CADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate Time Series Classification

arXiv:2609.35929 · cs.LG · Submitted 2026-09-28 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "CADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate Time Series Classification".

Jane: The gist: CADENCE presents a unified, CPU-native dual-expert architecture that decouples representation learning into two specialized pathways to achieve fast and accurate time series classification.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're talking about this paper called "CADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate Time Series Classification". It seems like they're tackling the big problem of getting high accuracy in time series classification without needing days or weeks of computing power.

Jane: Exactly. The main thing this paper claims is that by decoupling how we learn the representations, you can get both speed and accuracy, which is usually a trade-off. It suggests this new architecture, CADENCE, offers a way to do both simultaneously with one unified system.

Lu: The core idea here is splitting the representation learning into two distinct pathways instead of trying to jam everything into one model that has to do everything at once. This allows for specialized processing for different kinds of time series data.

Meng: So, what are these two pathways? Are they just two different models running side-by-side, or is it something more integrated? I'm curious about the practical engineering side of this decoupling.

Tom: It's a dual-expert architecture, so it splits things up into Branch A and Branch B. They use one expert for feature extraction and another for making the final decision, which is pretty clever.

Jane: Branch A is described as a Convolutional Linear Expert that uses ten thousand deterministic dilated features generated by length-nine integer kernels. It also involves using closed-form L2-regularized Woodbury ridge classification to get the actual class predictions.

Lu: And Branch B, the Distributional Interval Expert, is where things get interesting with competing dilated kernels called Hydra and thread-safe dyadic Cornish-Fisher moment approximations across things like raw series and FFT spectral energy bands.

Meng: Competing kernels sound computationally heavy. How do they manage that complexity while still being fast enough for a real application? I need to know how this translates from theory to something that runs on actual hardware without taking forever.

Tom: They use an entropy-based ExtraTrees ensemble here, which is used to fit those features extracted by Branch B. It’s about combining different kernel ideas in a way that seems robust.

Jane: And then there's this confidence-adaptive meta-router that sits on top of everything. This router decides whether to use pure expert routing or blend the predictions with confidence weighting based on how much the two experts disagree, using a difference threshold of zero point zero eight.

Lu: That routing mechanism is key because it allows the system to dynamically switch its strategy depending on which expert seems more confident about a particular data point, preserving rare classes in the process.

Meng: So if one expert is way more confident than the other, CADENCE leans on it instead of blending, and that decision-making happens in real time during inference. That sounds like something we could actually integrate into a fast pipeline.

Tom: It really does sound like they've managed to build a system that handles different types of data structures very well because it addresses the limitations you mentioned earlier, like phase-independent statistical distributions.

Paper summary: Jane: They show impressive results across the UCR Time Series Archive, achieving a grand mean accuracy of zero point eight eight six four across all one hundred nine datasets over thirty resamples. This puts it above HIVE-COTE one point zero by a statistically significant margin on that benchmark <ref:2609.35929#pg2>.

Lu: And they also showed specific wins on specialized sensor and kinematic datasets, like SemgHandMovementCh2, where they beat Hydra+MultiRocket with a gain of nine point nine six percentage points in accuracy.

Meng: That's a big number when you think about the computational cost we discussed earlier. They managed to achieve this level of performance while running in an average of seventeen point five three seconds per seed run, which is pretty fast for a monolithic ensemble approach.

Tom: It really highlights how this dual-expert approach establishes a better performance frontier compared to just stacking models together in one giant ensemble. The paper on CADENCE shows that combining complementary representations through this confidence-adaptive routing provides a scalable alternative to those heavy systems we've seen before.

Jane: So, looking at the authors, Onisa Mapunda is the lead researcher here, and they have made it open-source so the complete CPU-native pipeline is available for everyone to use. That’s a significant contribution for the community.

Lu: The potential for this architecture comes from how it handles uncertainty directly through that confidence routing layer and its ability to leverage both deterministic features and distributional approximations simultaneously. It opens up new avenues for building classifiers that are inherently more adaptive to the data structure they encounter on the fly.

Meng: From an engineering standpoint, having a CPU-native design is a big plus because it means lower latency and better efficiency on standard hardware, which is exactly what we need for deployment in many real-world applications.

Tom: The main implication for people listening to this show is that if you're trying to build a time series classifier that needs to be both fast and accurate, CADENCE shows you a path by separating the representation learning tasks into specialized paths instead of relying on one giant model.

Jane: It’s about moving away from monolithic ensembles toward architectures where different experts can handle different parts of the problem, guided by a smart router that knows when to trust which expert most.

Lu: The paper suggests that this confidence-adaptive approach is a viable way to balance the trade-off between speed and accuracy without needing the massive compute resources of those traditional meta-ensembles.

