@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.060_iaggad.yaml

@@ -0,0 +1,121 @@
+id: aisb.t3.060_iaggad
+name: 'IA-GGAD: Zero-shot Generalist Graph Anomaly Detection via Invariant and Affinity
+  Learning'
+version: 0.1.0
+one_line: Zero-shot generalist graph anomaly detection using invariant feature learning
+  and structure-insensitive affinity modules to handle cross-domain graph shifts.
+task_description: 'This packaged benchmark implements IA-GGAD, a zero-shot generalist
+  graph anomaly detection framework that detects anomalous nodes across diverse graph
+  domains without task-specific retraining or fine-tuning. The method addresses two
+  core domain-shift challenges: Feature Space Shift (FSS) through an anomaly-driven
+  graph invariant learning module that learns domain-invariant node representations,
+  and Graph Structure Shift (GSS) via a structure-insensitive affinity learning module
+  that captures cross-domain structural correspondences. Training uses source graphs
+  from social, citation, and e-commerce domains; evaluation measures anomaly detection
+  performance on held-out target graphs. The benchmark entry point is run_IA_GGAD.py
+  which executes 60-epoch training on source datasets and evaluation on target datasets,
+  reporting AUROC as the primary metric.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- auroc
+- anomaly_scores
+- graph_eval_report
+launch_profiles:
+- id: full_benchmark
+  label: Full Benchmark
+  description: 'Execute the full IA-GGAD benchmark pipeline: train on source datasets
+    (pubmed, Flickr, questions, YelpChi) and evaluate on target datasets (cora, citeseer,
+    ACM, BlogCatalog, Facebook, weibo, Reddit, Amazon). Default configuration uses
+    GCN backbone with 60 training epochs.
+
+    '
+dataset_download:
+  primary_method: direct_archive
+  sources:
+  - url: https://deepscientist.cc/AISB/060_iaggad
+    type: primary
+  notes:
+  - Archive contains complete source code, configuration files, and parameter sets.
+  - Graph datasets are referenced by config JSONs in params/ directory.
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: 3.8.14
+  cuda: '11.7'
+  pytorch: 1.13.1
+  flash_attn: null
+  key_packages:
+  - dgl==0.9.1
+  - torch==1.13.1
+  - numpy
+  - scikit-learn
+  notes:
+  - CUDA 11.7 build of DGL required (dgl-0.9.1+cu117).
+  - See bundled README/requirements.txt for full dependency set.
+  - GPU required for training and evaluation due to DGL graph operations.
+risk_flags:
+- requires_gpu
+- long_runtime
+risk_notes:
+- Single GPU training with 60 epochs may take several hours on minimum specs.
+- CPU-only execution not supported; DGL operations require CUDA.
+- No static code-audit execution performed; runtime validation required.
+recommended_when: 'Use this benchmark when you need to evaluate or optimize graph
+  anomaly detection methods that must generalize across domains without target-domain
+  fine-tuning. Ideal for evaluating zero-shot transfer capabilities of GAD models
+  on heterogeneous graph structures.
+
+  '
+not_recommended_when: 'Do not use if you require a CPU-only baseline, non-graph anomaly
+  detection task, or single-domain only evaluation. Not suitable when no GPU hardware
+  is available.
+
+  '
+paper:
+  title: 'IA-GGAD: Zero-shot Generalist Graph Anomaly Detection via Invariant and
+    Affinity Learning'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=Cggdvyt8ik
+  github: https://github.com/kg-cc/IA-GGAD/
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.060_iaggad.zip
+  archive_type: zip
+  local_dir_name: paper-60-IAGGAD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.060_iaggad.zip
+  sha256: 4a3697f8b28d81d5b8bb432759c92b99abc3ca5e02233241c1efcddf7a8137c4
+  size_bytes: 191680
+display:
+  palette_seed: emerald-slate-graph
+  art_style: anomaly-atlas
+  accent_priority: high
+image_path: ../image/060_aisb.t3.060_iaggad.jpg

