cs.LGApr 13, 2026
SaveExploring Concept Subspace for Self-explainable Text-Attributed Graph Learning
Organizations: Department of Computer Science, Stevens Institute of Technology, Hoboken, NJ, USA.
Abstract
We introduce Graph Concept Bottleneck (GCB) as a new paradigm for self-explainable text-attributed graph learning. GCB maps graphs into a subspace, concept bottleneck, where each concept is a meaningful phrase, and predictions are made based on the activation of these concepts. Unlike existing interpretable graph learning methods that primarily rely on subgraphs as explanations, the concept bottleneck provides a new form of interpretation. To refine the concept space, we apply the information bottleneck principle to focus on the most relevant concepts. This not only yields more concise and faithful explanations but also explicitly guides the model to "think" toward the correct decision. We empirically show that GCB achieves intrinsic interpretability with accuracy on par with black-box Graph Neural Networks. Moreover, it delivers better performance under distribution shifts and data perturbations, showing improved robustness and generalizability, benefitting from concept-guided prediction.
Figures & tables
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 50.00 (0.63) | 60.81 (0.50) | 38.44 (0.99) | 54.34 (1.10) | 35.42 (0.58) | 51.69 (0.26) | 16.38 (0.52) | 51.33 (0.51) | 54.31 (0.31) | 65.30 (0.35) |
| GCN | 55.10 (0.66) | 64.03 (1.05) | 46.52 (0.92) | 59.64 (0.92) | 36.98 (0.84) | 50.01 (0.70) | 12.91 (0.38) | 48.48 (0.21) | 59.04 (1.66) | 68.38 (1.73) |
| GAT | 51.30 (1.27) | 61.52 (1.15) | 45.62 (1.01) | 58.77 (0.87) | 33.31 (0.59) | 50.18 (0.41) | 12.93 (0.30) | 49.34 (0.18) | 57.05 (1.00) | 64.53 (0.85) |
| SAGE | 44.26 (1.78) | 53.95 (1.65) | 30.87 (0.41) | 48.42 (0.47) | 31.46 (0.20) | 48.27 (0.18) | 13.38 (0.17) | 49.18 (0.47) | 51.87 (1.24) | 62.06 (1.37) |
| GT | 38.26 (1.64) | 48.66 (1.36) | 28.38 (1.58) | 48.22 (0.77) | 30.90 (0.43) | 48.06 (0.37) | 12.64 (0.54) | 48.62 (0.26) | 54.06 (0.84) | 62.50 (0.77) |
Table 1: Node classification performance in OOD settings with upsampling ratio . The best-performing interpretable GNN is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 43.08 (0.45) | 56.55 (0.37) | 37.77 (0.53) | 53.77 (0.59) | 36.69 (0.45) | 52.57 (0.19) | 16.75 (0.84) | 51.11 (0.44) | 53.70 (0.32) | 64.95 (0.45) |
| GCN | 56.21 (1.25) | 66.49 (0.85) | 46.45 (1.42) | 58.14 (1.61) | 43.87 (3.28) | 54.11 (0.85) | 16.64 (0.77) | 50.43 (0.76) | 61.45 (0.38) | 65.64 (0.87) |
| GAT | 52.64 (1.33) | 61.70 (0.71) | 46.55 (0.87) | 60.54 (0.58) | 37.07 (1.09) | 52.38 (0.43) | 19.23 (1.76) | 52.20 (0.84) | 59.34 (0.90) | 66.75 (1.21) |
| SAGE | 53.15 (2.06) | 57.70 (2.21) | 39.37 (1.30) | 56.60 (0.90) | 38.45 (2.14) | 52.79 (0.90) | 16.51 (0.49) | 51.64 (0.39) | 61.58 (0.44) | 69.76 (0.49) |
| GT | 46.13 (2.23) | 55.15 (1.80) | 33.36 (1.69) | 52.65 (0.93) | 35.36 (0.78) | 51.70 (0.27) | 17.29 (0.50) | 51.31 (0.12) | 57.88 (0.95) | 62.61 (1.60) |
Table 2: Node classification performance in perturbation settings with pertubation ratio . The best-performing interpretable GNN is underlined , and the overall best-performing method is bolded .
Figure 1: Performance of the original GCB compared to its variant with random concepts (GCB-RC) across different concept set sizes, on regular splits (top row) and OOD splits (bottom row).
Figure 2: Performance of GCB across different concept sizes ( ) and training ratios (%) on regular splits and OOD splits.
