# NVIDIA Structured Data and Graph Models

## Instructions for AI Agents

- For clean Markdown of any page, append `.md` to the page URL
- For section-specific indexes, append `/llms.txt` to any section URL
- For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server

## Docs

- [Kumo Predict Overview](https://docs.nvidia.com/sdgm/platform-introduction.md): Get started with the Kumo Predict Platform - train and deploy models through Kumo Studio
- [Kumo Predict SDK](https://docs.nvidia.com/sdgm/fine-tuning/introduction.md): End-to-end fine-tuning walkthrough: connectors, tables, graphs, predictive queries, training
- [Prerequisites](https://docs.nvidia.com/sdgm/get-started/requirements.md): What you need before deploying the Kumo Predict Platform, regardless of deployment path.
- [Choose a Deployment Path](https://docs.nvidia.com/sdgm/get-started/choose-deployment-path.md): Choose a pip install, Docker Compose, or Helm (Kubernetes) deployment path for the Kumo Predict Platform.
- [Setup](https://docs.nvidia.com/sdgm/get-started/docker/prerequisites.md): Run Kumo from NVIDIA NGC with Docker Compose on one pre-provisioned Linux host.
- [GPU Sizing for Docker](https://docs.nvidia.com/sdgm/get-started/docker/gpu-sizing.md): How to size the GPU host for a Kumo Docker Compose deployment.
- [Setup](https://docs.nvidia.com/sdgm/get-started/helm/registry-and-image-access.md): Prerequisites and image pull setup for a Kumo Helm deployment on Kubernetes.
- [GPU Sizing for Helm](https://docs.nvidia.com/sdgm/get-started/helm/gpu-sizing.md): How to size GPU node pools for a Kumo Helm deployment on Kubernetes.
- [Quickstart: Train Your First Model](https://docs.nvidia.com/sdgm/quick-start/fine-tune.md)
- [Admin Setup Guide](https://docs.nvidia.com/sdgm/admin-setup-guide.md): Understand roles, groups, and projects in Kumo
- [Create and Manage Identities](https://docs.nvidia.com/sdgm/create-and-manage-identities.md): Create and manage reusable external identities for data connectors
- [Data Connectors](https://docs.nvidia.com/sdgm/data-connectors.md): Connect Kumo to Amazon S3, BigQuery, Snowflake, Databricks, Azure Data Lake Storage, and Google Cloud Storage, or upload local files.
- [Select Tables](https://docs.nvidia.com/sdgm/select-tables.md)
- [Choose Table Columns](https://docs.nvidia.com/sdgm/troubleshooting/column-selection.md): Select clean, relevant, and sufficiently complete columns for AutoML
- [Choose Table Relationships](https://docs.nvidia.com/sdgm/troubleshooting/table-linkages.md): Connect entity and event tables for an AutoML prediction task
- [Data Preparation Best Practices](https://docs.nvidia.com/sdgm/troubleshooting/data-quality-mistakes-and-how-to-catch-them.md): Choose and prepare relational data for Kumo AutoML
- [Column Preprocessing](https://docs.nvidia.com/sdgm/column-preprocessing.md)
- [Create Graph](https://docs.nvidia.com/sdgm/create-graph.md)
- [Understanding Predictive Query](https://docs.nvidia.com/sdgm/predictive-query.md)
- [Predictive Query Structure](https://docs.nvidia.com/sdgm/pquery-structure.md)
- [Task Types](https://docs.nvidia.com/sdgm/task-types.md)
- [Static vs. Temporal](https://docs.nvidia.com/sdgm/temporal-vs-static-pqueries.md)
- [Anchor Time and Data Leakage](https://docs.nvidia.com/sdgm/troubleshooting/what-is-anchor-time-and-why-is-it-important.md): Use temporal boundaries correctly when training and running AutoML models
- [Training](https://docs.nvidia.com/sdgm/training.md)
- [Large Scale Graph Learning](https://docs.nvidia.com/sdgm/how-distributed-training-works.md): How the Distributed Training Backend Works
- [How do I assign different weights to training samples?](https://docs.nvidia.com/sdgm/troubleshooting/can-i-perform-weighted-training.md): Learn how to use instance-level weights to influence model training in Kumo.