package/AISB/catalog/aisb.t3.060_iaggad.zh.yaml

@@ -0,0 +1,100 @@
+id: aisb.t3.060_iaggad
+name: 'IA-GGAD: Zero-shot Generalist Graph Anomaly Detection via Invariant and Affinity
+  Learning'
+version: 0.1.0
+one_line: |-
+  使用不变特征学习和结构不敏感亲和模块进行零样本通用图异常检测，以处理跨域图偏移。
+task_description: |-
+  此打包基准实现 IA-GGAD，这是一个零样本通用图异常检测框架，能够在无需任务特定重训练或微调的情况下检测跨多样化图域的异常节点。该方法解决两个核心域偏移挑战：通过异常驱动的图不变学习模块学习域不变节点表示来应对特征空间偏移（FSS），以及通过结构不敏感亲和学习模块捕获跨域结构对应关系来应对图结构偏移（GSS）。训练使用来自社交、引用和电子商务领域的源图；评估在保留的目标图上测量异常检测性能。基准入口点为 run_IA_GGAD.py，执行源数据集上的 60 轮训练和目标数据集上的评估，报告 AUROC 作为主要指标。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- auroc
+- anomaly_scores
+- graph_eval_report
+launch_profiles:
+- id: full_benchmark
+  label: Full Benchmark
+  label_zh: 完整基准测试
+  description: |-
+    执行完整的 IA-GGAD 基准测试流程：在源数据集（pubmed、Flickr、questions、YelpChi）上训练，并在目标数据集（cora、citeseer、ACM、BlogCatalog、Facebook、weibo、Reddit、Amazon）上评估。默认配置使用 GCN 主干网络，训练 60 轮。
+dataset_download:
+  primary_method: direct_archive
+  sources:
+  - url: https://deepscientist.cc/AISB/060_iaggad
+    type: primary
+  notes:
+  - 压缩包包含完整的源代码、配置文件和参数集。
+  - 图数据集由 params/ 目录中的配置 JSON 文件引用。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: 3.8.14
+  cuda: '11.7'
+  pytorch: 1.13.1
+  flash_attn: null
+  key_packages:
+  - dgl==0.9.1
+  - torch==1.13.1
+  - numpy
+  - scikit-learn
+  notes:
+  - 需要 CUDA 11.7 版本的 DGL（dgl-0.9.1+cu117）。
+  - 有关完整依赖项集合，请参阅捆绑的 README/requirements.txt。
+  - 由于 DGL 图操作需要 GPU 进行训练和评估。
+risk_flags:
+- requires_gpu
+- long_runtime
+risk_notes:
+  - 在最低配置上使用单 GPU 训练 60 轮可能需要数小时。
+  - 不支持纯 CPU 执行；DGL 操作需要 CUDA。
+  - 未执行静态代码审计；需要运行时验证。
+recommended_when: |-
+  当需要评估或优化必须在无需目标域微调的情况下跨域泛化的图异常检测方法时使用此基准。非常适合评估异构图结构上 GAD 模型的零样本迁移能力。
+not_recommended_when: |-
+  如果需要纯 CPU 基线、非图异常检测任务或仅单域评估，请勿使用。在没有 GPU 硬件可用时也不适用。
+paper:
+  title: 'IA-GGAD: Zero-shot Generalist Graph Anomaly Detection via Invariant and
+    Affinity Learning'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=Cggdvyt8ik
+  github: https://github.com/kg-cc/IA-GGAD/
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.060_iaggad.zip
+  archive_type: zip
+  local_dir_name: paper-60-IAGGAD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.060_iaggad.zip
+  sha256: 4a3697f8b28d81d5b8bb432759c92b99abc3ca5e02233241c1efcddf7a8137c4
+  size_bytes: 191680
+display:
+  palette_seed: emerald-slate-graph
+  art_style: anomaly-atlas
+  accent_priority: high
+image_path: ../image/060_aisb.t3.060_iaggad.jpg

