add TRACE paper

Author: gaukhar-n

app/projects/trace/assets/architecture.png

@@ -0,0 +1,65 @@
+import { Authors, Badges } from '@/components/utils'
+
+# TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval
+
+<Authors 
+    authors="Jialin Chen, Yale University; Ziyu Zhao, McGill University; Gaukhar Nurbek, University of Texas Rio Grande Valley; Aosong Feng, Yale University; Ali Maatouk, Yale University; Leandros Tassiulas, Yale University; Yifeng Gao, University of Texas Rio Grande Valley; Rex Ying, Yale University"
+/>
+
+<Badges
+  venue="NeurIPS 2025"
+  github="https://github.com/Graph-and-Geometric-Learning/TRACE-Multimodal-TSEncoder"
+  arxiv="https://arxiv.org/abs/2506.09114"
+  pdf="https://arxiv.org/pdf/2506.09114"
+/>
+
+## Introduction
+Time-series data is central to domains like healthcare, weather, and energy, yet it rarely exists alone. In real-world settings, it is often paired with rich textual context such as clinical notes or weather reports. This combination calls for models that can jointly understand time-series signals and text.
+As shown in figure below, a flash flood report describing heavy rainfall and strong winds can help retrieve historical time-series patterns with similar dynamics, supporting tasks like forecasting and disaster alerts. But existing approaches remain limited—they often ignore the textual context and struggle to align time-series and language representations effectively.
+![A Use Case of Text-to-Timeseries Retrieval|scale=0.7](./assets/use_case.png)
+## Method
+We introduce TRACE — a Time-series Retriever with Aligned Context Embedding. TRACE is the first multimodal retriever that learns semantically grounded time-series embeddings through fine-grained dual-level alignment. It uses a masked autoencoder with Channel Identity Tokens (CITs) to capture channel-specific behaviors and employs hierarchical hard negative mining to align time-series and textual representations effectively.
+TRACE serves two purposes:
+	1.	As a general-purpose retriever, it enhances foundation models via retrieval-augmented generation (RAG).
+	2.	As a standalone encoder, it achieves state-of-the-art performance on forecasting and classification benchmarks.
+![Overview of TRACE|scale=0.7](./assets/cover_fig.png)
+
+As shown in the figure below, TRACE first learns robust time-series representations through masked reconstruction with channel-aware attention. It then aligns each time-series channel with its corresponding text using fine-grained contrastive learning. Building on this, TRACE introduces a retrieval-augmented generation strategy that fetches relevant context for downstream tasks. This modular design delivers strong standalone performance while integrating seamlessly with existing time-series foundation models.
+![Architecture of TRACE|scale=0.7](./assets/architecture.png)
+
+## Results
+We evaluate TRACE from three perspectives:
+(1) its performance in cross-modal and time-series retrieval compared to strong baselines,
+(2) its effectiveness as a retriever in retrieval-augmented forecasting pipelines, and
+(3) its generalization as a standalone encoder for forecasting and classification.
+
+### Cross-modal Retrieval
+To assess retrieval performance, we replace TRACE’s encoder with several strong time-series foundation models that generate fixed-length embeddings. Each encoder is fine-tuned end-to-end with a lightweight projection layer and a contrastive learning objective for fair comparison.
+As shown in Table 1, TRACE achieves state-of-the-art results, with nearly 90% top-1 label matching and 44% top-1 modality matching. Its retrieval accuracy surpasses the classification performance of all models trained from scratch, underscoring the effectiveness of alignment-based supervision. Among baselines, Moment performs best, but TRACE’s fine-grained embeddings enable more precise cross-modal retrieval and semantic matching.
+
+![Table 1: Retrieval results on 2,000 bidirectional Text–Timeseries query pairs. “Random” indicates a non-informative retriever that ranks candidates uniformly at random. |scale=0.7](./assets/table-1.png)
+
+### Timeseries-to-Timeseries Retrieval
+We tested TRACE on a time-series-to-time-series retrieval task, where the goal is finding the most semantically similar series for each query.
+Table 2 shows TRACE outperforming all baselines—ED, DTW, SAX-VSM, and CTSR—across key metrics: Precision@1, Precision@5, and Mean Reciprocal Rank (MRR). It also maintained the lowest retrieval latency.
+The performance gap highlights a key difference. Methods like SAX-VSM and CTSR struggle to capture deeper temporal and semantic patterns. TRACE's alignment-aware training, by contrast, delivers accurate and efficient retrieval across multivariate signals while remaining scalable.
+
+![Table 2: TS-to-TS Retrieval performance comparison. Evaluation is conducted over 1000 randomly sampled weather time-series queries. |scale=0.3](./assets/table-2.png)
+
+### Retrieval-augmented Time Series Forecasting
+We used TRACE to find the most relevant time-series and text pairs from our dataset based on embedding similarity.
+Table 3 shows that retrieval augmentation improves forecasting performance across all models. The biggest gains came from combining time series with text (TS+Text), especially for decoder-only models like Timer-XL and Time-MoE.
+Interestingly, TRACE itself showed minimal improvement between TS-only and TS+Text retrieval. This isn't a weakness—it indicates TRACE's embeddings are already well-aligned across modalities. The model doesn't need much help because its multimodal space is already doing the work.
+This makes TRACE effective as a lightweight, general-purpose retriever for RAG pipelines.
+
+![Table 3: Forecasting performance on Weather dataset for next 24 steps under different retrieval-augmented generation settings.|scale=0.4](./assets/table-3.png)
+
+### Standalone Time Series Encoder
+We tested TRACE on forecasting and classification tasks, comparing it against traditional models trained from scratch and existing time series foundation models.
+The classification results (Table 4) revealed an interesting pattern: fine-tuned foundation models actually performed worse than simpler train-from-scratch models. The likely reason is over-generalization—their embeddings become too broad and lose domain-specific signals needed for accurate classification. TRACE took a different approach. It achieved significantly higher accuracy and F1 scores than baselines, both with and without retrieval-augmented generation (RAG). This suggests TRACE maintains discriminative structure while preserving semantic alignment.
+
+![Table 4: Weather Event Classification Results.|scale=0.2](./assets/table-4.png) 
+
+Table 5 shows TRACE outperforming baselines across datasets, particularly on longer prediction horizons where other models struggle. Traditional approaches show inconsistent performance as the forecasting window extends. TRACE's cross-modal design appears to be the key difference—it provides better semantic understanding and more context-aware predictions.
+
+![Table 5: Forecasting results (MAE and MSE) of full-shot models and time series foundation models on multi-variate (M) and univariate (U) datasets. Red: the best, Blue: the 2nd best.|scale=0.7](./assets/table-5.png)

