Description
1 Introduction
You may have heard that many honeybee colonies are failing. In the U.S., the number of hives that do not survive the
winter months has averaged 28.7% since 2006-2007. During the 2014 California almond bloom, between 15% and 25%
of beehives experienced failures ranging from complete hive collapse to dead and deformed brood. Typical culprits of
bee colony failures are bacteria and parasites (e.g., Varroa mites, foulbrood, chalkbrood) and bad weather. Other factors
contributing to bee colony failures include pesticides (e.g., neonicotinoids), monoculture (increasing emphasis on growing
large quantities of the same crop), and transportation stress caused by long hauls of large quantities of commercial
beehives on semitrucks. Since 1 out of every 3 crops have the honeybee as their sole pollinator, the nutritional diversity
of our food supply is in danger.
It’s not only the bees that are suffering, beekeepers are suffering, too. The average loss for beekeepers annually is
1 out of every 3 hives. The loss range varies from 25% to 50% per year. Nationally we’ve lost more than half of our
beekeepers since the arrival of the Varroa mite, a parasite that now afflicts many beehives. The cost of equipment, bee
packages, maintenance, and transportation is so high that profit margins for beekeepers are very small.
I believe that in the future significant practical and scientific benefits will come from transforming our bee yards into
smart worlds. Think of a beehive as an immobile robot that monitors the health of the bee colony inside, analyzes the
data from its sensors, and alerts the beekeeper of any deviations from the norm through the Internet of Things (IoT).
Why is this useful? It’s useful, because human beekeepers cannot monitor their hives continuously due to obvious
problems with logistics and fatigue. It will also reduce the number of invasive hive inspections that disrupt the life cycle
of bee colonies and put stress on the bees. The transportation costs will likely be lower, because many beekeepers now
drive long distances to their far flung bee yards.
In this assignment, we’ll use decision trees and random forests to classify audio data obtained from a deployed
electronic beehive monitor. I included three sample audio files: bee.wav, cricket.wav, noise.wav. You can quickly
listen to them to get an idea of what the original audio data is like. However, we won’t work with the raw data, because
audio signal processing is beyond the scope of this class. We’ll work with the processed data instead, i.e., feature vectors.
The zip archive for this assignment contains three directories: beefv, cricketfv, and noisefv. These directories contain
feature vectors obtained from wav audio files with bee buzzing, cricket chirping, and ambient noise, respectively.
Each audio file is turned into a feature vector by applying to it various audio filters (e.g., FFT, SFFT, MFCC, etc.) to
extract 193 features from it. Each feature is a float. After the features are extracted, each feature vector is turned into a
numpy array and persisted. Thus, the directory beefv contains 3,000 persisted numpy arrays with bee buzzing features;
the directory cricketfv contains 3,000 persisted numpy arrays of cricket chirping features; the directory noisefv contains
3,110 persisted numpy arrays with ambient noise features.
This ground truth classification was obtained from four human evaluators (three of my graduate students and myself)
who manually labeled 9,110 2-second audio samples from May and June 2017 by listening to each sample and placing
it into one of the three non-overlapping categories: bee buzzing, cricket chirping, and ambient noise. The bee buzzing
category consisted of the samples where the evaluators could hear the buzzing of bees. The cricket churping category
consisted of the files, captured at night, where the evaluators could hear the chirping of crickets. The noise category
included all samples where the evaluators could not clearly hear either bee buzzing or cricket churping. These were the
samples with sounds of static microphone noise, thunder, wind, rain, auto vehicles, human conversation, sprinklers, and
relative silence, i.e., absence of any sounds discernable to a human ear.
2 Loading Audio Data
Let’s see how we can read the data in. The file random_audio_forests.py with the starter code has the function
read_data_audio that takes the base directory where beefv, cricketfv, and noisefv are placed and returns two
numpy arrays: data and target.
>>> base_dir = ’/home/vladimir/audio_data/’
1
>>> data, target = read_audio_data(base_dir)
>>> data.shape
(9110, 193)
>>> target.shape
(9110,)
The data array is a 9110 x 193 numpy matrix where each row is a 193-element audio feature vector. The target
array is a 1D numpy array with 9110 labels. The first 3,000 elements of this array are 0’s, the second 3,000 elements are
1’s, and the rest are 2’s, where 0 denotes bee buzzing, 1 – cricket chirping, 2 – ambient noise.
3 Growing Decision Trees
Let’s see how to classify the processed audio data with decision trees. Implement the function train_test_split_eval_dtr().
def train_test_split_eval_dtr(dtr, data, target, test_size=0.4):
pass
This function takes a decision tree and the data and target arrays, uses train_test_split to split the data and
target arrays into the train and test data and train and test target arrays using the value of the keyword parameter
test_size, fits the decision tree to the train data and target, predicts the targets of the test data, computes the decision
tree’s accuracy (DTR accuracy), the classification report, and the confusion matrix, and prints all of them. Here is a test
run.
