Variable-output-space prediction machine learning models using contextual input embeddings
As described herein, various embodiments of the present invention use an output space refinement machine learning model to filter C of the B candidate predictive associations that are referred to herein as dynamically-preselected candidate predictive associations for each prediction input data object, with every prediction input data object being associated with a different subset of the dynamically-preselected candidate predictive associations, where C is less than B and is in some embodiments typically much less than B. As described in greater detail below, this approach reduces the number of computational operations that need to be performed by a final classification machine learning model (referred to herein as a variable-output-space prediction machine learning model), and leads to substantial computational efficiency advantages relative to naïve implementations.
1 . A computer-implemented method comprising:
providing, by one or more processors, a prediction input data object to a machine learning framework to receive a variable-output-space prediction, wherein the machine learning framework comprises:
(i) a multi-class classification machine learning model configured to generate a plurality of classification scores, respectively corresponding to a plurality of candidate predictive associations, based on an input vector corresponding to the prediction input data object and a plurality of candidate vectors respectively corresponding to the plurality of candidate predictive associations,
(ii) an isolated embedding model configured to generate an isolated input embedding based on the input vector,
(iii) a contextual embedding model configured to generate a contextual input embedding for the prediction input data object based at least in part on the isolated input embedding, wherein the contextual input embedding is generated based at least in part on a contextual input embedding subspace associated with a hypercube subspace geometric shape profile of a plurality of hypercube subspace geometric shape profiles, and
(iv) a deep learning model configured to generate the variable-output-space prediction based on (a) a vector subset of the plurality of candidate vectors respectively corresponding to an association subset of the plurality of candidate predictive associations respectively associated with a score subset of the plurality of classification scores that meets or exceeds a threshold and (b) the contextual input embedding, wherein the multi-class classification machine learning model performs less computationally intensive operations than the deep learning model due to an input space of the contextual input embedding; and
providing, by the one or more processors, the variable-output-space prediction as an output of the machine learning framework.
2 . The computer-implemented method of claim 1 , wherein a candidate predictive association feature set for a candidate predictive association of the plurality of candidate predictive associations comprises a historical engagement measure for the candidate predictive association with respect to a defined input profile.
3 . The computer-implemented method of claim 1 , further comprising:
generating a plurality of defined input profiles by:
identifying a group of historical prediction input data objects;
generating, using an input clustering machine learning model and based at least in part on a group of initial historical input feature sets for the group of historical prediction input data objects, a plurality of input clusters; and
generating the plurality of defined input profiles based at least in part on the plurality of input clusters.
4 . The computer-implemented method of claim 3 , wherein generating a defined input profile for the prediction input data object comprises:
determining a plurality of distance measures corresponding to the plurality of input clusters, wherein a distance measure of the plurality of distance measures indicates a distance from an initial historical input feature set for the prediction input data object to an initial historical input feature set of the plurality of input clusters; and
generating the defined input profile based at least in part on the plurality of distance measures.
5 . The computer-implemented method of claim 4 , wherein the machine learning framework is trained based on a plurality of ground-truth scores, wherein a ground-truth score of the plurality of ground-truth scores is based on the group of historical prediction input data objects.
6 . The computer-implemented method of claim 1 , wherein:
the isolated embedding model comprises machine learning sub-models configured to generate a plurality of dimensional regression outputs for the prediction input data object, and
a regression machine learning sub-model of the machine learning sub-models is configured to generate a dimensional regression output of the plurality of dimensional regression outputs based at least in part on a shared embedding feature set that is determined based at least in part on one or more embedding features for the prediction input data object.
7 . The computer-implemented method of claim 6 , wherein the isolated input embedding is generated based at least in part on the plurality of dimensional regression outputs.
8 . A system comprising:
one or more processors; and
at least one memory storing processor-executable instructions that, when executed by any one or more of the one or more processors, causes the one or more processors to perform operations comprising:
provide prediction input data object to a machine learning framework to receive a variable-output-space prediction, wherein the machine learning framework comprises:
(i) a multi-class classification machine learning model configured to generate a plurality of classification scores, respectively corresponding to a plurality of candidate predictive associations, based on an input vector corresponding to the prediction input data object and a plurality of candidate vectors respectively corresponding to the plurality of candidate predictive associations,
(ii) an isolated embedding model configured to generate an isolated input embedding based on the input vector,
(iii) a contextual embedding model configured to generate a contextual input embedding for the prediction input data object based at least in part on the isolated input embedding, wherein the contextual input embedding is generated based at least in part on a contextual input embedding subspace associated with a subspace geometric shape profile of a plurality of hypercube subspace geometric shape profiles, and
(iv) a deep learning model configured to generate the variable-output-space prediction based on (a) a vector subset of the plurality of candidate vectors respectively corresponding to an association subset of the plurality of candidate predictive associations respectively associated with a score subset of the plurality of classification scores that meets or exceeds a threshold and (b) the contextual input embedding, wherein the multi-class classification machine learning model performs less computationally intensive operations than the deep learning model due to an input space of the contextual input embedding; and
provide the variable-output-space prediction as an output of the machine learning framework.
