Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/pandaiscodingcpu/codeai_eda-and-ml-task

Repo for CodeAI task
https://github.com/pandaiscodingcpu/codeai_eda-and-ml-task

Last synced: 7 months ago
JSON representation

Repo for CodeAI task

Awesome Lists containing this project

README 

          
# CodeAI_EDA-and-ML-task  



## UPDATE: Unusual R2 for added predictions fixed!!  

***In the previous version, two seperate splits were used for `casual` predictions and `registered` predictions and when the total predictions was calculated it showed R2~1, to fix this issue, this time only one split was taken and based on that `alpha = 1.0` is kept for regularization, which resulted in R2 for added predictions to be 0.82297611687786.***



Repo for CodeAI task  



`dataset.csv`: this is the main dataset provided to us    



`primary-eda`: this file contains the fundamental EDA done on the dataset.    



`dataset02.csv` : this is the second dataset with which regression model is trained.    



`regression_model`: this is the file in which regression model is trained and compared  



`app.py`: this file contains the code for streamlit app, run `streamlit run app.py` in terminal.  



`Copy of Collab Notebook`: this file contains the solution to Tech AI Task 1



# Models Used  



`LinearRegression` : R2 Score: 0.8565676064919518 (while predicting directly `cnt`)



`RidgeRegression` : R2 Score:  0.8565611384382965 (while predicting directly `cnt`)



`LassoRegression` : R2 score:  0.8564998234510169 (while predicting directly `cnt`)



> **NOTE:** Colums like `instant`, `dteday`, `registered`, `casaul` were dropped during training, as they were giving unusual cross validation scoress and R2 scores



# Reason For model choice  

For the task. Linear Regression was used and then it was regularised using Ridge Regression, the reason behind Ridge regression is:-  

1. Ridge regression minimizes the cost function MSE(theta) by adding a regularization term

2. It reduces the coefficients for least related column with the target value

3. Before using Ridge Regression, the dataset were scaled as, these regularized models are sensitive to scale of input features

4. Also, the dataset is comparitively smaller with 731 rows and 16 columns, so using other regressors like RandomForest or GradientBoosting would increase the chances of model being overfit (though this can be avoided by fine tuning the hyperparameters)

5. Ridge Regression is chosen because it is better for small dataset like this and easy to implement.

6. Despite the R2 score of Linear Regression being more than RidgeRegression while directly predicting the `cnt`, Still RidgeRegression was chosen because it has Cross Validation score of 0.61311 where as Linear Regression has 0.60023



# Tuning Performed on Hyperparameters like alpha in case of RidgeRegression  

It is important to select perfect value of alpha so that the model performs well, so for that RidgeCV was performed on a list of alpha values ranging from 2.1 to 4.0.  



# Brownie Points  

1. Initially, the Ridge Regression model predict directly the `cnt` values with R2 score of 0.85656114 with `alpha = 2.7` and using `solver = cholesky`.

2. Then, two seperate Ridge models were created to predict the `casual ` rents and `registered` rents, the casual model gave R2 of 0.78936634 and the registered model gave R2 of 0.97004162.

3. Then using the predictions from both these models, `cnt_predictions` series was created by adding `reg_predictions` and `cas_predictions`.

4. Then when R2 was evaluated using `cnt_r2 = r2_score(y_test_cnt, cnt_predictions)` it was 0.9999998, this is because, in the dataset `cnt` is just the mathematical sum of `registered` and `casual` counts.

> **NOTE**: `alpha` for predicting casual counts and registered counts were different `4.0 and 3.8` respectively only to keep R2 < 1 for final count predictions after summing up.