- [Warm Start Training: Initialize from Existing Models](https://docs.nvidia.com/sdgm/troubleshooting/how-do-i-initialize-training-from-an-existing-model.md): Learn how to fine tune a model by initializing from a previously trained model's weights
- [Run Mode](https://docs.nvidia.com/sdgm/run-mode.md)
- [Model Plan](https://docs.nvidia.com/sdgm/model-planner.md)
- [Model Plan Intuition](https://docs.nvidia.com/sdgm/model-plan-intuition.md)
- [Evaluation](https://docs.nvidia.com/sdgm/evaluation.md)
- [Classification](https://docs.nvidia.com/sdgm/classification.md): Metrics used to evaluate binary, multiclass, and multilabel classification tasks in Kumo.
- [Regression](https://docs.nvidia.com/sdgm/regression.md): Metrics used to evaluate numerical prediction tasks in Kumo, including MAE, MSE, RMSE, MAPE, and SMAPE.
- [Link Prediction](https://docs.nvidia.com/sdgm/link-prediction.md): Metrics reported for link prediction tasks, including NDCG, MAP, MRR, and rank-at-K variants.
- [Baselines](https://docs.nvidia.com/sdgm/baselines.md)
- [Explainability](https://docs.nvidia.com/sdgm/explainability.md)
- [Debug Model Performance](https://docs.nvidia.com/sdgm/troubleshooting/debugging-poor-model-performance.md): Diagnose data, graph, training, and evaluation problems in Kumo AutoML
- [Compare AutoML with an External Model](https://docs.nvidia.com/sdgm/troubleshooting/how-can-i-compare-a-predictive-query-with-an-external-model.md): Compare an AutoML model with a historical baseline, heuristic, or external model on shared holdout data
- [How do I calibrate my model?](https://docs.nvidia.com/sdgm/troubleshooting/how-do-i-calibrate-my-model.md): Understand automatic classification calibration and available model controls
- [Batch Predictions](https://docs.nvidia.com/sdgm/batch-prediction.md)
- [Outputs](https://docs.nvidia.com/sdgm/batch-prediction-outputs.md)
- [Predict with a Different Graph](https://docs.nvidia.com/sdgm/troubleshooting/how-do-i-make-inference-on-a-different-graph.md): Use the Kumo Predict SDK to run a trained model against another compatible graph
- [Installation](https://docs.nvidia.com/sdgm/fine-tuning/installation.md): Install the Kumo Predict SDK for fine-tuning workflows
- [API Key Management](https://docs.nvidia.com/sdgm/fine-tuning/api-key-management.md): Create, use, rotate, and revoke API keys for the Kumo Predict SDK
- [Groups, Projects, and RBAC](https://docs.nvidia.com/sdgm/fine-tuning/groups-and-projects.md): Scope Kumo Predict SDK work to groups and projects in RBAC-enabled workspaces
- [Connectors](https://docs.nvidia.com/sdgm/fine-tuning/connectors.md): Connect to data sources and inspect source tables with the Kumo Predict SDK
- [Connector Reference](https://docs.nvidia.com/sdgm/fine-tuning/connector-reference.md): Reference for Kumo Predict SDK connector classes and table access patterns
- [Tables](https://docs.nvidia.com/sdgm/fine-tuning/tables.md): Create and configure Kumo Tables from source data with metadata inference
- [Graphs](https://docs.nvidia.com/sdgm/fine-tuning/graphs.md): Build a Kumo Graph by connecting tables through primary/foreign key relationships
- [Predictive Queries](https://docs.nvidia.com/sdgm/fine-tuning/predictive-queries.md): Define machine learning problems using Kumo's Predictive Query Language (PQL)
- [Training & Predictions](https://docs.nvidia.com/sdgm/fine-tuning/training.md): Train models and generate batch predictions with the Kumo Predict SDK
- [Compute Sessions](https://docs.nvidia.com/sdgm/fine-tuning/compute-sessions.md): Understand current Kumo Predict SDK session behavior and shared-compute limitations
