acceptodds
Under review as a conference paper at ICLR 2027

Fuzzy-MoE: Interpretable Regime-Conditioned Expert Routing for Non-Stationary Multivariate Time Series Forecasting

Abstract

In non-stationary multivariate time series, different variables and samples often exhibit heterogeneous latent dynamic states, while existing deep forecasting models usually compress them into a unified end-to-end mapping, leading to suboptimal modeling of time-varying dynamics and limited interpretability regarding which forecasting mechanism is activated under different latent states. To overcome these limitations, we reformulate time series forecasting as a unified framework of latent temporal state identification and interpretable expert routing, and propose Fuzzy-MoE, a fuzzy logic-based dynamic Mixture-of-Experts model. Fuzzy-MoE consists of multiple parallel expert mapping networks and a dual-view fuzzy router. By jointly exploiting local convolutional dynamics and global segmented statistics, the router infers latent temporal states and computes expert activation strengths through learnable Gaussian membership functions, enabling explicit IF-THEN rule-based expert selection. This fine-grained routing strategy allows different variables within the same sequence to activate different experts, effectively capturing heterogeneous temporal dynamics while improving model interpretability. Experimental results on multiple public time series benchmark datasets show that Fuzzy-MoE significantly outperforms mainstream forecasting methods in forecasting accuracy. Moreover, fuzzy memberships and rule activations provide interpretable routing diagnostics, demonstrating the effectiveness of the proposed framework in both forecasting performance and mechanism transparency. Unlike traditional MoE models that use black-box routing, Fuzzy-MoE’s routing is based on clear, interpretable fuzzy rules. This makes the expert selection transparent and traceable.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Fuzzy mixture-of-experts, Interpretable machine learning, Latent temporal states, Non-stationary dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yHCSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Fuzzy-MoE proposes a channel-level forecasting ensemble with local convolutional and global segmented features, learned Gaussian memberships, product aggregation, and a dual-view router over linear experts. Its MLP-gate comparison keeps the backbone fixed and supports the routing design. Two questions remain: whether the latent rules have the claimed physical or regime semantics, and whether input-dependent routing adds value beyond strongly smoothed or input-independent mixing.

Critical strengths
  1. The dual-view router combines local and global temporal features. Its Gaussian and product calculations make the routing computation inspectable in the learned latent space. Section 3.4 describes the linear expert design.

  2. Table 1 reports competitive results on six benchmarks. Appendix D.3 and Table 5 compare fuzzy and MLP gates with the same forecasting backbone. All twelve reported MSE comparisons favor the fuzzy gate. For example, ETTh2/96 has MSE 0.271 versus 0.303. These results support the complete fuzzy-gate design.

Critical weaknesses
  1. The claims about physical semantics and regime identification need independent support. The fuzzy antecedents use coordinates produced by a learned nonlinear projection. Their explicit formulas make routing traceable, but the connection to meaningful temporal states remains unvalidated. Appendix D.4 uses the same gate's argmax to label the t-SNE visualization. Separation under these labels does not independently validate recovered regimes. The rule-deletion experiment also has a narrower interpretation. Suppressing Gaussian firing strengths changes routing while leaving the associated expert with positive Softmax weight. The ablation measures sensitivity to that perturbation. It does not measure complete removal of the expert's predictive contribution or independently validate regime semantics.

  2. The experiments do not separate the benefit of input-dependent fuzzy routing from the benefit of bounded, strongly smoothed mixing. For the tau=5 implementation in Model.py, firing strengths in [0,1] bound the ratio between any pair of expert weights by exp(0.2), approximately 1.2214. With five experts in this implementation, the per-view weights lie at approximately 0.1699–0.2339. Averaging the two views preserves these bounds, apart from the negligible normalization epsilon. These conditional bounds should not be applied to runs using another temperature or to configurations with another expert count. Small weight changes may still matter when expert predictions differ. Uniform and learned input-independent gates would help resolve this attribution question. Table 5 supports the combined design but does not separate input dependence from smoothing or regularization.

The experimental entrypoint and effective temperatures also need clarification. Pipeline/OneMulEdit.py imports the absent model.CMoS. Model.py fixes tau=5 and does not read configs.tau. The ETTm2 configuration files specify tau=0.3, while the weather files specify tau=5. The package also includes initModel.py, which uses a different temperature-dependent gating calculation. Please identify the actual training/evaluation entrypoint, source version, and effective temperature for the main results. One or two representative saved prediction, metric, and configuration records should connect those results to the program used. A source update or textual explanation alone cannot document the existing runs retrospectively. This concern remains until matching records are provided.