9 . The system of claim 8 , wherein a candidate predictive association feature set for a candidate predictive association of the plurality of candidate predictive associations comprises a historical engagement measure for the candidate predictive association with respect to a defined input profile.
10 . The system of claim 8 , wherein the operations further comprise:
generating a plurality of defined input profiles by:
identifying a group of historical prediction input data objects;
generating, using an input clustering machine learning model and based at least in part on a group of initial historical input feature sets for the group of historical prediction input data objects, a plurality of input clusters; and
generating the plurality of defined input profiles based at least in part on the plurality of input clusters.
11 . The system of claim 10 , wherein generating a defined input profile for the prediction input data object comprises:
determining a plurality of distance measures corresponding to the plurality of input clusters, wherein a distance measure of the plurality of distance measures indicates a distance from an initial historical input feature set for the prediction input data object to an initial historical input feature set of the plurality of input clusters; and
generating the defined input profile based at least in part on the plurality of distance measures.
12 . The system of claim 10 , wherein the machine learning framework is trained based on a plurality of ground-truth scores, wherein a ground-truth score of the plurality of ground-truth scores is based on the group of historical prediction input data objects.
13 . The system of claim 8 , wherein:
the isolated embedding model comprises machine learning sub-models configured to generate a plurality of dimensional regression outputs for the prediction input data object, and
a regression machine learning sub-model of the machine learning sub-models is configured to generate a dimensional regression output of the plurality of dimensional regression outputs based at least in part on a shared embedding feature set that is determined based at least in part on one or more embedding features for the prediction input data object.
14 . The system of claim 13 , wherein the isolated input embedding is generated based at least in part on the plurality of dimensional regression outputs.
15 . One or more non-transitory computer-readable storage media including instructions that, when executed by one or more processors, cause the one or more processors to:
provide prediction input data object to a machine learning framework to receive a variable-output-space prediction, wherein the machine learning framework comprises:
(i) a multi-class classification machine learning model configured to generate a plurality of classification scores, respectively corresponding to a plurality of candidate predictive associations, based on an input vector corresponding to the prediction input data object and a plurality of candidate vectors respectively corresponding to the plurality of candidate predictive associations,
(ii) an isolated embedding model configured to generate an isolated input embedding based on the input vector,
(iii) a contextual embedding model configured to generate a contextual input embedding for the prediction input data object based at least in part on the isolated input embedding, wherein the contextual input embedding is generated based at least in part on a contextual input embedding subspace associated with a hypercube subspace geometric shape profile of a plurality of hypercube subspace geometric shape profiles, and
(iv) a deep learning model configured to generate the variable-output-space prediction based on (a) a vector subset of the plurality of candidate vectors respectively corresponding to an association subset of the plurality of candidate predictive associations respectively associated with a score subset of the plurality of classification scores that meets or exceeds a threshold and (b) the contextual input embedding, wherein the multi-class classification machine learning model performs less computationally intensive operations than the deep learning model due to an input space of the contextual input embedding; and
provide the variable-output-space prediction as an output of the machine learning framework.
16 . The one or more non-transitory computer-readable storage media of claim 15 , wherein a candidate predictive association feature set for a candidate predictive association of the plurality of candidate predictive associations comprises a historical engagement measure for the candidate predictive association with respect to a defined input profile.
17 . The one or more non-transitory computer-readable storage media of claim 15 , wherein the one or more processors are further caused to:
generate a plurality of defined input profiles by:
identifying a group of historical prediction input data objects;
generating, using an input clustering machine learning model and based at least in part on a group of initial historical input feature sets for the group of historical prediction input data objects, a plurality of input clusters; and
generating the plurality of defined input profiles based at least in part on the plurality of input clusters.
18 . The one or more non-transitory computer-readable storage media of claim 17 , wherein generating a defined input profile for the prediction input data object comprises:
determining a plurality of distance measures corresponding to the plurality of input clusters, wherein a distance measure of the plurality of distance measures indicates a distance from an initial historical input feature set for the prediction input data object to an initial historical input feature set of the plurality of input clusters; and
generating the defined input profile based at least in part on the plurality of distance measures.
19 . The one or more non-transitory computer-readable storage media of claim 15 , wherein:
the isolated embedding model comprises machine learning sub-models configured to generate a plurality of dimensional regression outputs for the prediction input data object, and
a regression machine learning sub-model of the machine learning sub-models is configured to generate a dimensional regression output of the plurality of dimensional regression outputs based at least in part on a shared embedding feature set that is determined based at least in part on one or more embedding features for the prediction input data object.
20 . The one or more non-transitory computer-readable storage media of claim 19 , wherein the isolated input embedding is generated based at least in part on the plurality of dimensional regression outputs.