import pandas as pd from sklearn.model_selection import train_test_split, KFold, cross_val_score, GridSearchCV from sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier from sklearn.svm import SVC from sklearn.metrics import accuracy_score, classification_report, confusion_matrix, roc_auc_score, f1_score, recall_score, precision_score from sklearn.preprocessing import StandardScaler from sklearn.pipeline import Pipeline from sklearn.decomposition import PCA import matplotlib.pyplot as plt pca_ccr = 0.98 # 主成分分析の累積寄与率の閾値 n_folds = 5 # クロスバリデーションの分割数 df = pd.DataFrame( columns=['DataType','Input', 'Target', 'Model', 'PCAdim', 'Accuracy', 'Recall', 'Precision', 'F1', 'AUC'] ) df_index = 0 def make_dataset(): """ データセットを作成する関数 """ # print("\nPreparing dataset...") all_data = pd.read_csv(f'Merged_{df.loc[df_index, 'DataType']}App.csv') input = df.loc[df_index, 'Input'] # 入力の種類を取得 if input == "shape": # print("Using shape features...") x = all_data.loc[:, 'shape-width':'shape-bottomRightY'] elif input == "color": # print("Using color features...") x = all_data.loc[:, 'chiu-lateral-L-min':'fiveClick-tip-b-kurtosis'] elif input == "texture": # print("Using texture features...") x = all_data.loc[:, 'chiu-lateral-contrast':'fiveClick-tip-correlation'] else: # print("Using all features...") x = all_data.loc[:, 'shape-width':'fiveClick-tip-b-kurtosis'] # 全データ scores = all_data.loc[:, 'A01':'C08'] # スコア invert_list = ['A01', 'A02', 'A03'] for invert in invert_list: scores[invert] = 5 - scores[invert] # 逆転スコア scores['A'] = scores.loc[:, 'A01':'A10'].sum(axis=1) # A群 scores['B'] = scores.loc[:, 'B07':'B29'].sum(axis=1) # B群 scores['C'] = scores.loc[:, 'C01':'C08'].sum(axis=1) # C群 scores['Total'] = scores.loc[:, 'A':'C'].sum(axis=1) # 総合スコア target = df.loc[df_index, 'Target'] # ターゲットスコア threshold = scores[target].median() # 中央値を閾値とする # print(f'Threshold for {target}: {threshold}') # scores['Total'].plot.hist(bins=20, edgecolor='black') # ヒストグラムを描画 # import matplotlib.pyplot as plt # plt.title('Total Score Distribution') # plt.xlabel('Total Score') # plt.ylabel('Frequency') # plt.show() # ヒストグラムを表示 scores['label'] = 0 scores.loc[scores[target] >= threshold, 'label'] = 1 # ラベル付け # print(scores.head(3)) return x, scores['label'] def cross_val(model, x, y): """ クロスバリデーションを行う関数 """ # print("\nStart cross-validation") pca_dim = calc_pca_dim(x) df.loc[df_index, 'PCAdim'] = pca_dim # 主成分分析の次元数を保存 kfold_cv = KFold(n_splits=n_folds, shuffle=True) pipe = Pipeline([("scaler", StandardScaler()), ("pca", PCA(n_components=pca_dim)), ("model", model)]) df.loc[df_index, 'Accuracy'] = cross_val_score(pipe, x, y, cv=kfold_cv, scoring="accuracy").mean() df.loc[df_index, 'Recall'] = cross_val_score(pipe, x, y, cv=kfold_cv, scoring="recall").mean() df.loc[df_index, 'Precision'] = cross_val_score(pipe, x, y, cv=kfold_cv, scoring="precision").mean() df.loc[df_index, 'F1'] = cross_val_score(pipe, x, y, cv=kfold_cv, scoring="f1").mean() df.loc[df_index, 'AUC'] = cross_val_score(pipe, x, y, cv=kfold_cv, scoring="roc_auc").mean() # print(f"Accuracy: {accuracy:.3f}, recall: {recall:.3f}, precision: {precision:.3f}, F1: {F1:.3f}, AUC: {auc:.3f}") def calc_pca_dim(x): """ 主成分分析の次元数を計算する関数 """ # print("\nCalculating PCA dimensions...") scaler = StandardScaler() x_scaled = scaler.fit_transform(x) pca = PCA(n_components=None) # 主成分分析の次元数を入力データの次元数に設定 pca.fit(x_scaled) ccr = 0 pca_dim = 0 for i in range(pca.n_components_): ccr += pca.explained_variance_ratio_[i] if ccr >= pca_ccr: pca_dim = i + 1 # 次元数は0から始まるので1を足す # print(f'Number of components to reach 95% variance: {pca_dim}') break return pca_dim def train_predict(x, y): """ モデルを訓練し、予測を行う関数 """ # データセットを訓練用とテスト用に分割 print("\nStarting model training and prediction...") x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=42) input_dim = x_train.shape[1] print(f"Input dim:{input_dim} Train:{x_train.shape[0]}, Test: {x_test.shape[0]}") pca_dim = calc_pca_dim(x_train) pipe = Pipeline([("scaler", StandardScaler()), ("pca", PCA(n_components=pca_dim)), ("model", RandomForestClassifier())]) pipe.fit(x_train, y_train) y_pred = pipe.predict(x_test) # scaler = StandardScaler() # x_train = scaler.fit_transform(x_train) # x_test = scaler.transform(x_test) # model = RandomForestClassifier() # model.fit(x_train, y_train) # y_pred = model.predict(x_test) accuracy = accuracy_score(y_test, y_pred) df['accuracy'] = accuracy print(f'Accuracy: {accuracy:.2f}') # print(classification_report(y_test, y_pred)) print('Confusion Matrix:') print(confusion_matrix(y_test, y_pred)) # メイン関数 if __name__ == "__main__": print("Starting analysis...") data_type_list = ["SmTIAS_Phone", "SmTIAS_Web", "HandyTCC_Phone", "HandyTCC_Web"] # データセットの種類 input_list = ["shape", "color", "texture", "all"] # 入力の種類 target_list = ["A", "B", "C", "Total"] # ターゲットスコア model_list = [RandomForestClassifier(), AdaBoostClassifier(), SVC(),] # モデルの種類 counts = len(data_type_list) * len(input_list) * len(target_list) * len(model_list) for data_type in data_type_list: for input in input_list: # 入力の種類を指定 for target in target_list: for model in model_list: model_name = model.__class__.__name__.replace("Classifier", "") # モデル名から"Classifier"を削除 print(f"\nAnalyzing {df_index+1}/{counts} {data_type}, {input}, {target}, {model_name}") df.loc[df_index,:] = [data_type, input, target, model_name, 0, 0, 0, 0, 0, 0] # 初期化 x, y = make_dataset() # 入力とターゲットを指定 # train_predict(x, y) # モデルの訓練と予測を実行 cross_val(model, x, y) # クロスバリデーションを実行 df_index += 1 print("\nAnalysis complete.") print(df) # 結果を表示 df.to_csv('analysis_results.csv', index=False) # 結果をCSVファイルに保存