Questions for authors
  1. What temperature and expert count were used for each main result? Can the authors provide uniform-gate and learned input-independent-gate comparisons with the same expert backbone and training protocol? Capacity- or entropy-matched MLP controls would further distinguish the fuzzy structure from smoothing.

  2. Is there independent regime annotation, a controlled temporal-feature intervention, or other externally grounded evidence connecting the learned antecedents to the claimed states? If this evidence is unavailable, can the claims be narrowed to latent-space routing traceability? The remaining structured-routing contribution would then need to be evaluated on that basis.

  3. Which actual entrypoint and model/source version produced the main results? How were the temperature sweeps implemented when Model.py fixes tau=5? Please clarify whether the configured tau=0.3 and tau=5 values took effect. Please provide one or two representative saved predictions, metric records, and effective configurations with an explicit mapping to the reported results. They should identify whether the results used Model.py, initModel.py, or another version. Please also connect Figure 5 to its original unrounded gate tensor, plotting script, source/configuration, checkpoint, and selected evaluation batch.

Additional presentation feedback Figure 2 labels the Prediction shape as B×L×C, whereas Section 3.4.2 and Eq. (11) define the final output as B×P×C. It labels the first local convolution as Conv 1×1, although that layer uses kernel_size=configs.kernel_size; the second convolution is 1×1. The architecture diagram can be prepared as an AI-assisted editable SVG, with every label and connection checked manually, or correctly redrawn. Figure 5 displays channel-0 weights of 0.198, 0.200, 0.196, 0.198, and 0.200, which sum to 0.992. The combined rounding allowance for five three-decimal values is 0.0025. Rounding alone therefore does not explain the discrepancy. Please provide the unrounded gate tensor, actual plotting script, and corresponding source/configuration, checkpoint, and evaluation batch. The numerical figure should be regenerated from those recorded values and its annotations corrected.

Reviewer_Yv99OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Fuzzy-MoE, a mixture-of-experts model for non-stationary multivariate time series forecasting. It introduces a dual-view fuzzy router that combines local convolutional features and global segment statistics, using learnable Gaussian membership functions to generate sample-channel-level expert weights. The authors evaluate the model on six forecasting benchmarks and provide routing visualizations and ablation studies to support its forecasting performance and interpretability.

Critical strengths
  1. The use of Gaussian fuzzy membership functions provides an explicit and inspectable routing mechanism, offering an interesting alternative to conventional MLP-based MoE gates.

  2. Fuzzy-MoE achieves promising results across six forecasting benchmarks. The fuzzy-gate comparisons, rule-deletion experiments, and routing visualizations provide preliminary support for the proposed approach.

Critical weaknesses
  1. The introduction and related work occupy too much space, which leads to insufficient space in the important part of the main text, namely the experimental section, with only the main experiment and ablation experiment, which will affect the overall integrity of the article. You can consider deleting or placing the related work in the Appendix, and putting the important experimental parts into the main text.

  2. There are multiple 1-2 words occupying a line in the article (such as line 84, 119, 148), which need to be adjusted and modified.

  3. The characters in Figure 2 are too small, which is not conducive to reading and understanding.

  4. The novelty of the method is limited. This framework mainly combines established components, including RevIN, convolutional feature extraction, Gaussian fuzzy membership functions, and linear experts. What is the difference between it and existing interpretable MoE and state-aware prediction methods?

  5. Eq. (7) directly inputs the activation strength of Gaussian fuzzy rule rk∈(0,1]r_k \in (0,1] into Softmax. When τ\tau is large, the weights of different experts tend to approach each other. Figure 5 on page 16 of Appendix shows that the expert weights of multiple channels on ETTh2 are mainly distributed between approximately 0.19-0.23, which is close to the uniform weight of 0.2 for five experts. Although Expert 1 has obtained a high weight, the overall difference is still limited. Does the prediction improvement of the model come from true dynamic expert selection or a soft weighted ensemble of multiple similar experts?

