This project uses a deep learning neural network to predict the likes-to-views ratio of a YouTube video.
YouTube is one of the largest video-sharing platforms, where content creators rely on audience engagement metrics such as likes, views, and comments to measure their success. Deep learning can analyze large-scale YouTube metadata, video features, and historical engagement trends to build a predictive model that estimates the number of likes a video will receive. This can help creators optimize their content strategy, improve audience engagement, and increase monetization opportunities. This is also a useful tool for brands to analyze and strategically place their YouTube ads.
Below is a summary and inference derived after implementing, training, and testing the model:
- The YouTube Likes Predictor model was trained on a cleaned and engineered dataset that combined video metadata, text features, and engagement metrics.
- After applying TF-IDF on video titles, descriptions, and tags, and reducing dimensionality using PCA, the final FFNN model was trained using L1 loss.
- During training, the model’s validation loss gradually decreased, showing stable learning.
- The final model achieved good performance in predicting the like-to-view ratio, with the lowest average validation loss of 0.039 using the Adam optimizer.
- The addition of text features (via TF-IDF) and new categorical features like language presence (e.g., Arabic, Hindi) helped improve prediction accuracy.
- However, handling highly skewed engagement data remained challenging, especially for videos with extreme outliers in views or comments.