AAI_2025_Capstone_Chronicles_Combined

Cinema Analytics and Prediction System

27

An LSTM model with pre-trained GloVe embeddings captured sequential patterns in

descriptions, performing well on dominant genres but showing low precision and recall for

minority classes. Training curves indicated no overfitting,

but performance plateaued due to class imbalance and

limited data for rare genres.From the best training results,

we can see (Figure 26) how the validation loss curves

demonstrate the success of the final LSTM model

architecture, which combined GloVe embeddings,

Figure 26: LSTM Training vs Validation Loss

metadata features, and oversampling to address class

imbalance. Both losses decline steadily in parallel,

converging to approximately 0.40 (training) and 0.45 (validation), indicating effective

optimization with minimal overfitting. This smooth convergence validates the model’s design

choices: the frozen-then-fine-tuned embedding strategy maintained stable feature extraction

during early training, while dropout regularization and carefully tuned LSTM units (64/128)

prevented overfitting despite the model’s complexity.

From the following prediction table (Figure 27), it is clearly shown high accuracy for

animation/family films with multiple genres. On the other hand, systematic underprediction of

horror/thriller genres and over-simplification of crime-drama hybrids. The model also has

occasional severe misclassifications (e.g., comedy -> animation). Predictions are most reliable

for genres with strong visual/thematic stereotypes and least reliable for mood-dependent or

adult-oriented genres.

194

Internal