A key assumption made by classic supervised text classification (or learning) is that classes appeared in the test data must have appeared in training, called the closed-world assumption (Fei and Liu, 2016; Chen and Liu, 2016)
. Although this assumption holds in many applications, it is violated in many others, especially in dynamic or open environments. For example, in social media, a classifier built with past topics or classes may not be effective in classifying future data because new topics appear constantly in social media(Fei et al., 2016). This is clearly true in other domains too, e.g., self-driving cars, where new objects may appear in the scene all the time.
Ideally, in the text domain, the classifier should classify incoming documents to the right existing classes used in training and also detect those documents that don’t belong to any of the existing classes. This problem is called open world classification or open classification (Fei and Liu, 2016). Such a classifier is aware what it does and does not know. This paper proposes a novel technique to solve this problem.
Problem Definition: Given the training data , where is the -th document, and is ’s class label, we want to build a model that can classify each test instance to one of the training or seen classes in or reject it to indicate that it does not belong to any of the training or seen classes, i.e., unseen. In other words, we want to build a -class classifier with the classes .
There are some prior approaches for open classification. One-class SVM (Schölkopf et al., 2001; Tax and Duin, 2004) is the earliest approach. However, as no negative training data is used, one-class classifiers work poorly. Fei and Liu (2016) proposed a Center-Based Similarity (CBS) space learning method (Fei and Liu, 2015)
. This method first computes a center for each class and transforms each document to a vector of similarities to the center. A binary classifier is then built using the transformed data for each class. The decision surface is like a “ball” encircling each class. Everything outside the ball is considered not belonging to the class. Our proposed method outperforms this method remarkably.Fei et al. (2016) further added the capability of incrementally or cumulatively learning new classes, which connects this work to lifelong learning (Chen and Liu, 2016) because without the ability to identify novel or new things and learn them, a system will never be able to learn by itself continually.
In computer vision,Scheirer et al. (2013) studied the problem of recognizing unseen images that the system was not trained for by reducing open space risk
. The basic idea is that a classifier should not cover too much open space where there are few or no training data. They proposed to reduce the half-space of a binary SVM classifier with a positive region bounded by two parallel hyperplanes. Similar works were also done in a probability setting byScheirer et al. (2014) and Jain et al. (2014). Both approaches use probability threshold, but choosing thresholds need prior knowledge, which is a weakness of the methods. Dalvi et al. (2013) proposed a multi-class semi-supervised method based on the EM algorithm. It has been shown that these methods are poorer than the method in (Fei and Liu, 2016).
The work closest to ours is that in (Bendale and Boult, 2016), which leverages an algorithm called OpenMax
to add the rejection capability by utilizing the logits that are trained via closed-world softmax function. One weak assumption of OpenMax is that examples with equally likely logits are more likely from the unseen or rejection class, which can be examples that are hard to classify. Another weakness is that it requires validation examples from the unseen/rejection class to tune the hyperparameters. Our method doesn’t make these weak assumptions and performs markedly better.
. Unlike traditional classifiers, DOC builds a multi-class classifier with a 1-vs-rest final layer of sigmoids rather than softmax to reduce the open space risk. It reduces the open space risk further for rejection by tightening the decision boundaries of sigmoid functions with Gaussian fitting. Experimental results show that DOC dramatically outperforms state-of-the-art existing approaches from both text classification and image classification domains.
2 The Proposed DOC Architecture
DOC uses CNN (Collobert et al., 2011; Kim, 2014) as its base and augments it with a 1-vs-rest final sigmoid layer and Gaussian fitting for classification. Note: other existing deep models like RNN (Williams and Zipser, 1989; Schuster and Paliwal, 1997) and LSTM (Hochreiter and Schmidhuber, 1997; Gers et al., 2002) can also be adopted as the base. Similar to RNN, CNN also works on embedded sequential data (using 1D convolution on text instead of 2D convolution on images). We choose CNN because OpenMax uses CNN and CNN performs well on text (Kim, 2014), which enables a fairer comparison with OpenMax.