  6. Table 1 does not reflect the results of Time MoE on the Electricity dataset.

  7. Inconsistent implementation details. Appendix C. 1 states that SGD with a learning rate of 2×10−52\times10^{-5} is used, while the same appendix specifies AdamW with StepLR and a completely different dataset specific learning rate. What configuration does the actual result of the paper correspond to?

Questions for authors

Please see Weaknesses.

1 comment
AuthorsOpenReview ↗

Comment: Dear reviewer Yv99,

Comment: Dear reviewer Yv99,

We greatly appreciate your constructive comments which substantially help us to improve the manuscript. We are hopeful that our responses help address your concerns and provide a clearer understanding of our contributions.

We sincerely apologize for our late reply and would be grateful if you could review our responses. If you find that our clarifications have adequately addressed your concerns, we would greatly appreciate your consideration of increasing the score.

Once again, we sincerely thank you for your time and efforts and wish you all the best.

Best regards,

All authors

Reviewer_n7FLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Fuzzy-MoE, an interpretable mixture-of-experts framework for non-stationary multivariate time-series forecasting. The main idea is to reformulate forecasting as a two-stage process consisting of latent temporal regime identification followed by regime-conditioned expert forecasting. Specifically, the model extracts complementary local convolutional features and global segment-level statistics, projects them into a low-dimensional fuzzy feature space, and uses learnable Gaussian membership functions together with product T-norm aggregation to generate expert routing weights. The routing is performed at the sample-channel level, allowing different variables within the same multivariate series to use different mixtures of forecasting experts. Experiments on six standard forecasting benchmarks show competitive or superior MSE/MAE compared with several recent forecasting baselines. The paper further provides rule-deletion experiments, temperature sensitivity analysis, t-SNE visualizations, routing heatmaps, and an ablation comparing the fuzzy gate against an MLP gate.

Critical strengths
  1. A simple and potentially useful channel-wise routing formulation for heterogeneous multivariate dynamics. The sample-channel-level routing mechanism is well motivated for multivariate forecasting, since different variables can exhibit different temporal behaviors even within the same observation window. The design is modular and could potentially be combined with different forecasting experts.
  2. Broad empirical gains on standard forecasting benchmarks. The method reports strong MSE/MAE performance across six commonly used datasets and multiple forecasting horizons. The fuzzy-gate-versus-MLP-gate experiment also provides preliminary evidence that the proposed routing mechanism can improve forecasting performance beyond a conventional learned gate.
Critical weaknesses
  1. The claimed interpretability is not sufficiently substantiated. The fuzzy variables are generated through a learned nonlinear projection, making their semantic meanings unclear. Although the Gaussian memberships and routing scores are traceable, the resulting IF-THEN rules cannot be directly linked to meaningful temporal properties such as trends, periodicity, or volatility. Moreover, these rules describe continuous fuzzy gating rather than explicit symbolic decisions. Thus, the method demonstrates routing traceability rather than convincing semantic interpretability.

  2. Insufficient validation of regime identification and experimental fairness. Although regime identification is central to the proposed framework, the experiments mainly focus on forecasting accuracy. The t-SNE visualizations and rule-deletion studies do not establish whether the learned routing corresponds to meaningful temporal regimes. Controlled experiments with known regime labels would strengthen this claim. Additionally, the extensive dataset- and horizon-specific hyperparameter tuning raises concerns about comparison fairness, as equivalent tuning budgets and input settings for baselines are not clearly established.

Questions for authors

Q1. What is the semantic meaning of each learned fuzzy variable z_m? Can the authors demonstrate that these dimensions consistently correspond to identifiable temporal properties, such as trend, volatility, periodicity, or local changes? If the dimensions cannot be assigned stable meanings across datasets or random seeds, how should the resulting IF-THEN rules be interpreted by a human user?

Q2. Can the authors directly evaluate the claimed temporal regime identification capability? For example, a controlled synthetic regime-switching experiment with known latent states could evaluate whether the learned expert assignments recover the underlying regimes using metrics such as NMI, ARI, or regime classification accuracy. This result would materially affect my assessment of the interpretability claim.

Q3. How strong is expert specialization in practice? In Figure 5, several routing weights appear relatively close to the uniform value 1/K. Could the authors report routing entropy, Top-1/Top-2 margins, per-expert utilization, and expert-output diversity across datasets? I would also be interested in whether hard or Top-k expert selection preserves the reported forecasting performance.

