Web links no longer always pull toward neighbors: ZeroPolar reads link polarity without labels, reaching a 73.99% eleven-graph mean with GPT-4o labels
Synopsis
Web platforms must classify new pages before anyone labels them, large language model labels per page are noisy, and links mean opposite things across sites: online stores often join same-category products, while university sites mostly join pages of different classes. ZeroPolar calls the sign with which links enter a prediction structural polarity, reads that sign without true labels from the excess agreement of LLM labels and text-prototype predictions along edges, and applies one signed propagation step. With one GPT-4o label per page, ZeroPolar-T reaches the highest eleven-graph mean, 73.99%, 6.01 points above LLM-BP.
Interpretation
On Web graphs, links now enter zero-shot classification with a measurable sign: on homophilous graphs pulling a node toward its neighbors helps, while on heterophilous hyperlink graphs that pull hurts, so a method must first decide whether to trust or invert the links. Earlier zero-shot methods either ignored links, pulled a page toward its neighbors as label propagation and graph neural networks do, or weighted propagation by an LLM-estimated homophily; on the WebKB hyperlink graphs one step toward the neighbors costs at least 9.27 points. Across eleven text-attributed graphs, one step toward the neighbors helps on the seven homophilous graphs and costs at least 15.42 points in the other direction, while on the four WebKB graphs the step toward the neighbors costs at least 9.27 points, so the better sign tracks the graph kind.
Structural polarity is readable without any true label, so a new graph needs neither a hand-written list of graph families nor extra LLM calls to estimate homophily. The released approximate LLM-BP reads the sign from a hand-written list of heterophilous graphs that does not include a new graph, and counting how often linked nodes share an LLM label mixes homophily with chance agreement that class counts and sizes inflate. The excess agreement after subtracting chance, taken as the larger of the GPT-4o-label and prototype-prediction statistics, is at least 0.360 on every homophilous graph and at most 0.078 on every WebKB graph, and both track the excess agreement of true labels.
One signed propagation step combines text, prototype scores and LLM labels, with two implementations matched to the labelling budget: training-free when labels are few, and a text classifier that denoises labels once at least half the nodes carry one. Propagating raw labels cannot repair errors that linked pages share and pulls pages toward wrong classes on graphs whose links join different classes, while a GNN trained on the labels needs them on most nodes and learns its aggregation from noisy labels. With one GPT-4o label per page, ZeroPolar-T reaches a 73.99% eleven-graph mean, 6.01 points above LLM-BP and 2.06 points above SAGE h64, and significantly above all eighteen baselines; with labels on 5% of nodes, training-free ZeroPolar reaches 67.85% against 57.05% for SAGE h64 trained on as many randomly chosen labels, a 10.80-point lead.
Perspective
Validation covers eleven text-attributed graphs, two additional graphs and three further Web platforms, Instagram, Reddit and Photo, with GPT-4o, Qwen3-8B, Llama-3.1-8B, Mistral-7B and Qwen3.8-27B as annotators and repeated under seven encoder settings; the intended readers are researchers and platform engineers working on zero-shot node classification over text-attributed graphs.
On setting sensitivity, ZeroPolar's eleven-graph mean stays between 72.05% and 72.47% across sixteen tested values, ZeroPolar-T keeps 73.99% for every tested value, and one propagation step works best for both implementations while a second step lowers the mean; in the few-shot true-label setting with ten labels per class, ZeroPolar-T reaches 65.36%, below few-shot LLM-BP at 68.07%.