Auto Data Analysis Regression

Motivation

One of the daily challenges for data scientists is selecting an optimal model. To select the optimal model, we need to work on data cleaning, such as handling missing values and outliers. Even if we carefully clean up the data, there is no promise of meeting the optimal model. On top of that, all datasets are unique and require different approaches to discover an optimal model. Several data scientists prefer to use Random Forest as their first choice because the model works well in general. However, the model is expensive, and there might be a better model if we tune it up well. The auto data analysis for regression suggests an optimal model, hyper-parameters, and proportion of train datasets.

1. Requirements

• python ==> 3.9
• py7zr
• Chrome Browser

Install:

pip install -r requirements.txt 

2. Execution

Analysis Regression

python main.py -I demo.csv -T AveragePrice -R 0 Date -D type region

For more details on parameters, please run:

python main.py --help

Re-Open Browser

Example:

python decompress.py -F demo.7z

For the more details of parameters, please run:

python decompress.py --help

3. Explain

demo.Auto.Data.Analysis.Regression.mp4

Missing Values

Auto Analysis detects missing values as its first step, and it visualizes them with a bar plot. If any missing values are detected, the model automatically applies KNN-imputation for numerical variables and mode for categorical variables.

Descriptive Statistics

Auto Analysis displays the descriptive statistics and the target behavior with a box plot quickly.

Correlation

Auto Analysis displays the entire correlation with a heatmap and then goes to a one-step deeper analysis. The bar plot shows the correlations between the target and other numerical variables. The model sets up a strong threshold with 0.7 and -0.7 and a moderate threshold with 0.5 and -0.5. The table shows the correlation values.

R-Squared with Dummy Variables

This is an optional step. The model can discover the relationships between the target and categorical variables if the model receives any value. However, the thresholds depend on the situations, and the model will drop off any categorical columns when it finds an optimal model.

Outliers

Using the side-by-side plots, the model visualizes outliers effectively and calculates the percentage of outliers.

If the percentage is over 5%, the model automatically applies Winsorizing to minimize number of outliers.

Suggesting The Best Model

To find the best model with given datasets, the model uses gridsearch. The model shows the best model, hyper-parameters, and the propotion of tain datasets. The selection is based on the scores, which are provided by sklearn as an attibute.

*If users want to include neural networks in the gridsearch, please provide "-N True", and it might take a lot of time to find the best one.
The best model is saved in a 7z file, and named 'best_best.pkl'

Model Selection

The model tests different regression models to find the optimal one. To obtain the best one, the model applies 4 different scalers, which include the standard scaler. Then, store the local best model to find a global best model to suggest it to users.

After

All results are saved in the local folder. If you want to check it again, please run the decompress.py file.  main.py -I demo.csv  <- analysis the dataset
decompress.py -F demo.7z <- show the result again

4. Future Study

This model suggests one of the best models. However, the best model would be different when you use different data preprocessing and analysis. For example, ayami-n found SVR is better than Random Forest when the author splits the demo datasets by region.