Integrated architecture searching system and method

1 . An integrated architecture searching system for a click-through rate prediction model, comprising:

a processor; and

a memory storing computer instructions which, when executed by the processor, the processor is configured to:

search for embedding vector dimensions of features and determine a matching embedding vector dimension for each pair of the features;

obtain a feature interaction result by searching for a feature interaction sub-network a feature interaction combination; and

obtain a click-through rate prediction value by incorporating the feature interaction result into a high-order implicit feature interaction search space and performing high-order implicit feature interaction on deep networks of different layers.

2 . The system according to claim 1 , wherein the processor is further configured to:

map feature combinations in different dimensions into a unified dimension space for interaction.

3 . The system according to claim 1 , wherein the processor is further configured to:

generate an integrated architecture of the click-through rate prediction model by generating an architecture of a current component based on structure selection of all search spaces before each of the search spaces.

4 . The system according to claim 1 , wherein the processor is further configured to:

protocol all sub-networks in the search spaces as a super network to perform joint optimization and update.

5 . The system according to claim 4 , wherein the joint optimization and update comprises:

in each round of training, training a sub-network with a largest number of parameters as a teacher network, and guiding training of remaining sub-networks with an output of the teacher network.

6 . An integrated architecture searching method for a click-through rate prediction model, comprising:

searching for embedding vector dimensions of features and determining a matching embedding vector dimension for each pair of the features;

obtaining a feature interaction result by searching for a feature interaction sub-network and a feature interaction combination; and

obtaining a click-through rate prediction value by incorporating the feature interaction result into a high-order implicit feature interaction search space and performing high-order implicit feature interaction on deep networks of different layers.

7 . The method according to claim 6 , further comprising:

mapping feature combinations in different dimensions into a unified dimension space for interaction.

8 . The method according to claim 6 , further comprising:

generating an integrated architecture of the click-through rate prediction model by generating an architecture of a current component based on structure selection of all search spaces before each search space.

9 . The method according to claim 8 , further comprising:

protocolling all sub-networks in the search spaces as a super network to perform joint optimization and update.

10 . The method according to claim 9 , wherein performing the joint optimization and update comprises:

in each round of training, training a sub-network with a largest number of parameters as a teacher network, and guiding training of remaining sub-networks with an output of the teacher network.

11 . A method for predicting a click-through rate, comprising:

obtaining network click-through rate data; and

obtaining a click-through rate value by inputting the network click-through rate data into a click-through rate prediction model obtained by:

searching for embedding vector dimensions of features and determining a matching embedding vector dimension for each pair of the features;

obtaining a feature interaction result by searching for a feature interaction sub-network and a feature interaction combination; and

obtaining a click-through rate prediction value by incorporating the feature interaction result into a high-order implicit feature interaction search space and performing high-order implicit feature interaction on deep networks of different layers.