Adaptive multi-model item selection systems and methods
An adaptive multi-model item selection method, comprising: receiving, from one of a plurality of client devices, a request including a client-side feature vector representing a state of the client device; determining, by an advocate model, a probability distribution of a plurality of specialist cluster models from the client-side feature vector; choosing, by a use case selector, a cluster corresponding to a use case from the probability distribution; and obtaining, by the use case selector based on the cluster (i.e., the cluster that was sampled by the user case selector), a specialist cluster model from the plurality of specialist cluster models.
1. An adaptive multi-model item selection method, comprising:
receiving, from a client device of a plurality of client devices, a request including a client-side feature vector representing a state of the client device;
determining, by an advocate model, a probability distribution of a plurality of specialist cluster models from the client-side feature vector;
choosing, by a use case selector, a cluster corresponding to a use case from the probability distribution; and
obtaining, by the use case selector based on the cluster, a specialist cluster model from the plurality of specialist cluster models.
2. The method according to claim 1 , further comprising:
selecting, by an item selector interface, an item from a database of items using the specialist cluster model.
3. The method according to claim 2 , further comprising:
wherein the item selector interface is a content item selector interface and the item is a content item.
4. The method according to claim 1 , further comprising:
scoring, by the advocate model, the client-side feature vector based on an estimate of a probability the client-side feature vector is associated with each cluster of a plurality of clusters.
5. The method according to claim 1 , wherein choosing the cluster corresponding to the use case from the probability distribution includes randomly sampling a cluster.
6. The method of claim 1 , further comprising:
receiving, from a server, a plurality of server-side client feature vectors, each server-side client feature vector representing any one of (i) a set of item features associated with a plurality of items available to a client device, (ii) a set of client features associated with at least one client device, or (iii) a combination of (i) and (ii);
storing, in a memory, a plurality of mini-specialist cluster models, each mini-specialist cluster model of the plurality of mini-specialist cluster models being (i) associated with a corresponding client-side feature vector and (ii) trained on the plurality of server-side client feature vectors;
clustering the plurality of mini-specialist cluster models to obtain a soft cluster assignment for each of the mini-specialist cluster models, thereby generating a plurality of soft cluster assignments; and
training, by a model trainer, each of the plurality of specialist cluster models by combining the plurality of mini-specialist cluster models based on the soft cluster assignments associated with the mini-specialist cluster models.
7. The method of claim 6 , wherein the client-side feature vector includes a first set of client context features and the server-side client feature vectors include a plurality of second sets of client context features.
8. A system for multi-model item selection, comprising:
a data communication device configured to receive, from one of a plurality of client devices, a request including a client-side feature vector representing a state of the client device;
a memory device configured to store an advocate model and one or more specialist cluster models;
an analyzer configured to determine, using the advocate model, a probability distribution of a plurality of specialist cluster models from the client-side feature vector; and
a use case selector configured to:
choose a cluster corresponding to a use case from the probability distribution, and
obtain, based on the cluster, a specialist cluster model from the plurality of specialist cluster models.
9. The system according to claim 8 , further comprising:
an item selector interface configured to select an item from a database of items using the specialist cluster model.
10. The system according to claim 9 , wherein the item selector interface is a content item selector interface and the item is a content item.
11. The system according to claim 8 , wherein the advocate model is further configured to score the client-side feature vector based on an estimate of a probability the client-side feature vector is associated with each cluster of a plurality of clusters.
12. The system according to claim 8 , wherein the use case selector is further configured to choose the cluster by randomly sampling a cluster from the probability distribution.
13. The system of claim 8 , wherein:
the analyzer is further configured to:
receive a plurality of server-side client feature vectors, each server-side client feature vector representing any one of (i) a set of item features associated with a plurality of items available to a client device, (ii) a set of client features associated with at least one client device, or (iii) a combination of (i) and (ii);
the memory device is further configured to:
store a plurality of mini-specialist cluster models, each mini-specialist cluster model of the plurality of mini-specialist cluster models being (i) associated with a corresponding client-side feature vector and (ii) trained on the plurality of server-side client feature vectors;
the analyzer is further configured to:
cluster the plurality of mini-specialist cluster models to obtain a soft cluster assignment for each of the mini-specialist cluster models, thereby generating a plurality of soft cluster assignments; and
further comprising:
a model trainer configured to train each of the plurality of specialist cluster models by combining the plurality of mini-specialist cluster models based on the soft cluster assignments associated with the mini-specialist cluster models.
14. The system of claim 13 , wherein the client-side feature vector includes a first set of client context features and the server-side client feature vectors include a plurality of second sets of client context features.
15. A computer system, comprising: one or more processors; and a non-transitory computer-readable storage medium storing instructions that when executed by the one or more processors cause the computer system to perform operations, comprising:
receiving, from a client device of a plurality of client devices, a request including a client-side feature vector representing a state of the client device;
determining, by an advocate model, a probability distribution of a plurality of specialist cluster models from the client-side feature vector;
choosing, by a use case selector, a cluster corresponding to a use case from the probability distribution; and
obtaining, by the use case selector based on the cluster, a specialist cluster model from the plurality of specialist cluster models.
16. The non-transitory computer-readable medium of claim 15 , further having stored thereon a sequence of instructions for causing the one or more processors to perform:
selecting, by an item selector, an item from a database of items using the specialist cluster model.
17. The non-transitory computer-readable medium of claim 16 , further comprising:
wherein the item selector interface is a content item selector interface and the item is a content item.
18. The non-transitory computer-readable medium of claim 15 , further having stored thereon a sequence of instructions for causing the one or more processors to perform:
scoring, by the advocate model, the client-side feature vector based on an estimate of a probability the client-side feature vector is associated with each cluster of a plurality of clusters.
19. The non-transitory computer-readable medium of claim 15 , wherein choosing the cluster corresponding to the use case from the probability distribution includes randomly sampling a cluster.
20. The non-transitory computer-readable medium of claim 15 , further having stored thereon a sequence of instructions for causing the one or more processors to perform:
receiving, from a server, a plurality of server-side client feature vectors, each server-side client feature vector representing any one of (i) a set of item features associated with a plurality of items available to a client device, (ii) a set of client features associated with at least one client device, or (iii) a combination of (i) and (ii);
storing, in a memory device, a plurality of mini-specialist cluster models, each mini-specialist cluster model of the plurality of mini-specialist cluster models being (i) associated with a corresponding client-side feature vector and (ii) trained on the plurality of server-side client feature vectors;
clustering the plurality of mini-specialist cluster models to obtain a soft cluster assignment for each of the mini-specialist cluster models, thereby generating a plurality of soft cluster assignments; and
training, by a model trainer, each of the plurality of specialist cluster models by combining the plurality of mini-specialist cluster models based on the soft cluster assignments associated with the mini-specialist cluster models.
21. The non-transitory computer-readable medium of claim 20 , wherein the client-side feature vector includes a first set of client context features and the server-side client feature vectors include a plurality of second sets of client context features.