Skip to content
All projects
AI/ML+ Data Analytics Published · ICAIII 2026

GenAI-Based Recommendation System

Modular Python backend for generating personalized recommendations, with data pipelines for cleaning, transformation, and building the user-item matrix. Ranking runs as a Learning-to-Rank ensemble over SBERT, ALS, and LightGBM, behind clean service interfaces a frontend can consume.

Key result

LTR ensemble— SBERT + ALS + LightGBM behind a service interface

01Problem

Recommenders tuned only for accuracy converge on the same popular items and give the user no reason to trust what they are shown. Making trust part of the ranking, rather than an afterthought, changes what the system needs to compute.

02Objective

Build a modular recommendation backend where ranking is a trust-aware Learning-to-Rank problem, and expose it behind service interfaces a frontend can consume rather than leaving it as research code.

03Architecture

  1. Amazon Dataset
  2. Data Processing
  3. SBERT
  4. ALS
  5. LightGBM
  6. Trust & Safety Filter
  7. Generative Explanation
Multi-stage retrieval, ranking, and generation pipeline

04Technology

  • Python
  • SBERT
  • ALS
  • LightGBM
  • Pandas
  • scikit-learn

05Implementation

Key technical decisions

  1. 01Built the backend as separate modules — data pipeline, embedding, ranking, serving — rather than one script.
  2. 02Combined SBERT semantic embeddings with ALS collaborative filtering so the system has both a content and a behavioral signal.
  3. 03Ran final ranking as a LightGBM Learning-to-Rank ensemble over those signals.
  4. 04Exposed the result behind clean service interfaces a frontend can consume.
  5. 05Measured the quality/performance trade-offs between approaches rather than picking one on intuition.

What was built

  • Data pipelines for cleaning, transformation, and building the user-item matrix.
  • SBERT embeddings over product text.
  • ALS collaborative filtering over the interaction matrix.
  • LightGBM Learning-to-Rank ensemble producing the final ordering.
  • Service interfaces exposing recommendations to a frontend.

06Data

Source corpus and preprocessing

  • Amazon product and review dataset
  • Cleaning and transformation pipeline ahead of modeling
  • User-item matrix built from the processed interactions
  • Product text encoded as SBERT embeddings

07Results

Modular

Pipeline, ranking, and serving separated

3 signals

SBERT + ALS + LightGBM ensemble

ICAIII 2026

Peer-reviewed and published

Evaluated on

  • Hit Rate@5
  • Hit Rate@10
  • Diversity
  • Novelty
  • Ranking latency
  • LLM explanation latency

Measured figures are reported in the published paper.

10Links

Publication (DOI)