Limited time offer 20% OFF StudySoup Subscription details

# Georgia Tech - CS 7646 - Class Notes - Week 3

### Created by: Monica Notetaker Elite Notetaker

> > > > Georgia Tech - CS 7646 - Class Notes - Week 3

Georgia Tech - CS 7646 - Class Notes - Week 3

This preview shows pages 1 - 3 of a 3 page document. to view the rest of the content
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner

This is the end of the preview. Please to view the rest of the content
Join more than 18,000+ college students at Georgia Institute of Technology who use StudySoup to get ahead
##### Description: 03-3 Assessing a Learning Algorithm 03-4 Ensemble Learners, Bagging and Boosting Decision Tree
3 Pages 59 Views 47 Unlocks
• Notes, Study Guides, Flashcards + More!

Unformatted text preview: 103-31 Assessing & Learning 21gorithm ( * 1 Assessing KNN: graph of scatter points gete mapped_to a cuve Hal chu xudden diepe het loole Uke steer A graph of KNW with Ka! will paul houch every single data point, there by will overtit -- A ce increase k eve are more likely to undert - Assessing Polynomial Modek : A graph of scatter pointe gek mapped to a smooth A trece graph of polynomial with d= / a straight line and may not fit as many points > As we increuve d we are more likely to over 17 Meme 1 : RMS Error Error can be measured as the absoluted digerence b/w the y value of the data point the model. RMSE EC Yteet - Uporedict) (Root Mean Squared Error) We generate thic model by kaining it on a kaining data set RMCE erabated on a testing data set will war ke aside, and not used while Kaining Thi RMKE is called out of cornpic root mean gured enor ++ We would expect out-of-sample essor to be lacgel, since the model has not seen point from test set s hmce may not* Cross Validation Data 7 Brenerally Train Text 60% 40%. for cross validato , test on 2011 - we kein on 80% date by sliding the data into 5 chunki - that then we Lolare these chunks such eath the chunk becomes the text dala at least one. And in 5 different tiak. Roll-forward Core Validation Croes Validation does to financial applications hey will as Kaining on Fuhure data and testing on hast dale ne essentially berrils beeking into the future This leads to unsealishe optimistic resulte htell- forward Ceass Validation engines that ou training data is always before our feeling dala We still hare multiple kale by Rolling out dala Forward. #71 we run out of data. Mekic & Cerrrelation => We can also che aves a model by observing the relationship between predicted and serial valua OLY this properly i measured quantitatively by using correlationIn most cases, as PMS error increases, correlation gou down * Overfitting Increase the degree of freedom As we let fine tune the model to fit the data Mort and more, we see that the in-cample and the out-of-comple rms pors botts Keep reducing te till a cirtain point. After thic point's the in-sample exor_continues to drop but the text ere tenets sting After this perint , hence, the model there is everfitting the data | Comparing KNM_&_Linear regression Perfosms beher + Space for saying_model - Linear Rey, > Compute time to train KNN - Compute time to query__- Linear Reg. > Ease to odd new data - KNN

×

I don't want to reset my password

Need help? Contact support

Need an Account? Is not associated with an account
We're here to help