Skip to content

SYSTEM Cited by 2 sources

LightGBM

LightGBM (Light Gradient Boosting Machine) is Microsoft Research's open-source gradient-boosting-on-decision-trees framework. Known for speed (histogram-based algorithms) and low memory footprint relative to XGBoost / CatBoost — a default choice for tabular ML and increasingly for time-series forecasting via wrappers like Nixtla MLForecast.

Stub page — minimum viable framing. Expand as deeper LightGBM internals are ingested.

Why Zalando picked it over deep learning

From the 2025-06-29 ZEOS post:

"After extensive experimentation with deep learning models like TFT and other machine learning approaches, we selected the LightGBM model integrated with Nixtla's MLForecast interface as the foundation of our demand forecasting pipeline. This stack enables significant advantages, including high-level abstractions for time series-specific feature generation with optimised performance, rapid prototyping through shorter feedback loops, and access to a robust, well-maintained open-source ecosystem."

And the ops payoff:

"Due to the ML model's lightweight training footprint, we bypass complexity, like for example not needing checkpointing, or separate infrastructure for inference."

Seen in

  • sources/2026-10-02-databricks-real-time-retail-intelligence-building-e-commerce-recommenda — the ranking/scoring model in an e-commerce recsys funnel. Databricks' 2026-10-02 retail reference architecture uses LightGBM as Stage 2 of the retrieval → ranking funnel: after AI Search returns 200–500 candidates, the endpoint builds per-pair feature vectors (user features + item features + request-derived context + cross features like user-category affinity and brand overlap) and LightGBM "predicts conversion probability in a single batch inference call without GPU resources." The gradient-boosted-trees-over-DL choice echoes Zalando's: cheap, fast, CPU-only tabular scoring is the right tool when features are structured — here it keeps the real-time serving path GPU-free while hitting a low-2-digit-ms budget. Trained/versioned via MLflow, retrained weekly under champion/challenger.

  • sources/2025-06-29-zalando-building-a-dynamic-inventory-optimisation-system-a-deep-dive — canonical wiki instance for time-series forecasting at scale. Trained with conformal inference via Nixtla MLForecast to produce 12-week probabilistic forecasts for 5M SKUs in under 2 hours end-to-end. The lightweight footprint drove a single-training-job-train-and-infer architectural collapse — LightGBM is the reason Zalando can skip SageMaker hosting endpoints.

Last updated · 766 distilled / 2,225 read