def evalOneMax_KNN(individual, df_test, base_estimator):
"""
This will have the code for testing the model on unseen data.
Parameters
----------
individual : list, required
List of dataframes with each dataframes last column the
dependent column.
df_test : DataFrame, required
The test dataframe against which you are evaluating your
model in the fitness function.
model : Estimator, required
The estimator to use in the model.
Side note: We can even start with just one partial part of the dataset, keep trying to
increase its cv score in fitness function. To get the optimal dataset. I am fearing,
it will make a dataset of exactly same values in the end. That will have the best cv error.
Therefore need to add a component of unseen data performance.
"""
total_mse = []
ensembles = []
centroids = []
for df_ in individual:
# clf = LinearRegression()
clf = base_estimator()
clf = clf.fit(df_.iloc[:, 0:-1], df_.iloc[:,-1])
ensembles.append(clf)
centroids.append(centroid_df(df_.iloc[:,0:-1]))
## for sum of mse return uncomment these
#y_pred = clf.predict(df_test[x_columns])
#mse = mean_squared_error(y_pred, df_test[y_column])
#total_mse.append(mse)
#print total_mse
y_preds_ensemble = []
for row in df_test.iloc[:, 0:-1].values:
distances = [distance(row, centroid) for centroid in centroids]
model_index = np.argmin(distances)
#todo: optional use the average of the 2 min distances ka prediction.
#todo: parameter.
y_pred = ensembles[model_index].predict(row)[0]
y_preds_ensemble.append(y_pred)
rmse_ensemble = np.sqrt(mean_squared_error(y_preds_ensemble, df_test.iloc[:,-1]))
return (rmse_ensemble),
def evalOneMax_KNN(individual, df_test, base_estimator):
"""
This will have the code for testing the model on unseen data.
Parameters
----------
individual : list, required
List of dataframes with each dataframes last column the
dependent column.
df_test : DataFrame, required
The test dataframe against which you are evaluating your
model in the fitness function.
model : Estimator, required
The estimator to use in the model.
Side note: We can even start with just one partial part of the dataset, keep trying to
increase its cv score in fitness function. To get the optimal dataset. I am fearing,
it will make a dataset of exactly same values in the end. That will have the best cv error.
Therefore need to add a component of unseen data performance.
"""
total_mse = []
ensembles = []
centroids = []
for df_ in individual:
# clf = LinearRegression()
clf = base_estimator()
clf = clf.fit(df_.iloc[:, 0:-1], df_.iloc[:, -1])
ensembles.append(clf)
centroids.append(centroid_df(df_.iloc[:, 0:-1]))
## for sum of mse return uncomment these
#y_pred = clf.predict(df_test[x_columns])
#mse = mean_squared_error(y_pred, df_test[y_column])
#total_mse.append(mse)
#print total_mse
y_preds_ensemble = []
for row in df_test.iloc[:, 0:-1].values:
distances = [distance(row, centroid) for centroid in centroids]
model_index = np.argmin(distances)
#todo: optional use the average of the 2 min distances ka prediction.
#todo: parameter.
y_pred = ensembles[model_index].predict(row)[0]
y_preds_ensemble.append(y_pred)
rmse_ensemble = np.sqrt(
mean_squared_error(y_preds_ensemble, df_test.iloc[:, -1]))
return (rmse_ensemble),
def eval_ensemble_oob_KNN_EG(individual, df, base_estimator, n_votes):
"""
Fitness is based on entire ensembles performance on Out of Bag samples
calculated using the union of the all samples used while training each child model.
Used in FEGO.
Parameters
----------
individual : list, required
List of dataframes with each dataframes last column the
dependent column.
df: DataFrame, required
The entire dataframe given to model to train - X and y combined.
model : Estimator, required
The estimator to use in the model.
#todo: not really needed. double check.
Side note: We can even start with just one partial part of the dataset, keep trying to
increase its cv score in fitness function. To get the optimal dataset. I am fearing,
it will make a dataset of exactly same values in the end. That will have the best cv error.
