Back

LLaGA: Large Language and Graph Assistant

Runjin ChenTong ZhaoAjay Kumar JaiswalNeil ShahZhangyang Wang

article2024ICML245 citations

Introduces a general-purpose framework that reorganizes graph topologies into structure-aware node sequences and maps them directly into large language model token embeddings, outperforming specialized graph neural networks across multiple tasks and unseen datasets without modifying the base model parameters.

Listen

Modern organizations increasingly rely on graph-structured data—such as social networks, citation databases, and e-commerce product catalogs—to extract strategic insights. While Large Language Models (LLMs) provide advanced reasoning and conversational abilities across diverse applications, adapting them to graph data remains difficult because translating complex topological relationships into plain text descriptions is often inefficient, verbose, and prone to losing structural nuance. Conversely, specialized graph neural networks typically suffer from poor task generalization and require separate tuning and custom classification heads for each new application.

The article introduces and evaluates the Large Language and Graph Assistant (LLaGA), a framework designed to adapt graph-structured data into LLM-compatible inputs without modifying the underlying language model's parameters. The core objective is to demonstrate that a single unified framework can achieve state-of-the-art accuracy across multiple standard graph tasks, generate natural language explanations for its predictions, and generalize effectively to entirely new datasets and domains in zero-shot settings.

To achieve this, the authors developed parameter-free structural templates that convert graph structures and neighborhood information into ordered node embedding sequences. They introduced two formats: the Neighborhood Detail Template, which captures local multi-hop subtrees with structural Laplacian positional embeddings, and the Hop-Field Overview Template, which summarizes broader neighborhoods via multi-hop message passing. These node sequences are translated into the token embedding space of a frozen language model (such as Vicuna-7B) using a lightweight, trainable projector. Training frames all tasks—specifically node classification, link prediction, and descriptive summarization—into a uniform question-answer conversational format. The approach was evaluated across four widely recognized benchmark datasets (Cora, Pubmed, ogbn-Arxiv, and ogbn-Products) against traditional graph neural networks, graph transformers, and language models.

The findings establish that LLaGA consistently matches or outperforms specialized graph baselines across single-task, multi-task, and zero-shot environments. In standard multi-task classification and link prediction settings, LLaGA achieved top performance across all four datasets, reaching up to 95.06% accuracy on Pubmed classification and 97.38% on product link prediction, whereas baseline models often suffered noticeable performance degradation when trained simultaneously across multiple datasets. In zero-shot transfer evaluations, where the model was tested on unseen citation and e-commerce graphs without additional training, LLaGA demonstrated massive gains; on zero-shot link prediction, it reached 87.35% and 92.99% accuracy on Cora and Products datasets respectively, vastly outperforming existing baselines that hovered around 50% to 68%. Furthermore, the model successfully generated coherent, accurate natural language explanations for node embeddings and maintained robust performance across different underlying base LLMs and text encoders.

These results demonstrate that organizations can deploy a single, general-purpose assistant to handle multiple graph-based machine learning workflows simultaneously. By freezing the base language model and training only the projector, the framework significantly reduces computational and fine-tuning overhead while eliminating the need for costly task-specific model architectures. The framework's explainability directly improves transparency in automated decision-making, while its out-of-domain transfer ability reduces the need for extensive data labeling in new operational environments.

Organizations evaluating this approach should consider adopting structural sequence translation over verbose textual graph prompting to achieve superior accuracy and lower latency. Practitioners should select templates based on domain requirements, leveraging neighborhood detail templates for localized, fine-grained tasks and hop-field overview templates for tasks needing broad receptive context. Continued development and ethical monitoring are recommended to validate performance across broader enterprise graphs and guard against automated decision biases. Confidence in these experimental findings is high given rigorous benchmarking against multiple competitive baselines, though users should note that performance was primarily validated on text-attributed citation and e-commerce benchmarks.