Meng: So for someone working on an AI system, this means you can potentially achieve near top accuracy while keeping inference times manageable, which is a huge practical win for deployment scenarios.

Tom: That's the gist of it then: CADENCE gives us a unified architecture that achieves zero point eight eight six four accuracy across the UCR datasets by smartly routing between two distinct representation methods <ref:2609.35929#pg2>. We'll take a quick break and come back to discuss what this means for time series classification in general after the next spot.

Conclusion: Tom: So we're wrapping up our look at CADENCE today, and we’re going to talk about what this whole paper actually means for time series classification.

Jane: Yeah, this paper is called "CADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate Time Series Classification." It’s essentially showing how you can get good results without needing massive amounts of compute time.

Lu: The authors are Onisa Mapunda and they’ve built a system that splits the work into two different experts, Branch A and Branch B.

Meng: So, to put it simply, they took one big classification problem and divided it up so two specialized parts could handle the heavy lifting separately.

Tom: Right, so instead of one giant model trying to do everything at once across all those different time series types, they use these two distinct pathways for learning the features.

Jane: That’s right. Branch A uses a convolutional linear expert to get some deterministic features, while Branch B uses something called a distributional interval expert that looks at various statistical moments and spectral energy bands.

Lu: The most interesting part is the confidence-adaptive meta-router they built on top of those two branches. This router decides whether to trust one expert more than the other based on how much they disagree.

Meng: So it’s not just running both models and averaging their results; it’s a dynamic system that knows when to lean heavily on one path or blend them softly depending on the data point.

Tom: Exactly. It’s about making intelligent decisions in real time during classification so you don't waste time guessing when the model isn't sure.

Jane: And the results show it hits an accuracy of zero point eight eight six four across all those datasets, which is quite competitive compared to what we see in other large models.

Lu: It opens up a new way to build these classifiers that can be fast enough for real-world use while still being accurate on complex time series data.

Meng: That speed and accuracy balance is what I’m really focused on—if this architecture can run in seconds instead of hours, that changes how we deploy these things.

Tom: It definitely does. It shows that you don't always have to build one huge, slow system when a smart split between two experts can give you a much better outcome.

Jane: And the authors have made the full pipeline open-source so other researchers can actually see how they did it and try to use it themselves.

Lu: It’s really encouraging because it validates this idea that specialized components working together can achieve performance levels that were previously only achievable with much more expensive setups.

Meng: So, the implication here is that we can start thinking about building these kind of dual-expert systems for things like industrial monitoring where speed matters a lot.

Tom: Absolutely. We’ll keep digging into how this specific routing mechanism works next, and we’ll see what other researchers think about this architecture.

cs.LG

Submitted: 2026-09-28

Updated: 2026-09-28

Comments: 7 pages, 5 figures. Code and benchmark evaluation scripts available at https://github.com/onisa-jr/CADENCE.git

Code: https://github.com/onisa-jr/CADENCE

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 82/100

The gist: The gist: CADENCE presents a unified, CPU-native dual-expert architecture that decouples representation learning into two specialized pathways to achieve fast and accurate time series classification.

Key concepts

Convolutional Linear Expert: This pathway extracts 10,000 deterministic features using specific integer kernels and pooling techniques. These features are then classified using a linear regression model with L2 regularization across various alpha values to determine class probabilities.
Distributional Interval Expert: This expert combines competing dilated kernels (Hydra) with quantile approximations derived from signal kinematics (X, ΔX, Δ²X) and FFT spectral energy bands. It uses an entropy-based ExtraTrees ensemble to fit these features for classification.
Confidence-Adaptive Meta-Router: An internal router decides whether to use the output from Expert A or Expert B based on a calculated difference in their predicted values. This routing is dynamic, using rare-class preservation, and can switch between pure expert routing and soft blending for optimal performance.
Cornish-Fisher Approximation: This mathematical tool estimates arbitrary quantiles (like those needed for interval experts) in linear time without needing to sort data. It is used here across raw series, velocity, acceleration, and FFT spectral energy bands to create robust features.

Summary

The gist: CADENCE presents a unified, CPU-native dual-expert architecture that decouples representation learning into two specialized pathways to achieve fast and accurate time series classification.

