Algorithm 2 Temporal Similarity Graph Learning

Input: refined tokens ${\mathbf{Z}}_m\in {\mathbb{R}}^{B\times T\times D_m}$. Output: classification logits $\widehat{\mathbf{y}}$. Graph node preparation:     Compress the feature dimension and obtain node features ${\mathbf{X}}_g\in {\mathbb{R}}^{B\times N\times D_g}$. Graph structure construction:     Build self-loop edges ${\mathbf{A}}_{\mathrm{self}}$ to preserve token identity.     Build temporal-chain edges ${\mathbf{A}}_{\mathrm{chain}}$ within window r to retain local continuity.     Compute cosine similarity $\mathbf{S}$ between normalized node features.     Apply temporal penalty ${\tilde{\mathbf{S}}}_{ij}={\mathbf{S}}_{ij}-\gamma |i-j|/(N-1)$.     Select top-k neighbors and form similarity edges ${\mathbf{A}}_{\mathrm{sim}}$.     Fuse edges: $\mathbf{A}=\mathrm{clip}({\mathbf{A}}_{\mathrm{self}}+{\mathbf{A}}_{\mathrm{chain}}+{\mathbf{A}}_{\mathrm{sim}},0,1)$. Hierarchical graph learning:     Compute assignment and embedding matrices with DenseSAGEConv blocks.     Apply DiffPool twice to obtain coarsened graph representations.     Refine the final graph with DenseSAGEConv and average node embeddings. return $\widehat{\mathbf{y}}=\mathrm{MLP}\left(\mathbf{g}\right)$.