package/AISB/catalog/aisb.t3.061_hsgkn.yaml

@@ -0,0 +1,136 @@
+schema_version: 1
+id: aisb.t3.061_hsgkn
+name: Hierarchical Shortest-Path Graph Kernel Network
+version: 0.1.0
+one_line: Optimize a graph-kernel network that fuses shortest-path structure with
+  learnable graph representations for graph classification.
+task_description: 'This packaged benchmark covers the Hierarchical Shortest-Path Graph
+  Kernel Network (HSP-GKN), a kernel-based graph learning framework that integrates
+  graph similarity estimation with downstream tasks within a unified optimization
+  framework. The approach designs a hierarchical shortest-path graph kernel that efficiently
+  preserves both semantic and structural information by transforming graphs into hierarchical
+  features. These features are matched against learnable hidden graph features to
+  produce a similarity vector serving as the graph embedding for end-to-end MLP training.
+  Unlike traditional graph kernel methods that isolate representation learning from
+  objective optimization, HSP-GKN enables task-specific representation learning with
+  computational complexity of O(n² log n) compared to O(n⁴) for standard SP kernels.
+
+  '
+capability_tags:
+- research_code_optimization
+- graph_learning
+- graph_kernels
+- classification
+- explainable_ml
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- imdb_binary_accuracy
+- graph_kernel_checkpoint
+- optuna_results
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run the packaged graph-classification benchmark route directly from
+    main.py. Uses default or provided dataset with GPU acceleration. Default dataset
+    is MUTAG. Example: python main.py --dataset MUTAG --device cuda:0
+
+    '
+- id: tune_and_eval
+  label: Tune + Eval
+  description: 'Run the packaged HSGKN training flow together with Optuna hyperparameter
+    search. Execute optuna_main.py to perform automated hyperparameter optimization
+    across learning rate, weight decay, hidden path count, normalization, cutoff,
+    dropout, and attribute normalization parameters.
+
+    '
+dataset_download:
+  primary_method: bundled
+  sources: []
+  notes:
+  - Graph datasets are automatically downloaded via DGL's TUDataset interface
+  - Supported datasets: PROTEINS, PROTEINS_full, ENZYMES, SYNTHETIC, BZR, COX2, NCI1,
+      MUTAG, IMDB-BINARY, IMDB-MULTI, COLLAB, DD, REDDIT-BINARY, REDDID-MULTI-5K,
+      REDDIT-MULTI-12K, PTC_MR, reddit_threads, SYNTHETICnew
+  - Cached datasets stored in cache/ directory after first download
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 1
+    gpu_vram_gb: 8
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: 2.2.1
+  flash_attn: null
+  key_packages:
+  - dgl==1.1.3
+  - optuna
+  - scikit-learn
+  - numpy
+  - torch
+  notes:
+  - Install dependencies via bundled requirements.txt before execution
+  - DGL (Deep Graph Library) required for graph data handling and TUDataset access
+  - Optuna used for hyperparameter optimization in tune_and_eval profile
+risk_flags: []
+risk_notes: []
+recommended_when: 'Use this benchmark when you need a graph-classification task that
+  combines kernel-style structural priors with trainable neural components. Particularly
+  suitable for researchers exploring the integration of explicit graph kernels with
+  end-to-end learning frameworks, or those benchmarking against traditional kernel
+  methods on TU graph datasets.
+
+  '
+not_recommended_when: 'Do not use this if you need a giant multimodal or LLM benchmark,
+  or if your dataset consists primarily of large-scale graphs where the O(n² log n)
+  kernel computation may become a bottleneck.
+
+  '
+paper:
+  title: Hierarchical Shortest-Path Graph Kernel Network
+  venue: NeurIPS 2025 Spotlight
+  year: 2025
+  url: https://openreview.net/forum?id=721bDIvjen
+  code_url: https://github.com/JXWANG-GRAPH/HSP-GKN
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.061_hsgkn.zip
+  archive_type: zip
+  local_dir_name: paper-61-HSGKN
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.061_hsgkn.zip
+  sha256: ce45c434498873e02198eeda9b546b4ac85920213e0c65e013302a6bac0a5f90
+  size_bytes: 109708
+display:
+  palette_seed: stone-cyan-kernel
+  art_style: graph-kernel
+  accent_priority: medium
+commercial:
+  annual_fee: null
+image_path: ../image/061_aisb.t3.061_hsgkn.jpg

