Methods and apparatus to perform multi-level hierarchical demographic classification

1. An apparatus to demographically classify an individual, the apparatus comprising:

a neural network structured to have an input layer, a first output layer, a second output layer subsequent to the first output layer, a sorting output layer, and a plurality of neural network modules interposed between the input layer and the first output layer, respective ones of the neural network modules including corresponding groups of interconnected neural layers, each group of interconnected neural layers having at least one connection between a corresponding input layer of the group of interconnected neural layers and corresponding output layer of the group of interconnected neural layers, the neural network structured to:

process inputs presented at the input layer to form first outputs at the first output layer, the inputs based on demographic information for the individual, the first outputs representing a plurality of first possible classifications of the individual according to a demographic classification system at a first hierarchical level;

process the first outputs at the sorting output layer to determine softmax outputs and sorted outputs, the sorted outputs different from the softmax outputs, the sorting output layer to process the first outputs based on a softmax operation to determine the softmax outputs, the sorting output layer to sort the first outputs separate from processing with the softmax operation to form groups of the first outputs that are grouped based on a plurality of second possible classifications of the individual according to the demographic classification system at a second hierarchical level different from the first hierarchical level, each one of the first outputs to have a one-to-one correspondence with a corresponding one of the plurality of first possible classifications, each one of the groups of the first outputs to have a one-to-one correspondence with a corresponding one of the plurality of second possible classifications;

process the groups of the first outputs to form second outputs at the second output layer, the second outputs representing possible combined classifications of the individual, the possible combined classifications corresponding to combinations of the plurality of first possible classifications and the plurality of second possible classifications; and

a processor to execute computer readable instructions to:

select one of the second outputs at the second output layer;

associate with the individual a respective one of the first possible classifications and a respective one of the second possible classifications corresponding to a respective one of the possible combined classifications represented by the selected second output;

compute a loss value based on a weighted combination of a first contribution determined from the softmax outputs of the sorting output layer of the neural network, a second contribution determined from the second outputs of the second output layer of the neural network, and a third contribution determined from coefficients of the neural network, the first contribution adjusted based on a first weight, the second contribution adjusted based on a second weight, and the third contribution adjusted based on a third weight; and

update one or more of the coefficients of the neural network based on the loss value.

2. The apparatus as defined in claim 1 , wherein the neural network is also structured to include:

a combining output layer structured to select, from each of the groups of the first outputs, a first one of the first outputs of the group having a greatest value to form third outputs,

wherein the second output layer is structured to convert the third outputs into probabilities to form the second outputs.

3. The apparatus as defined in claim 2 , wherein the plurality of neural network modules is structured to process the inputs presented at the input layer to form the first outputs at the first output layer.

4. The apparatus as defined in claim 1 , wherein the first contribution is based on a first cross-entropy of the softmax outputs and the second contribution is based on a second cross-entropy of the second outputs.

5. The apparatus as defined in claim 4 , wherein the processor is to update the one or more of the coefficients using a stochastic descent algorithm.

6. The apparatus as defined in claim 1 , wherein the processor is to:

query a database to obtain the demographic information for the individual; and

form the inputs based on contents of the demographic information.

7. The apparatus as defined in claim 6 , wherein the processor is to record the one of the first possible classifications and the one of the second possible classifications in the database in conjunction with the individual.

8. The apparatus as defined in claim 6 , wherein the demographic information includes a first value of a demographic characteristic associated with the individual, and the processor is to:

convert the first value to second value representative of a range of values of the demographic characteristic, the range of values including the first value; and

form a first one of the inputs based on the second value.

9. The apparatus as defined in claim 1 , wherein the second possible classifications represent demographic categories of the individual.

10. The apparatus as defined in claim 9 , wherein the first possible classifications represent demographic segments of the demographic categories.

11. A method of performing demographic classification of an individual, the method comprising:

obtaining data representative of demographic characteristics of an individual;

processing the data with a neural network to form first outputs at a first output layer of the neural network, the first outputs representing a plurality of first possible demographic classifications of the individual at a first hierarchical classification level, the neural network including an input layer, the first output layer, a second output layer subsequent to the first output layer, a sorting output layer, and a plurality of neural network modules interposed between the input layer and the first output layer, respective ones of the neural network modules including corresponding groups of interconnected neural layers, each group of interconnected neural layers having at least one connection between a corresponding input layer of the group of interconnected neural layers and corresponding output layer of the group of interconnected neural layers;

