output is not just a single value but a sum of contributions from specific input patches.
This allows users to visualize exactly which time segments of which variables (target or exogenous) drove the prediction.

3. Key Contributions

PatchDecomp Model: A novel TSF architecture that achieves high accuracy while providing inherent interpretability. It handles exogenous variables and decomposes predictions by input patch.
Rigorous Decomposition: Unlike methods that only visualize attention maps, PatchDecomp establishes a direct correspondence between input patches and output values through the entire processing pipeline (encoder to decoder), allowing for precise attribution.
Comprehensive Evaluation: The authors validated the model on multiple benchmark datasets (LTSF) and a complex Electricity Price Forecasting (EPF) task with exogenous variables.
Quantitative Interpretability Metric: The paper introduces the use of Comprehensiveness (AOPCR) to quantitatively prove that the model's identified important patches actually drive the prediction accuracy.

4. Experimental Results

Datasets & Baselines

Long-Term Forecasting (LTSF): Tested on 7 datasets (ETTh1/2, ETTm1/2, Weather, ECL, Traffic).
Electricity Price Forecasting (EPF): Tested on 5 markets (NP, PJM, BE, FR, DE) with exogenous variables (system load, generation).
Baselines: PatchTST, NBEATSx, NHITS, TFT, DLinear, TSMixer, Autoformer, iTransformer, TiDE.

Accuracy Performance

LTSF: PatchDecomp achieved competitive results, often ranking in the top tier. It performed exceptionally well on ETTh1 and ETTh2, likely due to the alignment between the data's periodicity and the patch length.
EPF: PatchDecomp and TFT achieved the highest accuracies among models handling exogenous variables.
Statistical Significance: Critical difference diagrams (Conover's test) confirmed that PatchDecomp belongs to the highest-ranking group for MSE and the second-highest for MAE, proving it does not sacrifice accuracy for interpretability.

Interpretability Analysis

Qualitative: Visualizations (e.g., Figure 4 in the paper) show that PatchDecomp can clearly identify dominant patches (e.g., "system load from the previous day" or "future generation forecasts") driving the price prediction. In contrast, TFT's attention weights were diffuse and harder to interpret.
Quantitative (AOPCR): When the top $k\%$ of important patches identified by PatchDecomp were removed, the prediction error (MAE) increased significantly more than when random patches were removed or when patches were removed based on TFT's importance scores. This proves PatchDecomp's explanations are comprehensive and truly reflect the model's decision-making logic.

5. Significance

Bridging the Gap: PatchDecomp successfully bridges the gap between high-performance deep learning and the need for transparency in industrial applications.
Actionable Insights: By providing patch-level explanations rather than point-wise or variable-level averages, the model offers granular insights (e.g., "the prediction was driven by the load spike 24 hours ago"). This is crucial for debugging models and validating forecasts in domains like energy trading.
Future Direction: The paper highlights that while static variables are currently embedded, their specific contribution decomposition is not yet visualized, identifying a clear path for future research.

In conclusion, PatchDecomp demonstrates that it is possible to design neural networks that are both highly accurate and rigorously interpretable, moving beyond the "black box" paradigm in time series forecasting.