package/AISB/catalog/aisb.t3.061_hsgkn.zh.yaml

@@ -0,0 +1,113 @@
+schema_version: 1
+id: aisb.t3.061_hsgkn
+name: 分层最短路径图核网络
+version: 0.1.0
+one_line: 优化一个融合最短路径结构与可学习图表示的图核网络，用于图分类。
+task_description: '本基准测试涵盖分层最短路径图核网络（HSP-GKN），这是一个基于核的图学习框架，将图相似度估计与下游任务整合在统一的优化框架中。该方法设计了一种分层最短路径图核，通过将图转换为分层特征来高效保留语义和结构信息。这些特征与可学习的隐藏图特征进行匹配，生成相似度向量作为图的嵌入表示，用于端到端的MLP训练。与传统图核方法将表示学习与目标优化分离不同，HSP-GKN实现了任务特定的表示学习，计算复杂度为O(n² log n)，而标准SP核为O(n⁴)。
+
+  '
+capability_tags:
+- research_code_optimization
+- graph_learning
+- graph_kernels
+- classification
+- explainable_ml
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- imdb_binary_accuracy
+- graph_kernel_checkpoint
+- optuna_results
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: '直接从main.py运行打包的图分类基准测试路由。使用默认或提供的数据集，配合GPU加速。默认数据集为MUTAG。示例：python main.py --dataset MUTAG --device cuda:0
+
+    '
+- id: tune_and_eval
+  label: 调优+评估
+  description: '运行打包的HSGKN训练流程，配合Optuna超参数搜索。执行optuna_main.py进行自动化超参数优化，涵盖学习率、权重衰减、隐藏路径数量、正则化、截断值、dropout和属性归一化等参数。
+
+    '
+dataset_download:
+  primary_method: bundled
+  sources: []
+  notes:
+  - 图数据集通过DGL的TUDataset接口自动下载
+  - 支持的数据集：PROTEINS、PROTEINS_full、ENZYMES、SYNTHETIC、BZR、COX2、NCI1、MUTAG、IMDB-BINARY、IMDB-MULTI、COLLAB、DD、REDDIT-BINARY、REDDIT-MULTI-5K、REDDIT-MULTI-12K、PTC_MR、reddit_threads、SYNTHETICnew
+  - 首次下载后，数据集缓存至cache/目录
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 1
+    gpu_vram_gb: 8
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: 2.2.1
+  flash_attn: null
+  key_packages:
+  - dgl==1.1.3
+  - optuna
+  - scikit-learn
+  - numpy
+  - torch
+  notes:
+  - 执行前通过捆绑的requirements.txt安装依赖
+  - DGL（深度图库）用于图数据处理和TUDataset访问
+  - Optuna用于tune_and_eval配置文件的超参数优化
+risk_flags: []
+risk_notes: []
+recommended_when: '当需要结合核风格结构先验与可训练神经组件的图分类任务时，使用此基准测试。特别适合探索显式图核与端到端学习框架整合的研究人员，或需要针对传统核方法对TU图数据集进行基准测试的场景。
+
+  '
+not_recommended_when: '如果需要超大规模多模态或LLM基准测试，或者数据集主要由大规模图构成（其中O(n² log n)核计算可能成为瓶颈），请勿使用此基准测试。
+
+  '
+paper:
+  title: Hierarchical Shortest-Path Graph Kernel Network
+  venue: NeurIPS 2025 Spotlight
+  year: 2025
+  url: https://openreview.net/forum?id=721bDIvjen
+  code_url: https://github.com/JXWANG-GRAPH/HSP-GKN
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.061_hsgkn.zip
+  archive_type: zip
+  local_dir_name: paper-61-HSGKN
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.061_hsgkn.zip
+  sha256: ce45c434498873e02198eeda9b546b4ac85920213e0c65e013302a6bac0a5f90
+  size_bytes: 109708
+display:
+  palette_seed: stone-cyan-kernel
+  art_style: graph-kernel
+  accent_priority: medium
+commercial:
+  annual_fee: null
+image_path: ../image/061_aisb.t3.061_hsgkn.jpg