Common use of Robustness Clause in Contracts

Robustness. We developed GAM with the goal of being able to handle graphs with “in- correct” edges (i.e. those that connect nodes with differing labels). We consider such edges “incorrect" under the label propagation assump- tion, despite the fact that they may refer to real- world connections between these nodes (e.g., citations between research articles on different topics). In ▇▇▇▇, Citeseer, and Pubmed, 19%, 26%, and 20% of the edges, respectively, are in- correct. To demonstrate the ability of GAM to MLP128 MLP128 + NGM MLP128 + GAM Accuracy (%) 20 30 40 50 60 70 74 handle these incorrect edges and perhaps even higher levels of noise, we performed a robust- ness analysis by introducing spurious edges to the graph, and testing whether our agreement model learns to ignore them. We added spuri- Figure 4: Robustness to noisy graphs. The x axis represents the percentage of correct edges remain- ing after adding wrong edges to the Citeseer dataset. ous edges by randomly sampling pairs of nodes with different true labels until the percentage of incorrect edges met a desired target. We tested the performance of GAM on a set of graphs created in this manner. MLPs are good base model candidates for testing this because they can only be affected by the graph quality through the GAM regularization terms (unlike GCN or GAT, where the graph is implicitly used in the model). The results are shown in Figure 4 on the Citeseer dataset (the hardest of the three datasets), for graphs containing between 5% and 74% correct edges. A plain MLP with 128 hidden units obtains 52.2% accuracy independent of the level of noise in the graph. Adding GAM to this MLP increases its accuracy by about 19%. This improvement persists even as the fraction of correct edges decreases. For example, the accuracy remains 70% even in the case where only 5% of the graph edges are correct. In contrast, the performance of NGM steadily decreases as the fraction of incorrect edges increases, to the point where it starts performing worse than the plain MLP (when the percent of correct edges ≤ 60%), and it is thus preferable not to use it.