Snap Inc.University of Texas at Austindirected graphsgeneral purposegraph neural networksLLaGA frameworktext-attributed graphlanguage models
arXiv: 2402.08170VITA-Group/LLaGA

No sufficiently relevant recommendations were found.

Cover for LLaGA: Large Language and Graph Assistant

Abstract

Graph Neural Networks (GNNs) have empowered the advance in graph-structured data analysis. Recently, the rise of Large Language Models (LLMs) like GPT-4 has heralded a new era in deep learning. However, their application to graph data poses distinct challenges due to the inherent difficulty of translating graph structures to language. To this end, we introduce the Large Language and Graph Assistant (LLaGA), an innovative model that effectively integrates LLM capabilities to handle the complexities of graph-structured data. LLaGA retains the general-purpose nature of LLMs while adapting graph data into a format compatible with LLM input. LLaGA achieves this by reorganizing graph nodes to structure-aware sequences and then mapping these into the token embedding space through a versatile projector. LLaGA excels in versatility, generalizability and interpretability, allowing it to perform consistently well across different datasets and tasks, extend its ability to unseen datasets or tasks, and provide explanations for graphs. Our extensive experiments across popular graph benchmarks show that LLaGA delivers outstanding performance across four datasets and three tasks using one single model, surpassing state-of-the-art graph models in both supervised and zero-shot scenarios. Our code is available at https://github.com/VITA-Group/LLaGA

Topics

directed graphsgeneral purposegraph neural networksgraph structuresinterpretabilitylanguage modellarge language modelsprojectorbipartite graphsweighted graph

Keywords

LLaGA frameworktext-attributed graphlanguage modelsself-supervised graph neural networksgraph transformernode classification taskslink prediction tasks

Table of Contents

  • 1. Introduction
  • 2. Methodology
  • 2.1. Notation
  • 2.2. Structure-Aware Graph Translation
  • 2.3. Alignment Tuning
  • 3. Experimental Results
  • 3.1. Setup
  • 3.2. Overall Performance Comparison (RQ1)
  • 3.3. Interpretation Ability Investigation (RQ2)
  • 3.4. Zero-Shot Ability Investigation (RQ3)
  • 3.5. Templates Ablation Study (RQ4)
  • 4. Related Work
  • 4.1. Graph Neural Networks
  • 4.2. Self-Supervised Learning for GNNs
  • 4.3. Large Language Models for Graphs
  • 5. Conclusion
  • Impact Statement
  • Acknowledgement
  • References
  • A. Dataset Statistics
  • B. More Metrics for Link Prediction Task
  • C. Extending to Non-TAGs
  • D. Flexibility with Text Encoding Methods
  • E. Integration with Various LLMs
  • F. Compared with Baseline Targets for TAGs
  • G. Experiment Variance

Knowls

  1. Knowl 1 — LLaGA aligns ordered graph-node representations with a frozen language model

    model/method

    LLaGA adapts a large language model (LLM) to graph tasks by representing each target node together with structural information from its neighborhood as an ordered sequence of node embeddings. A learned multilayer perceptron projector maps those embeddings into the LLM’s token-embedding space, where they replace the graph placeholder in a natural-language prompt. The LLM remains frozen; the projector is the only component tuned. A single projector is trained across datasets and tasks, allowing the same general-purpose LLM to answer graph questions without task-specific output heads.

  2. Knowl 2 — Neighborhood Detail Template creates a fixed-shape, ordered neighborhood sequence

    model/method

    For a target node vv, the Neighborhood Detail (ND) Template builds a fixed-shape sampled computational tree. Choose a neighbor sample size nin_i for each hop ii. At hop 1, sample up to n1n_1 neighbors of vv; if fewer are available, add placeholder nodes until there are exactly n1n_1. Recursively expand every real node at hop i−1i-1 with up to nin_i sampled neighbors, padding short neighborhoods to the required size. A placeholder’s descendants are all placeholders. Finally, traverse the tree in level order to form the node sequence. With depth kk, the sequence length is 1+n1+n1n2+⋯+∏j=1knj1+n_1+n_1n_2+\cdots+\prod_{j=1}^{k}n_j. Each sequence position corresponds to a specific position in the sampled tree, so the order preserves information about relative graph structure.

