From ddde80b4212875d321a91cd07b40ed781b9f09c7 Mon Sep 17 00:00:00 2001 From: Primoz Date: Thu, 24 Nov 2022 09:24:13 +0100 Subject: [PATCH] Add classification with clustering ml pipeline script. --- exploration/ml_pipeline_classification.py | 7 - ...pipeline_classification_with_clustering.py | 178 ++++++++++++++++++ 2 files changed, 178 insertions(+), 7 deletions(-) create mode 100644 exploration/ml_pipeline_classification_with_clustering.py diff --git a/exploration/ml_pipeline_classification.py b/exploration/ml_pipeline_classification.py index bde5c73..fc2fb81 100644 --- a/exploration/ml_pipeline_classification.py +++ b/exploration/ml_pipeline_classification.py @@ -51,13 +51,6 @@ import machine_learning.model # %% jupyter={"source_hidden": true} model_input = pd.read_csv("../data/intradaily_30_min_all_targets/input_JCQ_job_demand_mean.csv") -lime_cols = [col for col in model_input if col.startswith('limesurvey_demand')] -model_input['limesurvey_demand_control_ratio'].describe() -lime_cols - -# TODO: prek lime_cols ustvari klastre, ki jih nato kasneje ločeno preveriš z modeli (npr. k=5). Potrebno bo trikrat ponoviti spodnji postopek. -# Pomisli, če gre kaj zavizi v for loop (npr. modeli v seznamu) - # %% jupyter={"source_hidden": true} index_columns = ["local_segment", "local_segment_label", "local_segment_start_datetime", "local_segment_end_datetime"] model_input.set_index(index_columns, inplace=True) diff --git a/exploration/ml_pipeline_classification_with_clustering.py b/exploration/ml_pipeline_classification_with_clustering.py new file mode 100644 index 0000000..9d81b21 --- /dev/null +++ b/exploration/ml_pipeline_classification_with_clustering.py @@ -0,0 +1,178 @@ +# --- +# jupyter: +# jupytext: +# formats: ipynb,py:percent +# text_representation: +# extension: .py +# format_name: percent +# format_version: '1.3' +# jupytext_version: 1.13.0 +# kernelspec: +# display_name: straw2analysis +# language: python +# name: straw2analysis +# --- + +# %% jupyter={"source_hidden": true} +# %matplotlib inline +import datetime +import importlib +import os +import sys + +import numpy as np +import matplotlib.pyplot as plt +import pandas as pd +import seaborn as sns +from scipy import stats + +from sklearn import linear_model, svm, naive_bayes, neighbors, tree, ensemble +from sklearn.model_selection import LeaveOneGroupOut, cross_validate +from sklearn.dummy import DummyClassifier +from sklearn.impute import SimpleImputer +from lightgbm import LGBMClassifier +import xgboost as xg + +from sklearn.cluster import KMeans + +from IPython.core.interactiveshell import InteractiveShell +InteractiveShell.ast_node_interactivity = "all" + +nb_dir = os.path.split(os.getcwd())[0] +if nb_dir not in sys.path: + sys.path.append(nb_dir) + +import machine_learning.labels +import machine_learning.model + +# %% [markdown] +# # RAPIDS models + +# %% [markdown] +# ## PANAS negative affect + +# %% jupyter={"source_hidden": true} +model_input = pd.read_csv("../data/intradaily_30_min_all_targets/input_JCQ_job_demand_mean.csv") + +lime_cols = [col for col in model_input if col.startswith('limesurvey_demand')] +lime_col = 'limesurvey_demand_control_ratio' +model_input[lime_col].describe() + +# %% jupyter={"source_hidden": true} + +# Filter-out outlier rows by lime_col +model_input = model_input[(np.abs(stats.zscore(model_input[lime_col])) < 3)] + +uniq = model_input[[lime_col, 'pid']].drop_duplicates().reset_index(drop=True) +plt.bar(uniq['pid'], uniq[lime_col]) + +# %% jupyter={"source_hidden": true} +# Get clusters by lime col & and merge the clusters to main df +km = KMeans(n_clusters=5).fit_predict(uniq.set_index('pid')) +np.unique(km, return_counts=True) +uniq['cluster'] = km +uniq + +model_input = model_input.merge(uniq[['pid', 'cluster']]) + +# %% jupyter={"source_hidden": true} +index_columns = ["local_segment", "local_segment_label", "local_segment_start_datetime", "local_segment_end_datetime"] +model_input.set_index(index_columns, inplace=True) + +# %% jupyter={"source_hidden": true} + +for k in range(5): + model_input_subset = model_input[model_input["cluster"] == k].copy() + bins = [-10, -1, 1, 10] # bins for z-scored targets + model_input_subset.loc[:, 'target'] = \ + pd.cut(model_input_subset.loc[:, 'target'], bins=bins, labels=['low', 'medium', 'high'], right=False) #['low', 'medium', 'high'] + model_input_subset['target'].value_counts() + model_input_subset = model_input_subset[model_input_subset['target'] != "medium"] + model_input_subset['target'] = model_input_subset['target'].astype(str).apply(lambda x: 0 if x == "low" else 1) + + model_input_subset['target'].value_counts() + + + cv_method_str = 'logo' # logo, halflogo, 5kfold + if cv_method_str == 'halflogo': + model_input_subset['pid_index'] = model_input_subset.groupby('pid').cumcount() + model_input_subset['pid_count'] = model_input_subset.groupby('pid')['pid'].transform('count') + + model_input_subset["pid_index"] = (model_input_subset['pid_index'] / model_input_subset['pid_count'] + 1).round() + model_input_subset["pid_half"] = model_input_subset["pid"] + "_" + model_input_subset["pid_index"].astype(int).astype(str) + + data_x, data_y, data_groups = model_input_subset.drop(["target", "pid", "pid_index", "pid_half"], axis=1), model_input_subset["target"], model_input_subset["pid_half"] + else: + data_x, data_y, data_groups = model_input_subset.drop(["target", "pid"], axis=1), model_input_subset["target"], model_input_subset["pid"] + + # Treat categorical features + categorical_feature_colnames = ["gender", "startlanguage"] + additional_categorical_features = [] #[col for col in data_x.columns if "mostcommonactivity" in col or "homelabel" in col] + categorical_feature_colnames += additional_categorical_features + + categorical_features = data_x[categorical_feature_colnames].copy() + mode_categorical_features = categorical_features.mode().iloc[0] + + # fillna with mode + categorical_features = categorical_features.fillna(mode_categorical_features) + + # one-hot encoding + categorical_features = categorical_features.apply(lambda col: col.astype("category")) + if not categorical_features.empty: + categorical_features = pd.get_dummies(categorical_features) + + numerical_features = data_x.drop(categorical_feature_colnames, axis=1) + train_x = pd.concat([numerical_features, categorical_features], axis=1) + + # Establish cv method + cv_method = None # Defaults to 5 k-folds in cross_validate method + if cv_method_str == 'logo' or cv_method_str == 'half_logo': + cv_method = LeaveOneGroupOut() + cv_method.get_n_splits( + train_x, + data_y, + groups=data_groups, + ) + + n = 3 + + imputer = SimpleImputer(missing_values=np.nan, strategy='median') + + # Create dict with classification ml models + cmodels = { + 'dummy_classifier': DummyClassifier(strategy="most_frequent"), + 'logistic_regression': linear_model.LogisticRegression(), + 'support_vector_machine': svm.SVC(), + 'gaussian_naive_bayes': naive_bayes.GaussianNB(), + 'stochastic_gradient_descent_classifier': linear_model.SGDClassifier(), + 'knn': neighbors.KNeighborsClassifier(), + 'decision_tree': tree.DecisionTreeClassifier(), + 'random_forest_classifier': ensemble.RandomForestClassifier(), + 'gradient_boosting_classifier': ensemble.GradientBoostingClassifier(), + 'lgbm_classifier': LGBMClassifier(), + 'XGBoost_classifier': xg.sklearn.XGBClassifier() + } + + for model_title, model in cmodels.items(): + + classifier = cross_validate( + model, + X=imputer.fit_transform(train_x), + y=data_y, + groups=data_groups, + cv=cv_method, + n_jobs=-1, + error_score='raise', + scoring=('accuracy', 'average_precision', 'recall', 'f1') + ) + + print("\n-------------------------------------\n") + print("Current cluster:", k, end="\n") + print("Current model:", model_title, end="\n") + print("Acc", np.median(classifier['test_accuracy'])) + print("Precision", np.median(classifier['test_average_precision'])) + print("Recall", np.median(classifier['test_recall'])) + print("F1", np.median(classifier['test_f1'])) + print("Largest 5 ACC:", np.sort(-np.partition(-classifier['test_accuracy'], n)[:n])[::-1]) + print("Smallest 5 ACC:", np.sort(np.partition(classifier['test_accuracy'], n)[:n])) +# %%