Therefore need to add a component of unseen data performance.
"""
total_mse = []
ensembles = []
centroids = []
# All the indices used in each child model.
bag_idx = []
for eg_ in individual:
# clf = LinearRegression()
clf = eg_.estimator
ensembles.append(clf)
idx = eg_.X.index.tolist()
# print len(idx)
bag_idx.append(idx)
centroids.append(centroid_df(eg_.X))
## for sum of mse return uncomment these
#y_pred = clf.predict(df_test[x_columns])
#mse = mean_squared_error(y_pred, df_test[y_column])
#total_mse.append(mse)
#print total_mse
# flattening and converting to set.
# print bag_idx
bag_idx = set(sum(bag_idx, []))
# print len(bag_idx)
out_bag_idx = list(set(df.index.tolist()) - bag_idx)
print 'Size of OOB', len(out_bag_idx)
df_test = df.loc[out_bag_idx]
y_preds_ensemble = []
ensembles = np.array(ensembles)
for row in df_test.iloc[:, 0:-1].values:
distances = np.array([distance(row, centroid) for centroid in centroids])
# model_index = np.argmin(distances)
#todo: optional use the average of the 2 min distances ka prediction.
#todo: parameter.
model_ixs = distances.argsort()[:n_votes]
models = ensembles[model_ixs]
# mean of all predictions.
y_pred = np.mean([mdl.predict(row)[0] for mdl in models])
# y_pred = ensembles[model_index].predict(row)[0]
y_preds_ensemble.append(y_pred)
rmse_ensemble = np.sqrt(mean_squared_error(y_preds_ensemble, df_test.iloc[:,-1]))
return (rmse_ensemble),
def eval_ensemble_oob_KNN_EG(individual, df, base_estimator, n_votes):
"""
Fitness is based on entire ensembles performance on Out of Bag samples
calculated using the union of the all samples used while training each child model.
Used in FEGO.
Parameters
----------
individual : list, required
List of dataframes with each dataframes last column the
dependent column.
df: DataFrame, required
The entire dataframe given to model to train - X and y combined.
model : Estimator, required
The estimator to use in the model.
#todo: not really needed. double check.
Side note: We can even start with just one partial part of the dataset, keep trying to
increase its cv score in fitness function. To get the optimal dataset. I am fearing,
it will make a dataset of exactly same values in the end. That will have the best cv error.
Therefore need to add a component of unseen data performance.
"""
total_mse = []
ensembles = []
centroids = []
# All the indices used in each child model.
bag_idx = []
for eg_ in individual:
# clf = LinearRegression()
clf = eg_.estimator
ensembles.append(clf)
idx = eg_.X.index.tolist()
# print len(idx)
bag_idx.append(idx)
centroids.append(centroid_df(eg_.X))
## for sum of mse return uncomment these
#y_pred = clf.predict(df_test[x_columns])
#mse = mean_squared_error(y_pred, df_test[y_column])
#total_mse.append(mse)
#print total_mse
# flattening and converting to set.
# print bag_idx
bag_idx = set(sum(bag_idx, []))
# print len(bag_idx)
out_bag_idx = list(set(df.index.tolist()) - bag_idx)
print 'Size of OOB', len(out_bag_idx)
df_test = df.loc[out_bag_idx]
y_preds_ensemble = []
ensembles = np.array(ensembles)
for row in df_test.iloc[:, 0:-1].values:
distances = np.array(
[distance(row, centroid) for centroid in centroids])
# model_index = np.argmin(distances)
#todo: optional use the average of the 2 min distances ka prediction.
#todo: parameter.
model_ixs = distances.argsort()[:n_votes]
models = ensembles[model_ixs]
# mean of all predictions.
y_pred = np.mean([mdl.predict(row)[0] for mdl in models])
# y_pred = ensembles[model_index].predict(row)[0]
y_preds_ensemble.append(y_pred)
rmse_ensemble = np.sqrt(
mean_squared_error(y_preds_ensemble, df_test.iloc[:, -1]))
return (rmse_ensemble),