@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.083_starformer.yaml

@@ -0,0 +1,178 @@
+id: aisb.t3.083_starformer
+name: STaRFormer
+version: 0.1.0
+one_line: Transformer framework with dynamic attention-based regional masking and
+  semi-supervised contrastive learning for non-stationary, irregularly-sampled sequential
+  data.
+task_description: 'STaRFormer is a transformer-based framework designed for robust
+  sequential representation learning under real-world conditions where standard assumptions
+  of stationarity and regular sampling do not hold. The architecture employs an encoder-only
+  transformer backbone combined with Dynamic Attention-Based Regional Masking (DAReM),
+  which manipulates task-specific regions within input sequences to introduce synthetic
+  variations in statistical properties and simulate varying sampling frequencies.
+  The training process incorporates a novel semi-supervised contrastive learning approach
+  that maximizes agreement between batch-wise positive pairs (masked and unmasked
+  embeddings of the same input) and class-wise positive pairs (masked and unmasked
+  embeddings of the same class), producing two correlated latent representations that
+  enhance downstream task performance. Evaluation spans classification on spatiotemporal
+  and non-stationary benchmarks (DKT and Microsoft datasets), irregularly-sampled
+  benchmarks (3 public datasets), and regular time series (UEA, 30 datasets), plus
+  anomaly detection (2 datasets) and regression (19 datasets) against 6 and 15 baseline
+  models respectively.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: restricted
+primary_outputs:
+- pam_accuracy
+- latent_representations
+- dare_m_masks
+- ablation_reports
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: Review preserved sequential representation benchmark metadata, paper
+    content, and metric definitions.
+- id: restore_first
+  label: Restore First
+  description: Restore the missing STaRFormer implementation (transformer backbone,
+    DAReM module, semi-supervised CL loss) before attempting benchmark execution.
+- id: full_benchmark
+  label: Full Benchmark
+  description: Execute complete evaluation pipeline including classification, anomaly
+    detection, and regression benchmarks with ablation studies.
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: DKT (Digital Key Trajectory)
+    type: custom_restricted
+    description: BMW proprietary trajectory dataset from Digital Key use case
+  - name: Microsoft spatiotemporal
+    type: public
+    description: Public real-world dataset from Microsoft for non-stationary sequential
+      data
+  - name: Irregularly-sampled benchmark
+    type: public
+    description: Collection of 3 public datasets with irregular sampling patterns
+  - name: UEA time series
+    type: public
+    description: University of East Anglia benchmark with 30 public datasets for regular
+      time series classification
+  - name: Anomaly detection benchmark
+    type: public
+    description: Public benchmark with 2 datasets for anomaly detection evaluation
+  - name: Regression benchmark
+    type: public
+    description: Public benchmark with 19 datasets for regression evaluation
+  notes:
+  - The current local package does not preserve an executable benchmark route; treat
+    dataset setup as part of upstream restoration rather than a bundled run path.
+  - DKT dataset is proprietary BMW data requiring appropriate access credentials.
+  - Public datasets should be downloaded from respective original sources during restoration
+    phase.
+credential_requirements:
+  mode: restricted
+  items:
+  - BMW internal data access for DKT dataset
+  notes:
+  - DKT (Digital Key Trajectory) data is proprietary to BMW Group and requires internal
+    access authorization.
+  - Public benchmark datasets are freely accessible but must be sourced during restoration.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: 3.10.14
+  cuda: '11.8'
+  pytorch: 2.4.0
+  flash_attn: null
+  key_packages:
+  - transformers==4.50.3
+  notes:
+  - The current local package does not preserve executable setup files or runnable
+    source code; environment hints here should be treated as provisional until upstream
+    restoration.
+  - DAReM module requires attention weight extraction per transformer layer for regional
+    mask computation.
+  - Semi-supervised CL loss includes hyperparameter lambda_CL for balancing contrastive
+    and task losses.
+  - See the bundled README/requirements for the full dependency set.
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+risk_notes:
+- Current snapshot preserves paper and metadata only; no executable model code is
+  available.
