Home / Research

Research that survives real data, real constraints and real evaluation.

We develop recommendation, prediction and intelligent modelling systems for environments where users interact across multiple platforms, content categories and time windows. Our approach combines graph learning, sequential modelling, tabular learning, hybrid architectures and rigorous experiment design.

Research areas

Four directions, one system view.

01

Cross-Platform Recommendation & User Interest Modelling

Modern users move across websites, social platforms, marketplaces, apps and marketing channels. Their behaviour is not a static profile — it changes with time, context, content exposure, device and recent activity. We build models that represent both relational structure and temporal behaviour, treating users, content, categories, platforms and actions as a dynamic system rather than isolated events.

Interest evolutionTop-K predictionCold-start users & contentCross-platform adaptersCohort ranking analysis
02

Graph Neural Networks & Relationship Modelling

When data contains relationships among users, content, categories and platforms, graph modelling can reveal signals flat tables miss. We study graph-based recommendation methods such as LightGCN, GraphSAGE and GAT, plus hybrid graph-and-sequence architectures. The real challenge is not only whether graph models improve a metric — it is whether graph construction, edge filtering, neighbour sampling and memory use stay manageable as data grows.

LightGCNGraphSAGEGATNeighbour samplingDynamic graphs
03

Sequential & Transformer-Based Behaviour Modelling

User behaviour is time-dependent. A recent add-to-cart may matter more than a click from months ago; a sequence of category views can reveal intent before conversion. We evaluate sequential models such as GRU4Rec and SASRec to understand short-term preference shifts — and the limitations of sequence-only modelling.

GRU4RecSASRecShort-sequence usersTemporal interest shift
04

Hybrid AI Systems

The most useful system is often not one model, but a carefully designed hybrid. We combine graph embeddings with Transformer encoders, dynamic graph structures with time-decayed edges, and mixture-of-experts fusion across graph, sequence and tabular branches — then test whether the added complexity is justified by stability and interpretability.

GNN + TransformerTime-decayed edgesMixture-of-expertsAblation-driven
Model stack

We compare model families — not bet on one architecture.

Each complex model must earn its place against simpler, interpretable alternatives. That's what makes the research credible.

Collaborative Filtering

BPR-MF

A latent-factor reference point for user-content preference learning.

Tabular Learning

LightGBM

Tests whether structured business features explain the target without deep sequence or graph modelling.

Feature Interaction

DeepFM

Evaluates sparse feature crosses and high-order feature combinations.

Sequential Modelling

GRU4Rec · SASRec

Test short-term behaviour sequences and Transformer-style ranking.

Graph Learning

LightGCN · GraphSAGE-GAT

Test relationship structure across users, items, categories and platforms.

Hybrid Architecture

GNN-Transformer · MoE

Combine graph, sequence and tabular signals into a single research framework.

Metrics that matter

No single number decides a model.

Business, data and engineering teams each need different evidence. We evaluate every model from several angles.

AUC · classification LogLoss · probability quality NDCG@10 · ranking Recall@20 Precision@K HitRate@K MRR · first relevant Platform-level metrics Cohort analysis Cold-start performance GPU memory · feasibility Training time
Research & delivery method

A process built for traceability and practical decisions.

01 / Frame

Problem & scope

Clarify the business problem, data fields, anonymisation boundaries and evaluation requirements.

02 / Baseline

Compare fairly

Implement baselines, sanity-test, then design the hybrid architecture and experiment plan.

03 / Validate

Ablation & cohorts

Parameter search, ablation testing, platform/cohort and cold-start evaluation with reproducibility checks.

04 / Deliver

Report & prototype

Technical reports and working prototypes reviewable by both technical and business stakeholders.

We don't begin by assuming the most complex model is best. We compare simpler baselines, isolate each source of improvement, and examine trade-offs — using data versions, code commits, configuration files and logs so results stay reproducible.

Work with us

Have a modelling question that doesn't fit standard tools?

We can help evaluate the technical path before you commit to a direction.

Start a conversation →