Home / Projects

Cross-Platform User Interest Evolution Prediction.

A GNN-Transformer hybrid that models anonymised user behaviour sequences and user-content-platform relationships, then outputs interest scores, Top-K content/category preferences and confidence values — under realistic constraints: distributed data, inconsistent event definitions, sparse categories and short interaction histories.

Architecture

Six layers, from raw events to usable output.

GNN-Transformer Hybrid

Predicting how user interest drifts across platforms and time.

The system processes anonymised behaviour events and relationship structure, then produces business-readable outputs a downstream team can act on.

01 Behaviour Events
What it handles
Browsing, clicking, saving, add-to-cart, enquiries, orders, timestamps.
Why it matters
Captures observable user intent and conversion signals.
02 Relationship Graph
What it handles
User-item, user-category, user-platform and content-category edges.
Why it matters
Represents structure that sequence-only models miss.
03 Sequential Context
What it handles
Recent behaviour windows, order of actions, temporal interest shifts.
Why it matters
Separates historical preference from current intent.
04 Platform Context
What it handles
Platform identity, channel differences and platform-level adapters.
Why it matters
Prevents one platform's pattern from dominating predictions.
05 Hybrid Modelling
What it handles
Graph embeddings, Transformer encoders and gated mixture-of-experts fusion.
Why it matters
Combines relationship, temporal and tabular signals in one framework.
06 Output Layer
What it handles
Interest score, Top-K categories, confidence and offline/API outputs.
Why it matters
Makes findings usable by downstream business and product teams.
Project detail

From data inputs to delivered package.

ComponentProject detailPurpose
Data InputsAnonymised browsing, click, save, add-to-cart, enquiry and order behaviour; content category; timestamp; campaign source; device type; conversion flag.Represent the multi-platform behaviour environment without relying on personally identifiable data.
Baseline ModelsBPR-MF, LightGBM, DeepFM, GRU4Rec, SASRec, LightGCN and GraphSAGE-GAT.Create fair reference points before evaluating the hybrid architecture.
Hybrid ModelsGNN embedding + Transformer encoder; dynamic graph Transformer; gated mixture-of-experts fusion.Combine graph relationship signals, sequence dynamics and tabular business features.
EvaluationAUC, LogLoss, NDCG@10, Recall@20, Precision@K, MRR, HitRate@K, cohort metrics, cold-start performance, training time and GPU memory.Avoid relying on a single number and evaluate stability across user groups.
OutputsInterest score, Top-K categories, confidence value, offline batch scoring script, API prototype, logs and technical report.Translate research findings into a usable technical delivery package.
Case study

Seven baselines. Three hybrids. One disciplined comparison.

A client-facing algorithm research project focused on predicting evolving user interests from anonymised multi-platform behaviour data. It investigates seven baseline models and three GNN-Transformer hybrid approaches, using time-based validation, ablation studies and platform/cohort analysis.

The work includes data processing scripts, synthetic data generation, public-data sanity checks, baseline code, hybrid model experiments, result tables, API prototype design, technical reporting and consultation on future deployment. It reflects the type of work PWB Research is built for: technically uncertain, data-dependent, compute-intensive AI research that needs careful experimentation before production decisions are made.

Have a similar problem?

Multi-platform behaviour data that doesn't fit standard tools?

We can turn it into a research plan and a working technical path.

Start a conversation →