Feature extraction using multi-task learning

1. A method, comprising:

obtaining, by a deep neural network (DNN), training data, the deep neural network comprising at least one hidden layer;

obtaining, from a specified hidden layer of the at least one hidden layer, features of the training data, the specified hidden layer being connected respectively to a supervised classification network for classification tasks and an autoencoder based reconstruction network for reconstruction tasks;

training, based on the obtained features, the DNN, the supervised classification network and the reconstruction network as a whole, the training being guided by the classification tasks and the reconstruction tasks;

determining, based on labeled data in the training data, a classification loss of the supervised classification network;

adjusting, via back propagation (BP) of the classification loss, the supervised classification network and the DNN to obtain a first adjusted DNN and a first adjusted supervised classification network;

determining, based on all the training data and the first adjusted supervised classification network, a classification loss of the supervised classification network and reconstruction loss of the autoencoder based reconstruction network; and

adjusting, via BP of the classification loss and the reconstruction loss, the first adjusted supervised classification network, the first adjusted DNN, the autoencoder based reconstruction network to obtain a second adjusted DNN and a second adjusted supervised classification network and an adjusted autoencoder based reconstruction network.

2. The method of claim 1 , further comprising:

determining a feature output interface, which is selected from the group consisted of the following: one of the hidden layer(s) of the deep neural network, or an intermediated hidden layer of the reconstruction network.

3. The method of claim 2 , further comprising:

obtaining, by the deep neural network, testing data, the testing data being different from the training data;

obtaining, from the feature output interface, features of the testing data;

performing, based on the obtained features of the testing data, at least one operation including at least one of clustering and classification;

calculating at least one accuracy value including at least one of clustering accuracy and classification accuracy of the testing data; and

in response to the accuracy being larger than a predetermined threshold, determining that the training is complete, wherein a trained network comprises the trained deep neural network, the trained supervised classification network and the trained autoencoder based reconstruction network and an output of the trained network is the output of the feature output interface.

4. The method of claim 3 , wherein the training data include trip data, each trip comprising at least one segment, the method further comprising:

obtaining, by the DNN, at least one of the segments as the training data.

5. The method of claim 4 , wherein the testing data comprises trip data, each trip comprising at least one segment, the method further comprising:

obtaining, by the DNN, at least one of the segments as the testing data.

6. The method of claim 1 , wherein the classification loss is defined according to labels of the training data, and the reconstruction loss contains an error term determined based on the features outputted from the specified hidden layer.

7. The method of claim 6 , wherein, the reconstruction loss further comprises a regularization term, the regularization term is determined based on L1 norm.

8. A computer system, comprising:

a processor;

a computer-readable memory coupled to the processor, the memory comprising instructions that when executed by the processor perform actions of:

obtaining, by a deep neural network (DNN), training data, the deep neural network comprising at least one hidden layer;

obtaining, from a specified hidden layer of the at least one hidden layer, features of the training data, the specified hidden layer being connected respectively to a supervised classification network for classification tasks and an autoencoder based reconstruction network for reconstruction tasks;

training, based on the obtained features, the DNN, the supervised classification network and the reconstruction network as a whole, the training being guided by the classification tasks and the reconstruction tasks;

determining, based on labeled data in the training data, a classification loss of the supervised classification network;

adjusting, via back propagation (BP) of the classification loss, the supervised classification network and the DNN to obtain a first adjusted DNN and a first adjusted supervised classification network;

determining, based on all the training data and the first adjusted supervised classification network, a classification loss of the supervised classification network and reconstruction loss of the autoencoder based reconstruction network; and

adjusting, via BP of the classification loss and the reconstruction loss, the first adjusted supervised classification network, the first adjusted DNN, the autoencoder based reconstruction network to obtain a second adjusted DNN and a second adjusted supervised classification network and an adjusted autoencoder based reconstruction network.

9. The system of claim 8 , the processor also performing actions of:

determining a feature output interface, which is selected from the group consisted of the following: one of the hidden layer(s) of the deep neural network, or an intermediated hidden layer of the reconstruction network.

