PMGraph: Multimodal Supply-Chain Benchmark

• PMGraph is a public benchmark of supply-chain graphs linking manufacturers to products with multimodal, attributed, and heterogeneous data.
• It integrates textual, image, categorical, and numeric features to support robust manufacturer–product link prediction under real-world data imperfections.
• The benchmark enables noise-robust inference using staged multimodal fusion via the C-MAG architecture, emphasizing capability inference under incomplete evidence.

PMGraph is a public-use suite of supply-chain knowledge graphs centered on manufacturer–product relationships. Introduced alongside the Cascade Multimodal Attributed Graph (C-MAG), it is a benchmark family rather than a single flat graph, comprising bipartite manufacturer–product graphs, heterogeneous multimodal knowledge graphs, and related variants. The benchmark is designed to reflect real-world manufacturer data in which capabilities, certifications, locations, textual descriptions, and images are sparse, heterogeneous, noisy, and incompletely aligned across modalities, while the primary supervised task is manufacturer–product link prediction (Li et al., 11 Aug 2025).

1. Dataset identity and benchmark scope

PMGraph is defined as a public benchmark of supply-chain graphs linking manufacturers to products, with multimodal and attributed extensions. The paper makes a strict distinction between PMGraph and C-MAG: PMGraph is the benchmark or dataset suite, whereas C-MAG is the learning architecture proposed to operate on it. In its experimental instantiation, PMGraph contains 8,888 manufacturer nodes, 72,789 product nodes, 2,918 attribute nodes, and 29,178 image nodes, together with manufacturer–product, manufacturer–attribute, and manufacturer–image relations (Li et al., 11 Aug 2025).

The benchmark is simultaneously bipartite, heterogeneous, multimodal, and attributed. It is bipartite because its core downstream task is defined over manufacturer and product partitions. It is heterogeneous because it includes multiple node and edge types. It is multimodal because the node descriptions combine text, images, categorical metadata, and numeric metadata. It is attributed because manufacturer and product entities carry business descriptions, industry tags, process and material capabilities, certifications, NAICS codes, employee counts, geographic coordinates, and product-catalog text.

A central motivation of PMGraph is that supply-chain matching is not reducible to graph topology alone. The benchmark is intended to capture the practical setting in which one wants to reason about whether a manufacturer can produce a product, including potentially missing or future manufacturer–product relationships, especially for small and medium-sized enterprises. The benchmark therefore emphasizes capability inference under incomplete evidence rather than only reconstruction of already observed links.

2. Graph schema and data composition

A faithful formalization consistent with the paper is a graph $G=(V,E)$ with node partition

$V = V_M \cup V_P \cup V_A \cup V_I$

and edge partition

$E = E_{MP} \cup E_{MA} \cup E_{MI},$

where $V_M$ denotes manufacturer nodes, $V_P$ product nodes, $V_A$ attribute nodes, and $V_I$ image nodes; and where $E_{MP}$, $E_{MA}$, and $E_{MI}$ denote manufacturer$V = V_M \cup V_P \cup V_A \cup V_I$0product, manufacturer$V = V_M \cup V_P \cup V_A \cup V_I$1attribute, and manufacturer$V = V_M \cup V_P \cup V_A \cup V_I$2image edges, respectively. For downstream heteroGNN training, the paper states that reverse product$V = V_M \cup V_P \cup V_A \cup V_I$3manufacturer edges are added for bidirectional propagation (Li et al., 11 Aug 2025).

The benchmark statistics reported in the paper are as follows.

Category	Type	Count
Nodes	Manufacturer	8,888
Nodes	Product	72,789
Nodes	Attribute	2,918
Nodes	Image	29,178
Edges	manufacturer $V = V_M \cup V_P \cup V_A \cup V_I$4 product	112,597
Edges	manufacturer $V = V_M \cup V_P \cup V_A \cup V_I$5 attribute	83,105
Edges	manufacturer $V = V_M \cup V_P \cup V_A \cup V_I$6 image	29,178

Manufacturer features include textual descriptions, categorical metadata, numeric metadata, and, in the staged model setting, an aggregated attribute/image embedding. Product features include textual attributes and categorical data. Attribute nodes correspond explicitly to processes, certifications, materials, cities, and states. Image nodes are derived from product images scraped from manufacturer websites.

