GNN+Transformer
Hybrid architecture research
PWB Research builds recommendation, prediction and intelligent-decision systems for complex, multi-platform user behaviour — backed by H200 and B300-class GPU infrastructure so proper validation is actually possible.
We work closer to an applied research partner: we define the technical question, design the experiment, compare candidate models, and convert validated findings into usable systems.
Model design, experimental validation, ablation studies and technical reporting.
Explore research →Cross-platform ranking, user-interest prediction, cold-start handling and Top-K output.
See the project →H200 and B300-class resources for graph learning, Transformers and large parameter search.
View compute →Player modelling, matchmaking, churn prediction and in-game recommendation.
AI for games →A user may browse on one platform, save content on another, enquire through a campaign form and convert through a different channel. These signals are connected — but traditional systems treat them separately.
We model the relationship between users, content, platforms, categories and events — and study how those relationships evolve over time.
Every complex model is tested against simpler, interpretable baselines. We isolate each source of improvement before recommending a path.
Findings become interest scores, Top-K recommendations, confidence values, offline scoring scripts and API prototypes.
A GNN-Transformer hybrid that predicts how user interest drifts across platforms and time — outputting interest scores, Top-K preferences and confidence values under realistic, sparse-data constraints.
Read the project →User-item-category-platform edges.
Recent action windows & drift.
Structured signals & conversion.
Interest score + Top-K + confidence.
If your problem involves cross-platform user behaviour, personalisation, recommendation, ranking or prediction — we can turn it into a research plan, and the plan into a working technical path.
Start a conversation →