Holdout set¶

---

This example shows when and how to use ATOM's holdout set in an exploration pipeline.

The data used is a variation on the Australian weather dataset from Kaggle. You can download it from here. The goal of this dataset is to predict whether or not it will rain tomorrow training a binary classifier on target RainTomorrow.

Load the data¶

In [1]:

# Import packages
import pandas as pd
from atom import ATOMClassifier

# Import packages import pandas as pd from atom import ATOMClassifier

In [2]:

# Load data
X = pd.read_csv("./datasets/weatherAUS.csv")

# Let's have a look
X.head()

# Load data X = pd.read_csv("./datasets/weatherAUS.csv") # Let's have a look X.head()

Out[2]:

	Location	MinTemp	MaxTemp	Rainfall	Evaporation	Sunshine	WindGustDir	WindGustSpeed	WindDir9am	WindDir3pm	...	Humidity9am	Humidity3pm	Pressure9am	Pressure3pm	Cloud9am	Cloud3pm	Temp9am	Temp3pm	RainToday	RainTomorrow
0	MelbourneAirport	18.0	26.9	21.4	7.0	8.9	SSE	41.0	W	SSE	...	95.0	54.0	1019.5	1017.0	8.0	5.0	18.5	26.0	Yes	0
1	Adelaide	17.2	23.4	0.0	NaN	NaN	S	41.0	S	WSW	...	59.0	36.0	1015.7	1015.7	NaN	NaN	17.7	21.9	No	0
2	Cairns	18.6	24.6	7.4	3.0	6.1	SSE	54.0	SSE	SE	...	78.0	57.0	1018.7	1016.6	3.0	3.0	20.8	24.1	Yes	0
3	Portland	13.6	16.8	4.2	1.2	0.0	ESE	39.0	ESE	ESE	...	76.0	74.0	1021.4	1020.5	7.0	8.0	15.6	16.0	Yes	1
4	Walpole	16.4	19.9	0.0	NaN	NaN	SE	44.0	SE	SE	...	78.0	70.0	1019.4	1018.9	NaN	NaN	17.4	18.1	No	0

5 rows × 22 columns

Run the pipeline¶

In [3]:

# Initialize atom specifying a fraction of the dataset for holdout
atom = ATOMClassifier(X, n_rows=0.5, holdout_size=0.2, verbose=2)

# Initialize atom specifying a fraction of the dataset for holdout atom = ATOMClassifier(X, n_rows=0.5, holdout_size=0.2, verbose=2)

<< ================== ATOM ================== >>
Algorithm task: binary classification.

Dataset stats ==================== >>
Shape: (56877, 22)
Memory: 24.67 MB
Scaled: False
Missing values: 127041 (10.2%)
Categorical features: 5 (23.8%)
Duplicate samples: 20 (0.0%)
-------------------------------------
Train set size: 42658
Test set size: 14219
Holdout set size: 14219
-------------------------------------
|   |       dataset |         train |          test |
| - | ------------- | ------------- | ------------- |
| 0 |   44113 (3.5) |   33085 (3.5) |   11028 (3.5) |
| 1 |   12764 (1.0) |    9573 (1.0) |    3191 (1.0) |

In [4]:

# The test and holdout fractions are split after subsampling the dataset
# Also note that the holdout data set is not a part of atom's dataset
print("Length loaded data:", len(X))
print("Length dataset + holdout:", len(atom.dataset) + len(atom.holdout))

# The test and holdout fractions are split after subsampling the dataset # Also note that the holdout data set is not a part of atom's dataset print("Length loaded data:", len(X)) print("Length dataset + holdout:", len(atom.dataset) + len(atom.holdout))

Length loaded data: 142193
Length dataset + holdout: 71096

In [5]:

atom.impute()
atom.encode()

atom.impute() atom.encode()

