Method and device for cash advance recognition

1. A method for training a cash-return recognition model, the cash-return recognition model comprising a first Deep Q-Network (DQN) and a second DQN, the method comprising:

acquiring first transaction information of a first transaction and a first cash-return label value of the first transaction, the first cash-return label value indicating whether the first transaction is labeled as a cash-return transaction;

calculating a first cash-return predictive value for the first transaction by inputting the first transaction information to the first DQN, the first cash-return predictive value indicating whether the first transaction is predicted to be a cash-return transaction;

calculating a Q-value's label value corresponding to the first transaction information and the first cash-return predictive value by using a Q learning algorithm based on the first DQN and the first cash-return label value; and

training the second DQN based on the first transaction information, the first cash-return predictive value, and the Q-value's label value by adjusting parameters of the second DQN such that an absolute value of a difference between a first Q-value's predictive value output by the second DQN after the training and the Q-value's label value is smaller than that between the first Q-value's predictive value output by the second DQN before the training and the Q-value's label value, wherein the first Q-value's predictive value is output by inputting the first transaction information and the first cash-return predictive value to the second DQN after the training, and the first Q-value's predictive value corresponds to the first transaction information and the first cash-return predictive value.

2. The method for training a cash-return recognition model according to claim 1 , wherein calculating the first cash-return predictive value for the first transaction comprises:

inputting the first transaction information to the first DQN to obtain a two-dimensional vector of Q-value's predictive values; and

calculating the first cash-return predictive value according to a s-greedy policy based on the vector.

3. The method for training a cash-return recognition model according to claim 1 , wherein the first transaction information comprises a transaction amount, and wherein calculating a Q-value's label value corresponding to the first transaction information and the first cash-return predictive value by using a Q learning algorithm comprises:

acquiring second transaction information of a second transaction and a second cash-return label value of the second transaction, the second cash-return label value indicating whether the second transaction is labeled as a cash-return transaction;

determining a return value for the first transaction based on the first cash-return predictive value, the first cash-return label value, and the transaction amount;

inputting the second transaction information to the first DQN to obtain a second Q-value's predictive value corresponding to the second transaction information and the second cash-return label value; and

calculating the Q-value's label value corresponding to the first transaction information and the first cash-return predictive value based on the return value and the second Q-value's predictive value.

4. The method for training a cash-return recognition model according to claim 3 , wherein inputting the second transaction information to the first DQN to obtain the second Q-value's predictive value corresponding to the second transaction information and the second cash-return label value comprises:

inputting the second transaction information to the first DQN to obtain a two-dimensional vector of Q-value's predictive values corresponding to the second transaction information;

performing one-hot encoding on the second cash-return label value to obtain a two-dimensional vector corresponding to the second cash-return label value;

performing an element-wise multiplication on the vector of Q-value's predictive values and the vector corresponding to the second cash-return label value to obtain a two-dimensional joint characterization vector; and

selecting a non-zero value in the joint characterization vector as the second Q-value's predictive value.

5. The method for training a cash-return recognition model according to claim 3 , wherein determining the return value for the first transaction based on the first cash-return predictive value, the first cash-return label value, and the transaction amount comprises:

in response to that the first cash-return predictive value is consistent with the first cash-return label value, determining the return value for the first transaction to be the transaction amount; and

in response to that the first cash-return predictive value is inconsistent with the first cash-return label value, determining the return value for the first transaction to be the negative of the transaction amount.

6. The method for training a cash-return recognition model according to claim 3 , further comprising adding the return value for the first transaction to a total return value of the model after determining the return value.

7. The method for training a cash-return recognition model according to claim 1 , wherein the cash-return recognition model further comprises a sample queue including a number of samples in sequence, the method further comprises:

updating the sample queue by deleting a predetermined number of existing samples ranked in the front from the sample queue when receiving a predetermined number of new samples.

8. The method for training a cash-return recognition model according to claim 7 , wherein the method further comprises:

preparing the samples in the sample queue, each sample including a transaction's information, a cash-return predictive value of the transaction, and a Q-value's label value corresponding to the transaction's information and the cash-return predictive value.

9. The method for training a cash-return recognition model according to claim 1 , wherein the first transaction information comprises attribute information, a buyer's information, a seller's information, and logistics information of the first transaction.

10. The method for training a cash-return recognition model according to claim 1 , wherein the first DQN or the second DQN comprises one of the following neural networks: an Long Short-Term Memory (LSTM) network, a Deep Neural Network (DNN), a Convolutional Neural Network (CNN), and an Recurrent Neural Network (RNN).

11. The method for training a cash-return recognition model according to claim 1 , wherein training the second DQN based on the first transaction information, the first cash-return predictive value, and the Q-value's label value comprises training the second DQN by a stochastic gradient descent algorithm.

12. The method for training a cash-return recognition model according to claim 1 , further comprising setting parameters of the first DQN based on the parameters of the second DQN.

13. The method for training a cash-return recognition model according to claim 1 , wherein the first DQN and the second DQN are interchangeable.