2.1 CNN and Feed Forward Layers of DOC
The proposed DOC system (given in Fig. 1) is a variant of the CNN architecture (Collobert et al., 2011) for text classification (Kim, 2014)111https://github.com/alexander-rakhlin/CNN-for-Sentence-Classification-in-Keras. The first layer embeds words in document into dense vectors. The second layer performs convolution over dense vectors using different filters of varied sizes (see Sec. 3.4). Next, the max-over-time pooling layer selects the maximum values from the results of the convolution layer to form a -dimension feature vector . Then we reduce to a -dimension vector (
is the number of training/seen classes) via 2 fully connected layers and one intermediate ReLU activation layer:
where , , , and are trainable weights; is the output dimension of the first fully connected layer. The output layer of DOC is a 1-vs-rest layer applied to , which allows rejection. We describe it next.
2.2 1-vs-Rest Layer of DOC
Traditional multi-class classifiers (Goodfellow et al., 2016; Bendale and Boult, 2016) typically use softmax as the final output layer, which does not have the rejection capability since the probability of prediction for each class is normalized across all training/seen classes. Instead, we build a 1-vs-rest layer containing sigmoid functions for seen classes. For the -th sigmoid function corresponding to class , DOC takes all examples with as positive examples and all the rest examples as negative examples.
The model is trained with the objective of summation of all log loss of the sigmoid functions on the training data .
where is the indicator function and is the probability output from th sigmoid function on the th document’s th-dimension of .
During testing, we reinterpret the prediction of sigmoid functions to allow rejection, as shown in Eq. 3. For the -th sigmoid function, we check if the predicted probability is less than a threshold belonging to class . If all predicted probabilities are less than their corresponding thresholds for an example, the example is rejected; otherwise, its predicted class is the one with the highest probability. Formally, we have
Note that although multi-label classification (Huang et al., 2013; Zhang and Zhou, 2006; Tsoumakas and Katakis, 2006) may also leverage multiple sigmoid functions, Eq. 3 forbids multiple predicted labels for the same example, which is allowed in multi-label classification. DOC is also related to multi-task learning (Huang et al., 2013; Caruana, 1998), where each label is related to a 1-vs-rest binary classification task with shared representations from CNN and fully connected layers. However, Eq. 3 performs classification and rejection based on the outputs of these binary classification tasks.
Comparison with OpenMax
: OpenMax builds on the traditional closed-world multi-class classifier (softmax layer). It reduces the open space for each seen class, which is weak for rejecting unseen classes. DOC’s 1-vs-rest sigmoid layer provides a reasonable representation of all other classes (the rest of seen classes and unseen classes), and enables the 1 class forms a good boundary. Sec. 3.5 shows that this basic DOC is already much better than OpenMax. Below, we improve DOC further by tightening the decision boundaries more.
2.3 Reducing Open Space Risk Further
Sigmoid function usually uses the default probability threshold of = 0.5 for classification of each class . But this threshold does not consider potential open space risks from unseen (rejection) class data. We can improve the boundary by increasing . We use Fig. 2 to illustrate. The x-axis represents and y-axis is the predicted probability . The sigmoid function tries to push positive examples (belonging to the -th class) and negative examples (belonging to the other seen classes) away from the y-axis via a high gain around , which serves as the default decision boundary for with . As demonstrated by those 3 circles on the right-hand side of the y-axis, during testing, unseen class examples (circles) can easily fill in the gap between the y-axis and those dense positive (+) examples, which may reduce the recall of rejection and the precision of the -th seen class prediction. Obviously, a better decision boundary is at , where the decision boundary more closely “wrap” those dense positive examples with the probability threshold .
To obtain a better for each seen class
-th, we use the idea of outlier detection in statistics:
Assume the predicted probabilities of all training data of each class
follow one half of the Gaussian distribution (with mean), e.g., the three positive points in Fig. 2 projected to the y-axis (we don’t need ). We then artificially create the other half of the Gaussian distributed points (): for each existing point , we create a mirror point (not a probability) mirrored on the mean of 1.
In statistics, if a value/point is a certain number () of standard deviations away from the mean, it is considered an outlier. We thus set the probability threshold . The commonly used number for is 3, which also works well in our experiments.
Note that due to Gaussian fitting, different class can have a different classification threshold .
3 Experimental Evaluation
We perform evaluation using two publicly available datasets, which are exactly the same datasets used in (Fei and Liu, 2016).
