Reuters is a benchmark dataset for document classification. To be more precise, it is a multi-class (e.g. there are multiple classes), multi-label (e.g. each document can belong to many classes) dataset. It has 90 classes, 7769 training documents and 3019 testing documents. It is the ModApte (R(90)) subest of the Reuters-21578 benchmark (source).
The mean number of words per document, grouped by class, is between 93 and 1263 on the training set.
The training set has a vocabulary size of 35247. Even if you restrict it to words which appear at least 5 times and at most 12672 times in the training set, there are still 12017 words.
Classes and Labels
nr of documents mean number of
class name train test words in train set
1: earn : 2877 1087 104.4
2: acq : 1650 719 150.1
3: money-fx : 538 179 219.0
4: grain : 433 149 223.6
5: crude : 389 189 247.3
6: trade : 368 117 294.3
7: interest : 347 131 198.0
8: wheat : 212 71 225.6
9: ship : 197 89 203.7
10: corn : 181 56 259.1
11: money-supply : 140 34 170.5
12: dlr : 131 44 230.4
13: sugar : 126 36 247.2
14: oilseed : 124 47 277.9
15: coffee : 111 28 264.1
16: gnp : 101 35 372.8
17: gold : 94 30 188.5
18: veg-oil : 87 37 291.0
19: soybean : 78 33 347.9
20: livestock : 75 24 222.9
21: nat-gas : 75 30 257.7
22: bop : 75 30 288.1
23: cpi : 69 28 235.4
24: cocoa : 55 18 266.4
25: reserves : 55 18 216.1
26: carcass : 50 18 259.7
27: copper : 47 18 201.6
28: jobs : 46 21 232.7
29: yen : 45 14 282.8
30: ipi : 41 12 232.1
31: iron-steel : 40 14 220.5
32: cotton : 39 20 366.3
33: barley : 37 14 272.9
34: gas : 37 17 209.4
35: rubber : 37 12 274.5
36: alum : 35 23 180.5
37: rice : 35 24 359.8
38: palm-oil : 30 10 234.5
39: meal-feed : 30 19 387.1
40: sorghum : 24 10 511.3
41: retail : 23 2 324.8
42: zinc : 21 13 189.9
43: silver : 21 8 221.0
44: pet-chem : 20 12 204.9
45: wpi : 19 10 200.9
46: tin : 18 12 322.1
47: rapeseed : 18 9 168.6
48: orange : 16 11 169.6
49: strategic-metal : 16 11 205.6
50: housing : 16 4 207.2
51: hog : 16 6 162.3
52: lead : 15 14 216.7
53: soy-oil : 14 11 568.7
54: heat : 14 5 190.4
55: fuel : 13 10 194.6
56: soy-meal : 13 13 551.0
57: lei : 12 3 134.7
58: sunseed : 11 5 425.1
59: dmk : 10 4 212.1
60: lumber : 10 6 242.1
61: tea : 9 4 365.1
62: income : 9 7 286.4
63: nickel : 8 1 193.6
64: oat : 8 6 648.4
65: l-cattle : 6 2 298.0
66: instal-debt : 5 1 134.3
67: platinum : 5 7 174.8
68: groundnut : 5 4 258.0
69: rape-oil : 5 3 167.2
70: sun-oil : 5 2 201.9
71: coconut-oil : 4 3 471.4
72: jet : 4 1 109.6
73: coconut : 4 2 264.5
74: propane : 3 3 352.0
75: potato : 3 3 161.2
76: cpu : 3 1 133.2
77: rand : 2 1 345.7
78: palmkernel : 2 1 326.3
79: copra-cake : 2 1 495.0
80: dfl : 2 1 764.7
81: naphtha : 2 4 206.7
82: palladium : 2 1 93.0
83: nzdlr : 2 2 508.8
84: groundnut-oil : 1 1 277.5
85: castor-oil : 1 1 194.0
86: sun-meal : 1 1 153.0
87: lin-oil : 1 1 262.5
88: cotton-oil : 1 2 1262.7
89: rye : 1 1 383.0
90: nkr : 1 2 122.3
By far most documents have either one or two labels, but some have up to 15:
labelcount= 1, documentcount=9160
labelcount= 2, documentcount=1173
labelcount= 3, documentcount= 255
labelcount= 4, documentcount= 91
labelcount= 5, documentcount= 52
labelcount= 6, documentcount= 27
labelcount= 7, documentcount= 9
labelcount= 8, documentcount= 7
labelcount= 9, documentcount= 5
labelcount=10, documentcount= 3
labelcount=11, documentcount= 2
labelcount=14, documentcount= 2
labelcount=12, documentcount= 1
labelcount=15, documentcount= 1
Let's look at the relationship between the classes. Which classes occur often
together? Are there classes which can be used to predict the presence of other
classes? For example, wheat should imply grain.
