Method and system for learning transferable feature representations from a source domain for a target domain
The disclosed embodiments illustrate a domain adaptation method for learning transferable feature representations from a source domain for a target domain. The method includes receiving input data comprising a plurality of labeled instances of the source domain and a plurality of unlabeled instances of the target domain. The method includes learning common representation shared between the source domain and the target domain, based on the plurality of labeled instances of the source domain. The method includes labeling one or more unlabeled instances in the plurality of unlabeled instances of the target domain, based on the common representation. The method includes determining a target specific representation corresponding to the target domain. The method includes training a target specific classifier based on the target specific representation and the common representation to perform text classification on remaining one or more unlabeled instances of the plurality of unlabeled instances of the target domain.
1. A domain adaptation method for learning transferable feature representations from a source domain for a target domain, the domain adaptation method comprising:
formulating, by one or more processors, first and second neural networks, wherein the first neural network comprises a first hidden layer and the second neural network comprises a second hidden layer, and both the first and second hidden layers are partitioned;
receiving, by the one or more processors, real-time input data comprising a plurality of labeled instances of the source domain and a plurality of unlabeled instances of the target domain;
inputting, by the one or more processors, the plurality of labeled instances of the source domain into the first neural network;
learning, by the first neural network a, a common representation shared between the source domain and the target domain, based on the plurality of labeled instances of the source domain, wherein the first hidden layer comprises and is partitioned between the common representation and a source specific representation;
inputting, by the one or more processors, the unlabeled instances of the target domain into the second neural network;
labeling, by the second neural network, one or more unlabeled instances in the plurality of unlabeled instances of the target domain, based on the learned common representation;
determining, by the second neural network, a target specific representation corresponding to the target domain, based on the one or more labeled instances of the target domain, wherein the second hidden layer comprises and is partitioned between the learned common representation and the target specific representation; and
training, by the second neural network, a target specific classifier based on the determined target specific representation and the learned common representation to perform automatic text classification on remaining one or more unlabeled instances of the plurality of unlabeled instances of the target domain.
2. The domain adaptation method of claim 1 , wherein the determination of the target specific representation is an iterative process, wherein the target specific representation is updated by addition of new target specific features in each iteration in the iterative process.
3. The domain adaptation method of claim 2 , wherein the common representation comprises a plurality of common features shared between the source domain and the target domain, wherein the target specific representation comprises a plurality of target specific features corresponding to the target domain, wherein the plurality of target specific features is identified based on a positive contribution of the plurality of target specific features.
4. The domain adaptation method of claim 3 , further comprising merging, by the second neural network, a first target specific feature in the plurality of target specific features with a second target specific feature in the plurality of target specific features to obtain a single feature, based on a similarity score between the first target specific feature and the second target specific feature.
5. The domain adaptation method of claim 1 , wherein:
the learning of the common representation is based on a minimization of a domain divergence between the source domain and the target domain; and
the minimization of the domain divergence between the source domain and the target domain is based on a prediction of a likelihood that an instance is from the source domain or the target domain.
6. The domain adaptation method of claim 1 , further comprising:
learning, by the first neural network, the source specific representation corresponding to the source domain, based on the plurality of labeled instances of the source domain, wherein the source specific representation comprises a plurality of source specific features corresponding to the source domain; and
classifying, by the first neural network, a plurality of unlabeled instances of the source domain, based on the source specific representation and the common representation.
7. The domain adaptation method of claim 6 , wherein the classification of the plurality of unlabeled instances of the source domain is based on a minimization of a source classification error such that the source specific representation and the common representation contribute positively to the source classification error.
8. The domain adaptation method of claim 1 , wherein:
the first neural network comprises a first input layer, the first hidden layer, and a first output layer; and
the second neural network comprises a second input layer, the second hidden layer, and a second output layer.