>>> from sklearn import tree
>>> data, target = read_audio_data(base_dir)
>>> dtr = tree.DecisionTreeClassifier(random_state=random.randint(0, 100))
>>> train_test_split_eval_dtr(dtr, data, target)
DTR accuracy: 0.995334796926
Classification report for decision tree DecisionTreeClassifier(class_weight=None,
criterion=’gini’, max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
presort=False, random_state=47, splitter=’best’):
precision recall f1-score support
0 0.99 0.99 0.99 1178
1 1.00 1.00 1.00 1198
2 1.00 0.99 0.99 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1171 1 6]
[ 0 1198 0]
[ 10 0 1258]]
You may see and safely ignore a deprecation warning about rank in numpy.linalg.matrix.
4 Growing Random Forests
A random forest is a set of decision trees. Let’s generalize train_test_split_eval_dtr to random forests and implement
the function train_test_split_eval_rf().
def train_test_split_eval_rf(rf, data, target, test_size=0.4):
pass
This function takes a random forest rf and the data, target arrays, and the percent of of the data that should be
used for testing specified by test_size. The default value of test_size is 0.4, i.e., 40%. This function
1. uses train_test_split to split the data and target arrays into the train and test data using the value specified
by test_size;
2
2. fits the random forest to the train data and train target arrays;
3. predicts the targets of the test data;
4. computes the random forest’s accuracy (RF accuracy) and the confusion matrix;
5. prints the number of trees in the random forest, the random forest’s accuracy, the classification report, and the
confusion matrix.
>>> from sklearn.ensemble import RandomForestClassifier
>>> data, target = read_audio_data(’/home/vladimir/audio_data/’)
>>> rf = RandomForestClassifier(n_estimators=10, random_state=random.randint(0, 1000))
>>> train_test_split_eval_rf(rf, data, target)
# of trees in random forest: 10
RF accuracy: 0.997804610318
Classification report for decision tree RandomForestClassifier(bootstrap=True, class_weight=None,
criterion=’gini’,
max_depth=None, max_features=’auto’, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=10, n_jobs=1, oob_score=False, random_state=476,
verbose=0, warm_start=False):
precision recall f1-score support
0 0.99 1.00 1.00 1178
1 1.00 1.00 1.00 1198
2 1.00 1.00 1.00 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1176 0 2]
[ 0 1198 0]
[ 6 0 1262]]
5 Testing Trees and Forests
Are there any differences between decision trees and random forests on this data set? Let’s further generalize testing
trees and random forests. Implement the function train_test_split_dtr_range_eval().
def train_test_split_dtr_range_eval(n, data, target):
pass
This function takes the number of experiments n and the data and target arrays, creates n decision trees and calls
each train_test_split_eval_dtr, defined in Section 3, on each decision tree and the data and target arrays. Here is
a test run.
>>> from sklearn import tree
>>> data, target = read_audio_data(base_dir)
>>> train_test_split_dtr_range_eval(5, data, target)
Starting with train_test_split procedure
DTR accuracy: 0.995334796926
Classification report for decision tree DecisionTreeClassifier(class_weight=None,
criterion=’gini’, max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
presort=False, random_state=47, splitter=’best’):
precision recall f1-score support
0 0.99 0.99 0.99 1178
1 1.00 1.00 1.00 1198
2 1.00 0.99 0.99 1268
3
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1171 1 6]
[ 0 1198 0]
[ 10 0 1258]]
Starting with train_test_split procedure
DTR accuracy: 0.995883644347
Classification report for decision tree DecisionTreeClassifier(class_weight=None,
criterion=’gini’, max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
presort=False, random_state=50, splitter=’best’):
precision recall f1-score support
0 0.99 0.99 0.99 1178
1 1.00 1.00 1.00 1198
2 1.00 0.99 0.99 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1172 1 5]
[ 0 1198 0]
[ 9 0 1259]]
Starting with train_test_split procedure
DTR accuracy: 0.995060373216
Classification report for decision tree DecisionTreeClassifier(class_weight=None,
criterion=’gini’, max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
presort=False, random_state=72, splitter=’best’):
precision recall f1-score support
0 0.99 0.99 0.99 1178
1 1.00 1.00 1.00 1198
2 1.00 0.99 0.99 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1172 1 5]
[ 0 1198 0]
[ 11 1 1256]]
Starting with train_test_split procedure
DTR accuracy: 0.995609220637
Classification report for decision tree DecisionTreeClassifier(class_weight=None,
criterion=’gini’, max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
presort=False, random_state=25, splitter=’best’):
precision recall f1-score support
0 0.99 0.99 0.99 1178
4
1 1.00 1.00 1.00 1198
2 1.00 0.99 0.99 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1172 1 5]
[ 0 1198 0]
[ 8 2 1258]]
Starting with train_test_split procedure
DTR accuracy: 0.993962678375
Classification report for decision tree DecisionTreeClassifier(class_weight=None,
criterion=’gini’, max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
presort=False, random_state=3, splitter=’best’):
precision recall f1-score support
0 0.99 0.99 0.99 1178
1 0.99 1.00 1.00 1198
2 1.00 0.99 0.99 1268
avg / total 0.99 0.99 0.99 3644
Confusion matrix:
[[1171 1 6]
[ 0 1198 0]
[ 8 7 1253]]
Starting with train_test_split procedure
DTR accuracy: 0.996432491767
Classification report for decision tree DecisionTreeClassifier(class_weight=None,
criterion=’gini’, max_depth=None,
max_features=None, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
presort=False, random_state=83, splitter=’best’):
precision recall f1-score support
0 0.99 1.00 0.99 1178
1 1.00 1.00 1.00 1198
2 1.00 0.99 1.00 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1174 1 3]
[ 0 1198 0]
[ 9 0 1259]]
Finally, let’s implement the function train_test_split_rf_range_eval.
