AAI_2025_Capstone_Chronicles_Combined

Cinema Analytics and Prediction System

24

Figure 23: LSTM-based movie clusters table

The results indicate that while the LSTM + GloVe model captures some genre and

stylistic groupings, it struggles with deeper semantic coherence, especially for hybrid films.

Using five clusters ensured consistency with other models but may not fully reflect the

LSTM embedding structure. The project confirmed that transformer-based embeddings like

BERT outperform TF-IDF in capturing nuanced semantics. LSTM + GloVe adds supervised insight

but requires careful tuning. For deployment, BERT offers the best balance of robustness and

quality. Future work could fine-tune BERT on movie-specific data, combine embeddings with

user metadata, implement scalable embedding and fast similarity search (e.g., FAISS), and

integrate genre classification to enhance personalization. This work lays a strong foundation for

semantic-aware, personalized movie recommendation systems.

Genre classification

The genre classification task involved multi-label movie genre prediction from text.

Models tested included logistic regression with BERT embeddings, an LSTM with GloVe

191

Internal