14. A method for cash-return recognition, comprising:

acquiring first transaction information of a first transaction and a first cash-return label value of the first transaction, the first cash-return label value indicating whether the first transaction is labeled as a cash-return transaction;

calculating a first cash-return predictive value for the first transaction by inputting the first transaction information to a first Deep Q-Network (DQN), the first cash-return predictive value indicating whether the first transaction is predicted to be a cash-return transaction;

calculating a Q-value's label value corresponding to the first transaction information and the first cash-return predictive value by using a Q learning algorithm based on the first DQN and the first cash-return label value; and

training a second DON based on the first transaction information, the first cash-return predictive value, and the Q-value's label value by adjusting parameters of the second DQN such that a difference between a Q-value's predictive value output by the second DQN and the Q-value's label value decreases after the training;

assigning the parameters of the second DQN to the first DQN;

acquiring transaction information of a transaction; and

inputting the transaction information to the first DQN to obtain a cash-return predictive value for the transaction, the cash-return predictive value indicating whether the transaction is predicted to be a cash-return transaction.

15. The method for cash-return recognition according to claim 14 , wherein inputting the transaction information to the first DQN to obtain the cash-return predictive value for the transaction comprises:

inputting the transaction information to the first DQN to obtain a two-dimensional vector of Q-value's predictive values corresponding to the transaction information; and

calculating the cash-return predictive value for the transaction according to a s-greedy policy based on the vector.

16. A device for training a cash-return recognition model, the cash-return recognition model comprising a first Deep Q-Network (DQN) and a second DQN, wherein the device comprises one or more processors and one or more non-transitory computer-readable memories coupled to the one or more processors and configured with instructions executable by the one or more processors to cause the device to perform operations comprising:

acquiring first transaction information of a first transaction and a first cash-return label value of the first transaction, the first cash-return label value indicating whether the first transaction is labeled as a cash-return transaction;

calculating a first cash-return predictive value for the first transaction by inputting the first transaction information to the first DQN, the first cash-return predictive value indicating whether the first transaction is predicted to be a cash-return transaction;

calculating a Q-value's label value corresponding to the first transaction information and the first cash-return predictive value by using a Q learning algorithm based on the first DQN and the first cash-return label value; and

training the second DQN based on the first transaction information, the first cash-return predictive value, and the Q-value's label value by adjusting parameters of the second DQN such that an absolute value of a difference between a first Q-value's predictive value output by the second DQN after the training and the Q-value's label value is smaller than that between the first Q-value's predictive value output by the second DQN before the training and the Q-value's label value, wherein the first Q-value's predictive value is output by inputting the first transaction information and the first cash-return predictive value to the second DQN after the training, and the first Q-value's predictive value corresponds to the first transaction information and the first cash-return predictive value.

17. The device for training a cash-return recognition model according to claim 16 , wherein calculating the first cash-return predictive value for the first transaction comprises:

inputting the first transaction information to the first DQN to obtain a two-dimensional vector of Q-value's predictive values; and

calculating the first cash-return predictive value according to a s-greedy policy based on the vector.

18. The device for training a cash-return recognition model according to claim 16 , wherein the first transaction information comprises a transaction amount, and wherein calculating a Q-value's label value corresponding to the first transaction information and the first cash-return predictive value by using a Q learning algorithm comprises:

acquiring second transaction information of a second transaction and a second cash-return label value of the second transaction, the second cash-return label value indicating whether the second transaction is labeled as a cash-return transaction;

determining a return value for the first transaction based on the first cash-return predictive value, the first cash-return label value, and the transaction amount;

inputting the second transaction information to the first DQN to obtain a second Q-value's predictive value corresponding to the second transaction information and the second cash-return label value; and

calculating the Q-value's label value corresponding to the first transaction information and the first cash-return predictive value based on the return value and the second Q-value's predictive value.

19. The device for training a cash-return recognition model according to claim 18 , wherein inputting the second transaction information to the first DQN to obtain the second Q-value's predictive value corresponding to the second transaction information and the second cash-return label value comprises:

inputting the second transaction information to the first DQN to obtain a two-dimensional vector of Q-value's predictive values corresponding to the second transaction information;

performing one-hot encoding on the second cash-return label value to obtain a two-dimensional vector corresponding to the second cash-return label value;

performing an element-wise multiplication on the vector of Q-value's predictive values and the vector corresponding to the second cash-return label value to obtain a two-dimensional joint characterization vector; and

selecting a non-zero value in the joint characterization vector as the second Q-value's predictive value.

20. The device for training a cash-return recognition model according to claim 18 , wherein determining the return value for the first transaction based on the first cash-return predictive value, the first cash-return label value, and the transaction amount comprises:

in response to that the first cash-return predictive value is consistent with the first cash-return label value, determining the return value for the first transaction to be the transaction amount; and

in response to that the first cash-return predictive value is inconsistent with the first cash-return label value, determining the return value for the first transaction to be the negative of the transaction amount.