(2) 50-class reviews (Chen and Liu, 2014): The dataset has Amazon reviews of 50 classes of products. Each class has 1000 reviews. Although product reviews are used, we do not do sentiment classification. We still perform topic-based classification. That is, given a review, the system decides what class of product the review is about.
For every dataset, we keep a 20000 frequent word vocabulary. Each document is fixed to 2000-word length (cutting or padding when necessary).
3.2 Test Settings and Evaluation Metrics
For a fair comparison, we use exactly the same settings as in (Fei and Liu, 2016). For each class in each dataset, we randomly sampled 60% of documents for training, 10% for validation and 30% for testing. Fei and Liu (2016) did not use a validation set, but the test data is the same 30%. We use the validation set to avoid overfitting. For open-world evaluation, we hold out some classes (as unseen) in training and mix them back during testing. We vary the number of training classes and use 25%, 50%, 75%, or 100% classes for training and all classes for testing. Here using 100% classes for training is the same as the traditional closed-world classification. Taking 20 newsgroups as an example, for 25% classes, we use 5 classes (we randomly choose 5 classes from 20 classes for 10 times and average the results, as in (Fei and Liu, 2016)) for training and all 20 classes for testing (15 classes are unseen in training). We use macro -score over classes (1 for rejection) for evaluation. Please note that examples from unseen classes are dropped in the validation set.
We compare DOC with two state-of-the-art methods published in 2016 and one DOC variant.
cbsSVM: This is the latest method published in NLP (Fei and Liu, 2016). It uses SVM to build 1-vs-rest CBS classifiers for multiclass text classification with rejection option. The results of this system are taken from (Fei and Liu, 2016).
OpenMax: This is the latest method from computer vision (Bendale and Boult, 2016). Since it is a CNN-based method for image classification, we adapt it for text classification by using CNN with a softmax output layer, and adopt the OpenMax layer333https://github.com/abhijitbendale/OSDN for open text classification. When all classes are seen (100%), the result from softmax is reported since OpenMax layer always performs rejection. We use default hyperparameter values of OpenMax (Weibull tail size is set to 20).
DOC: This is the basic DOC . Gaussian fitting isn’t used to choose each .
Note that (Fei and Liu, 2016) compared with several other baselines. We don’t compare with them as it was shown that cbsSVM was superior.
3.4 Hyperparameter Setting
We use word vectors pre-trained from Google News444https://code.google.com/archive/p/word2vec/ (3 million words and 300 dimensions). For the CNN layers, 3 filter sizes are used . For each filter size, 150 filters are applied. The dimension of the first fully connected layer is 250.
|% of seen classes||25%||50%||75%||100%|
|% of seen classes||25%||50%||75%||100%|
3.5 Result Analysis
DOC is markedly better than OpenMax and cbsSVM in macro- scores for both datasets in the 25%, 50%, and 75% settings. For the 25% and 50% settings (most test examples are from unseen classes), DOC is dramatically better. Even for 100% of traditional closed-world classification, it is consistently better too. DOC() is better too.
For the 25% and 50% settings, DOC is also markedly better than DOC(), which shows that Gaussian fitting finds a better probability threshold than when many unseen classes are present. In the 75% setting (most test examples are from seen classes), DOC() is slightly better for 20 newsgroups but worse for 50-class reviews. DOC sacrifices some recall of seen class examples for better precision, while sacrifices the precision of seen classes for better recall. DOC() is also worse than cbsSVM for 25% setting for 50-class reviews. It is thus not as robust as DOC.
For the 25% and 50% settings, cbsSVM is also markedly better than OpenMax.
This paper proposed a novel deep learning based method, called DOC, for open text classification. Using the same text datasets and experiment settings, we showed that DOC performs dramatically better than the state-of-the-art methods from both the text and image classification domains. We also believe that DOC is applicable to images.
In our future work, we plan to improve the cumulative or incremental learning method in (Fei et al., 2016) to learn new classes without training on all past and new classes of data from scratch. This will enable the system to learn by self to achieve continual or lifelong learning (Chen and Liu, 2016). We also plan to improve model performance during testing (Shu et al., 2017).
This work was supported in part by grants from National Science Foundation (NSF) under grant no. IIS-1407927 and IIS-1650900.
