What if BlockBuster Had ML to Help Them Stock Their Shelves…?
Inspirit AI Project Presentation
2023-08-17
The Problem
Movie Genre Classification
Can we help create a fully automated blockbuster by classifying new movies into the correct genre section based only on their description?
We’ll use a data set of 54214 movies with their:
- genre (label)
- descriptions (input)
Custom-trained Word2Vec embeddings performed better than pretrained models
Embedding Models Used:
- Glove pre-trained “Wiki-gigaword 300” model (“glove”)
- Spacy pre-trained word2vec model (“SpCyw2v”)
- Word2Vec model trained on my movie description data (“myw2v”)
Baseline Models: Logistic Regression
Validation Accuracy
Over/under fitting depending on model complexity
Accuracy 
Next, I tried a series of models of increasing complexity
Note
Simpler, linear models underfit the data
More complex models overfit
Without further tweaking the simple models won out on validation set accuraacy for now
Shortening embedding vector lengths helped some models
I recreated the embeddings using vectors of length 100 instead of the original 300…which seem to help to increase validation accuracy for some models like RandomForest
What I’d do next:
Perhaps the overfitting can be overcome by full-scale hyperparameter tuning?