- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes ARISE, a retrieval-augmented forecasting framework for irregular multivariate time series (IMTS). ARISE operates in two stages on top of a frozen pretrained IMTS forecasting backbone (EB, DB). In the offline stage, Dynamic Window Construction evaluates C candidate left boundaries spaced proportionally across each sample's observed time span and learns a boundary selector and three view encoders (obs, dyn, val) under a composite objective (ℒcon = ℒforecast + λb ℒcompleteness + λl ℒlength) to populate a historical retrieval memory. In the online stage, Multi-Evidence Knowledge Retrieval retrieves top-K candidates per view, applies Maximal Marginal Relevance (MMR) to select R diverse histories, aligns each candidate's irregular future observations to the query's target timestamps via a Gaussian kernel (bandwidth σt), and applies multi-head cross-attention (with a learned null key K_∅) to produce a residual hidden-state correction Δh before the frozen backbone decoder DB.
The authors evaluate ARISE on 11 real-world IMTS datasets (PhysioNet, HumanActivity, USHCN, CESNET, Pamap2, EPA-Air, ClusterTrace, APTC, FNSPID, GDELT, Seabirds) across 3 horizons each against 9 non-retrieval IMTS baselines (TFMixer, ASTGI, APN, KAFNet, Hi-Patch, HyperIMTS, GNeuralFlow, GraFITi, Warpformer), reporting first-place average MAE on all 11 datasets and first-place average MSE on 9 of 11 datasets, along with component ablations and hyperparameter sensitivity plots on 4 datasets.
Critical strengths
Well-Motivated Problem Formulation and Target-Time Alignment for Irregular Retrieval: Extending retrieval-augmented forecasting from regular grids to irregular, asynchronous multivariate time series addresses a genuine practical gap in domains like clinical monitoring (PhysioNet) and environmental sensing (EPA-Air). In particular, using a Gaussian kernel over relative future offsets (τju vs. τqh) together with an explicit support mask Sqjhv and a learned null attention key K_∅ (Section 3.4.3) is a sensible way to handle asynchronous future timestamps and missing target support without interpolating the raw series onto a dense regular grid.
Critical weaknesses
Severe Technical Incompleteness and Unreproducible Core Formulations (Missing Equations and Unspecified Backbone): The paper omits the mathematical definitions of its core training losses and feature extractors, making the method impossible to reproduce from the manuscript: Identity of the Frozen Pretrained Backbone (EB, DB) Never Specified: Sections 3.4.2 and 3.4.4 state that ARISE encodes histories with a "frozen history encoder EB" and decodes hfused = hbase + Δh with a "pretrained backbone [that] remains fixed" (DB). Yet neither the main text nor the appendix ever identifies what architecture or pretrained model is used as (EB, DB)! Undefined Construction Losses in Eq. (5): Dynamic Window Construction—claimed as the first major contribution—is trained by ℒcon = ℒforecast + λb ℒcompleteness + λl ℒlength (Eq. 5). None of these three loss terms is ever defined mathematically anywhere in the main text or appendix; they are described only in three sentences of high-level prose (lines 254–260). Undefined View Descriptors and Encoders: The exact definitions of the boundary features ψim, the eligibility mask χselim, the three view feature maps ϕobs, ϕdyn, ϕval, the view encoder architecture Er, the content encoder uqjhv, and key hyperparameters (λb, λl, τcut, R, D, DA) are all missing. Internal Numerical Inconsistencies, Missing Retrieval Baselines, and Limited Novelty / Foundation-Model Alignment: Contradictions Between Tables 1–2 and Figures 3–4: In Table 2, w/o Attention Reader sets Δh = 0 (lines 427–428), which by Eq. (14) reduces ARISE to the standalone frozen pretrained backbone DB(EB(·)). However, its MSE on EPA-Air (1.57), CESNET (3.22), GDELT (2.97), and ClusterTrace (6.34) is dramatically worse than nearly all 2024–2026 baselines in Table 1 (e.g., TFMixer achieves 0.791, 2.327, 1.859, 4.199). Conversely, in the sensitivity plots (Figures 3–4), varying C, K, λMMR, σt across wide ranges barely changes MSE past the 3rd or 4th significant digit (e.g., CESNET stays within 2.1682–2.1692, a <0.05% variation, and λMMR = 0 is actually optimal on 2 of 4 datasets), whereas w/o Dynamic Window in Table 2 jumps from 2.16 to 2.71 (+25%). Furthermore, on GDELT, every sensitivity curve in Figures 3–4 lies between 1.910 and 1.923 (× 10-2), which contradicts the 1.819 (× 10-2) reported for ARISE in Tables 1, 2, and 3. No Retrieval-Augmented Baselines and No Variance / Latency Reporting: None of the 9 baselines in Table 1 is a retrieval-augmented method (e.g., RAFT, TS-RAG, Cross-RAG, SARAF, or RATL applied to the same IMTS backbone embeddings hbase). All results are single point estimates without standard deviations across random seeds, and there is no evaluation of memory or inference latency overhead despite running EB twice per query (hbase, hsel) and performing per-(horizon, variable) cross-attention. Novelty and Alignment with the LLM / Foundation Model Roadmap: Methodologically, ARISE combines heuristic summary descriptors with standard MMR (Carbonell & Goldstein, 1998), Gaussian kernel time alignment (mTAN, Shukla & Marlin, 2021), and history-key / future-value cross-attention (Cross-RAG, Lee et al., 2026). Rather than advancing unified continuous-time or event-tokenized foundation models that generalize zero-shot across irregular time series domains, ARISE requires multi-stage, dataset-specific training on handcrafted per-view statistics.
Questions for authors
Clarification / Core Assessment — What is the exact pretrained backbone (EB, DB) and why is w/o Attention Reader (Δh = 0) so weak in Table 2? Please specify the exact architecture and training recipe of the frozen backbone (EB, DB). Why does setting Δh = 0 (w/o Attention Reader in Table 2, which outputs the unaugmented frozen backbone prediction) yield MSEs of 1.57, 3.22, 2.97, 6.34—worse than almost every baseline in Table 1—and how can adding a retrieval residual Δh on top of such a weak frozen encoder beat state-of-the-art backbones like TFMixer and ASTGI? Clarification / Core Assessment — Exact mathematical formulations of ℒforecast, ℒcompleteness, ℒlength and ϕr: Please provide the exact mathematical equations for all three loss terms in Eq. (5), the eligibility mask χselim, and the three view feature maps ϕobs, ϕdyn, ϕval. Clarification / Core Assessment — Discrepancy on GDELT between Tables 1–3 (1.819) and Figures 3–4 (1.910–1.923), and comparison to RAG baselines: Why do all four sensitivity curves for GDELT in Figures 3 and 4 range between 1.9128 and 1.9203 (× 10-2), whereas Tables 1, 2, and 3 report 1.819 (× 10-2) for ARISE? In addition, how does ARISE compare against existing time-series RAG methods (such as RAFT, TS-RAG, or Cross-RAG) when applied directly in the latent space hbase of the same IMTS backbone?