Key Steps

1. Web Scraping:

   • Objective: Scrape movie metadata, including titles, genres, and poster image URLs, from IMDb or TMDb.

   • Implementation:

     • Used BeautifulSoup and requests to extract data such as titles and genres.

     • Downloaded poster images for vectorization and analysis.

   • Outcome: Collected a comprehensive dataset of movie metadata and poster images for further processing.

2. Data Storage:

   • Objective: Organize metadata and poster embeddings into a structured SQL database.

   • Implementation:

     • Designed a schema with three main tables:

       • Movies: Includes movie titles and poster images.

       • Genres: Stores unique genre names.

       • Movie-Genres Relationship: Links movies to their respective genres.

     • Stored poster embeddings (256-dimensional feature vectors) in a separate vectors table for efficient retrieval.

   • Outcome: A scalable database for storing and querying metadata and embeddings.

3. Image Vectorization:

   • Objective: Convert poster images into embeddings that capture visual features.

   • Implementation:

     • Fine-tuned a pre-trained ResNet34 model to generate 256-dimensional embeddings.

     • Incorporated data augmentation (resizing, normalization) to improve generalization.

   • Outcome: Generated embeddings for all poster images, representing them as high-dimensional feature vectors.

4. Genre Classification:

   • Objective: Train a neural network to predict movie genres based on poster embeddings.

   • Implementation:

     • Used a multi-label classification setup, as movies often belong to multiple genres.

     • Handled class imbalance by computing and applying appropriate weights during training.

     • Evaluated performance using metrics such as F1-score, precision, and recall.

   • Outcome: Developed a model capable of predicting genres with high accuracy.

5. Vector Search:

   • Objective: Perform similarity searches to classify or recommend movies based on poster vectors.

   • Implementation:

     • Stored embeddings in the SQL database for efficient querying.

     • Calculated similarities using Euclidean distance between vectors.

   • Outcome: Enabled genre classification and recommendations for new posters by querying the vector database.

Skills Demonstrated

• Web Scraping: Extracted structured data from unstructured web pages using BeautifulSoup and requests.

• Database Management: Designed and implemented a relational database schema, handling binary data storage for images.

• Deep Learning:

  • Fine-tuned a CNN for feature extraction and genre classification.

  • Implemented custom layers to generate embeddings for vector search.

• Data Engineering:

  • Managed large datasets of images and metadata.

  • Performed data augmentation and preprocessing to ensure model robustness.

• Machine Learning: Used metrics like F1-score and recall to assess multi-label classification performance.

• Optimization: Leveraged GPU acceleration for model training and batch processing for efficiency.

Key Outcomes

• Created a scalable pipeline for scraping, storing, and analyzing movie posters and metadata.

• Built a multi-label classification model for genre prediction with robust performance.

• Enabled vector-based similarity search for recommendations or classification of new posters.

Next Steps

• Expand Dataset: Include more diverse movies and genres to improve model generalization.

• Enhance Search: Integrate more advanced vector search frameworks like FAISS or Milvus for scalability.

• Deploy Solution: Build an interactive web interface for users to upload posters and get genre recommendations.

This project illustrates a practical application of machine learning in the entertainment domain, showcasing the potential of integrating data engineering with AI solutions.