Figure 3: Case study on cora . We visualize the concept activations of two instances produced by both the MLP encoder, which uses only node features, and the GCN encoder, which incorporates 2-hop neighborhood information.
Figure 4: Performance of GCB variations using different graph encoders.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | #Nodes | #Edges | Type | Domain | #Class |
|---|---|---|---|---|---|
| Computers | 87,229 | 721,081 | Co-purchase | E-commerce | 10 |
| PubMed | 19,717 | 44,338 | Citation | Biomedicine | 3 |
| Books-History | 41,551 | 358,574 | Co-purchase | E-commerce | 12 |
| Books-Children | 76,875 | 1,554,578 | Co-purchase | E-commerce | 24 |
| Sports-Fitness | 173,055 | 1,773,500 | Co-purchase | E-commerce | 13 |
| Cora | 2,708 | 5,429 | Citation | Computer Science | 7 |
Table 3: Summary statistics of source and target datasets.
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 68.00 (0.89) | 67.49 (0.76) | 63.81 (0.37) | 64.29 (0.34) | 53.78 (0.59) | 53.76 (0.58) | 53.34 (0.77) | 53.39 (0.76) | 69.22 (0.51) | 69.25 (0.57) |
| GCN | 72.38 (0.58) | 71.95 (0.49) | 62.47 (0.29) | 63.05 (0.33) | 53.63 (0.84) | 53.62 (0.82) | 54.49 (1.19) | 54.64 (1.03) | 67.45 (1.76) | 68.84 (1.82) |
| GAT | 74.58 (0.95) | 74.42 (0.90) | 63.92 (0.92) | 64.47 (0.91) | 55.71 (1.06) | 55.74 (1.10) | 56.29 (0.60) | 56.30 (0.59) | 67.54 (1.68) | 68.14 (1.59) |
| SAGE | 70.59 (0.68) | 70.66 (0.80) | 65.09 (0.62) | 65.52 (0.64) | 54.49 (0.46) | 54.50 (0.46) | 55.33 (0.38) | 55.33 (0.38) | 72.96 (0.30) | 72.74 (0.40) |
| GT | 72.36 (1.96) | 72.18 (1.56) | 64.40 (0.76) | 64.93 (0.68) | 54.79 (0.40) | 54.77 (0.39) | 56.15 (0.39) | 56.15 (0.39) | 72.27 (0.52) | 72.46 (0.56) |
Table 4: Node classification performance in regular settings . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 46.52 (0.59) | 57.50 (0.69) | 44.89 (0.70) | 60.49 (0.77) | 35.55 (0.59) | 51.54 (0.20) | 17.03 (0.51) | 51.28 (0.56) | 53.82 (0.42) | 63.23 (0.40) |
| GCN | 56.10 (0.46) | 64.04 (0.52) | 41.06 (0.43) | 56.08 (0.57) | 39.36 (3.25) | 52.54 (0.86) | 16.65 (0.73) | 50.74 (0.20) | 53.80 (0.55) | 60.37 (1.37) |
| GAT | 52.54 (1.48) | 63.29 (1.31) | 44.71 (0.68) | 60.63 (0.56) | 33.42 (0.37) | 51.49 (0.12) | 13.06 (0.28) | 49.78 (0.39) | 56.41 (2.24) | 64.32 (1.67) |
| SAGE | 40.39 (1.19) | 50.69 (1.08) | 40.99 (0.99) | 56.02 (0.78) | 35.71 (0.42) | 51.76 (0.31) | 15.97 (0.35) | 50.74 (0.51) | 49.65 (0.64) | 60.20 (0.83) |
| GT | 42.83 (1.35) | 51.62 (1.59) | 40.57 (1.22) | 56.93 (0.90) | 33.83 (0.60) | 51.47 (0.28) | 13.22 (0.33) | 50.14 (0.34) | 51.10 (0.69) | 59.51 (0.99) |
Table 5: Node classification performance in OOD settings with upsampling ratio . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 43.78 (0.52) | 54.85 (0.57) | 45.73 (0.59) | 58.72 (0.71) | 37.04 (0.67) | 52.33 (0.25) | 16.37 (0.61) | 51.37 (0.32) | 55.01 (0.33) | 65.33 (1.15) |