  3. Knowl 3 — Hop-Field Overview Template averages node features by neighborhood hop

    equation

    The Hop-Field Overview (HO) Template represents a target node’s surroundings with one embedding per hop, rather than separate embeddings for each sampled neighbor. Let xvx_v be the text feature of node vv, let ϕ\phi be a fixed text encoder, and let Nv1N_v^1 be the set of immediate neighbors of vv. Initialize hv(0)=ϕ(xv)h_v^{(0)}=\phi(x_v) and recursively compute, for nodes with nonempty immediate neighborhoods, hv(i)=1∣Nv1∣∑u∈Nv1hu(i−1)h_v^{(i)}=\frac{1}{|N_v^1|}\sum_{u\in N_v^1}h_u^{(i-1)}. This parameter-free neighbor-averaging message passing yields a sequence of hop embeddings for the central node. The experiments use four hop embeddings. The template summarizes each neighborhood field and can represent broader context, while retaining less individual-neighbor detail than the ND Template.

  4. Knowl 4 — Node-sequence embeddings combine encoded attributes with tree-position encodings

    equation

    In the ND Template, each position in the sampled tree receives a node-feature embedding and a structural position embedding. For text-attributed nodes, a fixed text encoder ϕ\phi maps the node text feature xvix_{v_i} to a vector; the paper uses encoders including SimTeG, SBERT, and RoBERTa. Let AtreeA_{\mathrm{tree}} be the adjacency matrix of the fixed-shape computational tree, DD its degree matrix, and II the identity matrix. The normalized Laplacian is L=I−D−1/2AtreeD−1/2=UTΛUL=I-D^{-1/2}A_{\mathrm{tree}}D^{-1/2}=U^{\mathsf T}\Lambda U, where UU contains Laplacian eigenvectors and Λ\Lambda contains the corresponding eigenvalues. If UiU_i is the position encoding for sequence position ii, the resulting embedding is hvi=0∥Uih_{v_i}=0\mathbin{\|}U_i for a placeholder and hvi=ϕ(xvi)∥Uih_{v_i}=\phi(x_{v_i})\mathbin{\|}U_i otherwise; ∥\| denotes concatenation, and the placeholder feature vector is zero with the text-embedding dimension. Because the sampled tree shape is fixed for a given template, its Laplacian position encodings are computed once and reused.

  5. Knowl 5 — One question-answer objective aligns the projector across three graph tasks

    model/method

    LLaGA trains the projector using node classification, link prediction, and node description in a shared question-answer format. Classification prompts ask for the target node’s class; link prompts ask whether two target nodes should be connected and expect yes or no; description prompts ask for a textual account of the central node. For text-attributed graphs, description targets are based on node features, with class information included in the target description. After tokenizing the prompt, the model inserts projected node embeddings at the graph-sequence position and trains to maximize pθ(Xanswer∣Xgraph,Xquestion,Xsystem)p_\theta(X_{\mathrm{answer}}\mid X_{\mathrm{graph}},X_{\mathrm{question}},X_{\mathrm{system}}). Here θ\theta denotes the trainable projector parameters, and the conditional probability is that of generating the answer given the graph input and prompt. The LLM is frozen, and the tasks share the projector rather than using separate task-specific losses or prediction heads.