- [Code Generation](https://docs.nvidia.com/sdgm/fine-tuning/code-generation.md): Generate reproducible Kumo Predict SDK Python code from existing Kumo entities
- [Artifact and Output Exports](https://docs.nvidia.com/sdgm/fine-tuning/artifact-exports.md): Configure prediction, training table, embedding, and model artifact exports with the Kumo Predict SDK
- [Online Serving](https://docs.nvidia.com/sdgm/fine-tuning/online-serving.md): End-to-end guide: train, export, and deploy your Kumo model for real-time inference
- [Notebooks](https://docs.nvidia.com/sdgm/examples/platform-notebooks.md): Notebook examples for Kumo Predict Platform
- [Predictive Query](https://docs.nvidia.com/sdgm/examples/predictive-query.md)
- [Business Operations](https://docs.nvidia.com/sdgm/examples/business-operations.md)
- [Demand Forecasting Solution](https://docs.nvidia.com/sdgm/examples/bizops-demand-forecast.md)
- [Shipment Delay Prediction](https://docs.nvidia.com/sdgm/examples/bizops-shipment-delay.md)
- [Entity Resolution](https://docs.nvidia.com/sdgm/examples/bizops-entity-resolution.md)
- [Fraud](https://docs.nvidia.com/sdgm/examples/fraud.md)
- [Chargeback Fraud Detection](https://docs.nvidia.com/sdgm/examples/fraud-chargeback-abuse.md)
- [Credit Card Fraud Detection Solution](https://docs.nvidia.com/sdgm/examples/fraud-credit-card.md)
- [Money Laundering Detection](https://docs.nvidia.com/sdgm/examples/fraud-money-laundering.md)
- [Payback Abuse Detection Solution](https://docs.nvidia.com/sdgm/examples/fraud-payback-abuse.md)
- [Fraud Detection Demo](https://docs.nvidia.com/sdgm/examples/fraud-detection-demo.md): Step-by-step demo for detecting fraudulent credit card transactions with Kumo
- [Growth and Marketing](https://docs.nvidia.com/sdgm/examples/growthmarketing.md): Kumo Predict Platform for growth and marketing — GNN-powered models for LTV, churn, lead scoring, and recommendations
- [Customer Churn Prediction](https://docs.nvidia.com/sdgm/examples/growth-churn.md)
- [Lead Scoring Optimization](https://docs.nvidia.com/sdgm/examples/growth-lead-scoring.md)
- [Customer Lifetime Value Prediction](https://docs.nvidia.com/sdgm/examples/growth-ltv.md)
- [Anti-Targeting Solution](https://docs.nvidia.com/sdgm/examples/growth-anti-targeting.md)
- [Best Time to Send Notification](https://docs.nvidia.com/sdgm/examples/best-time-to-send.md)
- [Feature Adoption Prediction](https://docs.nvidia.com/sdgm/examples/growth-feature-adoption.md): Predict which users will adopt new product features using a multi-label classification model trained on user telemetry data.
- [Personalization](https://docs.nvidia.com/sdgm/examples/personalization.md)
- [Buy-It-Again Recommendation](https://docs.nvidia.com/sdgm/examples/pzn-buy-it-again.md)
- [Cold Start Recommendation](https://docs.nvidia.com/sdgm/examples/pzn-cold-start.md)
- [Item-to-Item Recommendation](https://docs.nvidia.com/sdgm/examples/pzn-related-items.md)
- [Search Recommendations](https://docs.nvidia.com/sdgm/examples/pzn-search-browse-reranking.md)
- [Personalized Email Recommendations](https://docs.nvidia.com/sdgm/examples/pzn-email-recommendations.md)
- [Hybrid Graph Neural Networks](https://docs.nvidia.com/sdgm/examples/pzn-hybrid-graph-neural-networks.md): Why Recommendation Systems are Better Off Using Hybrid Graph Neural Networks
- [Try Something New](https://docs.nvidia.com/sdgm/examples/pzn-try-something-new.md)
- [Quotas and Limits](https://docs.nvidia.com/sdgm/quotas-and-limits.md): Maximum data scale limits and job execution quotas for Kumo deployments.