10. The system of claim 9 , the processor also performing actions of:

obtaining, by the deep neural network, testing data, the testing data being different from the training data;

obtaining, from the feature output interface, features of the testing data;

performing, based on the obtained features of the testing data, at least one operation including at least one of clustering and classification;

calculating at least one accuracy value including at least one of clustering accuracy and classification accuracy of the testing data; and

in response to the accuracy being larger than a predetermined threshold, determining that the training is complete, wherein a trained network comprises the trained deep neural network, the trained supervised classification network and the trained autoencoder based reconstruction network and an output of the trained network is the output of the feature output interface.

11. The system of claim 10 , wherein the training data include trip data, each trip comprising at least one segment, the method further comprising:

obtaining, by the DNN, at least one of the segments as the training data.

12. The system of claim 11 , wherein the testing data comprises trip data, each trip comprising at least one segment, the method further comprising:

obtaining, by the DNN, at least one of the segments as the testing data.

13. The system of claim 8 , wherein the classification loss is defined according to labels of the training data and the reconstruction loss contains an error term determined based on the features outputted from the specified hidden layer.

14. The system of claim 13 , wherein, the reconstruction loss further comprises a regularization term, the regularization term is determined based on L1 norm.

15. A computer program product, comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:

obtain, by a deep neural network (DNN), training data, the deep neural network comprising at least one hidden layer;

obtain, from a specified hidden layer of the at least one hidden layer, features of the training data, the specified hidden layer being connected respectively to a supervised classification network for classification tasks and an autoencoder based reconstruction network for reconstruction tasks; and

train, based on the obtained features, the DNN, the supervised classification network and the reconstruction network as a whole, the training being guided by the classification tasks and the reconstruction tasks;

determining, based on labeled data in the training data, a classification loss of the supervised classification network;

adjusting, via back propagation (BP) of the classification loss, the supervised classification network and the DNN to obtain a first adjusted DNN and a first adjusted supervised classification network;

determining, based on all the training data and the first adjusted supervised classification network, a classification loss of the supervised classification network and reconstruction loss of the autoencoder based reconstruction network; and

adjusting, via BP of the classification loss and the reconstruction loss, the first adjusted supervised classification network, the first adjusted DNN, the autoencoder based reconstruction network to obtain a second adjusted DNN and a second adjusted supervised classification network and an adjusted autoencoder based reconstruction network.

16. The computer program product of claim 15 , the program instructions executable by the processor also cause the processor to:

determine a feature output interface, which is selected from the group consisted of the following: one of the hidden layer(s) of the deep neural network, or an intermediated hidden layer of the reconstruction network.

17. The computer program product of claim 16 , the program instructions executable by the processor also cause the processor to:

obtain, by the deep neural network, testing data, the testing data being different from the training data;

obtain, from the feature output interface, features of the testing data;

perform, based on the obtained features of the testing data, at least one operation including at least one of clustering and classification;

calculate at least one accuracy value including at least one of clustering accuracy and classification accuracy of the testing data; and

in response to the accuracy being larger than a predetermined threshold, determining that the training is complete, wherein a trained network comprises the trained deep neural network, the trained classification network and the trained reconstruction network and an output of the trained network is the output of the feature output interface.

18. The computer program product of claim 17 , wherein the training data comprises trip data, each trip comprising at least one segment, the method further comprising:

obtaining, by the DNN, at least one of the segments as the training data.

19. The computer program product of claim 18 , wherein the testing data comprises trip data, each trip comprising at least one segment, the method further comprising:

obtaining, by the DNN, at least one of the segments as the testing data.

20. The computer program product of claim 15 , wherein the classification loss is defined according to labels of the training data and the reconstruction loss contains an error term determined based on the features outputted from the specified hidden layer.

21. The computer program product of claim 20 , wherein, the reconstruction loss further comprises a regularization term, the regularization term is determined based on L1 norm.