app/projects/trace/assets/cover_fig.png

@@ -19,6 +19,17 @@ export interface Publication {
 }
 
 export const publications: Publication[] = [
+  {
+    title: "TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval",
+    authors: "Jialin Chen, Ziyu Zhao, Gaukhar Nurbek, Aosong Feng, Ali Maatouk, Leandros Tassiulas, Yifeng Gao, Rex Ying",
+    venue: "NeurIPS 2025",
+    page: "trace",
+    code: "https://github.com/Graph-and-Geometric-Learning/TRACE-Multimodal-TSEncoder",
+    paper: "https://arxiv.org/pdf/2506.09114",
+    abstract: "We address the challenge of time-series retrieval, which remains largely underexplored as existing methods lack semantic grounding, struggle with heterogeneous modalities, and have limited capacity for multi-channel signals. We propose TRACE, a multimodal retriever that grounds time-series embeddings in aligned textual context.", 
+    impact: "TRACE enables fine-grained channel-level alignment and uses hard negative mining for semantically meaningful retrieval across flexible modes (Text-to-Timeseries and Timeseries-to-Text). Beyond retrieval, it functions as a standalone encoder that achieves state-of-the-art performance on forecasting and classification tasks.",
+    tags: [Tag.MultiModalFoundationModel],
+  },
   {
     title: "Mixture-of-Personas Language Models for Population Simulation",
     authors: "Ngoc Bui, Hieu Trung Nguyen, Shantanu Kumar, Julian Theodore, Weikang Qiu, Viet Anh Nguyen, Rex Ying",