+- Blocked-no-code status: no benchmark execution was performed in the packaging pass.
+- Restoration of transformer backbone, DAReM implementation, and CL training loop
+  required before execution.
+- DKT dataset access requires BMW Group authorization.
+recommended_when: Use this entry when you want paper-backed planning, upstream code
+  restoration for the STaRFormer sequential representation route, or analysis of the
+  DAReM methodology for non-stationary sequential data.
+not_recommended_when: Do not treat this local package as a ready-to-run benchmark;
+  the packaged snapshot does not preserve the executable model code. Do not expect
+  functional evaluation outputs without completing upstream restoration of the STaRFormer
+  implementation.
+paper:
+  title: 'STaRFormer: Semi-Supervised Task-Informed Representation Learning via Dynamic
+    Attention-Based Regional Masking for Sequential Data'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://neurips.cc/virtual/2025/poster/116860
+  authors:
+  - Maximilian Forstenhäusler (BMW Group, University of Glasgow)
+  - Daniel Külzer (BMW Group)
+  - Christos Anagnostopoulos (University of Glasgow)
+  - Shameem Puthiya Parambath (University of Glasgow)
+  - Natascha Weber (BMW Group)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.083_starformer.zip
+  archive_type: zip
+  local_dir_name: paper-83-STaRFormer
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.083_starformer.zip
+  sha256: 062e8be9653c9079088c40fde448d1ee016e45b85e622708798feff990ccff68
+  size_bytes: 8551
+display:
+  palette_seed: mint-violet-mask
+  art_style: sequential-masking
+  accent_priority: high
+image_path: ../image/083_aisb.t3.083_starformer.jpg
+capability_tags:
+- research_code_optimization
+- sequential_data
+- semi_supervised_learning
+- transformers
+- representation_learning
+- time_series
+- contrastive_learning
+- attention_mechanisms
+- non_stationary_analysis
+- anomaly_detection
+- regression
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track

package/AISB/catalog/aisb.t3.083_starformer.zh.yaml

@@ -0,0 +1,140 @@
+id: aisb.t3.083_starformer
+name: STaRFormer
+version: 0.1.0
+one_line: 面向非平稳与不规则采样序列数据的 Transformer 框架，结合动态注意力区域掩码与半监督对比学习。
+task_description: STaRFormer 是一个用于真实场景序列表征学习的 Transformer 框架，针对“非平稳+非规则采样”条件设计。其采用仅编码器 Transformer 主干，并结合动态注意力区域掩码（DAReM），对输入序列中任务相关区域进行操作，构造统计性质变化并模拟不同采样频率。训练中引入新的半监督对比学习：同时最大化批内正样本对（同一样本掩码/未掩码嵌入）与类内正样本对（同类样本掩码/未掩码嵌入）的一致性，得到相关潜在表示以提升下游性能。评测覆盖分类、异常检测与回归多类基准。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: restricted
+primary_outputs:
+- pam_accuracy
+- latent_representations
+- dare_m_masks
+- ablation_reports
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: Review preserved sequential representation benchmark metadata, paper content, and metric definitions.
+- id: restore_first
+  label: Restore First
+  description: Restore the missing STaRFormer implementation (transformer backbone, DAReM module, semi-supervised CL loss) before attempting benchmark execution.
+- id: full_benchmark
+  label: Full Benchmark
+  description: Execute complete evaluation pipeline including classification, anomaly detection, and regression benchmarks with ablation studies.
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: DKT (Digital Key Trajectory)
+    type: custom_restricted
+    description: BMW proprietary trajectory dataset from Digital Key use case
+  - name: Microsoft spatiotemporal
+    type: public
+    description: Public real-world dataset from Microsoft for non-stationary sequential data
+  - name: Irregularly-sampled benchmark
+    type: public
+    description: Collection of 3 public datasets with irregular sampling patterns
+  - name: UEA time series
+    type: public
+    description: University of East Anglia benchmark with 30 public datasets for regular time series classification
+  - name: Anomaly detection benchmark
+    type: public
+    description: Public benchmark with 2 datasets for anomaly detection evaluation
+  - name: Regression benchmark
+    type: public
+    description: Public benchmark with 19 datasets for regression evaluation
+  notes:
+  - The current local package does not preserve an executable benchmark route; treat dataset setup as part of upstream restoration rather than a bundled run path.