def train_test_split_rf_range_eval(nt_lower_bound, nt_upper_bound, data, target):
pass
This function generates random forests with the number of trees in [nt_lower_bound, nt_upper_bound] in increments of 10. For example, [10, 50], will create random forests with 10 trees, 20 trees, 30 trees, 40 trees, and 50 trees.
For each random forest rf, this function calls train_test_split_eval_rf on rf and the data and target arrays.
>>> train_test_split_rf_range_eval(10, 50, data, target)
5
Starting with train_test_split procedure
# of trees in random forest: 10
RF accuracy: 0.998353457739
Classification report for decision tree RandomForestClassifier(bootstrap=True,
class_weight=None, criterion=’gini’,
max_depth=None, max_features=’auto’, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=10, n_jobs=1, oob_score=False, random_state=780,
verbose=0, warm_start=False):
precision recall f1-score support
0 1.00 1.00 1.00 1178
1 1.00 1.00 1.00 1198
2 1.00 1.00 1.00 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1174 0 4]
[ 0 1198 0]
[ 2 0 1266]]
Starting with train_test_split procedure
# of trees in random forest: 20
RF accuracy: 0.997255762898
Classification report for decision tree RandomForestClassifier(bootstrap=True,
class_weight=None, criterion=’gini’,
max_depth=None, max_features=’auto’, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=20, n_jobs=1, oob_score=False, random_state=448,
verbose=0, warm_start=False):
precision recall f1-score support
0 0.99 1.00 1.00 1178
1 1.00 1.00 1.00 1198
2 1.00 0.99 1.00 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1177 0 1]
[ 0 1198 0]
[ 9 0 1259]]
Starting with train_test_split procedure
# of trees in random forest: 30
RF accuracy: 0.997530186608
Classification report for decision tree RandomForestClassifier(bootstrap=True,
class_weight=None, criterion=’gini’,
max_depth=None, max_features=’auto’, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=30, n_jobs=1, oob_score=False, random_state=849,
verbose=0, warm_start=False):
precision recall f1-score support
0 0.99 1.00 1.00 1178
1 1.00 1.00 1.00 1198
2 1.00 1.00 1.00 1268
6
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1175 0 3]
[ 0 1198 0]
[ 6 0 1262]]
Starting with train_test_split procedure
# of trees in random forest: 40
RF accuracy: 0.997804610318
Classification report for decision tree RandomForestClassifier(bootstrap=True,
class_weight=None, criterion=’gini’,
max_depth=None, max_features=’auto’, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=40, n_jobs=1, oob_score=False, random_state=569,
verbose=0, warm_start=False):
precision recall f1-score support
0 1.00 1.00 1.00 1178
1 1.00 1.00 1.00 1198
2 1.00 1.00 1.00 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1175 0 3]
[ 0 1198 0]
[ 5 0 1263]]
Starting with train_test_split procedure
# of trees in random forest: 50
RF accuracy: 0.998627881449
Classification report for decision tree RandomForestClassifier(bootstrap=True,
class_weight=None, criterion=’gini’,
max_depth=None, max_features=’auto’, max_leaf_nodes=None,
min_impurity_split=1e-07, min_samples_leaf=1,
min_samples_split=2, min_weight_fraction_leaf=0.0,
n_estimators=50, n_jobs=1, oob_score=False, random_state=609,
verbose=0, warm_start=False):
precision recall f1-score support
0 1.00 1.00 1.00 1178
1 1.00 1.00 1.00 1198
2 1.00 1.00 1.00 1268
avg / total 1.00 1.00 1.00 3644
Confusion matrix:
[[1177 0 1]
[ 0 1198 0]
[ 4 0 1264]]
What To Submit
Implement these functions in random_audio_forests.py and submit this file via Canvas. In a comment at the head of
the file, briefly answer the following question: Do you think that random forests buy us better performance on this data
set?
7
References
1. V. Kulyukin & S. Mukherjee. “Computer Vision in Low-Power Electronic Beehive Monitoring: In Situ Vision-Based
Bee Counting on Langstroth Hive Landing Pads.” Graphics, Vision, and Image Processing (GVIP), vol. 17, issue
1, pp. 25 – 37, June 2017, ISSN: 1687-398X.
2. V. Kulyukin & S. Reka. “Toward Sustainable Electronic Beehive Monitoring: Algorithms for Omnidirectional Bee
Counting from Images & Harmonic Analysis of Buzzing Signals.” Engineering Letters, vol. 24, no. 3, pp. 317-327,
Aug. 2016.
8
