Masked model training of a prediction network
In some embodiments, a method receives a first sequence of inputs for processing via a sub-model of a plurality of sub-model. The plurality of sub-models are part of a main model. An input in the sequence of inputs is masked with a masked value to generate a second sequence of inputs. The method processes the second sequence of inputs using the sub-model to generate a sequence of features that correspond to the second sequence of inputs and processes the sequence of features to generate a first output. The first output is processed to generate a second output of the main model. The sub-model is trained based on a feature in the sequence of features that corresponds to the masked input and the second output.
1. A method comprising:
receiving, by a computing device, a request for a recommendation of an item;
applying, by the computing device, the request to a trained main model to obtain a prediction relating to data in the request;
generating, by the computing device, a recommendation based on the prediction; and
sending, by the computing device, the recommendation as a response to the request, wherein the trained model comprises a trained sub-model of a plurality of sub-models, and wherein the sub-model is trained by:
masking an input in a sequence of inputs with a masked value in the sub-model, wherein the masked value is different from a value of the input;
generating, using the sub-model, a sequence of features using the sequence of inputs with the masked value;
generating a first value based on a comparison of an output of the main model to an expected output, wherein the output is determined using the sequence of features that is generated by the sub-model;
determining a second value that is based on a comparison of a feature in the sequence of features that corresponds to the masked input to an expected feature for the input; and
training a parameter of the sub-model using the second value and the first value, wherein the parameter is used by the sub-model to generate the sequence of features from the sequence of inputs.
2. The method of claim 1 , wherein:
the sequence of inputs comprises a first sequence of inputs,
masking the input in the first sequence of inputs with the masked value generates a second sequence of inputs, and
the sequence of features is generated using the second sequence of inputs.
3. The method of claim 2 , wherein training the sub-model comprises:
comparing a first output that corresponds to the masked input to the second sequence of inputs,
calculating a similarity between the first output and the second sequence of inputs based on the comparing; and
using the similarity to train the sub-model.
4. The method of claim 3 , wherein using the similarity to train the sub-model comprises:
adjusting the parameter of the sub-model to increase a similarity of the feature that corresponds to the masked input to the corresponding input in the sequence of inputs compared to a similarity for other inputs in the sequence of inputs.
5. The method of claim 1 , wherein the sequence of inputs is based on behavior of a user account that is using a video delivery system.
6. The method of claim 5 , wherein:
the masked value is based on information separate from the behavior of the user account.
7. The method of claim 1 , wherein the output represents a relevance of the item to the sequence of inputs and other input processed by the main model.
8. The method of claim 7 , wherein the relevance comprises a probability of selection of the item.
9. The method of claim 1 , wherein the sequence of features is a representation of the sequence of inputs.
10. The method of claim 1 , wherein the output comprises a first output, the method further comprising:
generating a second output from the sequence of features, wherein the second output is used to generate the first output of the main model.
11. The method of claim 1 , wherein masking the input comprises:
randomly masking one of the sequence of inputs.
12. The method of claim 1 , further comprising:
applying attention to the sequence of inputs based on a relationship between inputs, wherein the attention applies weights to the inputs differently.
13. The method of claim 1 , further comprising:
processing the sequence of inputs using multiple layers in the sub-model to generate the sequence of features.
14. The method of claim 1 , wherein the sequence of features is processed by one or more other sub-models in the plurality of sub-models to generate the output of the main model.
15. The method of claim 14 , wherein the one or more other sub-models receive input other than the sequence of inputs.
16. The method of claim 1 , wherein
the second value is a first loss based on the comparison of the feature in the sequence of features that corresponds to the masked input to the expected feature for the input;
the first value is a second loss based on the comparison of the output of the main model to the expected out.
17. The method of claim 1 , wherein the sub-model comprises a self-attention sub-model that applies attention to inputs in the sequence of inputs based on relationships between the inputs.
18. A non-transitory computer-readable storage medium having stored thereon computer executable instructions, which when executed by a computing device, cause the computing device to be operable for:
receiving a request for a recommendation of an item;
applying the request to a trained main model to obtain a prediction relating to data in the request;
generating a recommendation based on the prediction; and
sending the recommendation as a response to the request, wherein the trained model comprises a trained sub-model of a plurality of sub-models, and wherein the sub-model is trained by:
masking an input in a sequence of inputs with a masked value in the sub-model, wherein the masked value is different from a value of the input;
generating, using the sub-model, a sequence of features using the sequence of inputs with the masked value;
generating a first value based on a comparison of an output of the main model to an expected output, wherein the output is determined using the sequence of features that is generated by the sub-model;
determining a second value that is based on a comparison of a feature in the sequence of features that corresponds to the masked input to an expected feature for the input; and
training a parameter of the sub-model using the second value and the first value, wherein the meter is used by the sub-model to generate the sequence of features from the sequence of inputs.
19. The non-transitory computer-readable storage medium of claim 18 , wherein:
the sequence of inputs comprises a first sequence of inputs,
masking the input in the first sequence of inputs with the masked value generates a second sequence of inputs, and
the sequence of features is generated using the second sequence of inputs.
20. An apparatus comprising:
one or more computer processors; and
a computer-readable storage medium comprising instructions for controlling the one or more computer processors to be operable for:
receiving a request for a recommendation of an item;
applying the request to a trained main model to obtain a prediction relating to data in the request;
generating a recommendation based on the prediction; and
sending the recommendation as a response to the request, wherein the trained model comprises a trained sub-model of a plurality of sub-models, and wherein the sub-model is trained by:
masking an input in a sequence of inputs with a masked value in the sub-model, wherein the masked value is different from a value of the input;
generating, using the sub-model, a sequence of features using the sequence of inputs with the masked value;
generating a first value based on a comparison of an output of the main model to an expected output wherein the output is determined using the sequence of features that is generated by the sub-model;
determining a second value that is based on a comparison of a feature in the sequence of features that corresponds to the masked input to an expected feature for the input; and
training a parameter of the sub-model using the second value and the first value, wherein the parameter is used by the sub-model to generate the sequence of features from the sequence of inputs.