+  - DKT dataset is proprietary BMW data requiring appropriate access credentials.
+  - Public datasets should be downloaded from respective original sources during restoration phase.
+credential_requirements:
+  mode: restricted
+  items:
+  - BMW internal data access for DKT dataset
+  notes:
+  - DKT (Digital Key Trajectory) data is proprietary to BMW Group and requires internal access authorization.
+  - Public benchmark datasets are freely accessible but must be sourced during restoration.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: 3.10.14
+  cuda: '11.8'
+  pytorch: 2.4.0
+  flash_attn: null
+  key_packages:
+  - transformers==4.50.3
+  notes:
+  - The current local package does not preserve executable setup files or runnable source code; environment hints here should be treated as provisional until upstream restoration.
+  - DAReM module requires attention weight extraction per transformer layer for regional mask computation.
+  - Semi-supervised CL loss includes hyperparameter lambda_CL for balancing contrastive and task losses.
+  - See the bundled README/requirements for the full dependency set.
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+risk_notes:
+- Current snapshot preserves paper and metadata only; no executable model code is available.
+- Blocked-no-code status: no benchmark execution was performed in the packaging pass.
+- Restoration of transformer backbone, DAReM implementation, and CL training loop required before execution.
+- DKT dataset access requires BMW Group authorization.
+recommended_when: 当你需要基于论文进行方法规划、恢复 STaRFormer 序列表征代码路线，或分析 DAReM 在非平稳序列中的作用时可使用本条目。
+not_recommended_when: 不要将当前本地包作为可直接运行基准；当前快照未保留可执行模型代码。未完成上游恢复前无法得到有效评估输出。
+paper:
+  title: 'STaRFormer: Semi-Supervised Task-Informed Representation Learning via Dynamic Attention-Based Regional Masking for Sequential Data'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://neurips.cc/virtual/2025/poster/116860
+  authors:
+  - Maximilian Forstenhäusler (BMW Group, University of Glasgow)
+  - Daniel Külzer (BMW Group)
+  - Christos Anagnostopoulos (University of Glasgow)
+  - Shameem Puthiya Parambath (University of Glasgow)
+  - Natascha Weber (BMW Group)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.083_starformer.zip
+  archive_type: zip
+  local_dir_name: paper-83-STaRFormer
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.083_starformer.zip
+  sha256: 062e8be9653c9079088c40fde448d1ee016e45b85e622708798feff990ccff68
+  size_bytes: 8551
+display:
+  palette_seed: mint-violet-mask
+  art_style: sequential-masking
+  accent_priority: high
+image_path: ../image/083_aisb.t3.083_starformer.jpg
+capability_tags:
+- research_code_optimization
+- sequential_data
+- semi_supervised_learning
+- transformers
+- representation_learning
+- time_series
+- contrastive_learning
+- attention_mechanisms
+- non_stationary_analysis
+- anomaly_detection
+- regression
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track

package/AISB/catalog/aisb.t3.084_ift.yaml

@@ -0,0 +1,139 @@
+id: aisb.t3.084_ift
+name: Towards Accurate Time Series Forecasting via Implicit Decoding
+version: 0.1.0
+one_line: Paper-backed time-series decoder optimization using implicit frequency-based
+  forecasting modules, currently requiring upstream source restoration.
+task_description: 'This benchmark entry covers the Implicit Forecaster (IF) method
+  from a NeurIPS 2025 paper, which replaces the output MLP layers of existing time-series
+  transformer encoders with a novel frequency-domain decoding module. Instead of independent
+  pointwise predictions, IF globally forms forecasts by predicting constituent waves
+  parameterized by frequency, amplitude, and phase, enabling better handling of long-short-term
+  dynamics and multi-periodicity patterns. The approach is architecture-agnostic and
+  can boost mainstream TSF models including Informer, Autoformer, PatchTST, and iTransformer.