CADENCE Methodology

CADENCE decouples representation learning into two specialized pathways: (i) a Convolutional Linear Expert pairing 10,000 deterministic dilated features with closed-form L2-regularized Woodbury ridge classification, and (ii) a Distributional Interval Expert pairing competing dilated kernels (Hydra) with thread-safe dyadic Cornish-Fisher moment approximations across signal kinematics (X, ΔX, Δ2X) and FFT spectral energy bands, fitted with an entropy-based ExtraTrees ensemble<ref:2609.35929#pg2>

Branch A: Convolutional Linear Expert

Branch A extracts D A = 10,000 features using length-9 integer kernels ω ∈ (-1, 2) 9 (84 unique configurations with 3 weights equal to 2 and 6 weights equal to-1) and Proportion of Positive Values (PPV) pooling<ref:2609.35929#pg5> Standardized features F A are classified via multiclass Ridge regression with closed-form L2 regularization across 15 log-spaced candidate alphas in [10-3, 10 4]<ref:2609.35929#pg5> Posterior class probabilities are derived via softmax over raw decision values<ref:2609.35929#pg5>

Branch B: Distributional Interval Expert

Branch B combines 16 groups of 8 competing Hydra kernels (128 features) [3], dyadic intervals up to depth 5 with Cornish-Fisher [9] quantile approximations across raw series (X), velocity (ΔX), and acceleration (Δ2X) (1,701 features), and real FFT spectral quantiles (22 features)<ref:2609.35929#pg5> The Cornish-Fisher expansion estimates arbitrary quantiles q α in linear time O(M) without sorting<ref:2609.35929#pg5> Features are fitted with 100 Extremely Randomized Trees (ExtraTrees) [10] using Shannon entropy and 10% random feature subsampling per split<ref:2609.35929#pg5>

Confidence-Adaptive Meta-Router

CADENCE incorporates an internal validation meta-router with rare-class preservation that dynamically selects between pure expert routing and confidence-weighted soft blending, followed by a full refit on 100% of training data<ref:2609.35929#pg5> The routing decision is governed by empirical difference Δ = Val A - Val B with threshold τ = 0.08<ref:2609.35929#pg5> In the competitive regime, predictions are smoothly blended with weight w A<ref:2609.35929#pg5>

Experimental Results and Benchmark Rankings

CADENCE achieves a grand mean accuracy of 0.8864 across all 109 datasets of the standard equal-length UCR Time Series Archive [6] over 30 resamples (3,270 total evaluations)<ref:2609.35929#pg2> This ranks among evaluated classifiers across the archive, surpassing HIVE-COTE 1.0 by a statistically significant margin<ref:2609.35929#pg5> CADENCE executes in an average of 17.53 seconds per seed run, establishing a superior Pareto frontier compared to monolithic ensembles<ref:2609.35929#pg5>

Breakthroughs and Component Ablation

CADENCE demonstrates decisive wins on specialized sensor and kinematic datasets over Hydra+MultiRocket, including SemgHandMovementCh2 (0.8713 vs 0.7717, +9.96 pp)<ref:2609.35929#pg5> It also directly surpasses HIVE-COTE 2.0 on 39 datasets (e.g., DistalPhalanxTW +7.29 pp)<ref:2609.35929#pg5> The controlled component ablation shows that the full system, CADENCE (Full System), achieves a mean accuracy of 0.8864<ref:2609.35929#pg5>

Routing Analysis and Computational Efficiency

Across all 109 evaluated datasets, 82 datasets (75.2%) trigger Regime II (Competitive Blending, e.g., Phoneme, ArrowHead, ECG200), 18 datasets (16.5%) trigger Regime III (Pure Branch A Dominance), and 9 datasets (8.3%) trigger Regime I (Pure Branch B Dominance)<ref:2609.35929#pg5> In total, 27 datasets (24.8%) trigger dominance regimes where test inference skips one branch completely<ref:2609.35929#pg5> The entire 3,270-run benchmark completed in 15.92 CPU hours (7.5 hours wall-clock time via bidirectional execution on an Intel Core i7-7600U)<ref:2609.35929#pg5> CADENCE establishes a superior Pareto frontier, nearing HC2 accuracy while running orders of magnitude faster<ref:2609.35929#pg5>

CADENCE provides a fast, CPU-native classifier achieving 0.8864 accuracy across the 109 UCR datasets, closing the gap to HIVE-COTE 2.0 by just 0.31 percentage points while running in seconds<ref:2609.35929#pg5> The paper concludes that combining complementary representations through confidence-adaptive routing provides a scalable alternative to monolithic ensembles<ref:2609.35929#pg5> Limitations include testing on univariate benchmarks and relying on softmax calibration<ref:2609.35929#pg5> The work is open-source, with the complete CPU-native pipeline publicly available at https://github.com/onisa-jr/CADENCE.git<ref:2609.35929#pg5>

--- Page 1 ---

CADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate Time Series Classification

Onisa Mapunda

onisajr@gmail.com

Abstract Time series classification (TSC) has long been characterized by a sharp trade-off between classification accuracy and computational scalability<ref:2609.35929#pg2> Heterogeneous meta-ensembles like HIVE-COTE 2.0 [5] achieve state-of-the-art accuracy by combining representations across temporal, frequency, shapelet, and dictionary domains, but require days or weeks of compute<ref:2609.35929#pg2> Conversely, ultra-fast random convolutional transforms such as MiniRocket [2] and Hydra [3] provide orders-of-magnitude speedups, but struggle with phase-independent statistical distributions, kinematic transitions, and catastrophic decision tree fragmentation on datasets with large class counts when combined with tree heads<ref:2609.35929#pg2>