processing the first outputs with the neural network to determine softmax outputs and sorted outputs at the sorting layer, the sorted outputs different from the softmax outputs, the sorting layer to process the first outputs based on a softmax operation to determine the softmax outputs, the sorting layer to sort the first outputs separate from processing with the softmax operation to form groups of the first outputs that are grouped based on a plurality of second possible demographic classifications of the individual at a second hierarchical classification level different from the first hierarchical classification level, each one of the first outputs to have a one-to-one correspondence with a corresponding one of the plurality of first possible demographic classifications, each one of the groups of the first outputs to have a one-to-one correspondence with a corresponding one of the plurality of second possible demographic classifications;

processing the groups of the first outputs with the neural network to form second outputs at the second output layer of the neural network, the second outputs representing possible combined demographic classifications of the individual corresponding to combinations of the plurality of first possible demographic classifications and the plurality of second possible demographic classifications;

computing a loss value based on a weighted combination of a first contribution determined from the softmax outputs of the sorting output layer of the neural network, a second contribution determined from the second outputs of the second output layer of the neural network, and a third contribution determined from coefficients of the neural network, the first contribution adjusted based on a first weight, the second contribution adjusted based on a second weight, and the third contribution adjusted based on a third weight; and

updating one or more of the coefficients of the neural network based on the loss value.

12. The method as defined in claim 11 , wherein the processing of the groups of the first outputs to form the second outputs includes:

identifying, for each of the groups of the first outputs, a first one of the first outputs of the group having a greatest value to form third outputs; and

converting the third outputs into probabilities to form the second outputs.

13. The method as defined in claim 11 , wherein the second possible demographic classifications represent demographic categories of the individual.

14. The method as defined in claim 13 , wherein the first possible demographic classifications represent demographic segments of the demographic categories.

15. The method as defined in claim 11 , wherein the first contribution is based on a first cross-entropy of the softmax outputs and the second contribution is based on a second cross-entropy of the second outputs.

16. The method as defined in claim 15 , wherein the updating of the one or more of the coefficients includes using a stochastic descent algorithm.

17. A tangible computer-readable storage medium comprising instructions that, when executed, cause a machine to at least:

obtain data representative of demographic characteristics of an individual;

process the data with a neural network to form first outputs at a first output layer of the neural network, the first outputs representing a plurality of first possible demographic classifications of the individual at a first hierarchical classification level, the neural network including an input layer, the first output layer, a second output layer subsequent to the first output layer, a sorting output layer, and a plurality of neural network modules interposed between the input layer and the first output layer, respective ones of the neural network modules including corresponding groups of interconnected neural layers, each group of interconnected neural layers having at least one connection between a corresponding input layer of the group of interconnected neural layers and corresponding output layer of the group of interconnected neural layers;

process the first outputs with the neural network to determine softmax outputs and sorted outputs at the sorting output layer, the sorted outputs different from the softmax outputs, the sorting output layer to process the first outputs based on a softmax operation to determine the softmax outputs, the sorting output layer to sort the first outputs separate from processing with the softmax operation to form groups of the first outputs that are grouped based on a plurality of second possible demographic classifications of the individual at a second hierarchical classification level different from the first hierarchical classification level, each one of the first outputs to have a one-to-one correspondence with a corresponding one of the plurality of first possible demographic classifications, each one of the groups of the first outputs to have a one-to-one correspondence with a corresponding one of the plurality of second possible demographic classifications;

process the groups of the first outputs with the neural network to form second outputs at the second output layer of the neural network, the second outputs representing possible combined demographic classifications of the individual corresponding to combinations of the plurality of first possible demographic classifications and the plurality of second possible demographic classifications;

compute a loss value based on a weighted combination of a first contribution determined from the softmax outputs of the sorting output layer of the neural network, a second contribution determined from the second outputs of the second output layer of the neural network, and a third contribution determined from coefficients of the neural network, the first contribution adjusted based on a first weight, the second contribution adjusted based on a second weight, and the third contribution adjusted based on a third weight; and

update one or more of the coefficients of the neural network based on the loss value.

18. The tangible computer-readable storage medium as defined in claim 17 , wherein the instructions, when executed, cause the machine to process the groups of the first outputs to form the second outputs by:

identifying, for each of groups of the first outputs, a first one of the first outputs of the group having a greatest value to form third outputs; and

converting the third outputs into probabilities to form the second outputs.

19. The tangible computer-readable storage medium as defined in claim 17 , wherein the first contribution is based on a first cross-entropy of the softmax outputs and the second contribution is based on a second cross-entropy of the second outputs.