  6. Knowl 6 — Evaluation uses four graph datasets and fixed projector-training settings

    experimental setup

    Experiments use Cora (2,708 nodes; 5,429 edges), Pubmed (19,717; 44,338), ogbn-Arxiv (169,343; 1,166,243), and ogbn-Products (2,449,029; 61,859,140). The first three are citation graphs; Products is an e-commerce graph. The primary LLM is Vicuna-7B-v1.5-16K, the default node-text encoder is SimTeG, and the projector is trained for one epoch with learning rate 2×10−52\times10^{-5} and batch size 16. Cora training examples are replicated three times to compensate for the smaller training set. ND uses two sampled hops with 10 neighbors per hop; HO uses four hop embeddings. LLaGA-ND-7B and LLaGA-HO-7B denote the two template variants. The paper compares single-focus models trained on one dataset and task, task experts trained across datasets for one task, classification experts trained for classification and link prediction across datasets, and a general model trained across all three tasks and datasets. Accuracy is reported for classification and link prediction; node descriptions are evaluated using SBERT similarity and description-label accuracy.

  7. Knowl 7 — The general LLaGA model outperforms reported language-model baselines across its evaluated tasks

    empirical result

    In the general-model setting, with dataset order Arxiv, Products, Pubmed, Cora, node-classification accuracy (%) is 74.29, 82.21, 92.42, 87.82 for LLaGA-ND-7B and 75.01, 82.07, 94.45, 87.82 for LLaGA-HO-7B. Link-prediction accuracy (%) in the same order is 90.53, 96.82, 86.31, 81.91 for ND and 92.04, 86.80, 89.81, 84.41 for HO. GPT-3.5-Turbo’s reported general-model classification accuracies are 55.00, 75.25, 88.00, 71.75, and its link-prediction accuracies are 63.80, 60.30, 68.70, 65.74. In the paper’s comparison with GraphGPT, LLaGA’s general-model Arxiv classification accuracies are 74.29 (ND) and 75.01 (HO), versus GraphGPT Mix 64.76 and Std 63.90; Pubmed classification is 92.42 and 94.45 versus GraphGPT Mix 74.16; Pubmed link prediction is 86.31 and 89.81 versus GraphGPT Mix 58.86 and Std 80.26. LLaGA’s general model was trained on 12 dataset-task combinations, while the compared GraphGPT model covered three.

  8. Knowl 8 — LLaGA transfers link prediction and node classification to unseen datasets without tuning

    empirical result

    For zero-shot link prediction, models trained on Arxiv and Pubmed were tested on Cora; accuracy was 58.97 for GCN, 67.68 for GraphSAGE, 50.74 for GraphGPT-7B, 86.47 for LLaGA-ND-7B, and 87.35 for LLaGA-HO-7B. When trained on Arxiv, Pubmed, and Cora and tested on the out-of-domain Products graph, the corresponding accuracies were 56.73, 58.92, 50.74, 92.65, and 92.99. For zero-shot node classification, models trained on node-description tasks in the training datasets were tested on unseen class sets. On Cora’s 7 classes, using only node-sequence embeddings gave 8.30% for GraphGPT-7B and 34.69% for LLaGA-7B; adding the central node’s text attributes gave 44.65% and 59.59%. On Products’ 47 classes, the corresponding accuracies were 1.40% and 13.89% with embeddings alone, and 18.84% and 43.79% with text attributes. Thus, in these tested transfers, LLaGA exceeded the GraphGPT baseline, and adding node text improved zero-shot classification accuracy for both models.

  9. Knowl 9 — Generated descriptions provide semantic and label-level evidence of interpretability

    empirical result

    The node-description evaluation reports SBERT similarity to ground-truth descriptions and the accuracy of labels inferred from generated descriptions. In dataset order Arxiv, Products, Pubmed, Cora, the base SBERT similarities between randomly selected ground-truth descriptions are 0.2231, 0.1513, 0.4869, and 0.3221. LLaGA-ND-7B’s SBERT scores are 0.6023, 0.4952, 0.6847, and 0.6465, with description-label accuracies of 74.64%, 83.18%, 92.27%, and 86.72%. LLaGA-HO-7B’s SBERT scores are 0.6228, 0.5193, 0.6934, and 0.6545, with label accuracies of 75.49%, 84.60%, 94.27%, and 86.90%. The generated explanations can make plausible interpretations even when a predicted label differs from the ground-truth label.