Q4. Are all baseline results obtained under the same input lengths, data splits, training budgets, and hyperparameter-selection protocol as Fuzzy-MoE? The appendix indicates substantial dataset- and horizon-specific tuning for Fuzzy-MoE. A controlled comparison using identical look-back lengths and comparable tuning budgets would help determine whether the performance improvements are attributable to the proposed routing design.

1 comment
AuthorsOpenReview ↗

Comment: Dear reviewer n7FL,

Comment: Dear reviewer n7FL,

We greatly appreciate your constructive comments which substantially help us to improve the manuscript. We are hopeful that our responses help address your concerns and provide a clearer understanding of our contributions.

We sincerely apologize for our late reply and would be grateful if you could review our responses. If you find that our clarifications have adequately addressed your concerns, we would greatly appreciate your consideration of increasing the score.

Once again, we sincerely thank you for your time and efforts and wish you all the best.

Best regards,

All authors

Comments on the paper

2 comments
Reviewer_yHCSOpenReview ↗

Response to authors

Thank you for the clarifications and the new comparisons. The fuzzy rules make the routing easier to inspect, but the static and matched non-fuzzy gates perform nearly as well. I’m still not convinced that dynamic fuzzy routing offers a meaningful forecasting advantage on real data, so I’ll keep my original overall score.

AuthorsOpenReview ↗

Response to Reviewer yHCS (third round)

Response to Reviewer yHCS (third round)

We thank the reviewer for the continued engagement and for acknowledging that the fuzzy rules make the routing easier to inspect. We respect the reviewer's assessment. We summarise what Fuzzy-MoE offers on real data beyond the average error, with the evidence for each point.

1. The explanation is the computation. Every expert weight is a product of Gaussian memberships with explicit centres and widths, so the rule is read directly from the model and no post-hoc surrogate is needed. The trained model uses these rules:

  • permuting the rule centres and widths across experts, with all other weights fixed, raises the Weather MAE by up to 26.1% (Table 14);
  • the learned rules alone (largest summed firing strength, no fitted model) reproduce the router's dominant expert in 87–97% of the (window, variable) cases on Weather and in 96–99.6% on the synthetic benchmark (App. F.11).

2. The rule conditions have a stable meaning.

  • The fuzzy variables correlate most with trend and volatility (up to |ρ| = 0.49), with cross-seed stability 0.58–0.92 over all 24 settings (Table 15).
  • On Weather, physically related variables (temperature, moisture, radiation, wind, precipitation) are routed alike without any variable labels: permutation test p ≤ 0.001 in all 12 runs.
  • The rain variable is routed differently after rainfall in 10 of 12 runs (Table 20).

3. The bound gives a guarantee, not a tuned behaviour.

  • Because firing strengths lie in (0, 1], the ratio of any two expert weights is at most exp(1/τ) (Sec. 3.3.3), and τ sets how selective the router may be.
  • No expert collapses, without any load-balancing loss: removing even the least-used expert raises the error on every dataset (Table 4, right).
  • In all 16 real-data settings the router stays within 0.2% of uniform mixing or below it. An unbounded router, as used in existing MoE forecasters, is up to 10.2% worse (Tables 2, 12).

4. No fixed alternative is reliable across data types.

  • The static gate is slightly better where the identity of the variable decides the expert (Weather, Electricity), but 16.5% behind when regimes change within a variable.
  • The matched non-fuzzy gate is an ablation of Fuzzy-MoE: it keeps the bound and projection of the fuzzy router and drops the rules, so its accuracy confirms that the gain comes from the fuzzy formulation.
  • Which case applies is not known before training. Over all 17 router settings, Fuzzy-MoE with the per-variable prior is never more than 3.4% from the best router, with a mean gap of 0.9% (Table 23). The prior also shows that per-variable knowledge enters the rule framework as a multiplicative term, without changing its structure (App. F.8).

5. Accuracy and size.

  • Fuzzy-MoE is the most accurate method in 20 of 24 settings against the published baselines.
  • It has 2.7K–0.69M parameters, more than 70 times fewer than Time-MoE; the rule layer adds only 2·K·M parameters per view.
  • Its rule form follows TSK/ANFIS systems (Gaussian antecedents, product T-norm), a standard form in energy and weather applications that require auditable decisions.

We thank the reviewer again for the two rounds of detailed comments, which have made the claims of the paper more precise.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.