- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper argues that existing decomposition-based spatio-temporal forecasters model the aperiodic (event-induced) component with one shared parameterization, although the effect of an event depends on when, where and what kind it is. It proposes HEAM, a module inserted before every layer of a backbone. At each layer the hidden representation Z is split into a "context" part Zc, produced by an MLP from time-of-day, day-of-week and node embeddings and trained to reconstruct stop-gradient(Z), and an "aperiodic" residual Za = Z − Zc. Zc is soft-assigned by cosine similarity to G pool prototypes ("Situate"); a projection of Za is soft-assigned to K prototypes within each pool ("Identify"); prototypes are updated by EMA. The two assignments mix a G×K table of learnable vectors into a modulation vector m, and Za is rescaled as (1 + γ·m) ⊙ Za, concatenated with Z and passed to the (widened) next layer. HEAM is added to GraphWaveNet and to a patch-Transformer variant of VisiFold, and evaluated on PEMS08, METR-LA and Solar-AL at horizons of 24/36/48 steps with five seeds, plus a peak/off-peak breakdown, an ablation at horizon 48 and visualizations of the routing.
Critical strengths
-
Backbone-controlled evaluation with multiple seeds. HEAM is compared with its own host backbone under the same optimizer, schedule and epoch budget, over five seeds with standard deviations (Table 4), on two architecturally different hosts. For MAE at the 36- and 48-step horizons the improvements (e.g. PEMS08 16.36 → 16.05, METR-LA 5.01 → 4.72) are several standard deviations wide. This is more careful than most plug-in papers and is the main reason I do not rate the paper lower.
-
A sensible problem statement and a reasonably light module. Conditioning the treatment of the non-regular residual on spatio-temporal context is a plausible refinement of decomposition-based forecasters, and the module adds no cross-node or cross-token computation.
Critical weaknesses
-
The paper never shows that HEAM improves forecasting of aperiodic events; what is evaluated is rush hour.
- There is no definition, detection or annotation of events anywhere. The "event-dense periods" used in RQ2 are fixed clock intervals (06:30–08:00 and 16:00–17:30 on PEMS08, 07:30–09:00 and 16:00–17:30 on METR-LA, 12:00–15:00 on Solar-AL). Morning and evening peaks are the most periodic structure in traffic data, so a larger gain there is evidence about recurring high-variance periods, not about incidents or extreme weather. At horizon 24 (Figure 8) the peak/off-peak difference is negligible on Solar (−2.2% vs −2.0%) and reversed on METR-LA (−2.7% vs −2.9%), where HEAM-V is 0.6% worse than its backbone in peak hours.
- The "aperiodic representation" is whatever part of a hidden state is not predictable from (time-of-day, day-of-week, node id). That residual contains all input-dependent information, including ordinary level and trend, so calling it event-induced is an interpretation that is not verified. The evidence for "latent event types" is one node on one day (Figure 5b).
- A direct test is available and missing: stratify test error by the magnitude of the deviation from the historical time-of-week profile, or use incident logs, and show the gain concentrates in high-deviation windows.
-
The gains are small, confounded with added capacity, and the baseline table is not reliable enough to support the state-of-the-art claim.
- HEAM concatenates the modulated residual with Z and doubles the input width of every consuming layer, in addition to per-layer embedding tables, prototypes and modulators. No parameter count, FLOPs or run-time is reported, and no width-matched backbone is included, so capacity is a live alternative explanation. The ablation variants "w/ Decompose" and "Concat" are the relevant controls but are shown only as bars spanning about 0.2–0.3 MAE at a single horizon and backbone, without numbers or deviations; the hyperparameter study shows the same narrow range.
- At 24 steps HEAM-V is worse than its backbone on PEMS08 RMSE (25.61 → 25.66) and on METR-LA MAE and MAPE (4.42 → 4.43, 11.69 → 11.79), and on PEMS08 it is behind STAEformer in MAE at 24 steps (14.60 vs 14.91) and behind VisiFold in RMSE at 48 steps (27.76 vs 27.79). The text reports only the favourable 48-step MAE comparison.
- Several baseline rows are implausible: on METR-LA, DCRNN has MAE 19.34 / MAPE 52% at 24 steps but 8.16 at 36 steps; on PEMS08, DualCast jumps from MAE 19.66 (36 steps) to 26.09 (48 steps). Baselines use "configurations adapted from BasicTS" with input length equal to horizon and a single run; they appear untuned for this setting.
- Scope is narrow for a general plug-in claim: three datasets, two hosts, one of which had to be re-architected to accept the module. The mechanism (prototype routing into a pool of channel-scaling vectors) is close to memory- and prototype-based heterogeneity models (e.g. MegaCRN, which targets incidents and is neither cited nor compared) and to DPRNet, from which the orthogonality regularizer is taken. The learned gate γ, which determines how much modulation is applied at all, is never reported.
Questions for authors
Clarifications that could change my assessment:
- Please stratify test error by a data-driven event measure (e.g. deciles of |x − historical time-of-week mean| in the input or target window, or incident records for METR-LA / PEMS). Do HEAM's gains concentrate in the high-deviation strata?
- Report parameters, FLOPs and training/inference time for each backbone with and without HEAM, and add a backbone whose layers are widened to the same parameter count.
- Give the ablation of Figure 4b as a table with mean ± std, for both backbones and all three horizons.
- What values does γ take after training, per layer and dataset? What is the distribution of the resulting scale (1 + γ·m)?
- Please check the DCRNN (METR-LA) and DualCast (PEMS08) rows and state how baseline hyperparameters were chosen for 24–48-step inputs.
- Does HEAM help under the standard 12-step protocol and on larger benchmarks (e.g. LargeST), and when plugged into a decomposition-based host such as STDN or HyperD, which is the family the paper criticises?
Suggestion for future work (does not affect my rating): comparison with memory-based heterogeneity models such as MegaCRN.