- [Deployment Modes](https://docs.nvidia.com/sdgm/deployment-modes.md): Overview of Kumo Predict Platform deployment options: VPC Kubernetes (Helm) and Simplified VPC (Docker Compose).
- [Virtual Private Cloud](https://docs.nvidia.com/sdgm/virtual-private-cloud.md): Deploy Kumo fully inside your own cloud account with customer-controlled infrastructure, private network access, enterprise identity, and strong team separation.
- [Simplified VPC Deployment](https://docs.nvidia.com/sdgm/simplified-vpc-deployment.md)
- [Configure Single Sign-On](https://docs.nvidia.com/sdgm/configuring-sso.md)
- [Browser](https://docs.nvidia.com/sdgm/browsers.md)
- [kumoai](https://docs.nvidia.com/sdgm/sdk/kumoai.md): Root module — initialization, logging, and core type enumerations
- [kumoai.connector](https://docs.nvidia.com/sdgm/sdk/kumoai-connector.md): Connector and SourceTable interfaces for accessing raw data
- [kumoai.graph](https://docs.nvidia.com/sdgm/sdk/kumoai-graph.md): Graph, Table, and Column — the relational schema definitions for Kumo
- [kumoai.pquery](https://docs.nvidia.com/sdgm/sdk/kumoai-pquery.md): PredictiveQuery, training tables, and prediction tables
- [kumoai.encoder](https://docs.nvidia.com/sdgm/sdk/kumoai-encoder.md): Encoder overrides for column-level preprocessing
- [kumoai.trainer](https://docs.nvidia.com/sdgm/sdk/kumoai-trainer.md): Trainer, ModelPlan, training jobs, batch prediction, and online serving
- [Kumo AI SaaS Security White Paper](https://docs.nvidia.com/sdgm/security-and-governance.md)
- [Privacy Policy](https://docs.nvidia.com/sdgm/privacy-policy.md)
- [Consumer Privacy](https://docs.nvidia.com/sdgm/consumer-privacy.md)
- [Model Risk Management](https://docs.nvidia.com/sdgm/model-risk-management.md): Kumo features can help you with Model Risk Management
- [Data Processing Addendum](https://docs.nvidia.com/sdgm/data-processing-addendum.md): Kumo's Data Processing Addendum governing the processing of Customer Personal Data
- [Quotas and Limits](https://docs.nvidia.com/sdgm/platform-v2/quotas-and-limits.md)
- [Browser](https://docs.nvidia.com/sdgm/platform-v2/browsers.md)
- [AWS S3](https://docs.nvidia.com/sdgm/aws-s3.md)
- [AWS Glue Catalog](https://docs.nvidia.com/sdgm/aws-glue.md)
- [Google Cloud BigQuery](https://docs.nvidia.com/sdgm/google-cloud-bigquery.md)
- [Snowflake](https://docs.nvidia.com/sdgm/snowflake-connector.md): Connect your Snowflake data warehouse to Kumo using the Direct Connector or Secure Data Sharing.
- [Databricks](https://docs.nvidia.com/sdgm/databricks-connector.md)
- [Azure Data Lake Storage](https://docs.nvidia.com/sdgm/adls.md): Connect Kumo to Azure Data Lake Storage Gen2 for reading Parquet and CSV files.
- [Local Upload](https://docs.nvidia.com/sdgm/local-data-upload.md)
- [Deployment Modes](https://docs.nvidia.com/sdgm/platform-v2/deployment-modes.md)
- [SaaS](https://docs.nvidia.com/sdgm/platform-v2/saas.md)
- [Virtual Private Cloud](https://docs.nvidia.com/sdgm/platform-v2/virtual-private-cloud.md)
- [Simplified VPC Deployment](https://docs.nvidia.com/sdgm/platform-v2/simplified-vpc-deployment.md)
- [Snowflake Native App Fine-Tuning with Virtual Private Cloud Online Serving](https://docs.nvidia.com/sdgm/platform-v2/spcs-train-vpc-serving.md)
- [Private Link](https://docs.nvidia.com/sdgm/platform-v2/private-link.md)
- [Kumo as a Snowflake Native Application](https://docs.nvidia.com/sdgm/platform-v2/spcs/snowflake-native-application.md)
- [Installing the Snowflake Native App](https://docs.nvidia.com/sdgm/platform-v2/spcs/installing-kumo-on-spcs.md)
- [Snowflake SPCS Setup](https://docs.nvidia.com/sdgm/fine-tuning/snowflake-spcs-setup.md): Run the Kumo Fine-Tune SDK inside Snowpark Container Services notebooks
- [Architecture and Security](https://docs.nvidia.com/sdgm/platform-v2/spcs/spcs-security.md)
- [Schedule Kumo Native App and Batch Predictions](https://docs.nvidia.com/sdgm/platform-v2/spcs/kumo-app-schedule-guide.md): This guide walks through setting up Snowflake Tasks and Procedures to automatically: run a Snowflake Notebook at the scheduled time, start the Kumo App and trigger a Batch Prediction job, and after the job finishes, stop the Kumo App.
- [Enabling Parallel Job Concurrency in Kumo SPCS](https://docs.nvidia.com/sdgm/platform-v2/spcs/spcs-job-concurrency.md)
- [Estimating Snowflake Resource Usage](https://docs.nvidia.com/sdgm/platform-v2/spcs/spcs-estimating-resource-usage.md)
- [Troubleshooting and Support](https://docs.nvidia.com/sdgm/platform-v2/spcs/spcs-troubleshooting-and-support.md)
- [Databricks Native Application](https://docs.nvidia.com/sdgm/platform-v2/databricks-native-application.md)
- [Installing the Databricks Native App](https://docs.nvidia.com/sdgm/platform-v2/installing-the-databricks-native-app.md)
- [Architecture and Security](https://docs.nvidia.com/sdgm/platform-v2/databricks-security.md)
- [SSO Configuration Guide](https://docs.nvidia.com/sdgm/platform-v2/configuring-sso.md)
- [Kumo AI SaaS Security White Paper](https://docs.nvidia.com/sdgm/platform-v2/security-and-governance.md)
- [Privacy Policy](https://docs.nvidia.com/sdgm/platform-v2/privacy-policy.md)
- [Consumer Privacy](https://docs.nvidia.com/sdgm/platform-v2/consumer-privacy.md)
- [Model Risk Management](https://docs.nvidia.com/sdgm/platform-v2/model-risk-management.md): Kumo features can help you with Model Risk Management
- [Data Processing Addendum](https://docs.nvidia.com/sdgm/platform-v2/data-processing-addendum.md): Kumo's Data Processing Addendum governing the processing of Customer Personal Data
- [Kumo Platform Release Notes](https://docs.nvidia.com/sdgm/releases/platform.md): New updates and improvements for the Kumo Platform
- [2025 Product Updates](https://docs.nvidia.com/sdgm/releases/releases-2025.md): New updates and improvements
- [2024 Product Updates](https://docs.nvidia.com/sdgm/releases/releases-2024.md): New updates and improvements
- [2023 Product Updates](https://docs.nvidia.com/sdgm/releases/releases-2023.md): New updates and improvements
- [Overview](https://docs.nvidia.com/sdgm/rfm/overview.md): Introduction to Kumo Relational - the relational foundation model for predictive AI on relational data
- [How Kumo Relational Works](https://docs.nvidia.com/sdgm/rfm/introduction.md): Understand how Kumo Relational turns relational context and PQL into predictions through NVIDIA NIM
- [Deploy, Install, and Connect](https://docs.nvidia.com/sdgm/rfm/sdk-getting-started.md): Deploy a Kumo Relational NIM, install the NVIDIA Kumo Relational Client, and verify the connection
- [Quickstart: Kumo Relational](https://docs.nvidia.com/sdgm/quick-start/rfm.md): Build a relational graph and run your first prediction against a Kumo Relational NIM
