Merge pull request #106 from flyingdutchman23/feature-importance

Add feature importance plot

mypy.ini

@@ -53,3 +53,6 @@ ignore_missing_imports = True
 
 [mypy-dtaidistance.*]
 ignore_missing_imports = True
+
+[mypy-shap.*]
+ignore_missing_imports = True

pyproject.toml

@@ -21,6 +21,7 @@ lightgbm = "^3.1.1"
 pyts = "^0.11.0"
 imbalanced-learn = "^0.8.0"
 dtaidistance = "^2.2.5"
+shap = "^0.39.0"
 
 [tool.poetry.dev-dependencies]
 pytest = "^6.1.0"

src/thesis/classify.py

@@ -13,6 +13,7 @@
 from lightgbm import LGBMClassifier
 import numpy as np
 import pandas as pd
+import shap
 from sklearn import metrics
 from sklearn.base import BaseEstimator
 from sklearn.metrics import (
@@ -72,6 +73,10 @@
 tensorflow.random.set_seed(SEED)
 
 
+def is_pipeline_finger(pipeline: Pipeline) -> bool:
+    return is_data_finger(list(pipeline.named_steps.keys())[0])
+
+
 def combine(dataPart: DataPart, metric_name: str):
     return f"{dataPart}_{metric_name}"
 
@@ -192,7 +197,7 @@ def __init__(self, config):
             index=build_index(self.measurements),
             dtype=np.int8,
         )
-        self.defects: Final = sorted(set(self.y))
+        self.defects: Final = [data.Defect(i) for i in sorted(set(self.y))]
         self.defect_names: Final = data.get_names(self.defects)
         self.cv_splits: Final = self._generate_cv_splits()
 
@@ -350,7 +355,7 @@ def _train(
                     (get_data_transformer(pipeline).transform(X_val), y_val),
                     (get_data_transformer(pipeline).transform(X_train), y_train),
                 ]
-            if is_data_finger(list(pipeline.named_steps.keys())[0]):
+            if is_pipeline_finger(pipeline):
                 feature_name, categorical_feature = get_categorical_features_info(
                     get_data_transformer(pipeline), X
                 )
@@ -505,8 +510,30 @@ def _save_models(
         X = self.get_X(model_name)
         self._train(pipeline, X, range(0, len(X)), range(0), False)
 
-        if isinstance(get_classifier(pipeline), LGBMClassifier):
-            lightgbm.plot_importance(get_classifier(pipeline))
+        if (
+            is_pipeline_finger(pipeline)
+            # @note: LGBMClassifier is chosen to make the tests pass the CI,
+            #        but all classifiers except k-NN also work.
+            and isinstance(get_classifier(pipeline), LGBMClassifier)
+            and "finger_all" not in pipeline.named_steps
+        ):
+            explainer = shap.Explainer(
+                pipeline.named_steps["classifier"],
+                feature_names=get_feature_names(pipeline[0]),
+                output_names=self.defect_names,
+            )
+            X_tr = pd.DataFrame(
+                data=pipeline[:-1].transform(X),
+                index=X.index,
+                columns=get_feature_names(pipeline[0]),
+            )
+            shap.summary_plot(
+                explainer.shap_values(X_tr),
+                X_tr,
+                class_names=self.defect_names,
+                max_display=X.shape[1],
+                show=self.config["general"]["show_plots"],
+            )
             util.finish_plot(
                 "feature_importance", model_folder, self.config["general"]["show_plots"]
             )

src/thesis/fingerprint.py

@@ -24,10 +24,10 @@
 from .data import CLASS, get_defects, PATH, PD, TIME_DIFF, VOLTAGE_SIGN
 
 
-PD_ID = "PD-Value"
-PD_DIFF_ID = "PD-Diff"
-TD_ID = "TimeDiff"
-CORR_ID = "Correlate"
+PD_ID = "A"
+PD_DIFF_ID = "\u0394A"
+TD_ID = "\u0394t"
+CORR_ID = "Corr"
 
 
 class Group(Enum):
@@ -41,46 +41,45 @@ def __str__(self):
 
 
 # @note: parameter in TU Graz fingerprint
-PD_VAR = f"{PD_ID} Variance"
-PD_SKEW = f"{PD_ID} Skewness"
-PD_KURT = f"{PD_ID} Kurtosis"
-PD_WEIB_A = f"{PD_ID} Weibull A"
-PD_WEIB_B = f"{PD_ID} Weibull B"
+PD_VAR = f"{PD_ID}-Var"
+PD_SKEW = f"{PD_ID}-Skew"
+PD_KURT = f"{PD_ID}-Kurt"
+PD_WEIB_A = f"{PD_ID}-Weib-\u03B1"
+PD_WEIB_B = f"{PD_ID} Weib-\u03B2"
 