+  The current snapshot preserves only metadata and metric contracts; executable code,
+  launcher scripts, and README are absent. This entry is suitable for paper-backed
+  planning and upstream restoration planning, but not for direct local execution until
+  the upstream GitHub source (https://github.com/rakuyorain/Implicit-Forecaster) is
+  rehydrated into the package.
+
+  '
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: 'Review the preserved metric contracts, catalog metadata, and paper-backed
+    design rationale. No executable code is present in the current snapshot. Suitable
+    for route planning.
+
+    '
+- id: restore_and_validate
+  label: Restore and Validate
+  description: 'Restore upstream GitHub source tree before treating this as a runnable
+    benchmark. Requires fetching the external repository and rehydrating code, scripts,
+    and documentation into the current package snapshot.
+
+    '
+dataset_download:
+  primary_method: manual_or_not_explicit
+  sources:
+  - type: upstream_github
+    url: https://github.com/rakuyorain/Implicit-Forecaster
+    note: Linked in the paper abstract; not preserved in current snapshot
+  notes:
+  - The current local package contains no dataset scripts, README, or setup instructions.
+  - Dataset setup is an upstream restoration step, not a bundled one-click download.
+  - The paper reports experiments on 14 real-world time-series datasets commonly used
+    in TSF research.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No explicit dataset or API credential variables were detected in the bundled snapshot.
+  - Any credential needs must be determined during upstream source restoration.
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: '2.0'
+  flash_attn: null
+  key_packages: []
+  notes:
+  - Paper reports experiments on a single NVIDIA RTX 4090 GPU.
+  - Implementation built on PyTorch 2.0 per the paper.
+  - Current local snapshot preserves no dependency files or environment specifications.
+  - Environment details sourced from paper rather than executable artifacts.
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+risk_notes:
+- Current snapshot contains no executable code files; metric surface is blocked.
+- Code, launcher scripts, and README are absent; upstream rehydration required.
+- Top-level README absent in the current snapshot.
+- No benchmark execution was performed in this packaging pass.
+- Treat as a paper-backed planning entry pending source restoration.
+recommended_when: 'Use this entry when planning paper-backed research on decoder-side
+  optimization for time-series forecasting, studying frequency-domain implicit decoding
+  design, or executing upstream source restoration. Appropriate for route planning
+  and benchmark rehydration projects.
+
+  '
+not_recommended_when: 'Do not use as a ready-to-run local benchmark. The current snapshot
+  lacks executable anchors for any staged metrics. Execution requires upstream GitHub
+  restoration before any benchmark run is attempted.
+
+  '
+paper:
+  title: Towards Accurate Time Series Forecasting via Implicit Decoding
+  authors:
+  - Xinyu Li (University of Melbourne)
+  - Yuchen Luo (University of Melbourne)
+  - Hao Wang (Zhejiang University)
+  - Haoxuan Li (Peking University, MBZUAI)
+  - Liuhua Peng (University of Melbourne)
+  - Feng Liu (University of Melbourne)
+  - Yandong Guo (AI2 Robotics)
+  - Kun Zhang (Carnegie Mellon University, MBZUAI)
+  - Mingming Gong (University of Melbourne, MBZUAI)
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=gqoeQPhQcE
+  github: https://github.com/rakuyorain/Implicit-Forecaster
+display:
+  palette_seed: amber-sky-implicit
+  art_style: decoding-wave
+  accent_priority: high
+image_path: ../../../AISB/image/084_aisb.t3.084_ift.jpg
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.084_ift.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.084_ift.zip
+  archive_type: zip
+  sha256: d71123fe55a046305a8aca0aec7ee5a5205ef1742c296c01e77b62f5d7542b7c
+  size_bytes: 41911

package/AISB/catalog/aisb.t3.084_ift.zh.yaml

@@ -0,0 +1,111 @@
+id: aisb.t3.084_ift
+name: Towards Accurate Time Series Forecasting via Implicit Decoding
+version: 0.1.0
+one_line: 基于论文的时间序列解码器优化路线，采用隐式频域预测模块；当前需先恢复上游源码。
+task_description: 该条目覆盖 NeurIPS 2025 的 Implicit Forecaster（IF）方法。它将现有时间序列 Transformer 编码器的输出 MLP 解码层替换为新的频域解码模块。不同于逐点独立预测，IF 通过预测由频率、振幅、相位参数化的组成波形来全局形成预测，更好处理长短期动态与多周期性模式。该方法与架构无关，可提升 Informer、Autoformer、PatchTST、iTransformer 等主流 TSF 模型。当前快照仅保留元数据与指标契约，不含代码、启动脚本和 README；适合基于论文的规划与恢复，不适合直接本地执行。
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: 'Review the preserved metric contracts, catalog metadata, and paper-backed design rationale. No executable code is present in the current snapshot. Suitable for route planning.
