PATRA: Bridging the Gap Between Temporal Dynamics and LLM Reasoning
PATRA: Pattern-Aware Alignment and Balanced Reasoning for Time Series Question Answering
PATRA is a novel framework for Time Series Question Answering (TSQA) that achieves SOTA performance by introducing pattern-aware alignment and reinforcement learning. It decomposes time series into trend and seasonal components to deeply align with textual queries, outperforming models like GPT-4o and ChatTS in complex temporal reasoning.
TL;DR
Time Series Question Answering (TSQA) requires more than just "reading" numbers; it requires an intuitive understanding of trends and cycles. PATRA (Pattern-Aware Alignment and Balanced Reasoning) shifts the paradigm by explicitly decomposing time series into physical patterns (trend/seasonality) and using a balanced Reinforcement Learning strategy to ensure the model doesn't just "guess" the easy answers but actually reasons through the data.
Problem: The "Shallow Alignment" Trap
Most current approaches to TSQA fall into two camps:
- Textualization: Converting numbers to text tokens (e.g., "1.2, 1.5..."). This loses the "shape" of the data and overwhelms the LLM's context window.
- Shallow Multimodal Fusion: Treating time series like image patches. Unlike a cat in a photo, a data point in a time series only makes sense in the context of its underlying Trend and Seasonality.
Moreover, training an LLM on multiple tasks (e.g., "What is the max value?" vs. "Why did this crash happen?") often fails because the model takes the path of least resistance, mastering the easy "Max Value" questions while ignoring the complex logical "Why" questions.
Methodology: Deep Pattern Alignment
PATRA introduces a Pattern-Aware Alignment Mechanism. Instead of feeding raw embeddings to the LLM, it performs Latent Decomposition.
1. Latent Decomposition
The model splits the time series embeddings into three distinct streams:
- Trend: Captures the long-term direction.
- Seasonality: Captures periodic fluctuations.
- Full Component: Captures the holistic raw signal.
2. Multi-Level Cross-Modal Interaction
PATRA uses Learnable Alignment Tokens (LATs) to extract query-specific information from the text and align it with these three temporal patterns. This ensures that when a user asks about a "cycle," the model specifically looks at the seasonal latent features.
Figure: The PATRA framework, showcasing the interaction between the Pattern-Aware Alignment module and the LLM Backbone.
Reinforcement Learning with Balanced Rewards
To solve the optimization imbalance, PATRA utilizes Group Relative Policy Optimization (GRPO). The key innovation here is the Balanced Task Reward.
- For Labeled Tasks: A stage-wise reward (checking format -> range -> correctness).
- For Generative Tasks: A ROUGE-L based semantic score.
- Normalization: Both rewards are mapped to a unified scale [0, 2] to prevent the model from "hacking" the easy accuracy-based tasks at the expense of complex reasoning.
Experimental Performance
PATRA was tested against heavyweights like GPT-4o and DeepSeek-R1.
| Task | PATRA (7B) | ChatTS (7B) | Qwen2.5 (7B) |
|---|---|---|---|
| Comprehension (Acc) | 56.03% | 44.83% | 42.24% |
| Reasoning (Acc) | 44.59% | 22.97% | 36.48% |
| Prescience (Acc) | 52.78% | 25.92% | 26.85% |
Beyond standard benchmarks, PATRA demonstrated incredible Out-of-Domain (OOD) capabilities. In financial trend prediction (Finance 7-days), it achieved an MSE of 0.191, significantly better than GPT-4o's 0.365.
Table: PATRA outperforms both textual and multimodal baselines across all TSQA dimensions.
Critical Insights: Why it Works
The ablation studies prove that Multi-Pattern Alignment is superior to single-pattern alignment. By visualizing the latent space, the authors proved that the "Latent Trend" extracted by the model has a 0.986 Pearson correlation with the actual physical trend. This confirms the model isn't just seeing noise; it is seeing the "physics" of the data.
Limitations
Despite its strength, the model can sometimes suffer from Attention Bias in highly volatile data. If a massive, non-representative spike occurs, the pattern-aware module might over-index on that local feature, leading to a "False Positive" in trend detection.
Conclusion
PATRA marks a move toward "expert-level" AI for time series. By moving from shallow patch-based vision-mimicry to deep, pattern-aware decomposition, it allows LLMs to ground their reasoning in the actual dynamics of the physical world. For industries like finance and energy, this is a significant step toward reliable, automated decision-support systems.