Time-series forecasting tasks are central to a broad range of application domains, including stock pricepredictions [1,2],servicedemandforecasting [3,4],andmedicalprognoses[5-7].

One ofthe most important strategies ofTransformer isposition encoding, which has been proven crucial when applying Transformer tomanycomputer vision tasks [22,34,35].

Anchorbased strategies have been treated as effective ways to alleviate such efficiency problems by propagation on representative entities instead of the whole graph.

Moreover, the problem of finding worst-case perturbations is non-convex and underparameterized, both ofwhich engender anon-favorable optimization landscape.

Right: measurement overthefirst60epochs. Update step adaptation (see Section 4.3)isapplied.