Deep and wide machine learned model for job recommendation
In an example, features in a boosting decision tree model are initialized to zero, the boosting decision tree model located in a GLMM and connected to a deep neural network collaborative filtering model via a prediction layer. While the features in the boosting decision tree model remain zero, the deep neural network collaborative filtering model is trained. One or more trees in the boosting decision tree model are boosted using logits produced by the training of the deep neural network collaborative filtering model as a margin. The prediction layer is trained using features from the deep neural network collaborative filtering model and features from the boosting decision tree model. It is then determined whether a set of convergence criteria is met. If not, then the deep neural network collaborative filtering model is retrained using the features and the process is repeated until the set of convergence criteria is met.
1. A system comprising:
a memory storing a generalized linear mixed model (GLMM), the GLMM comprising a deep neural network collaborative filtering model and a boosting decision tree model connected via a prediction layer; and
a computer-readable medium having instructions stored thereon, which, when executed by a processor, cause the system to:
initialize features in the boosting decision tree model to set values for the features to zero;
while the features in the boosting decision tree model remain zero, train the deep neural network collaborative filtering model;
boost one or more trees in the boosting decision tree model using logits produced by the training of the deep neural network collaborative filtering model as a margin;
train the prediction layer using features from the deep neural network collaborative filtering model and features from the boosting decision tree model;
determine whether a set of convergence criteria is met; and
in response to a determination that the set of convergence criteria is not met:
retrain the deep neural network collaborative filtering model using the features in the boosting decision tree model; and
repeat the boosting, training, determining, and retraining until the set of convergence criteria is met.
2. The system of claim 1 , wherein the GLMM takes member information and job title information from a social networking service and outputs a prediction value as to whether a member corresponding to the member information would apply to a job posting corresponding to the job title information.
3. The system of claim 1 , wherein the boosting decision tree model has at least two trees, one tree corresponding to member information and one tree corresponding to job title information.
4. The system of claim 2 , wherein the instructions further cause the system to:
pass the member information and job title information corresponding to a plurality of job postings in the social networking service to the GLMM to obtain a plurality of prediction values; and
display information regarding job postings having job title information that produce highest of the prediction values to the member via a computer display.
5. The system of claim 1 , wherein the deep neural network collaborative filtering model has two identity embedding networks, one for member information and one for job information.
6. The system of claim 5 , wherein the identity embedding network for the member information has an identical structure to the identity embedding network for the job information.
7. The system of claim 6 , wherein the identity embedding network for the member information has an input layer that takes sparse categorical features, an embedding layer that projects each categorical feature into a dense latent vector, one or more hidden layers concatenating and connecting the dense latent vectors, an entity representation layer in which the one or more hidden layers are mapped, and a Hadamard product layer that performs a Hadamard product on entities in the entity representation layer.
8. A computerized method comprising:
initializing features in a boosting decision tree model to set values for the features to zero, the boosting decision tree model located in a GLMM and connected to a deep neural network collaborative filtering model via a prediction layer;
while the features in the boosting decision tree model remain zero, training the deep neural network collaborative filtering model;
boosting one or more trees in the boosting decision tree model using logits produced by the training of the deep neural network collaborative filtering model as a margin;
training the prediction layer using features from the deep neural network collaborative filtering model and features from the boosting decision tree model;
determining whether a set of convergence criteria is met; and
in response to a determination that the set of convergence criteria is not met:
retraining the deep neural network collaborative filtering model using the features in the boosting decision tree model; and
repeating the boosting, training, determining, and retraining until the set of convergence criteria is met.
9. The method of claim 8 , wherein the GLMM takes member information and job title information from a social networking service and outputs a prediction value as to whether a member corresponding to the member information would apply to a job posting corresponding to the job title information.
10. The method of claim 8 , wherein the boosting decision tree model has at least two trees, one tree corresponding to member information and one tree corresponding to job title information.
11. The method of claim 9 , further comprising:
passing the member information and job title information corresponding to a plurality of job postings in the social networking service to the GLMM to obtain a plurality of prediction values; and
displaying information regarding job postings having job title information that produce highest of the prediction values to the member via a computer display.
12. The method of claim 8 , wherein the deep neural network collaborative filtering model has two identity embedding networks, one for member information and one for job information.
13. The method of claim 12 , wherein the identity embedding network for the member information has an identical structure to the identity embedding network for the job information.
14. The method of claim 13 , wherein the identity embedding network for the member information has an input layer that takes sparse categorical features, an embedding layer that projects each categorical feature into a dense latent vector, one or more hidden lavers concatenating and connecting the dense latent vectors, an entity representation layer in which the one or more hidden layers are mapped, and a Hadamard product layer that performs a Hadamard product on entities in the entity representation layer.
15. A non-transitory machine-readable storage medium comprising instructions, which when implemented by one or more machines, cause the one or more machines to perform operations comprising:
initializing features in a boosting decision tree model to set values for the features to zero, the boosting decision tree model located in a GLMM and connected to a deep neural network collaborative filtering model via a prediction layer;
while the features in the boosting decision tree model remain zero, training the deep neural network collaborative filtering model;
boosting one or more trees in the boosting decision tree model using logits produced by the training of the deep neural network collaborative filtering model as a margin;
training the prediction layer using features from the deep neural network collaborative filtering model and features from the boosting decision tree model;
determining whether a set of convergence criteria is met; and
in response to a determination that the set of convergence criteria is not met:
retraining the deep neural network collaborative filtering model using the features in the boosting decision tree model; and
repeating the boosting, training, determining, and retraining until the set of convergence criteria is met.
16. The non-transitory machine-readable storage medium of claim 15 , wherein the GLMM takes member information and job title information from a social networking service and outputs a prediction value as to whether a member corresponding to the member information would apply to a job posting corresponding to the job title information.
17. The non-transitory machine-readable storage medium of claim 15 , wherein the boosting decision tree model has at least two trees, one tree corresponding to member information and one tree corresponding to job title information.
18. The non-transitory machine-readable storage medium of claim 16 , wherein the operations further comprise:
passing the member information and job title information corresponding to a plurality of job postings in the social networking service to the GLMM to obtain a plurality of prediction values; and
displaying information regarding job postings having job title information that produce highest of the prediction values to the member via a computer display.
19. The non-transitory machine-readable storage medium of claim 15 , wherein the deep neural network collaborative filtering model has two identity embedding networks, one for member information and one for job information.
20. The non-transitory machine-readable storage medium of claim 19 , wherein the identity embedding network for the member information has an identical structure to the identity embedding network for the job information.