The paper also states that all six heterograph variants used in the experiments share the same manufacturer–product topology and compress node features to 64 dimensions via truncated SVD. This indicates that PMGraph is not only a raw entity-relation collection but also a benchmark family with controlled feature and topology variants for ablation and comparative modeling.

3. Data sources and benchmark construction

PMGraph is assembled by merging two source resources: SUDOKN SC-KG and the manufacturer directory with associated ontology of Li et al. The resulting corpus is anchored on 8,888 unique manufacturer URLs, each associated with multiple products and rich metadata such as employee counts, descriptions, industry tags, capabilities, certifications, locations, NAICS codes, and product catalogs (Li et al., 11 Aug 2025).

The image-construction pipeline is comparatively detailed. The benchmark begins from a curated set of over 8,000 manufacturer websites, normalizes target URLs, retrieves webpage content, parses HTML, identifies intra-domain hyperlinks, and keeps links matching product-related lexical cues such as “product,” “item,” “catalog,” “gallery,” and “prod,” while discarding links associated with non-product sections such as “about,” “contact,” “blog,” “news,” “login,” and “signup.” If too few product pages are detected, additional internal pages are explored. Image elements are then extracted, their URLs resolved, binaries downloaded, and files stored by domain.

After scraping, the images are filtered by a multimodal LLM-based quality filter specified in the appendix as Gemini-2.0-Flash. The prompt returns “Yes” only if an image is certainly one of the listed products, and “No” otherwise. This filtering step removes off-topic images, uncertain images, and low-resolution content. The paper reports that 145,888 images remain after filtering, but the benchmark uses a 20% random sample for the main experiments, yielding the 29,178 image nodes reported in the released PMGraph statistics.

The attribute construction is narrower and more explicitly enumerated than the broader manufacturer metadata. The Stage-1 graph used by C-MAG includes attribute nodes for processes, certifications, materials, cities, and states. By contrast, the paper does not provide a detailed ontology schema, taxonomy hierarchy, or explicit normalization rules for certification vocabularies, capability ontologies, or product taxonomy reconciliation. This suggests that PMGraph is concretely specified at the node/edge and feature level, but not exhaustively documented as a formal ontology resource.

4. Benchmark task and evaluation protocol

The main benchmarked task on PMGraph is manufacturer–product link prediction, defined over held-out manufacturer–product edges. Positive examples are observed manufacturer–product edges, and negative examples are sampled non-edges at a 1:1 negative sampling ratio. The train/validation/test split is 80/10/10% over edges (Li et al., 11 Aug 2025).

The evaluation uses ROC-AUC and PR-AUC. Precision and recall are defined in the paper as

$V = V_M \cup V_P \cup V_A \cup V_I$7

The paper also defines PR-AUC as the area under the precision–recall curve. The supervised models optimize Weighted BCE for link prediction, while Stage 1 pretraining uses binary cross-entropy over manufacturer–attribute and manufacturer–image edges.

An important protocol detail is that PMGraph supports both flat-graph and heterogeneous-graph ablations. In the flat-graph setting, the full heterograph may still be used during training, but validation and test evaluation are restricted to manufacturer–product links, with attribute and image edges removed from validation and test graphs. This keeps evaluation aligned with the intended downstream task rather than conflating auxiliary modality relations with the target relation.

The benchmark is therefore not simply a static graph dump. It is a graph-learning evaluation environment with fixed node and edge types, specified splits, explicit negative sampling, modality-aware ablations, and threshold-independent link-prediction metrics.

