The feature selection cleaning was well thought out. Features were left out if they were believed to not include relevant information for movie rating prediction or were believed to skew the distribution. In one such feature selection, the crew was deemed to probably not have an impact on the movie's rating. I imagine that there is a possibility that the crew might have an impact on the movie's rating, so it would be interesting to compare models that include this feature to models that do not, with all other factors held equal. This could also have been analyzed by preliminary data analysis by directly looking at the correlation between the movie rating and this feature.
I would expect to have a high impact in the prediction of movie rating, and the method to transform the actors feature was creative in that creation of new features included selecting the top actors and considering frequency of actors, which is a good way to encode the information without using a one-
hot encoding transformation that may prove uninformative given the high amount of different actors across the different movies.
The project implements a variety of models and compares the results. The models selected were well chosen in that both decision trees and regression models were implemented, which have different behaviors in prediction learning. For the decision tree model, it is mentioned that the model tended to overfit the data, which is typical of decision trees. The group thus implements random forests to remedy the variance of the model. It would have been interesting to also try an ordinary decision tree model with regularization via pruning or setting a threshold cutoff of splits or final nodes created, and comparing the results against the random forest model. Overall, the models created illustrate a good understanding and implementation of the concepts taught in lecture. It is interesting and makes sense that the decision tree model outperforms the SVM models, as I intuitively believe movie rating to be a problem that would be learned well by a decision tree approach.
Good work and analysis!
The feature selection cleaning was well thought out. Features were left out if they were believed to not include relevant information for movie rating prediction or were believed to skew the distribution. In one such feature selection, the crew was deemed to probably not have an impact on the movie's rating. I imagine that there is a possibility that the crew might have an impact on the movie's rating, so it would be interesting to compare models that include this feature to models that do not, with all other factors held equal. This could also have been analyzed by preliminary data analysis by directly looking at the correlation between the movie rating and this feature.
I would expect to have a high impact in the prediction of movie rating, and the method to transform the actors feature was creative in that creation of new features included selecting the top actors and considering frequency of actors, which is a good way to encode the information without using a one-
hot encoding transformation that may prove uninformative given the high amount of different actors across the different movies.
The project implements a variety of models and compares the results. The models selected were well chosen in that both decision trees and regression models were implemented, which have different behaviors in prediction learning. For the decision tree model, it is mentioned that the model tended to overfit the data, which is typical of decision trees. The group thus implements random forests to remedy the variance of the model. It would have been interesting to also try an ordinary decision tree model with regularization via pruning or setting a threshold cutoff of splits or final nodes created, and comparing the results against the random forest model. Overall, the models created illustrate a good understanding and implementation of the concepts taught in lecture. It is interesting and makes sense that the decision tree model outperforms the SVM models, as I intuitively believe movie rating to be a problem that would be learned well by a decision tree approach.
Good work and analysis!