Deep Learning on Graphs (Yao Ma, Jiliang Tang) (Z-Library)

# Deep Learning on Graphs — Reading Guide ## 【One-Line Pitch】 A comprehensive, self-contained textbook that takes you from graph and deep learning fundamentals through graph neural network (GNN) methods, applications, and cutting-edge research — ideal for students, practitioners, and researchers entering the field of graph representation learning. ## 【Book Arc】 - **Opening (~0%–8%)**: Introduces the book's four-part structure (Foundations, Methods, Applications, Advances), explains why deep learning on graphs matters, and positions the work within the three generations of graph representation learning — from traditional embedding to modern embedding to GNNs. - **Early (~17%–33%)**: Builds the technical foundation — graph representations, properties, spectral theory, and complex graph types (heterogeneous, bipartite, signed, hypergraphs, dynamic) — followed by a crash course in deep learning essentials: feedforward networks, CNNs, RNNs, autoencoders, and training techniques. - **Middle (~33%–50%)**: Covers the core methods: graph embedding techniques for simple and complex graphs, the general GNN framework (graph filters, pooling, parameter learning), robustness against adversarial attacks, scalability via sampling methods, GNNs for complex graphs, and deep models beyond GNNs (autoencoders, VAEs, GANs on graphs). - **Late (~50%–75%)**: Shifts to applications across five domains — natural language processing (semantic role labeling, machine translation, QA, knowledge graphs), computer vision (visual QA, action recognition, image classification, point clouds), data mining (social networks, recommender systems, traffic prediction, fake news detection), and biochemistry/healthcare (drug discovery, protein prediction, disease prediction). - **Ending (~75%–92%)**: Discusses advanced topics — deeper GNN architectures (Jumping Knowledge, DropEdge, PairNorm), self-supervised learning on graphs, expressiveness via the Weisfeiler–Lehman test — plus emerging applications in combinatorial optimization, program representation, and physics, with further reading for each chapter. ## 【Key Takeaways】 - **Graph representation learning has evolved in three generations** (Early): from traditional dimension reduction (IsoMap, LLE) to word2vec-inspired modern embedding to deep learning on graphs — understanding this lineage helps contextualize why GNNs represent a paradigm shift. - **Graphs come in many flavors beyond simple structures** (Early): heterogeneous, bipartite, multidimensional, signed, hypergraph, and dynamic graphs each require specialized treatment — a recurring theme throughout the methods chapters. - **GNNs follow a general framework of filters and pooling** (Middle): spectral-based and spatial-based graph filters form the core, while flat and hierarchical pooling handle graph-level tasks — mastering this framework lets you understand most GNN variants. - **Robustness and scalability are practical bottlenecks** (Middle): adversarial attacks (white-box, gray-box, black-box) threaten GNN reliability, and sampling methods (node-wise, layer-wise, subgraph-wise) are essential for scaling to large graphs. - **GNNs shine across diverse application domains** (Late): from semantic role labeling and knowledge graph completion in NLP to skeleton-based action recognition in vision, and from recommender systems to drug-target binding prediction — the same core ideas transfer widely. - **Going deeper isn't always better** (Ending): techniques like Jumping Knowledge, DropEdge, and PairNorm address the over-smoothing problem in deep GNNs, while the Weisfeiler–Lehman test provides a theoretical lens on expressiveness. - **The field is genuinely interdisciplinary** (Ending): GNNs now extend into combinatorial optimization, program representation learning, and physics — signaling fertile ground for cross-domain contributions. ## 【Reading Tips】 - **Skim Part I if you have ML background**: Chapters 2–3 cover standard graph theory and deep learning basics; if you're comfortable with CNNs, RNNs, and spectral methods, jump straight to Part II. - **Deep-read Chapters 4–5 for the core**: Graph embedding and the general GNN framework are the heart of the book — spend time here to build mental models that make later chapters easier. - **Use Part III as a reference, not a linear read**: Application chapters (10–13) are organized by domain; pick the ones relevant to your work and treat others as examples of how GNNs generalize. - **Watch for the complex graph thread**: Heterogeneous, signed, and dynamic graphs appear in both embedding (Ch. 4) and GNN (Ch. 8) chapters — reading them together gives a complete picture. - **Check "Further Reading" sections**: Each chapter ends with curated pointers for deeper dives — valuable if you're using the book as a research starting point. ## 【Coverage Limits】 This guide is based on the book's table of contents, preface, and front matter; detailed technical content from individual chapters (e.g., specific algorithms, formulas, experimental results) is not covered here. Excerpts do not include worked examples or code implementations. ##