-PD_DIFF_WEIB_B = f"{PD_DIFF_ID} Weibull B"
+PD_DIFF_WEIB_B = f"{PD_DIFF_ID}-Weib-\u03B2"
 
-TD_MAX = f"{TD_ID} Max"
-TD_MEAN = f"{TD_ID} Mean"
-TD_MIN = f"{TD_ID} Min"
-TD_VAR = f"{TD_ID} Variance"
-TD_SKEW = f"{TD_ID} Skewness"
-TD_KURT = f"{TD_ID} Kurtosis"
-TDIFF_NORM_WEIB_A = f"{TD_ID} Sorted Normed Weibull A"
-TDIFF_NORM_WEIB_B = f"{TD_ID} Sorted Normed Weibull B"
+TD_MAX = f"{TD_ID}-Max"
+TD_MEAN = f"{TD_ID}-Mean"
+TD_MIN = f"{TD_ID}-Min"
+TD_VAR = f"{TD_ID}-Var"
+TD_SKEW = f"{TD_ID}-Skew"
+TD_KURT = f"{TD_ID}-Kurt"
+TDIFF_NORM_WEIB_A = f"{TD_ID}-Norm-Weib-\u03B1"
+TDIFF_NORM_WEIB_B = f"{TD_ID}-Norm-Weib-\u03B2"
 
 
 # @note: parameter in Lukas fingerprint
-PDS_PER_SEC = "Number of PDs/sec"
+PDS_PER_SEC = "PDs/Sec"
 
-PD_MEAN = f"{PD_ID} Mean"
-PD_MAX = f"{PD_ID} Max"
-PD_CV = f"{PD_ID} std/mean"
-PD_SUM = f"{PD_ID} Sum"
+PD_MEAN = f"{PD_ID}-Mean"
+PD_MAX = f"{PD_ID}-Max"
+PD_CV = f"{PD_ID}-Std/Mean"
+PD_SUM = f"{PD_ID}-Sum"
 
 PD_DIFF_MEAN = f"{PD_DIFF_ID} Mean"
 PD_DIFF_SKEW = f"{PD_DIFF_ID} Skewness"
 PD_DIFF_VAR = f"{PD_DIFF_ID} Variance"
 PD_DIFF_KURT = f"{PD_DIFF_ID} Kurtosis"
 PD_DIFF_WEIB_A = f"{PD_DIFF_ID} Weibull A"
 
-TD_MEDIAN = f"{TD_ID} Median"
+TD_MEDIAN = f"{TD_ID}-Median"
 
-CORR_PD_DIFF_TO_PD_BINS = f"{CORR_ID} {PD_DIFF_ID} - PD Bins"
-CORR_NEXT_PD_TO_PD_BINS = f"{CORR_ID} Next PD - PD Bins"
-CORR_NEXT_PD_TO_PD = f"{CORR_ID} Next PD - PD"
-CORR_PD_DIFF_TO_PD = f"{CORR_ID} {PD_DIFF_ID} - PD"
-CORR_PD_DIFF_TO_TD = f"{CORR_ID} PD - {PD_DIFF_ID}"
-CORR_PD_TO_TD = f"{CORR_ID} PD - {TD_ID}"
+CORR_PD_DIFF_TO_PD_BINS = f"{CORR_ID}-{PD_DIFF_ID}-{PD_ID}-Bins"
+CORR_NEXT_PD_TO_PD_BINS = f"Auto-{CORR_ID}-{PD_ID}-Bins"
+CORR_NEXT_PD_TO_PD = f"Auto-{CORR_ID}-{PD_ID}-1"
+CORR_PD_DIFF_TO_PD = f"{CORR_ID}-{PD_DIFF_ID}-{PD_ID}"
+CORR_PD_TO_TD = f"{CORR_ID}-{PD_ID}-{TD_ID}"
 
