Choose Model

Choose Model

Purpose

Select the right ML/DL algorithm based on current SOTA benchmarks, problem type, data characteristics, and production constraints.

How It Works

Master Decision Framework (2025)

Tabular Data — Still the Most Common

| Data Size | Best Models | Notes | |-----------|------------|-------| | 1M | LightGBM (fastest), CatBoost, XGBoost | LightGBM: histogram-based, memory efficient | | High cardinality cats | CatBoost (ordered target encoding) | Handles categoricals natively | | Many features (>500) | LightGBM + LASSO pre-filter, or FT-Transformer | Feature selection critical | | Multi-modal (tab+text+img) | Deep learning (multimodal fusion) | Embeddings + tabular features |

The uncomfortable truth: On pure tabular data, well-tuned gradient boosting still beats deep learning on most benchmarks (Grinsztajn et al., 2022). Deep learning wins when you have multi-modal data, can pre-train, or need online learning.

Text Data

| Task | Best Models (2025) | |------|-------------------| | Classification | Fine-tuned DeBERTa-v3, ModernBERT, SetFit (few-shot) | | Generation | Llama 3.1/3.2, Mistral, Qwen 2.5 (fine-tuned) | | Embeddings | E5-Mistral, BGE-M3, Nomic-embed | | NER | GLiNER (zero-shot), fine-tuned DeBERTa | | Sentiment | Fine-tuned transformer or zero-shot LLM | | Summarization | Fine-tuned BART/T5, or prompted LLM |

Image Data

| Task | Best Models (2025) | |------|-------------------| | Classification | DINOv2 + linear probe, ConvNeXt V2, EfficientNet V2 | | Object detection | YOLOv10, RT-DETR, Co-DETR | | Segmentation | SAM 2, Mask2Former, OneFormer | | Generation | FLUX, Stable Diffusion 3, DALL-E 3 | | OCR/Document | Donut, LayoutLMv3, Florence-2 |

Time Series

| Task | Best Models (2025) | |------|-------------------| | Forecasting | TimesFM (Google), Chronos (Amazon), Lag-Llama, PatchTST | | Classification | InceptionTime, ROCKET, Hydra-MR | | Anomaly detection | TimesNet, Anomaly Transformer | | Foundation models | TimesFM, Moment, Timer |

New paradigm: Foundation models for time series (zero-shot forecasting) — similar to how LLMs changed NLP. TimesFM and Chronos can forecast without training on your specific series.

Production Constraints Matrix

| Constraint | Best Models | |-----------|-------------| | Must explain predictions | EBM (Explainable Boosting), logistic regression, SHAP on XGBoost | | <10ms latency | Linear models, distilled models, ONNX-optimized | | Runs on CPU | LightGBM, sklearn models, ONNX Runtime | | Runs on edge/mobile | TFLite, Core ML, ONNX with quantization | | Handles concept drift | Online learning: River, Vowpal Wabbit | | Minimal labeled data | SetFit, TabPFN, zero-shot LLMs, active learning | | Real-time streaming | River, Flink ML, online gradient boosting |

AutoML (When to Use)

| Tool | Strength | When | |------|----------|------| | AutoGluon | Best overall, ensembling | Production baselines | | FLAML | Fast, resource-aware | Quick experiments | | Optuna | Flexible HPO | Custom search spaces | | H2O AutoML | Enterprise, explainability | Regulated industries | | Auto-sklearn | Meta-learning warm starts | Research |

Pro tip: Start with AutoGluon for a strong baseline, then beat it with domain knowledge and feature engineering.

Usage Examples

"I have 50K rows of customer data with 30 features, 15% categorical.
I need to predict churn with explainable predictions. What model?"
→ XGBoost with SHAP. Baseline with AutoGluon. Consider EBM if
  full glass-box interpretability is required.

"I need to forecast 500 product SKUs daily, with only 6 months of history
per SKU. Some SKUs have very few data points."
→ TimesFM or Chronos (zero-shot foundation model) as baseline.
  LightGBM with lag features for the high-volume SKUs.

Output Format

• Primary Recommendation: Model + rationale with expected performance range
• Alternatives: 2-3 options with trade-off analysis
• Baseline Strategy: Quick baseline (AutoGluon or simple model)
• Benchmark References: Papers/competitions proving the recommendation
• Starter Code: Production-ready implementation with proper CV

---

Key References

• Grinsztajn et al. (2022) — Why do tree-based models still outperform deep learning on tabular data?
• McElfresh et al. (2023) — When Do Neural Nets Outperform Boosted Trees on Tabular Data?
• Erickson et al. (2020) — AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data
• Das et al. (2024) — A Decoder-Only Foundation Model for Time-Series Forecasting (TimesFM)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ericwang915
• Source: ericwang915/data-scientist-skills
• License: MIT

Install and usage instructions live in the source repository linked above.