- [Data Requirements](https://docs.nvidia.com/sdgm/rfm/data-requirements.md): Prepare relational tables, keys, timestamps, and history for Kumo Relational
- [Data Types and Semantic Types](https://docs.nvidia.com/sdgm/rfm/data-types.md): Understand physical data types and semantic column meaning in Kumo Relational
- [Time and End-Time Columns](https://docs.nvidia.com/sdgm/rfm/time-columns.md): Choose table time and end-time metadata and understand how Kumo Relational uses it during sampling
- [Best Practices](https://docs.nvidia.com/sdgm/rfm/best-practices.md): Best practices for data preparation with Kumo Relational
- [Snowflake Connector](https://docs.nvidia.com/sdgm/rfm/connectors/snowflake.md): Connect Kumo Relational to data stored in Snowflake
- [SQLite Connector](https://docs.nvidia.com/sdgm/rfm/connectors/sqlite.md): Connect Kumo Relational to data stored in SQLite
- [Databricks Connector](https://docs.nvidia.com/sdgm/rfm/connectors/databricks.md): Connect Kumo Relational to data stored in a Databricks SQL warehouse
- [DuckDB Connector](https://docs.nvidia.com/sdgm/rfm/connectors/duckdb.md): Connect Kumo Relational to data stored in a DuckDB database
- [Amazon S3](https://docs.nvidia.com/sdgm/rfm/connectors/s3.md): Read CSV and Parquet data from Amazon S3 for a Kumo Relational graph
- [Create a Graph](https://docs.nvidia.com/sdgm/rfm/graph-creation.md): Create a Kumo Relational graph from prepared DataFrames or a supported data source
- [Configure Table Metadata](https://docs.nvidia.com/sdgm/rfm/table-definitions.md): Review and correct Kumo Relational's inferred keys, timestamps, and semantic types
- [Define Relationships](https://docs.nvidia.com/sdgm/rfm/graph-definitions.md): Review and correct relationships between tables in a Kumo Relational graph
- [Validate the Graph](https://docs.nvidia.com/sdgm/rfm/graph-validation.md): Inspect and validate a Kumo Relational graph before making predictions
- [Make Predictions](https://docs.nvidia.com/sdgm/rfm/make-predictions.md): Run a Kumo Relational prediction through the NVIDIA Kumo Relational Client
- [Query Kumo Relational](https://docs.nvidia.com/sdgm/rfm/querying-rfm.md): Understand the target, entity, horizon, and current Kumo Relational PQL constraints
- [Prediction Types](https://docs.nvidia.com/sdgm/rfm/prediction-types.md): Task taxonomy: regression, classification, forecasting, ranking, and static tasks
- [Write Predictive Queries](https://docs.nvidia.com/sdgm/rfm/writing-predictive-queries.md): Translate a business question into a complete, testable Kumo Relational PQL query
- [Filters and Operators](https://docs.nvidia.com/sdgm/rfm/filters-and-operators.md): Use entity selection, context filters, anchor times, and assumptions in Kumo Relational
- [Forecasting](https://docs.nvidia.com/sdgm/rfm/forecasting.md): Run single-entity multi-timeframe forecasts with Kumo Relational
- [Predict with Custom Context](https://docs.nvidia.com/sdgm/rfm/custom-context.md): Supply labeled context examples directly to a Kumo Relational prediction task
- [Prediction Results](https://docs.nvidia.com/sdgm/rfm/prediction-results.md): Understand the pandas DataFrame returned by a Kumo Relational prediction
- [Prediction Explainability](https://docs.nvidia.com/sdgm/rfm/prediction-explainability.md): Configure and request Kumo Relational prediction explanations
- [Understand Explanations](https://docs.nvidia.com/sdgm/rfm/understand-explanations.md): Interpret Kumo Relational cohort, subgraph, context-example, and summary outputs
- [Configuration](https://docs.nvidia.com/sdgm/rfm/configuration.md): Configure neighborhood sampling, run modes, timing, and inference behavior for Kumo Relational
- [Batch Prediction](https://docs.nvidia.com/sdgm/rfm/batch-prediction.md): Score Kumo Relational entities with built-in bounded batching and retries