 # @note: parameter in own fingerprint
 DURATION = "duration"
@@ -89,37 +88,37 @@ def __str__(self):
 PD_MIN = f"{PD_ID} min"
 PD_MEDIAN = f"{PD_ID} Median"
 PD_STD = f"{PD_ID} std"
-PD_VAR = f"{PD_ID} var"
-PD_NUM_PEAKS_50 = f"{PD_ID} num peaks 50"
-PD_NUM_PEAKS_10 = f"{PD_ID} num peaks 10"
-PD_NUM_PEAKS_5 = f"{PD_ID} num peaks 5"
-PD_NUM_PEAKS_100 = f"{PD_ID} num peaks 100"
+PD_VAR = f"{PD_ID}-Var"
+PD_NUM_PEAKS_50 = f"{PD_ID}-Num-peaks-50"
+PD_NUM_PEAKS_10 = f"{PD_ID}-Num-peaks-10"
+PD_NUM_PEAKS_5 = f"{PD_ID}-Num-peaks-5"
+PD_NUM_PEAKS_100 = f"{PD_ID}-Num-peaks-100"
 PD_RATIO = f"{PD_ID} ratio"
 PD_PERC_REOCCUR = f"{PD_ID} percentage reocurring"
-PD_COUNT_ABOVE_MEAN = f"{PD_ID} count above mean"
-PD_COUNT_BELOW_MEAN = f"{PD_ID} count below mean"
-PD_CHANGE_QUANTILES = f"{PD_ID} ChangeQuantiles"
-PD_NORM_WEIB_A = f"{PD_ID} Weibull normed sorted A"
-PD_NORM_WEIB_B = f"{PD_ID} Weibull normed sorted B"
-
-TD_LONGEST_STRIKE_BELOW_MEAN = f"{TD_ID} longest strike below mean"
-TD_CHANGE_QUANTILES = f"{TD_ID} ChangeQuantiles"
-TD_SUM = f"{TD_ID} Sum"
-
-CORR_2ND_NEXT_PD_TO_PD = f"{CORR_ID} Auto 2nd Next {PD_ID}"
-CORR_3RD_NEXT_PD_TO_PD = f"{CORR_ID} Auto 3rd Next {PD_ID}"
-CORR_5TH_NEXT_PD_TO_PD = f"{CORR_ID} Auto 5th Next {PD_ID}"
-CORR_10TH_NEXT_PD_TO_PD = f"{CORR_ID} Auto 10th Next {PD_ID}"
-
-AUTOCORR_NEXT_TD = f"{CORR_ID} Auto Next {TD_ID}"
-AUTOCORR_2ND_NEXT_TD = f"{CORR_ID} Auto 2nd Next {TD_ID}"
-AUTOCORR_3RD_NEXT_TD = f"{CORR_ID} Auto 3rd Next {TD_ID}"
-AUTOCORR_5TH_NEXT_TD = f"{CORR_ID} Auto 5th Next {TD_ID}"
-AUTOCORR_10TH_NEXT_TD = f"{CORR_ID} Auto 10th Next {TD_ID}"
+PD_COUNT_ABOVE_MEAN = f"{PD_ID}-Num->-mean"
+PD_COUNT_BELOW_MEAN = f"{PD_ID}-Num-<-mean"
+PD_CHANGE_QUANTILES = f"{PD_ID}-change-quantiles"
+PD_NORM_WEIB_A = f"{PD_ID}-norm-Weib-\u03B1"
+PD_NORM_WEIB_B = f"{PD_ID}-norm-Weib-\u03B2"
+TD_LONGEST_STRIKE_BELOW_MEAN = f"{PD_ID}-max-strike-<-mean"
+
+TD_CHANGE_QUANTILES = f"{TD_ID}-Change-quantiles"
+TD_SUM = f"{TD_ID}-Sum"
+
+CORR_2ND_NEXT_PD_TO_PD = f"Auto-{CORR_ID}-{PD_ID}-2"
+CORR_3RD_NEXT_PD_TO_PD = f"Auto-{CORR_ID}-{PD_ID}-3"
+CORR_5TH_NEXT_PD_TO_PD = f"Auto-{CORR_ID}-{PD_ID}-5"
+CORR_10TH_NEXT_PD_TO_PD = f"Auto-{CORR_ID}-{PD_ID}-10"
+
+AUTOCORR_NEXT_TD = f"Auto-{CORR_ID}-{TD_ID}-1"
+AUTOCORR_2ND_NEXT_TD = f"Auto-{CORR_ID}-{TD_ID}-2"
+AUTOCORR_3RD_NEXT_TD = f"Auto-{CORR_ID}-{TD_ID}-3"
+AUTOCORR_5TH_NEXT_TD = f"Auto-{CORR_ID}-{TD_ID}-5"
+AUTOCORR_10TH_NEXT_TD = f"Auto-{CORR_ID}-{TD_ID}-10"
 
 # @note: further parameters
-PD_BY_TD_WEIB_A = f"{PD_ID} / {TD_ID} Weibull A"
-PD_BY_TD_WEIB_B = f"{PD_ID} / {TD_ID} Weibull B"
+PD_BY_TD_WEIB_A = f"{PD_ID}/{TD_ID}-Weib-\u03B1"
+PD_BY_TD_WEIB_B = f"{PD_ID}/{TD_ID}-Weibu-\u03B2"
 
 
 def get_parameter_group(df: pd.DataFrame, group: Group) -> pd.DataFrame: