KAN-ADIA 2026 Abstracts

Area 1 - KAN-ADIA
Full Papers Short Papers

Area 1 - KAN-ADIA

Full Papers
Paper Nr: 6
Title:

Time-Series Forecasting with NBEATS and TimeKAN Models Using KAN Variants
Authors:

Ali Bayeh and Samira Sadaoui
Abstract: This paper presents a comprehensive benchmark of leading Kolmogorov-Arnold Network (KAN) variants for time-series forecasting. Despite the proliferation of KAN models and their promising results, a systematic, comparative evaluation on standardized real-world forecasting tasks is still lacking. To address this gap, we conduct an empirical study that integrates several KAN variants, such as baseline KAN, ChebyKAN, EfficientKAN, SineKAN and TruKAN (with fixed and learnable knots, denoted TruKAN‑F and TruKAN‑L respectively), into two advanced deep‑learning architectures: NBEATS and TimeKAN. NBEATS employs a deep stack of progressively refining blocks, while TimeKAN uses parallel blocks operating across multiple frequency components. These designs make these architectures well suited to modeling complex temporal dependencies of time-series data. We evaluate and compare the performance of all these models on four diverse, real‑world benchmark datasets (ETTh1, ETTm1, Electricity, and Weather) across multiple prediction horizons, ranging from short‑term to long‑term. The empirical results demonstrate that integrating robust temporal modeling architectures with flexible function approximators leads to strong predictive performance.
Paper Nr: 7
Title:

KANGuard: Residual Backdoor Risk Validation for Machine Unlearning Using Kolmogorov-Arnold Networks
Authors:

Abbas Yazdinejad
Abstract: Machine unlearning has been proposed as a principled mechanism for removing the influence of poisoned training data from deployed models. A critical weakness of current evaluation practice is that a reduction in attack success rate (ASR) is treated as the primary evidence of successful forgetting. We demonstrate empirically that four out of five approximate unlearning methods leave substantial residual ASR between 68% and 99% for high-impact attacks while simultaneously reporting near-baseline clean accuracy, exposing a fundamental gap between standard classification-level metrics and genuine behavioural normalisation. We formalise this gap through the concept of behavioural fidelity, defined as the statistical closeness of the post-unlearning model’s output distribution to a trusted clean reference, and cast its detection as a binary hypothesis test under the Neyman–Pearson framework. We propose KANGuard, a post-unlearning validation framework in which Kolmogorov–Arnold Network (KAN) classifiers act as the decision boundary over a structured feature space of logit vectors, predictive entropy, confidence margins, energy scores, and penultimate embeddings. KANs are chosen for their per-feature spline transformations, which are architecturally suited to the heterogeneous anomaly structures present across these feature groups. Evaluated on CIFAR-10 and the safety-critical German Traffic Sign Recognition Benchmark, KANGuard demonstrates consistent residual-risk discrimination across multiple unlearning strategies and poison rates. Under sparse trigger coverage (γ = 0.1), KAN-Full achieves AUROC ≈ 0.83 while unsupervised baselines collapse to near chance. In the genuinely hard reduced-separability regime, AUROC remains at 0.89–0.92. KAN-Full additionally maintains well-calibrated uncertainty estimates (ECE < 0.003) without post-hoc correction, enabling interpretable deployment-gating decisions.
Short Papers
Paper Nr: 8
Title:

Improving Kolmogorov-Arnold Networks Using Nature-Inspired Optimization Metaheuristics
Authors:

Haifa Zaidi, Malek Mouhoub and Samira Sadaoui
Abstract: Kolmogorov-Arnold Networks (KANs) have recently became a strong alternative to traditional neural networks because of their effectiveness in function approximation. However, KANs use a single global representation of the problem space, which can limit their ability to model functions with local complexity or discontinuities. To address this issue, X-KAN was proposed by integrating KAN into the eXtended Classifier System for Function approximation (XCSF), an evolutionary rule-based learning framework. In this paper, we improve X-KAN by exploring two nature-inspired optimization methods: Genetic Algorithms (GAs) and Particle Swarm Optimization (PSO). We introduce new evolutionary operators, leading to new variants called X-KAN1 and X-KAN2. These operators are designed to improve the search process by balancing exploration and exploitation, with the goal of producing rule populations that achieve better predictive performance while reducing model complexity. Empirical results on several regression datasets and synthetic benchmark functions show that the two variants are effective compared with the baseline KAN.