Ecommerce ops · Production

Trendyol ensembles Mixture-of-Experts neural networks and LTR GBDTs for e-commerce search ranking

The problem

Trendyol's single-model ranking architecture could not simultaneously capture both engagement-specific behavioral signals and high-precision tabular features, leaving gaps that neither model family alone could fill.

Workflow diagram · grounded in source
1
User search triggers pipeline
trigger
“our search journey is the primary bridge between millions of users and an ever-changing catalog”
2
MoE NN scores engagement signals
ai_action
“MoE handles the engagement specific signals of the journey”
3
LTR GBDT scores tabular features
ai_action
“LTR GBDT enforces the learned hierarchial scoring by exploiting the structured data input”
4
Async concurrent model execution
integration
“we fire requests to both NN and GBDT models concurrently with leveraging the native capabilities of Torch and Triton. The total latency effectively becomes the latency of the slowest model, rather than the sum”
5
Normalize scores to percentile ranks
ai_action
“we normalize them into percentile ranks before blending. This ensures one model doesn't drown out the other due to scale differences”
6
Rank blend produces final order
output
“Rank Blend: Mean of percentile ranks from both models”
Reported outcome

The rank-blend mean ensemble was the clear winner in A/B tests across all platforms, delivering statistically significant uplifts across key user actions and downstream conversion objectives.

Reported metrics
Core metric improvementsconsistent improvements across our core metrics
User actions and conversion upliftstatistically significant uplifts across key user actions and downstream conversion objectives
Reported stack
Mixture-of-Experts NNsLearning-to-Rank GBDTsTorchTriton
Source
https://medium.com/trendyol-tech/better-together-ensembling-neural-networks-and-gbdts-for-e-commerce-search-ranking-505a9b05bf19
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The rank-blend mean ensemble was the clear winner in A/B tests across all platforms, delivering statistically significant uplifts across key user actions and downstream conversion objectives.

What tools did this team use?

Mixture-of-Experts NNs, Learning-to-Rank GBDTs, Torch, Triton.

What results were reported?

Core metric improvements: consistent improvements across our core metrics; User actions and conversion uplift: statistically significant uplifts across key user actions and downstream conversion objectives (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

User search triggers pipeline → MoE NN scores engagement signals → LTR GBDT scores tabular features → Async concurrent model execution → Normalize scores to percentile ranks → Rank blend produces final order.

Related ecommerce ops cases