9. A domain adaptation system for learning transferable feature representations from a source domain for a target domain, the domain adaptation system comprising:
one or more processors configured to:
formulate first and second neural networks, wherein the first neural network comprises a first hidden layer and the second neural network comprises a second hidden layer, and both the first and second hidden layers are partitioned,
receive real-time input data comprising a plurality of labeled instances of the source domain and a plurality of unlabeled instances of the target domain,
input the plurality of labeled instances of the source domain into the first neural network, and
input the unlabeled instances of the target domain into the second neural network;
the first neural network configured to learn a common representation shared between the source domain and the target domain, based on the plurality of labeled instances of the source domain, wherein the first hidden layer comprises and is partitioned between the common representation and a source specific representation; and
the second neural network configured to:
label one or more unlabeled instances in the plurality of unlabeled instances of the target domain, based on the learned common representation,
determine a target specific representation corresponding to the target domain, based on the one or more labeled instances of the target domain, wherein the second hidden layer comprises and is partitioned between the learned common representation and the target specific representation, and
train a target specific classifier based on the determined target specific representation and the learned common representation to perform automatic text classification on remaining one or more unlabeled instances of the plurality of unlabeled instances of the target domain.
10. The domain adaptation system of claim 9 , wherein the determination of the target specific representation is an iterative process, wherein the target specific representation is updated by addition of new target specific features in each iteration in the iterative process.
11. The domain adaptation system of claim 10 , wherein the common representation comprises a plurality of common features shared between the source domain and the target domain, wherein the target specific representation comprises a plurality of target specific features corresponding to the target domain, wherein the plurality of target specific features is identified based on a positive contribution of the plurality of target specific features.
12. The domain adaptation system of claim 11 , wherein the second neural network is further configured to merge a first target specific feature in the plurality of target specific features with a second target specific feature in the plurality of target specific features to obtain a single feature, based on a similarity score between the first target specific feature and the second target specific feature.
13. The domain adaptation system of claim 9 , wherein the first neural network is further configured to:
learn the source specific representation corresponding to the source domain, based on the plurality of labeled instances of the source domain, wherein the source specific representation comprises a plurality of source specific features corresponding to the source domain; and
classify a plurality of unlabeled instances of the source domain, based on the source specific representation and the common representation.
14. The domain adaptation system of claim 9 , wherein:
the first neural network comprises a first input layer, the first hidden layer, and a first output layer; and
the second neural network comprises a second input layer, the second hidden layer, and a second output layer.
15. A computer program product for use with a computer, the computer program product comprising a non-transitory computer readable medium, wherein the non-transitory computer readable medium stores a computer program code for domain adaptation for learning transferable feature representations from a source domain for a target domain, wherein the computer program code is executable by one or more processors to:
formulate first and second neural networks, wherein the first neural network comprises a first hidden layer and the second neural network comprises a second hidden layer, and both the first and second hidden layers are partitioned;
receive real-time input data comprising a plurality of labeled instances of the source domain and a plurality of unlabeled instances of the target domain;
input the plurality of labeled instances of the source domain into the first neural network, and
input the unlabeled instances of the target domain into the second neural network;
learn, by the first neural network, a common representation shared between the source domain and the target domain, based on the plurality of labeled instances of the source domain, wherein the first hidden layer comprises and is partitioned between the common representation and a source specific representation;
label, by the second neural network, one or more unlabeled instances in the plurality of unlabeled instances of the target domain, based on the learned common representation;
determine, by the second neural network, a target specific representation corresponding to the target domain, based on the one or more labeled instances of the target domain, wherein the second hidden layer comprises and is partitioned between the learned common representation and the target specific representation; and
train, by the second neural network, a target specific classifier based on the determined target specific representation and the learned common representation to perform automatic text classification on remaining one or more unlabeled instances of the plurality of unlabeled instances of the target domain.
16. The computer program product of claim 15 , wherein:
the first neural network comprises a first input layer, the first hidden layer, and a first output layer; and
the second neural network comprises a second input layer, the second hidden layer, and a second output layer.