--- Page 2 ---

CADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate TSC

Time series classification (TSC) is a core problem in applied machine learning, underlying critical applications in electrocardiology [6], industrial anomaly detection, seismology, wearable motion analytics, and acoustic speech processing [7]<ref:2609.35929#pg2> The 109 univariate, equal-length datasets of the UCR Time Series Archive [6] exhibit tremendous structural diversity: series lengths range from dozens to thousands of time points, sample sizes vary from tens to thousands of instances, and class cardinalities span binary discrimination (K=2) to fine-grained categorization (K=60)<ref:2609.35929#pg2> Historically, achieving top accuracy demanded heavy heterogeneous meta-ensembles, most notably HIVE-COTE 2.0 (HC2) [5], which combines STC, TDE, DrCIF, and Arsenal<ref:2609.35929#pg2> While HC2 achieves an unmatched published benchmark accuracy of 0.8895, its computational cost is substantial, requiring an average of approximately 3 hours per dataset (over 340 CPU hours across the 112 archive datasets)<ref:2609.35929#pg2> In response, the field shifted toward randomized convolutional transforms, initiated by ROCKET [1], refined by MiniRocket [2], MultiRocket [4], and Hydra [3]<ref:2609.35929#pg2> By projecting time series into high-dimensional linear spaces via dilated convolutions pooled with Proportion of Positive Values (PPV), these methods run in seconds<ref:2609.

Improvements for AI systems

No monolithic ensemble reliance: CADENCE decouples representation learning into two specialized pathways (Branch A and Branch B), allowing it to leverage both a Convolutional Linear Expert (MiniRocket) for deterministic features and a Distributional Interval Expert (Hydra) for kinematic/spectral analysis.
Adaptive expert selection: CADENCE incorporates an internal validation meta-router with rare-class preservation that dynamically selects between pure expert routing and confidence-weighted soft blending, which allows the system to choose the most appropriate decision strategy based on empirical data, as described by its routing logic: The routing decision is governed by empirical difference Δ = Val A - Val B with threshold τ = 0.08.
Enhanced robustness against class imbalance: The meta-router implements a mechanism that allows for Conditional Inference-Time Branch Pruning which Completely skips the unselected branch during dominance regimes, reducing test-time overhead and improving performance when one expert clearly dominates, as noted in the analysis: In total, 27 datasets (24.8%) trigger dominance regimes where test inference skips one branch completely.
Superior performance on specialized tasks: CADENCE achieves decisive wins on specialized sensor and kinematic datasets over Hydra+MultiRocket, specifically demonstrating improvements like +9.96 pp on SemgHandMovementCh2, indicating high accuracy in complex physical monitoring applications.

Abstract

Time series classification (TSC) exhibits a sharp trade-off between accuracy and computational scalability. Meta-ensembles like HIVE-COTE 2.0 reach state-of-the-art accuracy but require extensive compute, whereas ultra-fast random convolutional transforms (e.g., MiniRocket, Hydra) run in seconds but struggle with phase-independent distributions, signal kinematics, and decision tree fragmentation on large class counts. In this work, we present CADENCE (Confidence-Adaptive Dual-Expert Network for time series Classification Excellence), a unified, CPU-native dual-expert architecture. CADENCE decouples representation learning into two specialized pathways: (i) a Convolutional Linear Expert pairing 10,000 deterministic dilated features with closed-form L2-regularized Woodbury ridge classification, and (ii) a Distributional Interval Expert pairing competing dilated kernels (Hydra) with dyadic Cornish-Fisher moment approximations across signal kinematics and FFT spectral bands, fitted with an ExtraTrees ensemble. An internal validation meta-router with rare-class preservation dynamically selects between pure expert routing and confidence-weighted soft blending, followed by a full refit on 100% of training data. Evaluated across all 109 equal-length UCR Archive datasets over 30 resamples (3,270 total runs), CADENCE achieves a grand mean accuracy of 0.8864. This ranks #2 across the archive, surpassed only by HIVE-COTE 2.0 (0.8895, p Holm = 0.295, no statistically significant difference), while outperforming Hydra+MultiRocket (0.8818), MultiRocket (0.8797), and HIVE-COTE 1.0 (0.8786, p Holm = 0.048). CADENCE closes the gap to HIVE-COTE 2.0 to 0.31 percentage points while taking an average of only 17.53 seconds per dataset on a dual-core CPU. Source code and evaluation scripts: https://github.com/onisa-jr/CADENCE.git