| GCN | 57.28 (1.35) | 67.86 (0.98) | 41.59 (0.68) | 56.90 (0.80) | 37.73 (3.06) | 51.58 (0.59) | 16.00 (0.35) | 49.23 (0.48) | 54.99 (1.28) | 61.65 (2.33) |
| GAT | 52.81 (1.43) | 60.66 (1.28) | 43.09 (1.45) | 58.16 (1.39) | 34.77 (1.08) | 51.50 (0.38) | 13.85 (0.44) | 49.44 (0.28) | 56.99 (0.10) | 65.85 (0.84) |
| SAGE | 51.40 (1.77) | 62.46 (1.54) | 36.25 (1.39) | 52.63 (1.14) | 35.04 (0.55) | 51.47 (0.27) | 14.36 (0.17) | 48.96 (0.40) | 52.43 (0.78) | 60.92 (0.97) |
| GT | 47.70 (1.31) | 58.26 (1.19) | 36.12 (1.62) | 53.10 (1.23) | 33.16 (0.47) | 50.15 (0.37) | 14.18 (0.15) | 49.59 (0.21) | 54.83 (0.89) | 62.58 (1.07) |
Table 6: Node classification performance in OOD settings with upsampling ratio . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 47.58 (0.44) | 59.14 (0.61) | 41.44 (0.42) | 56.87 (0.70) | 35.38 (0.66) | 51.29 (0.43) | 16.08 (0.42) | 50.69 (0.34) | 52.72 (0.50) | 62.79 (0.50) |
| GCN | 62.08 (1.59) | 69.26 (1.83) | 43.58 (0.45) | 54.79 (0.34) | 47.15 (4.10) | 55.23 (1.48) | 17.35 (0.97) | 49.91 (0.17) | 62.47 (1.15) | 64.89 (1.38) |
| GAT | 60.32 (1.56) | 68.94 (1.15) | 48.46 (0.66) | 61.74 (0.80) | 35.80 (0.67) | 51.70 (0.30) | 15.35 (1.35) | 51.40 (0.38) | 57.17 (1.59) | 60.82 (1.69) |
| SAGE | 50.49 (1.06) | 57.96 (1.12) | 35.75 (1.49) | 53.22 (0.90) | 37.94 (0.94) | 52.14 (0.40) | 16.70 (0.73) | 52.08 (0.22) | 62.88 (0.18) | 72.40 (1.02) |
| GT | 47.31 (2.61) | 56.27 (1.92) | 30.80 (1.22) | 50.68 (0.82) | 37.51 (0.46) | 52.90 (0.16) | 17.33 (0.79) | 51.58 (0.24) | 62.18 (0.80) | 68.79 (1.87) |
Table 7: Node classification performance in OOD settings with upsampling ratio . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 46.79 (0.88) | 58.83 (0.84) | 38.16 (0.41) | 53.71 (0.37) | 37.69 (0.44) | 52.53 (0.15) | 16.19 (0.46) | 51.58 (0.44) | 53.96 (0.34) | 64.83 (0.25) |
| GCN | 58.93 (1.32) | 67.16 (1.38) | 46.96 (0.95) | 58.48 (1.13) | 42.47 (4.66) | 52.11 (1.09) | 15.95 (0.86) | 50.46 (0.68) | 62.44 (0.37) | 68.39 (0.49) |
| GAT | 55.01 (1.93) | 61.57 (1.94) | 43.59 (1.57) | 57.78 (1.21) | 34.32 (1.37) | 51.16 (0.53) | 17.03 (0.95) | 51.28 (0.42) | 58.93 (2.81) | 66.43 (2.35) |
| SAGE | 53.45 (2.23) | 57.21 (2.34) | 42.45 (1.44) | 57.74 (0.95) | 40.93 (6.08) | 52.32 (0.67) | 16.00 (0.61) | 51.42 (0.61) | 61.58 (0.22) | 68.96 (1.30) |
| GT | 38.25 (2.04) | 45.19 (1.18) | 39.80 (3.32) | 55.30 (2.16) | 34.43 (1.15) | 51.44 (0.30) | 14.35 (0.74) | 51.02 (0.27) | 56.30 (1.16) | 64.35 (1.49) |
Table 8: Node classification performance in adversarial settings with perturbation ratio . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 45.04 (1.20) | 57.94 (0.91) | 41.94 (0.33) | 57.38 (0.21) | 34.65 (0.60) | 51.57 (0.30) | 18.09 (0.43) | 51.09 (0.58) | 54.13 (0.30) | 64.72 (0.64) |