Here are the 50 strongest predictors:
castor-oil => cotton-oil (0.999742566611)
castor-oil => groundnut-oil (0.999742566611)
castor-oil => lin-oil (0.999742566611)
castor-oil => nkr (0.999742566611)
castor-oil => rye (0.999742566611)
castor-oil => sun-meal (0.999742566611)
copra-cake => palmkernel (0.999742566611)
cotton-oil => castor-oil (0.999742566611)
cotton-oil => groundnut-oil (0.999742566611)
cotton-oil => lin-oil (0.999742566611)
cotton-oil => nkr (0.999742566611)
cotton-oil => rye (0.999742566611)
cotton-oil => sun-meal (0.999742566611)
groundnut-oil => castor-oil (0.999742566611)
groundnut-oil => cotton-oil (0.999742566611)
groundnut-oil => lin-oil (0.999742566611)
groundnut-oil => nkr (0.999742566611)
groundnut-oil => rye (0.999742566611)
groundnut-oil => sun-meal (0.999742566611)
lin-oil => castor-oil (0.999742566611)
lin-oil => cotton-oil (0.999742566611)
lin-oil => groundnut-oil (0.999742566611)
lin-oil => nkr (0.999742566611)
lin-oil => rye (0.999742566611)
lin-oil => sun-meal (0.999742566611)
nkr => castor-oil (0.999742566611)
nkr => cotton-oil (0.999742566611)
nkr => groundnut-oil (0.999742566611)
nkr => lin-oil (0.999742566611)
nkr => rye (0.999742566611)
nkr => sun-meal (0.999742566611)
palmkernel => copra-cake (0.999742566611)
rye => castor-oil (0.999742566611)
rye => cotton-oil (0.999742566611)
rye => groundnut-oil (0.999742566611)
rye => lin-oil (0.999742566611)
rye => nkr (0.999742566611)
rye => sun-meal (0.999742566611)
sun-meal => castor-oil (0.999742566611)
sun-meal => cotton-oil (0.999742566611)
sun-meal => groundnut-oil (0.999742566611)
sun-meal => lin-oil (0.999742566611)
sun-meal => nkr (0.999742566611)
sun-meal => rye (0.999742566611)
castor-oil => copra-cake (0.999613849916)
castor-oil => dfl (0.999613849916)
castor-oil => naphtha (0.999613849916)
castor-oil => nzdlr (0.999613849916)
castor-oil => palladium (0.999613849916)
castor-oil => palmkernel (0.999613849916)
Multi-label Scoring
I have to work on the scoring; it seems as if I made a mistake in my experiments when I calculated the score.
I've never been in a multi-label context, so it was not directly clear to me which scoring is used. Thanks to Chirag Nagpal who pointed me in the right direction.
Let \(Y_i \in \{0, 1\}^k\) be the set of correct labels for document \(i\) and \(Z_i \in \{0, 1\}^k\) be the set of predicted labels:
- Binary Accuracy: For each document, one has to make a decision for each possible category. Hence \(\text{acc} = \frac{1}{n}\sum_{i=1}^n \frac{|Y_i \cap Z_i|}{k}\). As most documents belong to one or two categories, the simplest classifier simply decides all the time that the document does not belong to any category. For the used dataset, this leads to an accuracy of 0.986. Hence accuracy is not suitable.
