Codes

CODE 0

import seaborn as sns
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np

plt.style.use(‘winter’)
pd.set_option(‘max_columns’, None)

CODE 1

from sklearn.impute import SimpleImputer
imp = SimpleImputer(missing_values=np.nan, strategy=‘mean’)
imp.fit_transform(X)

General

1. Import Stuff using [CODE0]
2. Train Test split
3. Create and get error of a baseline model (like average or median)

Data Exploration

1. Open the dataset
2. Data Understanding
   • DataFrame .shape
   • .head() and .tail()
   • .dtypes
   • .describe()
   • .info()
3. Data Preparation
   • Drop irrelevant columns and rows
     • use .columns to obtain list of columns
     • copy and paste column list
     • Remove unnecessary columns
     • Set a variable to be the dataset indexed with the remaining columns, add .copy() at the end
     • you can also use .drop([‘cols you wanna drop’], axis=1, inplace=True)
   • change necessary column type, using dataset[column] = pd.to_TheFinalDataType(dataset[column])
     • You can use .to_numeric, .to_datetime, etc.
   • Identify duplicate columns
     • use .duplicated()
     • use .loc[.duplicated()] to look at all duplicate rows
     • use .duplicated().sum() to get number of duplicated rows
     • use .duplicated(subset=[]) to get duplicated of specific columns
     • To remove duplicated use slicing with ~ in front for inverse
     • use .reset_index(drop=True) to reset index
   • Rename Columns
     • To rename, .rename(columns={’old_name’: ‘new_name’}, inplace=True)
   • Feature Creation
   • Remove or do some other action to NaN values
     • Using pandas
       • Identify using .isna()
       • Get total using .isna().sum()
       • Replace using .fillna(value or mean, median, mode, etc)
     • Using Scikit-Learn
       • Replace using mean, most_frequent or constant for strategy, use [CODE1]
4. Feature Understanding (Univariate Analysis) use ax = to save any plot into a matplotlib plot
   • Value Count use [column].value_counts() to see number of unique values and their count
   • Histogram, use .hist(title=“Plot0”) or .plot(kind=‘hist’) on any pd df
   • KDE, use .plot(kind=‘kde’)
   • Bar Graph, use .plot(kind=‘bar’, title=“Plot1”) or barh instead of bar for horizontal plot
   • Box Plot
5. Feature Relationship
   • Scatter Plot
     • For basic, use .plot(kind=‘scatter’, x=’column’, y=’column too’)
     • For Pro, use sns.scatterplot(data=df, x=‘column’, y=‘column 2’), you can use hue= property to use other columns as the color
   • Pair Plot
     • Use sns.pairplot(data=df, vars=[a very long array of stuff])
     • Use sns.pairplot(data=df, vars_x=[…], vars_y=[…]) to get only specific stuff
     • use can also use hue= for this also.
   • Correlation, use .corr(), you can pass in the type of [pearson, kendall, spearman or callable]
     • Use sns.heatmap() to get better visualization use annot=True to see the values inside

Feature Selection + Data Preparation

1. Use sklearn.compose.make_column_transformer to apply feature selection for individual columns.
2. Use sklearn.pipeline.make_pipeline to make pipelines to automate stuff. Column transform can be added to it as a process.
3. Dimensionality Reduction
   • PCA: sklearn.decomposition.PCA
   • t-SNE: sklean.mainfold.TSNE
   • UMAP: gotta install umap-learn, umap.umap_ as umap and use as if it was t-SNE or PCA
4. Scaling
   • Normalizer: sklearn.preprocessing.Normalizer
   • StandardSalar: sklearn.preprocessing.StandardScalar
   • MinMaxSalar: sklearn.preprocessing.MinMaxScalar
   • RobustScalar: sklean.preprocessing.RobustScalar
5. Feature expansion:
   • PolynomialFeatures: sklearn.preprocessing.PolynomialFeatures
6. Use Categorical Encoding
   • One Hot Encoding: sklearn.preprocessing.OneHotEncoder
   • Ordinal Encoding: sklearn.preprocessing.OrdinalEncoder
   • LabelEncoder (similar to ordinal but only for target): sklearn.preprocessing.Labelencoder
   • Embeddings
7. Missing Values
   • SimpleImputer: sklearn.impute.SimpleImputer
   • KNNImputer: sklearn.impute.KNNImputer
8. Data Transformation / Mapping
   • KBinsDiscretizer: sklearn.preprocessing.KBinsDiscretizer
   • FunctionTransformer: sklearn.preprocessing.FunctionTransformer
   • Binarizer (Thresholding (turns numbers into 0/1)): sklean.preprocessing.Binarizer
9. Clustering
   • KMeans: use sklearn.cluster.KMeans, to create new feature, it would be a categorical feature. KMeans takes in n_clusters. Use dummies with this to get final features.
10. Text Transforms
    • CountVectorizer (bag of words): sklearn.feature_extraction.text.CountVectorizer
    • Tf-IDF Vectorizer: sklearn.feature_extraction.text.TfidfVectorizer
    • tfidf and bow: sklearn.feature_extraction.text.TfidfTransformer
    • Hashing Vectorizer: sklearn.feature_extraction.text.HashingVectorizer
11. Feature Selection
    • Keep the top k features: sklearn.feature_selection.SelectKBest() / sklearn.feature_selection.SelectPercentile()
    • Recursive feature elimination: sklearn.feature_selection.RFE() / sklearn.feature_selection.RFECV()
    • concatenate all the features obtained previously to get a new dataset. For concatenation use pd.concat([], axis=1)

Other very cool stuff

1. sklearn.compose.make_column_selector: Automatically select numerical/categorical columns.
2. sklearn.compose.make_column_selector: Automatically select columns for column transformer.
3. [any model].check_is_fitted(): check if the model is trained.
4. sklearn.model_selection.cross_val_score: cross validation in one line.

Model Selection

Use this flow chart ![[ml_map.svg]]