Fitting Imputer...
Imputing missing values...
 --> Dropping 264 samples due to missing values in feature MinTemp.
 --> Dropping 133 samples due to missing values in feature MaxTemp.
 --> Dropping 610 samples due to missing values in feature Rainfall.
 --> Dropping 24309 samples due to missing values in feature Evaporation.
 --> Dropping 27066 samples due to missing values in feature Sunshine.
 --> Dropping 3827 samples due to missing values in feature WindGustDir.
 --> Dropping 3799 samples due to missing values in feature WindGustSpeed.
 --> Dropping 4123 samples due to missing values in feature WindDir9am.
 --> Dropping 1572 samples due to missing values in feature WindDir3pm.
 --> Dropping 558 samples due to missing values in feature WindSpeed9am.
 --> Dropping 1085 samples due to missing values in feature WindSpeed3pm.
 --> Dropping 715 samples due to missing values in feature Humidity9am.
 --> Dropping 1461 samples due to missing values in feature Humidity3pm.
 --> Dropping 5563 samples due to missing values in feature Pressure9am.
 --> Dropping 5557 samples due to missing values in feature Pressure3pm.
 --> Dropping 21479 samples due to missing values in feature Cloud9am.
 --> Dropping 22836 samples due to missing values in feature Cloud3pm.
 --> Dropping 368 samples due to missing values in feature Temp9am.
 --> Dropping 1106 samples due to missing values in feature Temp3pm.
 --> Dropping 610 samples due to missing values in feature RainToday.
Fitting Encoder...
Encoding categorical columns...
 --> LeaveOneOut-encoding feature Location. Contains 26 classes.
 --> LeaveOneOut-encoding feature WindGustDir. Contains 16 classes.
 --> LeaveOneOut-encoding feature WindDir9am. Contains 16 classes.
 --> LeaveOneOut-encoding feature WindDir3pm. Contains 16 classes.
 --> Ordinal-encoding feature RainToday. Contains 2 classes.

In [6]:

# Unlike train and test, the holdout data set is not transformed until used for predictions
atom.holdout

# Unlike train and test, the holdout data set is not transformed until used for predictions atom.holdout

Out[6]:

	Location	MinTemp	MaxTemp	Rainfall	Evaporation	Sunshine	WindGustDir	WindGustSpeed	WindDir9am	WindDir3pm	...	Humidity9am	Humidity3pm	Pressure9am	Pressure3pm	Cloud9am	Cloud3pm	Temp9am	Temp3pm	RainToday	RainTomorrow
0	Launceston	0.3	16.0	0.0	NaN	NaN	NNW	30.0	SW	NW	...	74.0	39.0	NaN	NaN	2.0	NaN	6.4	15.1	No	0
1	Mildura	12.4	25.0	0.0	7.6	10.2	S	24.0	SSE	SSE	...	60.0	37.0	1020.2	1016.5	1.0	1.0	17.0	24.5	No	0
2	BadgerysCreek	16.5	27.2	2.0	NaN	NaN	SE	39.0	SSW	SE	...	91.0	63.0	1024.4	1022.6	NaN	NaN	21.6	26.3	Yes	1
3	Bendigo	4.4	18.9	1.4	NaN	NaN	W	24.0	SSW	SSW	...	71.0	41.0	1025.2	1022.5	NaN	NaN	12.8	18.4	Yes	0
4	Albany	15.8	23.3	0.0	3.6	5.2	NaN	NaN	NaN	SSW	...	91.0	74.0	1015.3	1016.3	7.0	5.0	21.0	22.7	No	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
14214	Sydney	15.7	20.4	0.0	10.6	4.6	NaN	NaN	S	SSE	...	57.0	69.0	1019.2	1019.3	5.0	7.0	18.4	18.2	No	0
14215	Brisbane	18.3	25.5	0.4	6.6	3.1	ENE	28.0	ESE	ENE	...	73.0	62.0	1021.5	1020.2	7.0	6.0	24.2	25.2	No	1
14216	Albany	12.2	19.4	0.6	1.0	NaN	NaN	NaN	W	NaN	...	92.0	NaN	1020.5	1017.5	8.0	NaN	13.6	NaN	No	0
14217	Sydney	13.5	25.6	0.0	5.8	10.4	NaN	NaN	WNW	WNW	...	57.0	31.0	1015.0	1012.5	4.0	1.0	16.3	24.9	No	0
14218	SydneyAirport	16.0	26.0	1.6	5.4	9.9	NE	39.0	NW	ENE	...	70.0	48.0	1025.4	1022.2	3.0	3.0	19.5	24.9	Yes	0

14219 rows × 22 columns

In [7]:

atom.run(models=["GNB", "LR", "RF"])

atom.run(models=["GNB", "LR", "RF"])

Training ========================= >>
Models: GNB, LR, RF
Metric: f1


Results for Gaussian Naive Bayes:
Fit ---------------------------------------------
Train evaluation --> f1: 0.6121
Test evaluation --> f1: 0.5929
Time elapsed: 0.031s
-------------------------------------------------
Total time: 0.031s


Results for Logistic Regression:
Fit ---------------------------------------------
Train evaluation --> f1: 0.6272
Test evaluation --> f1: 0.6
Time elapsed: 0.078s
-------------------------------------------------
Total time: 0.078s


Results for Random Forest:
Fit ---------------------------------------------
Train evaluation --> f1: 1.0
Test evaluation --> f1: 0.6046
Time elapsed: 2.907s
-------------------------------------------------
Total time: 2.907s


Final results ==================== >>
Duration: 3.016s
-------------------------------------
Gaussian Naive Bayes --> f1: 0.5929
Logistic Regression  --> f1: 0.6
Random Forest        --> f1: 0.6046 ~ !

