Sequential Recommender System for Video Recommendations

Role

AI Engineer

Keywords

Recommender Systems

Machine Learning

pytorch

API

FastAPI

Year

2026

← Home Sequential Recommender System for Video Recommendations Table of Contents About Introduction Objective Models Retriever A — Item2Vec Ranker A — LACLRec Combined B — RetrieverRankerRec Other Projects Let’s Work Together

About

Github

GitHub - maithtruong/recommender-systems-bootcamp: A collection of Recommender Systems for different tasks for the final project at Cole's Bootcamp.

A collection of Recommender Systems for different tasks for the final project at Cole's Bootcamp. - maithtruong/recommender-systems-bootcamp

https://github.com/maithtruong/recommender-systems-bootcamp

Tech Stack

Python

pytorch

FastAPI

Introduction

This project implements a production-style sequential recommendation pipeline on the
MovieLens dataset. The system follows the two-stage retrieve-then-rank architecture
standard in industry, where a fast retrieval model narrows the full item catalogue
to a candidate set, and a precise ranking model reorders those candidates for the
final recommendation.

Results (MovieLens-small, @K=10):

Model	Recall@10	NDCG@10
Retriever A (Item2Vec + FAISS)	0.0139	0.0064
Ranker A (LACLRec)	0.1125	0.0557
Combined B (Retriever → Ranker)	0.0769	0.0360

The standalone LACLRec ranker achieved the best performance.

Objective

The goal is to simulate a real-world production recommendation pipeline by implementing
every stage end-to-end:

Data preprocessing — convert raw interaction logs into padded user sequences

Retrieval model — Item2Vec (skip-gram) trained on interaction sequences

Approximate nearest neighbour search — FAISS index over item embeddings

Ranking model — LACLRec, a transformer trained with self-supervised contrastive learning

REST API serving — FastAPI server with endpoints for each model and the combined pipeline

Offline evaluation — Recall@K and NDCG@K with ablation across all three configurations

The pipeline covers all mandatory requirements from the bootcamp specification.
The optional LLM reranker stage is not implemented in this version.

Models

This project uses multiple model architectures to bechmark. They are described below.

Retriever A — Item2Vec

A Word2Vec model trained on user interaction sequences. Each item is treated as a
"word" and each user session as a "sentence". At inference time, the mean-pooled
embedding of the input sequence is used to query a FAISS index for fast approximate
nearest-neighbour retrieval.

Ranker A — LACLRec

A transformer-based sequential recommender trained with a self-supervised contrastive
learning objective. The model produces two augmented views of each sequence (one
learned, one random) and aligns their representations using NT-Xent loss, alongside
standard next-item cross-entropy loss.

Combined B — RetrieverRankerRec

A two-stage pipeline that composes Retriever A and Ranker A. No additional training
is required — the two models are loaded from their existing checkpoints and wired
together at inference time.