LabelKAN -- Kolmogorov-Arnold Networks for Inter-Label Learning: Avian Community Learning

TL;DR

LabelKAN introduces a modular, interpretable inter-label learning mechanism based on Kolmogorov-Arnold Networks that sits atop existing multi-label predictors to learn species co-occurrence patterns in ecological communities. By training a per-label KAN head on the base model's predictions, the approach captures higher-order label dependencies and improves both presence-absence and presence-only predictions, with notable gains for rare species. Across NN, DMVP, and VAE baselines in avian JSDMs, LabelKAN yields consistent improvements (approximately 2–3% in micro and macro AUPRC), and the VAE + LabelKAN combination achieves the best overall performance while enabling interpretability of learned label interactions. The work demonstrates practical conservation insights, including how declines in a sentinel species like the Great Blue Heron can cascade through associated bird communities, informing targeted management under frameworks like GBF.

Abstract

Global biodiversity loss is accelerating, prompting international efforts such as the Kunming-Montreal Global Biodiversity Framework (GBF) and the United Nations Sustainable Development Goals to direct resources toward halting species declines. A key challenge in achieving this goal is having access to robust methodologies to understand where species occur and how they relate to each other within broader ecological communities. Recent deep learning-based advances in joint species distribution modeling have shown improved predictive performance, but effectively incorporating community-level learning, taking into account species-species relationships in addition to species-environment relationships, remains an outstanding challenge. We introduce LabelKAN, a novel framework based on Kolmogorov-Arnold Networks (KANs) to learn inter-label connections from predictions of each label. When modeling avian species distributions, LabelKAN achieves substantial gains in predictive performance across the vast majority of species. In particular, our method demonstrates strong improvements for rare and difficult-to-predict species, which are often the most important when setting biodiversity targets under frameworks like GBF. These performance gains also translate to more confident predictions of the species spatial patterns as well as more confident predictions of community structure. We illustrate how the LabelKAN leads to qualitative and quantitative improvements with a focused application on the Great Blue Heron, an emblematic species in freshwater ecosystems that has experienced significant population declines across the United States in recent years. Using the LabelKAN framework, we are able to identify communities and species in New York that will be most sensitive to further declines in Great Blue Heron populations.

Figures (6)

• Figure 1: (Top) Example 3-layer Kolmogorov-Arnold Network with learned activation functions. (Bottom) LabelKAN model abstraction with KAN heads.
• Figure 2: Comparison between LabelKAN (top) and a Neural Network label layer (bottom) as a function of the mean change in error versus the species prevalence. On average, the LabelKAN model performs better at all prevalence levels over its base model, with only $7.5\%$ of species performing worse, with the worst performance reduction of only $0.4\%$. In contrast, for the Neural Network, nearly $25\%$ of species perform worse than the base line model, with a worst performing species degradation of $1.16\%$ average error.
• Figure 3: Predictions for the American Robin, a common species, on August 1, 2019. LabelKAN (left) results in more confident predictions (closer to the 0/1 value) across the spatial extent than the base VAE (right).
• Figure 4: Subset of the learned functions and edge connections from the LabelKAN model for the American Robin (top) and Northern Shoveler (bottom), the most and least prevalent species respectively, showing the top 5 contributing species. Each plot has, for the first layer, the average relative strength kans. For the American Robin, most of the signal comes from itself, whereas for the Northern Shoveler, the signal strength from the input species logits is more distributed across co-occurring species.
• Figure 5: Predictions for the Least Flycatcher, a rare species, on August 1, 2019 for the VAE + LabelKAN (left) and base VAE (right). Despite the KAN only receiving information about the predicted community from the VAE, it significantly improves the species predictions in the Adirondacks and southwestern New York compared to the base model (validation by experts at the Cornell Lab of Ornithology), showing the value of learning from co-occurring species.