Sec. 2 provides more experimental details on Few-shot Learning for our GraphAdapter. Sec. 3 describes more details about datasets and implementation. Sec. 4 visualizes the textual graph nodes used for classification before and after utilizing our Sec. Notably, the TaskRes* exploits the enhanced base classifier. We present the numerical results of "Figure 3 in the main text" as Table 2.

GraphAdapter, which performs the textual adapter by explicitly modeling the dual-modality structure knowledge ( i.e., the correlation of different semantics/classes

See Fig. task-specificanylinear high-AS edefines T e ={t | =( e, l), l 2Rd}. InordertoT e in Eq. andsample l ateach average, discovery e that risetotheT e (see Different (e.g., 32 0.68compared discovered sampling 6

Inmanycases,thenumber ofdimensions in the data is much larger than the number of actual data samples.