  10. Knowl 10 — Ablations show that neighborhood structure, sequence order, and Laplacian encodings contribute to performance

    empirical result

    The template ablation is evaluated in the classification-expert setting. Values below are accuracies in dataset order Arxiv, Products, Pubmed, Cora. For node classification, using only the center-node embedding (None) gives 73.92, 80.45, 94.60, 84.50; ND with shuffled neighbor order gives 74.35, 82.87, 94.93, 86.16; ND without Laplacian encoding gives 75.53, 82.77, 94.70, 86.35; full ND gives 75.85, 83.58, 95.06, 87.64; and HO gives 75.99, 83.32, 94.80, 89.30. For link prediction, the corresponding values are 89.98, 91.73, 78.19, 83.97 (None); 90.16, 96.31, 81.32, 80.00 (ND without order); 90.59, 96.26, 84.48, 85.88 (ND without Laplacian encoding); 90.81, 96.56, 92.36, 87.35 (full ND); and 94.30, 96.05, 88.64, 88.53 (HO). Relative to using only the central node embedding, the templates generally improve results, particularly for link prediction; the ablations also show that preserving sequence order and using Laplacian encodings can matter.

Coverage note — Supplementary results on non-text-attributed graphs, alternate text encoders and LLM backbones, the Patton comparison, run-to-run variance, and the full baseline results for every supervised training regime are omitted because they are supporting robustness or comparison analyses rather than the core method and generalization findings.