- [Improve Model Performance](https://docs.nvidia.com/sdgm/rfm/improve-model-performance.md): Evaluate and tune Kumo Relational with key-aligned historical backtests
- [Notebooks](https://docs.nvidia.com/sdgm/examples/notebooks.md): Follow an example Kumo Relational workflow in a notebook
- [Python SDK API](https://docs.nvidia.com/sdgm/rfm/python-api-reference.md): Public Python interfaces for connecting to and running predictions with Kumo Relational
- [Migrate to Kumo Relational Client 1.0](https://docs.nvidia.com/sdgm/rfm/migrate-to-kumo-relational.md): Update package names, imports, model identifiers, and removed client APIs for Kumo Relational Client 1.0
- [NIM REST API](https://docs.nvidia.com/sdgm/rfm/nim-rest-api.md): Send synchronous prediction requests to Kumo Relational through its NIM-compatible REST API
- [Glossary](https://docs.nvidia.com/sdgm/rfm/glossary.md): Definitions of data, graph, prediction, and configuration terms used by Kumo Relational
- [Research Overview](https://docs.nvidia.com/sdgm/research.md): Research from Kumo on relational learning, graph models, and predictive AI
- [Relational Deep Learning](https://docs.nvidia.com/sdgm/research/relational-deep-learning.md): How relational deep learning represents databases as temporal heterogeneous graphs and learns directly from their structure
- [Introduction to Graph Neural Networks](https://docs.nvidia.com/sdgm/research/gnn-introduction.md): An introduction to graphs, graph neural networks, message passing, and the move to graph transformers
- [Hybrid GNNs for Recommendations](https://docs.nvidia.com/sdgm/research/hybrid-graph-neural-networks.md): How a hybrid graph neural network combines repeat and exploratory behavior for recommendations
- [Introduction to Graph Transformers](https://docs.nvidia.com/sdgm/research/graph-transformers-intro.md): How attention mechanisms adapt to graph-structured data, why standard transformers fall short, and the innovations behind graph-aware architectures
- [Relational Graph Transformers](https://docs.nvidia.com/sdgm/research/relational-graph-transformers.md): How graph-aware attention overcomes message-passing limits for relational databases and supports relational foundation models
- [KumoRFM: A Relational Foundation Model](https://docs.nvidia.com/sdgm/research/kumorfm-paper.md): Why KumoRFM was needed, how its relational foundation-model architecture works, and what the RelBench results show
- [RelBench](https://docs.nvidia.com/sdgm/research/relbench.md): A guide to the RelBench benchmark for deep learning on multi-table relational databases
- [PluRel: Synthetic Relational Data](https://docs.nvidia.com/sdgm/research/plurel.md): How synthetic relational databases make scaling-law research possible for relational foundation models
- [Churn Prediction](https://docs.nvidia.com/sdgm/research/churn-prediction.md): How graph neural networks improve churn prediction by learning from relational customer data
- [Recommender Systems with LLMs](https://docs.nvidia.com/sdgm/research/recsys-with-llms.md): How graph transformers and encoder-based LLM embeddings combine for personalized recommendations
- [Time Series Forecasting](https://docs.nvidia.com/sdgm/research/time-series-forecasting.md): How Graph Transformers use relational context, temporal signals, and diffusion models for time series forecasting
- [PyG Practitioner's Guide](https://docs.nvidia.com/sdgm/research/pyg-practitioners-guide.md): A practical introduction to PyTorch Geometric, from a first GNN to architecture selection and production graph learning
- [Optimize PyG with torch.compile](https://docs.nvidia.com/sdgm/research/pyg-compile-guide.md): A practical guide to optimizing PyG-based GNN training by eliminating re-compilations, graph breaks, and host-device synchronization
- [Graph Construction](https://docs.nvidia.com/sdgm/research/graph-construction-guide.md): Three production optimizations for faster graph construction: ingestion skip, eager loading, and deterministic edge reuse.
