Efficient recommendation services

Patent metadata

Field	Value
Publication number	US-10262299-B2
Application number	US-201715433741-A
Country	US
Kind code	B2
Filing date	Feb 15, 2017
Priority date	Feb 15, 2017
Publication date	Apr 16, 2019
Grant date	Apr 16, 2019

Abstract

The disclosed subject matter involves identifying clusters and segments of a population of data for use in a recommendation service. Clusters of members or items are formed, where the clusters, or partitions are close to being equal in size, items are distributed based on similarities identified with matrix factorization. A matrix used in the matrix factorization is customized based on the recommendation type. The items are formed into clusters based on the similarities and the clusters are used in training of a generalized linear mixed model treating the clusters as random-level effects. The trained model may be used in the recommendation service. Other embodiments are described and claimed.

First claim

What is claimed is:

1. A computer readable storage medium having instructions for providing a recommendation using balanced clusters of items in a dataset, the instructions stored thereon, the instructions when executed on a machine cause the machine to:
receive a new item;:
generate an embedding for the new item;
apply cluster balancing to the new item embedding to assign the new item to a cluster, the cluster present in a trained model, wherein the trained model is stored in memory, when in operation is communicatively coupled to the machine, and wherein the trained model comprises cluster-level random effects in a generalized linear mixed modeling of the dataset, the trained model based on matrix factorization of a custom matrix, the custom matrix having custom factors that correlate items in the dataset, wherein the items in the dataset include at least two disparate item categories, the two disparate item categories correlated with each other based on at least one factor; and
provide a recommendation corresponding to the new item and other items in the dataset.

2. The medium as recited in claim 1 , wherein the custom factors include global factors to correlate items in the two disparate item categories, and personalization factors specific to an item in an item category.

3. The medium as recited in claim 1 , wherein the recommendation is a recommendation in a professional social network and the two disparate item categories comprise a member category and a job category, and further comprising instructions to provide a job recommendation, a top similar member recommendation, a top similar job recommendation, or a member recommendation for a job.

4. The medium as recited in claim 3 , wherein the custom matrix for the job recommendation comprises rows of m members, M m , and columns of jobs and member-specific personalization factors, wherein a job is positively correlated with a member M i when the member M i has previously viewed or applied for the job, and wherein the member M i is positively correlated to a member-specific personalization factor F j when the factor F j is identified in a profile associated with member M i .

5. The medium as recited in claim 3 , wherein the custom matrix for the top similar member recommendation comprises rows of m members, M m , and columns of M′ n members, wherein a member M i is positively correlated with a member M′ j when the members M i and M′ j have both been previously viewed by third member M k .

6. The medium as recited in claim 3 , wherein the custom matrix for the top similar job recommendation comprises rows of m jobs, J m , and columns of jobs J′ i and job-specific personalization factors, wherein a job J i is positively correlated with a job J′ j when the jobs J i and J′ j have been previously viewed or applied for by a same member, and wherein job J i is positively correlated with a job-specific personalization factor when the job-specific personalization factor is identified in a profile associated with job J i .

7. The medium as recited in claim 3 , wherein the custom matrix for the member recommendation comprises rows of m jobs, J m , and columns of n members M n , wherein a job J i is positively correlated with a member M j when a user viewed, saved or contacted member M j for job J i .

8. A computer readable storage medium having instructions stored thereon, the instructions when executed on a machine cause the machine to:
generate coefficient vectors for a first set of items in a dataset, and to generate coefficient vectors for a second set of items in the dataset for matrix factorization;
cluster balance each set of items to assign each item in the dataset into a corresponding cluster of a plurality of clusters, wherein the item clusters are substantially similar in size within a size variance threshold;
perform training to use clusters of the first set of items or clusters of the second set of item as a cluster-level random effect in a generalized linear mix modeling of the dataset, to generate a trained item model and store the trained item model in a data storage device, when in operation, the data storage device accessible by the machine; and
provide a recommendation service to users, the recommendation service comprising a plurality of recommendation types, wherein the training is arranged to use a matrix customized for use with a selected recommendation type.

9. The medium as recited in claim 8 , wherein first and second set of items comprise two disparate item categories, and wherein the Item categories include a member category and a job category, and wherein items in the job category are each assigned to a corresponding job cluster, and items in the member category are each assigned to a corresponding member cluster.

10. The medium as recited in claim 9 , wherein the instructions to provide the recommendation service further comprise instructions to:
retrieve the trained model from the data storage device; and
provide a recommendation based on the cluster-level random effect corresponding to either the job cluster or the member cluster.

11. The medium as recited in claim 8 , wherein the instructions to cluster balance further comprise instructions to:
bisect a segment in a graph into new segments, the segment including items corresponding to an item category, the bisection to be performed along a hyperplane corresponding to a selected item in the segment of the graph; and
continue to bisect or combine the segment and resulting new segments until each resulting new segment includes a number of items corresponding to the item category, wherein the number of items fall within a size threshold, and wherein each resulting new segment is assigned a unique clusterID.

12. The medium as recited in claim 9 , wherein the recommendation service is a recommendation service in a professional social, network, and wherein the recommendation service comprises instructions to provide a job recommendation, a top similar member recommendation, a top similar job recommendation, or a member recommendation for a job.

13. The medium as recited in claim 12 , wherein the matrix customized for use with the job recommendation comprises rows of m members, M m , and columns of jobs and member-specific personalization factors, wherein a job is positively correlated with a member M i when the member M i has previously viewed or applied for the job, and wherein the member M i is positively correlated to a member-specific personalization factor F j when the factor F j is identified in a profile associated with member M i .

14. The medium as recited in claim 12 , wherein the matrix customized for use with the top similar member recommendation comprises rows of m members, M m , and columns of M′ n members, wherein a member M i is positively correlated with a member M′ j when the members M i and M′ j have both been previously viewed by third member M k .

15. The medium as recited in claim 12 , wherein the matrix customized for use with the top similar job recommendation comprises rows of m jobs, J m , and columns of jobs J′ i and job-specific personalization factors, wherein a job J i is positively correlated with a job J′ j when the jobs J i and J′ j have been previously viewed or applied for by a same member, and wherein job J i is positively correlated with a job-specific personalization factor when the job-specific personalization factor is identified in a profile associated with job J i .

16. The medium as recited in claim 12 , wherein the matrix customized for use with the member recommendation for a job comprises rows of m j

Assignees

• Linkedin Corp
• Microsoft Technology Licensing Llc

Inventors

• Li Yi
• Chen Victor
• Kenthapadi Krishnaram
• Venkataraman Ganesh

Classifications

• G06Q10/40

  Business processes related to social networking or social networking services · CPC title

• G06N20/00

  Machine learning · CPC title

• G06Q10/1053 Primary

  Employment or hiring · CPC title

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06N99/005

  Physics · mapped topic