References

  1. 1.Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.
  2. 2.Chai, Z., Zhang, T., Wu, L., Han, K., Hu, X., Huang, X., and Yang, Y. Graphllm: Boosting graph reasoning ability of large language model. arXiv preprint arXiv:2310.05845, 2023.
  3. 3.Chen, J., Ma, T., and Xiao, C. Fastgcn: fast learning with graph convolutional networks via importance sampling. arXiv preprint arXiv:1801.10247, 2018.
  4. 4.Chen, J., Gao, K., Li, G., and He, K. Nagphormer: A tokenized graph transformer for node classification in large graphs. In The Eleventh International Conference on Learning Representations, 2022.
  5. 5.Chen, Z., Mao, H., Li, H., Jin, W., Wen, H., Wei, X., Wang, S., Yin, D., Fan, W., Liu, H., et al. Exploring the potential of large language models (llms) in learning on graphs. arXiv preprint arXiv:2307.03393, 2023a.
  6. 6.Chen, Z., Mao, H., Wen, H., Han, H., Jin, W., Zhang, H., Liu, H., and Tang, J. Label-free node classification on graphs with large language models (llms). arXiv preprint arXiv:2310.04668, 2023b.
  7. 7.Chiang, W.-L., Liu, X., Si, S., Li, Y., Bengio, S., and Hsieh, C.-J. Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 257–266, 2019.
  8. 8.Chiang, W.-L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J. E., Stoica, I., and Xing, E. P. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023. URL https://lmsys.org/blog/2023-03-30-vicuna/.
  9. 9.Defferrard, M., Bresson, X., and Vandergheynst, P. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016.
  10. 10.Duan, K., Liu, Q., Chua, T.-S., Yan, S., Ooi, W. T., Xie, Q., and He, J. Simteg: A frustratingly simple approach improves textual graph learning. arXiv preprint arXiv:2308.02565, 2023.
  11. 11.Dwivedi, V. P. and Bresson, X. A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699, 2020.
  12. 12.Dwivedi, V. P., Liu, Y., Luu, A. T., Bresson, X., Shah, N., and Zhao, T. Graph transformers for large graphs. arXiv preprint arXiv:2312.11109, 2023.
  13. 13.Fatemi, B., Halcrow, J., and Perozzi, B. Talk like a graph: Encoding graphs for large language models. arXiv preprint arXiv:2310.04560, 2023.
  14. 14.Gao, H., Wang, Z., and Ji, S. Large-scale learnable graph convolutional networks. Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018.
  15. 15.Guo, J., Du, L., and Liu, H. Gpt4graph: Can large language models understand graph structured data? an empirical evaluation and benchmarking. arXiv preprint arXiv:2305.15066, 2023.
  16. 16.Hamilton, W., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017.
  17. 17.Hassani, K. and Khasahmadi, A. H. Contrastive multi-view representation learning on graphs. In International conference on machine learning, pp. 4116–4126. PMLR, 2020.
  18. 18.He, X., Bresson, X., Laurent, T., Perold, A., LeCun, Y., and Hooi, B. Harnessing explanations: Llm-to-lm interpreter for enhanced text-attributed graph representation learning. arXiv preprint arXiv:2305.19523, 2023.
  19. 19.Hou, Z., Liu, X., Cen, Y., Dong, Y., Yang, H., Wang, C., and Tang, J. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 594–604, 2022.
  20. 20.Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., and Leskovec, J. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems, 33:22118–22133, 2020.
  21. 21.Huang, J., Zhang, X., Mei, Q., and Ma, J. Can llms effectively leverage graph structural information: when and why. arXiv preprint arXiv:2309.16595, 2023.
  22. 22.Jin, B., Zhang, W., Zhang, Y., Meng, Y., Zhang, X., Zhu, Q., and Han, J. Patton: Language model pretraining on text-rich networks. arXiv preprint arXiv:2305.12268, 2023.
  23. 23.Ju, M., Zhao, T., Wen, Q., Yu, W., Shah, N., Ye, Y., and Zhang, C. Multi-task self-supervised graph neural networks enable stronger task generalization. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=1tHAZRqftM.
  24. 24.Kipf, T. and Welling, M. Semi-supervised classification with graph convolutional networks. ArXiv, abs/1609.02907, 2017.
  25. 25.Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2016.
  26. 26.Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–1216, Stanford, CA, 2000. Morgan Kaufmann.
  27. 27.Liu, H., Feng, J., Kong, L., Liang, N., Tao, D., Chen, Y., and Zhang, M. One for all: Towards training one graph model for all classification tasks. arXiv preprint arXiv:2310.00149, 2023a.
  28. 28.Liu, H., Li, C., Wu, Q., and Lee, Y. J. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023b.
  29. 29.Liu, J., Yang, C., Lu, Z., Chen, J., Li, Y., Zhang, M., Bai, T., Fang, Y., Sun, L., Yu, P. S., et al. Towards graph foundation models: A survey and beyond. arXiv preprint arXiv:2310.11829, 2023c.
  30. 30.Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.
  31. 31.Mao, H., Chen, Z., Tang, W., Zhao, J., Ma, Y., Zhao, T., Shah, N., Galkin, M., and Tang, J. Graph foundation models, 2024.
  32. 32.Perez, E., Ringer, S., Lukosiˇ ut¯ e, K., Nguyen, K., et al. ˙ Discovering language model behaviors with model-written evaluations, 2022. URL https://arxiv.org/abs/2212.09251.
  33. 33.Reimers, N. and Gurevych, I. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 3982–3992, 2019.
  34. 34.Rozemberczki, B., Allen, C., and Sarkar, R. Multi-scale attributed node embedding, 2019.
  35. 35.Sun, C., Gu, H., and Hu, J. Scalable and adaptive graph neural networks with self-label-enhanced training. arXiv preprint arXiv:2104.09376, 2021.
  36. 36.Tang, J., Yang, Y., Wei, W., Shi, L., Su, L., Cheng, S., Yin, D., and Huang, C. Graphgpt: Graph instruction tuning for large language models. arXiv preprint arXiv:2310.13023, 2023.
  37. 37.Thekumparampil, K. K., Wang, C., Oh, S., and Li, L.-J. Attention-based graph neural network for semi-supervised learning. arXiv preprint arXiv:1803.03735, 2018.
  38. 38.Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Roziere, B., Goyal, N., Hambro, E., `Azhar, F., et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.
  39. 39.Velickovi ˇ c, P., Cucurull, G., Casanova, A., Romero, A., ´ Lio, P., and Bengio, Y. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017.
  40. 40.Velickovi ˇ c, P., Fedus, W., Hamilton, W. L., Li ´ o, P., Bengio, ` Y., and Hjelm, R. D. Deep graph infomax. 2019.
  41. 41.Velickovi ˇ c, P., Cucurull, G., Casanova, A., Romero, A., ´ Lio, P., and Bengio, Y. Graph attention networks. In ` International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=rJXMpikCZ.
  42. 42.Wang, W., Chen, Z., Chen, X., Wu, J., Zhu, X., Zeng, G., Luo, P., Lu, T., Zhou, J., Qiao, Y., et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023.
  43. 43.Wu, F., Souza, A., Zhang, T., Fifty, C., Yu, T., and Weinberger, K. Simplifying graph convolutional networks. In International conference on machine learning, pp. 6861–6871. PMLR, 2019.
  44. 44.Wu, Q., Zhao, W., Li, Z., Wipf, D. P., and Yan, J. Node-former: A scalable graph structure learning transformer for node classification. Advances in Neural Information Processing Systems, 35:27387–27401, 2022.
  45. 45.Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In International Conference on Learning Representations, 2018.
  46. 46.Yang, Z., Cohen, W., and Salakhudinov, R. Revisiting semi-supervised learning with graph embeddings. In International conference on machine learning, pp. 40–48. PMLR, 2016.
  47. 47.Ye, R., Zhang, C., Wang, R., Xu, S., and Zhang, Y. Natural language is all a graph needs. arXiv preprint arXiv:2308.07134, 2023.
  48. 48.Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., and Liu, T.-Y. Do transformers really perform badly for graph representation? Advances in Neural Information Processing Systems, 34:28877–28888, 2021.
  49. 49.You, Y., Chen, T., Wang, Z., and Shen, Y. L2-gcn: Layer-wise and learned efficient training of graph convolutional networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2124–2132, 2020.
  50. 50.Yun, S., Jeong, M., Kim, R., Kang, J., and Kim, H. J. Graph transformer networks. Advances in neural information processing systems, 32, 2019.
  51. 51.Zhang, H., Wu, Q., Yan, J., Wipf, D., and Yu, P. S. From canonical correlation analysis to self-supervised graph neural networks. Advances in Neural Information Processing Systems, 34:76–89, 2021.
  52. 52.Zhao, T., Shah, N., and Ghazizadeh, E. Learning from graphs beyond message passing. In The Second Tiny Papers Track at ICLR 2024.