- [Adaptive Graph Sampling](https://docs.nvidia.com/sdgm/research/adaptive-sampling-guide.md): How metapath-aware graph sampling redistributes neighbor budgets to improve graph learning
- [When Should I Use Kumo Predict or Kumo Relational?](https://docs.nvidia.com/sdgm/faq.md): Choose between training a dedicated AutoML model and using Kumo Relational for immediate predictions
- [How does Kumo handle missing values in my dataset?](https://docs.nvidia.com/sdgm/troubleshooting/how-does-kumo-handle-missing-values-in-my-dataset.md)
- [What types of data can Kumo ingest?](https://docs.nvidia.com/sdgm/troubleshooting/what-types-of-data-can-kumo-ingest.md): Supported AutoML data sources, file formats, and column types
- [Can I use an external model to embed a column?](https://docs.nvidia.com/sdgm/troubleshooting/how-can-i-use-an-external-model-for-embedding-a-column.md): Provide precomputed embeddings as an AutoML input column
- [How do I use a feature store with Kumo?](https://docs.nvidia.com/sdgm/troubleshooting/how-do-i-use-a-feature-store.md)
- [What does a MISSING column score mean?](https://docs.nvidia.com/sdgm/troubleshooting/does-a-missing-score-field-indicate-that-the-column-was-excluded-from-training.md): Interpret a missing column contribution score
- [How can I detect data leakage?](https://docs.nvidia.com/sdgm/troubleshooting/what-mechanisms-does-kumo-provide-to-detect-data-leakage.md): Use evaluation, explanations, and source-data timing to identify leakage
- [How does Kumo handle timezones for timestamp values?](https://docs.nvidia.com/sdgm/troubleshooting/how-does-kumo-handle-timezones-for-timestamp-values.md)
- [Can I use specific date and time values in my PQuery filters?](https://docs.nvidia.com/sdgm/troubleshooting/can-i-use-specific-date-and-time-values-in-my-pquery-filters.md)
- [How can I reduce model training time?](https://docs.nvidia.com/sdgm/troubleshooting/what-is-the-recommended-way-to-reduce-model-training-time.md): Choose a smaller AutoML search and development dataset for faster iteration
- [How do I improve model performance?](https://docs.nvidia.com/sdgm/troubleshooting/model-improvement.md): A concise checklist for improving a Kumo AutoML model
- [Can I tune model hyperparameters?](https://docs.nvidia.com/sdgm/troubleshooting/can-i-tune-model-hyperparameters.md)
- [How do I specify the train/validation/test splits?](https://docs.nvidia.com/sdgm/troubleshooting/how-do-i-specify-the-train-validation-test-splits.md): Suggest Edits
- [How do I perform backtesting on a holdout dataset?](https://docs.nvidia.com/sdgm/troubleshooting/how-do-i-perform-backtesting-on-a-holdout-dataset.md)
- [How does Kumo handle the cold start problem in ML?](https://docs.nvidia.com/sdgm/troubleshooting/how-does-kumo-handle-the-cold-start-problem-in-ml.md)
- [How do I generate predictions on new data?](https://docs.nvidia.com/sdgm/troubleshooting/how-do-i-generate-predictions-on-new-data-using-a-previously-trained-model.md): Run a trained AutoML model after refreshing its graph data
- [How can I diagnose problems with my data pipeline at batch prediction time?](https://docs.nvidia.com/sdgm/troubleshooting/how-can-i-diagnose-problems-with-my-data-pipeline-at-batch-prediction-time.md)
- [What happens if the predictions per entity value in batch predictions differs from the initial pQuery’s RANK TOP K value?](https://docs.nvidia.com/sdgm/troubleshooting/what-happens-if-the-predictions-per-entity-value-in-batch-predictions-differs-from-the-initial-pquerys-rank-top-k-value.md)
- [What are embedding outputs?](https://docs.nvidia.com/sdgm/troubleshooting/embedding-outputs.md): Export learned entity representations from an AutoML batch prediction
- [How Is Kumo Relational Different from an LLM or Text-to-SQL?](https://docs.nvidia.com/sdgm/faq/kumorfm-vs-llm.md): Understand when to use relational prediction, language generation, or database retrieval