| GCN | 65.19 (1.66) | 67.62 (1.61) | 46.43 (1.13) | 58.40 (1.17) | 38.56 (1.38) | 51.96 (0.60) | 18.17 (1.20) | 50.51 (0.67) | 63.15 (2.44) | 68.98 (2.14) |
| GAT | 63.56 (1.59) | 68.88 (1.40) | 43.79 (1.41) | 58.11 (0.90) | 35.94 (2.11) | 51.83 (0.55) | 17.30 (1.53) | 51.83 (0.72) | 58.79 (1.66) | 67.83 (1.26) |
| SAGE | 47.78 (0.92) | 55.17 (0.52) | 40.60 (0.80) | 56.65 (0.63) | 38.86 (1.28) | 52.61 (0.62) | 16.93 (0.48) | 51.67 (0.34) | 59.90 (0.67) | 66.69 (0.61) |
| GT | 40.32 (1.93) | 46.77 (1.27) | 27.14 (1.44) | 46.19 (1.04) | 35.26 (0.73) | 51.95 (0.35) | 16.80 (1.02) | 51.26 (0.49) | 60.67 (0.99) | 67.86 (1.05) |
Table 9: Node classification performance in adversarial settings with perturbation ratio . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 49.30 (0.81) | 59.94 (0.97) | 42.01 (0.43) | 57.97 (0.22) | 37.57 (3.23) | 51.27 (0.41) | 15.37 (0.09) | 51.51 (0.29) | 52.92 (0.41) | 61.84 (0.70) |
| GCN | 60.24 (0.83) | 68.81 (1.01) | 47.16 (1.29) | 58.67 (1.02) | 37.70 (2.16) | 51.52 (0.53) | 17.24 (1.82) | 50.60 (0.84) | 59.73 (0.71) | 62.88 (0.35) |
| GAT | 57.88 (2.24) | 64.39 (1.85) | 44.69 (1.35) | 58.83 (1.13) | 37.82 (1.08) | 52.08 (0.46) | 14.93 (1.04) | 50.50 (0.56) | 58.14 (2.12) | 62.40 (2.18) |
| SAGE | 50.26 (1.95) | 57.82 (1.43) | 29.81 (2.27) | 49.99 (1.55) | 36.98 (0.47) | 52.00 (0.08) | 17.10 (0.40) | 50.99 (0.43) | 62.87 (0.63) | 70.36 (0.31) |
| GT | 51.12 (2.34) | 56.14 (2.39) | 32.63 (1.41) | 51.08 (0.71) | 35.34 (0.86) | 51.61 (0.50) | 16.19 (0.68) | 50.84 (0.28) | 60.47 (0.47) | 66.23 (0.77) |
Table 10: Node classification performance in adversarial settings with perturbation ratio . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .
| Cora | Citeseer | WikiCS | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) | F1 (%) | BACC (%) |
| MLP | 47.76 (0.43) | 58.52 (0.43) | 44.65 (0.39) | 58.45 (0.41) | 35.26 (0.57) | 51.72 (0.39) | 16.78 (0.31) | 50.46 (0.59) | 55.24 (0.16) | 65.36 (1.10) |
| GCN | 52.19 (1.00) | 61.01 (0.95) | 46.07 (1.23) | 56.61 (1.12) | 42.37 (1.34) | 53.07 (0.85) | 16.20 (1.08) | 50.73 (0.48) | 65.21 (0.63) | 68.76 (1.02) |
| GAT | 54.25 (2.86) | 63.55 (1.63) | 46.58 (0.47) | 60.72 (0.81) | 37.55 (1.31) | 52.63 (0.43) | 18.61 (0.96) | 51.83 (0.37) | 59.33 (2.20) | 65.94 (1.80) |
| SAGE | 44.04 (1.28) | 50.19 (0.89) | 32.53 (2.83) | 50.50 (1.70) | 36.20 (0.84) | 52.24 (0.19) | 16.29 (0.46) | 51.15 (0.41) | 62.65 (0.42) | 70.73 (0.50) |
| GT | 42.25 (1.81) | 52.13 (2.09) | 31.04 (2.55) | 49.44 (2.11) | 35.67 (0.87) | 51.02 (0.17) | 13.85 (0.83) | 50.82 (0.54) | 62.95 (1.19) | 70.35 (1.05) |
Table 11: Node classification performance in adversarial settings with perturbation ratio . The best-performing interpretable GNN on each dataset is underlined , and the overall best-performing method is bolded .