01
User Preference Analysis & Dynamic Modelling
Builds a profile from onboarding, updates it from views, saves, selections and dislikes, and expands it with a multi-label Logistic Regression model.
Ramanayake A.R.M.C.N.K
IT4010 Research Project 2026
Smart Fashion Assistant: An Intelligent Personalized and Trend-Aware Fashion Recommendation System
OutfitIQ
A multi-engine intelligent fashion recommendation system with dynamic preference learning and trend awareness.
Online fashion has no shortage of clothes. The hard part is choosing. OutfitIQ learns what each user likes, follows what the Sri Lankan market is moving towards, and turns a single garment into a complete, compatible outfit, all in one mobile app.
The four intelligent enginesR26-IT-052
-
User Preference LearningOnboarding + interactions + ML expansion01
-
Hybrid RecommendationStructured match + semantic ranking02
-
Trend AnalysisMRTF + GRU forecasting03
-
Outfit GenerationRules + Transformer compatibility04
Recommendations, outfits and trend insights in one mobile app
85.29%
Semantic ranking accuracy after fashion fine-tuning (up from 71.60%)
0.4276
Trend forecast Information Coefficient across 95 temporal cutoffs
0.728
Outfit compatibility ROC-AUC on the Polyvore test set
20
Sri Lankan fashion retailers observed daily
The problem
Choosing clothes online is still hard
The same kind of garment shows up across many retailers, each described with different categories, styles and colours. A user's taste changes over time, new users have no history at all, and a product that matches someone's taste may not work with the rest of their outfit.
Existing work treats personalisation, trend forecasting and outfit compatibility as separate problems.
Our approach
Four engines, one workflow
OutfitIQ keeps four kinds of intelligence separate, then connects them. It asks what this user is likely to prefer, which products match that taste, which fashion attributes are gaining market interest, and which garments work together as a complete look.
Read the methodology
System components
Meet the team
The intelligent engines
Each engine is owned by one member of the research team.
Ramanayake A.R.M.C.N.K
02
Product Data Collection & Hybrid Recommendation
Normalises a multi-retailer catalogue and ranks products with structured matching, a fine-tuned SentenceTransformer, and a product-quality score.
Gunathilake T.M.P.G.K.N
03
Trend Analysis
Combines the Market Restock–Turnover Factor (MRTF) with a GRU-based TrendNet to score which fashion attributes are rising in the market.
G.D.Elladeniya
04
Outfit Compatibility & Generation
Completes an outfit from one selected garment using interpretable fashion rules and an Outfit Transformer over cached FashionCLIP embeddings.
Rajapaksha P.D.S.S
Architecture
How the system fits together
User signals, retailer catalogues and market observations flow through the four engines into recommendations, outfits and trend insights in the mobile app.
Project resources