AuthorsChaoxi Niu, Guansong Pang, Ling Chen
TL;DR
HimNet uses hierarchical node and graph memory modules inside a graph autoencoder to detect anomalous graphs, reaching 80.6% AUC on DD vs 70.6% for PK-iF (+10.0 points).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Graph-level anomaly detection must catch graphs that are abnormal in part or in whole, i.e., locally-anomalous or globally-anomalous graphs.
Existing GLAD methods focus on discriminative embeddings and may not preserve primary graph semantics, making them ineffective when rich structural and attribute information is needed.
HOW IT WORKS
HimNet introduces a GNN Encoder, Node Memory Module, Graph Memory Module, and Graph Decoder to learn hierarchical node-to-graph normal patterns jointly.
You can think of HimNet like a two-level cache: node memory stores detailed local templates, while graph memory stores global prototypes, similar to RAM plus a library of canonical blueprints.
By reconstructing graphs only from these memory blocks, HimNet exposes graphs that cannot be well expressed by normal patterns, something a plain context-limited autoencoder cannot reliably do.
DIAGRAM
This diagram shows how HimNet processes a graph at inference time to compute reconstruction and approximation based anomaly scores.
DIAGRAM
This diagram shows how HimNet is trained on normal graphs and then evaluated against baselines on 16 datasets.
PROCESS
Graph Autoencoder
HimNet first uses the Graph Autoencoder with the GNN Encoder and Graph Decoder to learn node and graph representations that preserve structural and attribute semantics.
Node Memory Module
The Node Memory Module stores P memory blocks and approximates node embeddings so the Graph Decoder reconstructs graphs only from combinations of normal node patterns.
Graph Memory Module
The Graph Memory Module stores Q graph memory blocks and approximates graph embeddings, enforcing that each graph is expressed as a mixture of normal global prototypes.
Training and Inference
HimNet jointly minimizes reconstruction, approximation, and entropy losses during training, then uses the combined Lrec and Lapp value as the anomaly score during inference.
KEY CONTRIBUTIONS
Hierarchical node to graph memory network HimNet
HimNet is the first memory based GLAD framework that jointly uses the Node Memory Module and Graph Memory Module inside a graph autoencoder for graph-level anomaly detection.
Three dimensional node memory module
HimNet introduces a three dimensional Node Memory Module Mn ∈ R^{P×N×D} with multiple two dimensional blocks, each capturing one type of normal pattern across all nodes.
Joint learning with reconstruction and approximation
HimNet jointly minimizes graph reconstruction error and graph approximation error, plus an entropy term, to learn hierarchical normal patterns and detect both locally and globally anomalous graphs.
RESULTS
AUC
80.6%
+10.0 over PK-iF on DD
AUC
68.6%
+14.1 over OCGCN on NCI1
AUC
78.0%
-0.2 vs GLocalKD on REDDIT
AUC
71.1%
+6.7 over GLocalKD on PPAR-gamma
These AUC scores come from 16 real world GLAD datasets, including DD, NCI1, REDDIT, and PPAR-gamma, showing that HimNet consistently improves anomaly detection over both two step and end to end baselines.
BENCHMARK
These AUC scores come from 16 real world GLAD datasets, including DD, NCI1, REDDIT, and PPAR-gamma, showing that HimNet consistently improves anomaly detection over both two step and end to end baselines.
BENCHMARK
AUC scores for DD, comparing HimNet against representative two step and end to end GLAD baselines.
KEY INSIGHT
On REDDIT, HimNet jumps from 21.8% AUC for GAE to 78.0% AUC, a +56.2 point improvement despite using the same encoder backbone.
This is surprising because simply inserting the Node Memory Module and Graph Memory Module into the Graph Autoencoder radically changes performance, contradicting the intuition that reconstruction based methods are inherently weak on large graphs.
WHY IT MATTERS
HimNet shows that explicitly storing hierarchical normal patterns in memory blocks makes reconstruction based GLAD practical even on large, complex graph datasets.
Builders can now design graph anomaly detectors that distinguish local and global irregularities using memory guided reconstruction, instead of relying solely on contrastive or distillation based embeddings.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…