+
+    '
+- id: restore_and_validate
+  label: Restore and Validate
+  description: 'Restore upstream GitHub source tree before treating this as a runnable benchmark. Requires fetching the external repository and rehydrating code, scripts, and documentation into the current package snapshot.
+
+    '
+dataset_download:
+  primary_method: manual_or_not_explicit
+  sources:
+  - type: upstream_github
+    url: https://github.com/rakuyorain/Implicit-Forecaster
+    note: Linked in the paper abstract; not preserved in current snapshot
+  notes:
+  - The current local package contains no dataset scripts, README, or setup instructions.
+  - Dataset setup is an upstream restoration step, not a bundled one-click download.
+  - The paper reports experiments on 14 real-world time-series datasets commonly used in TSF research.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No explicit dataset or API credential variables were detected in the bundled snapshot.
+  - Any credential needs must be determined during upstream source restoration.
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: '2.0'
+  flash_attn: null
+  key_packages: []
+  notes:
+  - Paper reports experiments on a single NVIDIA RTX 4090 GPU.
+  - Implementation built on PyTorch 2.0 per the paper.
+  - Current local snapshot preserves no dependency files or environment specifications.
+  - Environment details sourced from paper rather than executable artifacts.
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+risk_notes:
+- Current snapshot contains no executable code files; metric surface is blocked.
+- Code, launcher scripts, and README are absent; upstream rehydration required.
+- Top-level README absent in the current snapshot.
+- No benchmark execution was performed in this packaging pass.
+- Treat as a paper-backed planning entry pending source restoration.
+recommended_when: 当你规划时间序列预测中解码侧优化研究、分析频域隐式解码设计，或执行上游源码恢复时可使用本条目。适用于路线规划与基准再水化。
+not_recommended_when: 不要将其作为可直接运行的本地基准。当前快照对所有分阶段指标都缺少可执行锚点，必须先恢复上游 GitHub 源码后再运行。
+paper:
+  title: Towards Accurate Time Series Forecasting via Implicit Decoding
+  authors:
+  - Xinyu Li (University of Melbourne)
+  - Yuchen Luo (University of Melbourne)
+  - Hao Wang (Zhejiang University)
+  - Haoxuan Li (Peking University, MBZUAI)
+  - Liuhua Peng (University of Melbourne)
+  - Feng Liu (University of Melbourne)
+  - Yandong Guo (AI2 Robotics)
+  - Kun Zhang (Carnegie Mellon University, MBZUAI)
+  - Mingming Gong (University of Melbourne, MBZUAI)
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=gqoeQPhQcE
+  github: https://github.com/rakuyorain/Implicit-Forecaster
+display:
+  palette_seed: amber-sky-implicit
+  art_style: decoding-wave
+  accent_priority: high
+image_path: ../../../AISB/image/084_aisb.t3.084_ift.jpg
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.084_ift.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.084_ift.zip
+  archive_type: zip
+  sha256: d71123fe55a046305a8aca0aec7ee5a5205ef1742c296c01e77b62f5d7542b7c
+  size_bytes: 41911