- Subset Accuracy: This is calculated by
sklearn. The set of predicted labels must be exactly the same as the true labels. - F1 score: See user manual
- Micro/macro averaged ROC or Precision/Recall curve:
- Micro: Calculate metrics globally by counting the total true positives, false negatives and false positives.
- Macro: Calculate metrics for each label, and find their unweighted mean. This does not take label imbalance into account.
- Coverage error: See user manual
A nice overview is given by A Literature Survey on Algorithms for Multi-label Learning.
n-gram Features
See N-Gram-Based Text Categorization: Categorizing Text With Python.
I will give this a try when I find some time. If you make it before, please share the results. You could adjust my tf-idf code.
Classifier comparison (tf-idf)
The following are the accuracies as well as the training and test times. All classifiers got the same tf-idf features.
Used vocabulary size = 26147
Classifier Acc F1
--------------------------------------------------------------------------------
MLP : 82.61% 85.56% in 676.44s train / 0.91s test
LinearSVC : 81.05% 84.04% in 27.00s train / 6.45s test
Logistic Regression (C=1000) : 80.79% 84.10% in 35.48s train / 6.46s test
k nn 5 : 72.97% 76.07% in 9.89s train / 1092.29s test
k nn 3 : 72.28% 75.43% in 9.90s train / 1080.20s test
Logistic Regression (C=1) : 67.47% 67.21% in 30.22s train / 6.44s test
Random Forest (200 estimators): 65.75% 64.36% in 57.56s train / 4.15s test
Random Forest (50 estimators) : 64.79% 63.69% in 14.82s train / 1.30s test
Decision Tree : 55.75% 53.23% in 28.43s train / 0.22s test
Naive Bayes : 43.86% 47.98% in 214.75s train / 88.79s test
SVM, linear : 33.55% 29.67% in 6326.33s train / 2397.51s test
There are a couple of things to notice here:
- Speed:
- Naive Bayes and k-nn is slow
- Random forests and MLP are fast
- SVM depends extremely on the implementation (see What is the difference between LinearSVC and SVC(kernel=“linear”)?)
- Prediction Quality:
- LinearSVC, logistic regression and MLP are accurate
- Achieving high binary accuracy seems to be easier than achieving high F1 scores. It's no surprise that high subset accuracy is hard to achieve.
The MLP has a reasonable prediction quality and test time.
Multilayer Perceptron
When training a multilayer perceptron for a multi-label classification task, there are two important things to keep in mind:
- Output layer: Do not use softmax, as the normalization does not make sense in this case.
- Loss: Use
`binary_crossentropy
When you print precision, recall, F1-score and accuracy you note the following:
- Binary accuracy gets to 98% in the first epoch and over 99% in the second. It stays that high.
- Precision is at about 4% in the first epoch and over 97% in the second. It stays that high.
- Recall needs about 15 epochs of steady progress to get over 98%.
As a consequence, the F1 score steadily increases. Hence by using other metrics one can see that the classifier makes great improvements, although the accuracy is pretty high from the beginning.
Code
See GitHub.
If you use it, please cite this article or link to this blog post:
@Misc{Thom2017-reuters,
Title = {The Reuters Dataset},
Author = {Martin Thoma},
Month = jul,
Year = {2017},
Url = {https://martin-thoma.com/nlp-reuters}
}
Data loading
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
Utility file for the Reuters text categorization benchmark dataset.
See also
--------
http://www.vision.caltech.edu/Image_Datasets/Caltech101/
"""
from nltk.corpus import reuters
from nltk.corpus import stopwords
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.feature_extraction.text import TfidfVectorizer
n_classes = 90
labels = reuters.categories()
def load_data(config={}):
"""
Load the Reuters dataset.
Returns
-------
Tuple of Numpy arrays: `(x_train, y_train), (x_test, y_test)`.