In [8]:

atom.plot_prc()

atom.plot_prc()

In [9]:

# Based on the results on the test set, we select the best model for further tuning
atom.run("lr_tuned", n_calls=10, n_initial_points=5)

# Based on the results on the test set, we select the best model for further tuning atom.run("lr_tuned", n_calls=10, n_initial_points=5)

Training ========================= >>
Models: LR_tuned
Metric: f1


Running BO for Logistic Regression...
| call             | penalty |       C |  solver | max_iter | l1_ratio |      f1 | best_f1 |    time | total_time |
| ---------------- | ------- | ------- | ------- | -------- | -------- | ------- | ------- | ------- | ---------- |
| Initial point 1  |      l2 |  0.0961 |     sag |      965 |      --- |  0.5951 |  0.5951 |  0.531s |     0.547s |
| Initial point 2  |      l2 |  0.0051 | libli.. |      558 |      --- |   0.612 |   0.612 |  0.500s |     1.047s |
| Initial point 3  |      l2 |  1.2006 |   lbfgs |      195 |      --- |  0.6261 |  0.6261 |  0.500s |     1.547s |
| Initial point 4  |    none |     --- |     sag |      416 |      --- |  0.6293 |  0.6293 |  0.625s |     2.172s |
| Initial point 5  |      l2 |  0.0778 | newto.. |      617 |      --- |  0.6323 |  0.6323 |  0.516s |     2.688s |
| Iteration 6      |    none |     --- | newto.. |      631 |      --- |  0.6176 |  0.6323 |  0.531s |     3.438s |
| Iteration 7      |      l2 |  0.0739 | newto.. |      607 |      --- |  0.6425 |  0.6425 |  0.516s |     4.188s |
| Iteration 8      |      l2 |  0.1852 | newto.. |      588 |      --- |  0.6242 |  0.6425 |  0.516s |     4.954s |
| Iteration 9      |      l2 |  1.7612 |   lbfgs |      545 |      --- |  0.6157 |  0.6425 |  0.484s |     5.797s |
| Iteration 10     |      l2 | 98.7461 | newto.. |      389 |      --- |  0.6211 |  0.6425 |  0.531s |     6.594s |
Bayesian Optimization ---------------------------
Best call --> Iteration 7
Best parameters --> {'penalty': 'l2', 'C': 0.0739, 'solver': 'newton-cg', 'max_iter': 607}
Best evaluation --> f1: 0.6425
Time elapsed: 6.876s
Fit ---------------------------------------------
Train evaluation --> f1: 0.6271
Test evaluation --> f1: 0.5969
Time elapsed: 0.109s
-------------------------------------------------
Total time: 6.985s


Final results ==================== >>
Duration: 6.985s
-------------------------------------
Logistic Regression --> f1: 0.5969

Analyze the results¶

We already used the test set to choose the best model for futher tuning, so this set is no longer truly independent. Although it may not be directly visible in the results, using the test set now to evaluate the tuned LR model would be a mistake, since it carries a bias. For this reason, we have set apart an extra, indepedent set to validate the final model: the holdout set. If we are not going to use the test set for validation, we might as well use it to train the model and so optimize the use of the available data. Use the full_train method for this.

In [10]:

# Re-train the model on the full dataset (train + test) 
atom.lr_tuned.full_train()

# Re-train the model on the full dataset (train + test) atom.lr_tuned.full_train()

Model LR_tuned successfully retrained.

In [11]:

# Evaluate on the holdout set
atom.lr_tuned.evaluate(dataset="holdout")

# Evaluate on the holdout set atom.lr_tuned.evaluate(dataset="holdout")

Out[11]:

accuracy             0.856179
average_precision    0.725656
balanced_accuracy    0.742188
f1                   0.624304
jaccard              0.453810
matthews_corrcoef    0.549459
precision            0.746120
recall               0.536683
roc_auc              0.882404
Name: LR_tuned, dtype: float64

In [12]:

atom.lr_tuned.plot_prc(dataset="holdout")

atom.lr_tuned.plot_prc(dataset="holdout")