I have just finished the Predicting Bike Rentals guided project and I would like to get your feedback.
I have used RMSE, R Square, and the difference between training error and validation error to figure out which algorithm predicts and fits better. I am not sure if my analysis and result are correct so I eagerly waiting for your feedback.
I said in my conclusion: “Among the above models Random Forest with min_samples_leaf = 6 and max_depth = 8 has shown the best result.” Do you agree?
I have used this forum post a lot and I highly appreciate it.
This is the link to my project.
Hello @jafarinasim, thanks for sharing your work. I’ve carefully gone through your work, I’ve very impressed with how you’ve presented your work. It is easy to follow along, easy to understand what you are doing. Everything is awesome.
It would be great if you do Hyperparameter tuning to arrive at the best model parameters.
Thank you @info.victoromondi for your feedback and also good resources. I read the tutorial and it is really helpful. Thanks again.
Hi @jafarinasim, thanks for sharing your project. It was very interesting to see your approach to the problem.
I have one question for you. I’m not sure how the R Square error works.
Did you learn about it in this course? If so, where? I might have missed it.
I’d appreciate any info about how it works.
Hi @jafarinasim, your work looks great. I was wondering about columns with
< 0.3 correlation coefficient. I recall previous lessons had shown us that excluding
weakly correlated features can actually improve model performance. But in this case, my model actually did worst without them. Do you see the same issue on your end?
Thanks for sharing this article. It was very helpful for me (granted, after reading it around 3 times and experimenting with Predicting Bike Rentals for a whil).
I’d like to ask you something about it.
During the article the author analyzes the evolution of the MSE as the size of the training set increases.
My question is: would it be as appropriate to come to conclusions about bias and variance if we modified another variable of the model?
Let me explain myself.
Let’s say I train my model using a different number of features each time (i.e. [1,3,5,7,9]). Or a different number of k neighbors each time.
With the results I get I’d be able to build similar plots to the ones in the article.
But, does this count as analyzing the variance and bias as well?
Or, in order to achieve that I have to do it just like in the article, analyzing the performance over different train set sizes?
I hope the question makes sense.
Thanks so much for your help!
Hi @scoodood, thanks for sharing with me your idea and also the good question. Yes, you are right the errors are increased by excluding the features with <0.3 correlation coefficient. It seems the heatmap shows the linear relationships so excluding these features is suitable for the linear model, not non-linear classifiers. Please have a look at this.
After I looked at this project again I figured out that
- By creating ‘time_lable’ feature, it is better to remove ‘hr’.
- Scaling features may also be helpful.
Hi @jafarinasim, thanks for the link, it was a great read. Thanks