Data retrieval using reinforced co-learning for semi-supervised ranking
A computer-implement method comprises: training a classifier with labeled data from a dataset; classifying, by the trained classifier, unlabeled data from the dataset; providing, by the classifier to a policy gradient, a reward signal for each data/query pair; transferring, by the classifier to a ranker, learning; training, by the policy gradient, the ranker; ranking data from the dataset based on a query; and retrieving data from the ranked data in response to the query.
1. A computer-implemented method, comprising:
training a computer device classifier with labelled data and unlabeled data from a dataset;
providing, by the trained computer device classifier to a policy gradient function, a reward signal for data/query pairs based on the labelled and unlabeled data, wherein the reward signal is a combination of a normalized discounted cumulative gain and a discriminative score output by the trained computer device classifier;
transferring, by the trained computer device classifier to a ranker, learning from the trained computer device classifier;
providing, by the policy gradient function, a gradient based on the reward signal to the ranker to improve ranking;
ranking, by the ranker, data from the dataset based on a query including predicted relevance;
receiving a user selection of the ranked data; and
retrieving the user selected ranked data from a remote computer device.
2. The method of claim 1 , wherein the ranker includes a neural network.
3. The method of claim 1 , wherein the trained computer device classifier includes feature sharing.
4. The method of claim 3 , wherein the transferred learning includes an intermediate feature.
5. The method of claim 1 , wherein the training the computer device classifier comprises iteratively training with labeled data from the dataset.
6. The method of claim 5 , wherein the training the computer device classifier comprises using both positive and negative samples from the dataset.
7. The method of claim 6 , wherein the negative samples include negative classified data.
8. The method of claim 1 , wherein the ranking the data from the dataset based on the query comprises a Markov Decision Process.
9. A machine-readable storage device embodying instructions that, when executed by a machine, cause the machine to perform operations comprising:
one or more processors of a machine;
a memory storing instruction that, when executed by the one or more processors, cause the machine to perform operations comprising:
training a computer device classifier with labelled data and unlabeled data from a dataset;
providing iteratively, by the trained computer device classifier to a policy gradient function, a reward signal for data/query pairs based on the labelled and unlabeled data, wherein the reward signal is a combination of a normalized discounted cumulative gain and a discriminative score output by the trained computer device classifier;
transferring, by the trained computer device classifier to a ranker, learning from the trained computer device classifier;
providing, by the policy gradient function, a gradient based on the reward signal to the ranker to improve ranking;
ranking, by the ranker, data from the dataset based on a query including predicted relevance
receiving a user selection of the ranked data; and
retrieving the user selected ranked data from a remote computer device.
10. A system, comprising:
one or more processors of a machine;
a memory storing instruction that, when executed by the one or more processors, cause the machine to perform operations comprising:
training a computer device classifier with labelled data and unlabeled data from a dataset;
providing, by the trained computer device classifier to a policy gradient function, a reward signal for data/query pairs based on the labelled and unlabeled data, wherein the reward signal is a combination of a normalized discounted cumulative gain and a discriminative score output by the trained computer device classifier;
transferring, by the trained computer device classifier to a ranker, learning from the trained computer device classifier;
providing, by the policy gradient function, a gradient based on the reward signal to the ranker to improve ranking;
ranking, by the ranker, data from the dataset based on a query including predicted relevance
receiving a user selection of the ranked data; and
retrieving the user selected ranked data from a remote computer device.
11. The system of claim 10 , wherein the classifier and ranker are neural networks.
12. The system of claim 10 , wherein the computer device classifier is a neural network with feature sharing.
13. The system of claim 12 , wherein the transferred learning includes an intermediate feature.
14. The system of claim 10 , wherein the the training the computer device classifier comprises iteratively training with labeled data from the dataset.
15. The system of claim 14 , wherein the training the computer device classifier comprises using both positive and negative samples from the dataset.
16. The system of claim 15 , wherein the negative samples include negative classified data.
17. The system of claim 10 , wherein the ranking the data from the dataset based on the query comprises a Markov Decision Process.