"""
stop_words = stopwords.words("english")
vectorizer = TfidfVectorizer(stop_words=stop_words)
mlb = MultiLabelBinarizer()
documents = reuters.fileids()
test = [d for d in documents if d.startswith("test/")]
train = [d for d in documents if d.startswith("training/")]
docs = {}
docs["train"] = [reuters.raw(doc_id) for doc_id in train]
docs["test"] = [reuters.raw(doc_id) for doc_id in test]
xs = {"train": [], "test": []}
xs["train"] = vectorizer.fit_transform(docs["train"]).toarray()
xs["test"] = vectorizer.transform(docs["test"]).toarray()
ys = {"train": [], "test": []}
ys["train"] = mlb.fit_transform([reuters.categories(doc_id) for doc_id in train])
ys["test"] = mlb.transform([reuters.categories(doc_id) for doc_id in test])
data = {
"x_train": xs["train"],
"y_train": ys["train"],
"x_test": xs["test"],
"y_test": ys["test"],
"labels": globals()["labels"],
}
return data
if __name__ == "__main__":
config = {}
data = load_data(config)
print("len(data['x_train'])={}".format(len(data["x_train"])))
print("data['x_train'].shape={}".format(data["x_train"].shape))
MLP
#!/usr/bin/env python
"""Train and evaluate a MLP."""
import time
from keras.layers import Activation, Input, Dropout
from keras.layers import Dense
from keras.models import Model
from keras.optimizers import Adam
import reuters
from keras import backend as K
from scoring import get_tptnfpfn, get_accuracy, get_f_score
def create_model(nb_classes, input_shape):
"""Create a MLP model."""
input_ = Input(shape=input_shape)
x = input_
x = Dense(256, activation="relu")(x)
x = Dropout(0.5)(x)
x = Dense(256, activation="relu")(x)
x = Dense(nb_classes)(x)
x = Activation("sigmoid")(x)
model = Model(inputs=input_, outputs=x)
return model
def recall(y_true, y_pred):
"""
Recall metric.
Only computes a batch-wise average of recall.
Computes the recall, a metric for multi-label classification of
how many relevant items are selected.
"""
true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)))
recall = true_positives / (possible_positives + K.epsilon())
return recall
def precision(y_true, y_pred):
"""
Precision metric.
Only computes a batch-wise average of precision.
Computes the precision, a metric for multi-label classification of
how many selected items are relevant.
Source
------
https://github.com/fchollet/keras/issues/5400#issuecomment-314747992
"""
true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))
predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1)))
precision = true_positives / (predicted_positives + K.epsilon())
return precision
def f1(y_true, y_pred):
"""Calculate the F1 score."""
p = precision(y_true, y_pred)
r = recall(y_true, y_pred)
return 2 * ((p * r) / (p + r))
def get_optimizer(config):
"""Return an optimizer."""
lr = config["optimizer"]["initial_lr"]
optimizer = Adam(lr=lr) # Using Adam instead of SGD to speed up training
return optimizer
def main(data_module):
"""Load data, train model and evaluate it."""
data = data_module.load_data()
model = create_model(data_module.n_classes, (data["x_train"].shape[1],))
print(model.summary())
optimizer = get_optimizer({"optimizer": {"initial_lr": 0.001}})
model.compile(
loss="binary_crossentropy",
optimizer=optimizer,
metrics=[precision, recall, f1, "accuracy"],
)
t0 = time.time()
model.fit(
data["x_train"],
data["y_train"],
batch_size=32,
epochs=20,
validation_data=(data["x_test"], data["y_test"]),
shuffle=True,
# callbacks=callbacks
)
t1 = time.time()
res = get_tptnfpfn(model, data)
t2 = time.time()
print(
(
"{clf_name:<30}: {acc:0.2f}% {f1:0.2f}% in {train_time:0.2f}s "
"train / {test_time:0.2f}s test"
).format(
clf_name="MLP",
acc=(get_accuracy(res) * 100),
f1=(get_f_score(res) * 100),
train_time=(t1 - t0),
test_time=(t2 - t1),
)
)
if __name__ == "__main__":
main(reuters)
See also
- NLTK: Other datasets
- Reuters-21578
- Publications:
- Blog posts:
- Miguel Malvarez: Classifying Reuters-21578 collection with Python: Representing the data
- Miguel Malvarez: Classifying Reuters-21578 collection with Python