5. C-MAG as the principal modeling architecture on PMGraph

C-MAG is the model architecture introduced to operate on PMGraph, and its structure is tightly matched to the benchmark’s layered heterogeneity. The model is explicitly two-stage. In Stage 1, it builds a manufacturer-centered multimodal attributed graph over manufacturers, attributes, and images. Manufacturer textual descriptions, attribute labels, and product images are embedded into a shared 768-D Jina-CLIP-v1 space, then compressed to 32-D by truncated SVD. Manufacturer embeddings are pretrained through unsupervised link prediction on manufacturer–attribute and manufacturer–image edges using a two-layer GraphSAGE encoder with dimensions $V = V_M \cup V_P \cup V_A \cup V_I$8; the paper also validates Stage 1 with a two-layer R-GCN (Li et al., 11 Aug 2025).

In Stage 2, the downstream manufacturer–product graph is initialized with both the Stage-1 manufacturer embedding and additional text, categorical, and numeric features. Manufacturer and product textual attributes are encoded with Jina-CLIP, compressed, combined with one-hot encoded categorical metadata and standardized numeric metadata, and further compressed via SVD. For manufacturer nodes, the paper states that the 32-D Stage-1 embedding is concatenated with a 64-D Stage-2 metadata embedding, yielding 96-D, then compressed to 64-D. Product nodes retain 64-D embeddings derived from textual and categorical data.

The downstream propagation stage uses two-layer HeteroSAGE or HeteroGAT. The reported architecture details are two layers, projection layers with 128 hidden units, dropout 0.5, and, for HeteroGAT, 4 attention heads. The link score is described through the evaluation section as a dot-product score between node embeddings. Conceptually, C-MAG treats PMGraph’s manufacturer-side modalities as an intermediate aggregation problem before performing manufacturer–product reasoning, rather than collapsing all modalities into a single flat graph at the outset.

This staged design is the paper’s main modeling claim about PMGraph. It implies that the benchmark was constructed not only to test whether multimodal information helps, but also to test whether modality-aware staged fusion outperforms naïve heterogeneous fusion.

6. Empirical behavior, limitations, and naming ambiguities

The benchmark results show that the cascade architectures are the strongest variants overall. Under HeteroSAGE, $V = V_M \cup V_P \cup V_A \cup V_I$9 reaches 70.58 ROC-AUC / 66.09 PR-AUC. Under HeteroGAT, $E = E_{MP} \cup E_{MA} \cup E_{MI},$0 achieves the best ROC-AUC at 75.46, while $E = E_{MP} \cup E_{MA} \cup E_{MI},$1 achieves the best PR-AUC at 74.30. The paper interprets this as evidence that staged multimodal fusion outperforms flat fusion on PMGraph, and that the cascade variants occupy the top two positions across both heteroGNN backbones. The image-sampling ablation further reports that performance peaks at a 20% image sampling ratio, with larger image inclusion giving only marginal gains and sometimes suffering from residual image noise (Li et al., 11 Aug 2025).

The same experiments also define the benchmark’s main limitations. PMGraph is intentionally realistic, and realism here includes substantial data imperfections: sparse and heterogeneous public manufacturer data, multimodal misalignment, incomplete modality coverage, and residual image noise. The paper explicitly notes that adding images does not consistently improve link prediction, that multimodal integration can sometimes degrade performance, and that evaluation was limited to a relatively small set of heteroGNN architectures. PMGraph is therefore not a benchmark in which more modalities are automatically beneficial; rather, it is a benchmark for noise-robust multimodal supply-chain inference.

A recurrent source of confusion is the name itself. In the cited supply-chain paper, PMGraph denotes the benchmark, while C-MAG denotes the model. This should not be conflated with PGM, the “Principled Graph Management” algorithm for Graph Generation restricted master problems, whose paper explicitly does not use the term PMGraph (Yarkony et al., 2022). It should also not be conflated with MPG, the acronym used by “Meta-Property Graphs,” whose paper consistently uses Meta-Property Graphs and MetaGPML rather than PMGraph (Sadoughi et al., 2024). Within current arXiv usage, PMGraph most directly denotes the multimodal supply-chain benchmark introduced with C-MAG, publicly released at https://huggingface.co/datasets/shawntzx/PMGraph (Li et al., 11 Aug 2025).