Citation

MLA
Chen, R., et al. “LLaGA: Large Language and Graph Assistant”. arXiv, 2024, http://arxiv.org/abs/2402.08170v3.
APA
Chen, R., Zhao, T., Jaiswal, A., Shah, N., & Wang, Z. (2024). LLaGA: Large Language and Graph Assistant. arXiv. http://arxiv.org/abs/2402.08170v3
Chicago
Chen, R., T. Zhao, A. Jaiswal, N. Shah, and Z. Wang. 2024. “LLaGA: Large Language and Graph Assistant”. arXiv. http://arxiv.org/abs/2402.08170v3.
Harvard
Chen, R. et al. (2024) “LLaGA: Large Language and Graph Assistant”, arXiv [Preprint]. Available at: http://arxiv.org/abs/2402.08170v3.
Vancouver
1. Chen R, Zhao T, Jaiswal A, Shah N, Wang Z (2024) LLaGA: Large Language and Graph Assistant. arXiv

BibTeX

@article{chen2024llaga,
  title = {LLaGA: Large Language and Graph Assistant},
  author = {Chen, Runjin and Zhao, Tong and Jaiswal, Ajay and Shah, Neil and Wang, Zhangyang},
  year = {2024},
  journal = {arXiv},
  url = {http://arxiv.org/abs/2402.08170v3},
  eprint = {2402.08170}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: https://creativecommons.org/licenses/by/4.0/