TrendVerse 📈

An end-to-end MLOps-powered fashion trend forecasting and recommendation platform that combines time-series forecasting with multimodal recommendation systems to identify emerging fashion trends and deliver personalized recommendations.

Built as a university project with a strong emphasis on implementing a complete machine learning lifecycle using modern MLOps tools and practices.

Overview

TrendVerse analyzes historical fashion trend data to forecast future trends using the Prophet forecasting model. In addition, it leverages CLIP embeddings generated from Instagram fashion page metadata to build a personalized recommendation system based on user interests.

The project demonstrates the integration of machine learning, workflow orchestration, experiment tracking, data versioning, containerization, and cloud deployment within a reproducible MLOps pipeline.

Key Features

Fashion Trend Forecasting

• Forecast future fashion trends from historical trend datasets.
• Time-series analysis using Prophet.
• Automated trend prediction pipeline.

Personalized Recommendations

• Recommendation engine based on user interests.
• CLIP-powered embedding generation from Instagram fashion page metadata.
• Similarity-based recommendation retrieval.

End-to-End MLOps Pipeline

• Automated workflows using Apache Airflow.
• Experiment tracking with MLflow.
• Dataset and model versioning using DVC.
• Centralized project management through DagsHub.
• Containerized services with Docker.
• Cloud deployment on AWS EC2.

Tech Stack

Machine Learning

• Prophet
• CLIP
• Scikit-learn
• Pandas
• NumPy

MLOps & Infrastructure

• Apache Airflow
• MLflow
• DVC
• DagsHub
• Docker
• Git
• GitHub
• AWS EC2

Backend

• Flask

Project Architecture

TrendVerse
│
├── airflow/          # Workflow orchestration pipelines
├── data/             # Datasets and processed data
├── flask/            # Flask application
├── model/            # Saved models and artifacts
├── mlproject/        # MLflow project and experiment tracking
├── requirements.txt
└── ...

Machine Learning Pipeline

Trend Forecasting Pipeline

1. Load historical fashion trend datasets.
2. Preprocess and clean trend data.
3. Train Prophet forecasting models.
4. Generate future trend predictions.
5. Store and version forecasting artifacts.

Recommendation Pipeline

1. Collect Instagram fashion page metadata.
2. Generate CLIP embeddings.
3. Compute similarity between user interests and content embeddings.
4. Retrieve and rank recommendations.
5. Serve recommendations through the Flask application.

MLOps Workflow

The project was designed around modern MLOps principles:

• Data versioning using DVC
• Experiment tracking using MLflow
• Pipeline orchestration using Apache Airflow
• Reproducibility through Docker containers
• Collaborative development using Git and GitHub
• Artifact and repository management through DagsHub
• Deployment on AWS EC2

Dataset

The project utilizes publicly available fashion and trend-related datasets sourced from Kaggle.

Deployment

The application was deployed on AWS EC2 and containerized using Docker as part of the end-to-end MLOps workflow.

Learning Outcomes

This project provided hands-on experience with:

• Building production-oriented machine learning pipelines
• Time-series forecasting with Prophet
• Multimodal representation learning using CLIP
• Recommendation system development
• Workflow orchestration with Apache Airflow
• Experiment tracking and reproducibility
• Data and model versioning
• Cloud deployment and infrastructure management
• End-to-end MLOps practices

Repository

DagsHub Repository:

https://dagshub.com/l230915/trendverse

Future Improvements

• Real-time trend ingestion from social media platforms
• Enhanced recommendation ranking models
• Interactive analytics dashboard
• Automated CI/CD pipelines
• Advanced deep learning forecasting models
• User feedback-driven recommendation refinement