Bendale and Boult (2016)
Abhijit Bendale and Terrance E Boult. 2016.
Towards open set deep networks.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pages 1563–1572.
- Caruana (1998) Rich Caruana. 1998. Multitask learning. In Learning to learn, Springer, pages 95–133.
- Chen and Liu (2014) Zhiyuan Chen and Bing Liu. 2014. Mining topics in documents: standing on the shoulders of big data. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, pages 1116–1125.
Chen and Liu (2016)
Zhiyuan Chen and Bing Liu. 2016.
Lifelong Machine Learning. Morgan & Claypool Publishers.
- Collobert et al. (2011) Ronan Collobert, Jason Weston, Léon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. 2011. Natural language processing (almost) from scratch. Journal of Machine Learning Research 12(Aug):2493–2537.
- Dalvi et al. (2013) Bhavana Dalvi, William W Cohen, and Jamie Callan. 2013. Exploratory learning. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, pages 128–143.
- Fei and Liu (2015) Geli Fei and Bing Liu. 2015. Social media text classification under negative covariate shift. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP-2015).
- Fei and Liu (2016) Geli Fei and Bing Liu. 2016. Breaking the closed world assumption in text classification. In Proceedings of NAACL-HLT. pages 506–514.
- Fei et al. (2016) Geli Fei, Shuai Wang, and Bing Liu. 2016. Learning cumulatively to become more knowledgeable. In Proceedings of SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2016).
- Gers et al. (2002) Felix A Gers, Nicol N Schraudolph, and Jürgen Schmidhuber. 2002. Learning precise timing with lstm recurrent networks. Journal of machine learning research 3(Aug):115–143.
- Goodfellow et al. (2016) Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016. Deep Learning. MIT Press. http://www.deeplearningbook.org.
- Hochreiter and Schmidhuber (1997) Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation 9(8):1735–1780.
Huang et al. (2013)
Yan Huang, Wei Wang, Liang Wang, and Tieniu Tan. 2013.
Multi-task deep neural network for multi-label learning.In Image Processing (ICIP), 2013 20th IEEE International Conference on. IEEE, pages 2897–2900.
- Jain et al. (2014) Lalit P Jain, Walter J Scheirer, and Terrance E Boult. 2014. Multi-class open set recognition using probability of inclusion. In European Conference on Computer Vision. Springer, pages 393–409.
- Kim (2014) Yoon Kim. 2014. Convolutional neural networks for sentence classification. arXiv preprint arXiv:1408.5882 .
- Rennie (2008) Jason Rennie. 2008. 20 newsgroup dataset.
- Scheirer et al. (2013) Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. 2013. Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(7):1757–1772.
- Scheirer et al. (2014) Walter J Scheirer, Lalit P Jain, and Terrance E Boult. 2014. Probability models for open set recognition. IEEE transactions on pattern analysis and machine intelligence 36(11):2317–2324.
- Schölkopf et al. (2001) Bernhard Schölkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. 2001. Estimating the support of a high-dimensional distribution. Neural computation 13(7):1443–1471.
- Schuster and Paliwal (1997) Mike Schuster and Kuldip K Paliwal. 1997. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11):2673–2681.
- Shu et al. (2017) Lei Shu, Hu Xu, and Bing Liu. 2017. Lifelong learning crf for supervised aspect extraction. In Proceedings of Annual Meeting of the Association for Computational Linguistics (ACL-2017, short paper).
- Tax and Duin (2004) David MJ Tax and Robert PW Duin. 2004. Support vector data description. Machine learning 54(1):45–66.
- Tsoumakas and Katakis (2006) Grigorios Tsoumakas and Ioannis Katakis. 2006. Multi-label classification: An overview. International Journal of Data Warehousing and Mining 3(3).
- Williams and Zipser (1989) Ronald J Williams and David Zipser. 1989. A learning algorithm for continually running fully recurrent neural networks. Neural computation 1(2):270–280.
- Zhang and Zhou (2006) Min-Ling Zhang and Zhi-Hua Zhou. 2006. Multilabel neural networks with applications to functional genomics and text categorization. IEEE transactions on Knowledge and Data Engineering 18(10):1338–1351.