# US 12,223,274 B2 · Representation learning with side information > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,223,274 B2 - Document ID: US12223274B2 - Status: Granted - Kind: B2 - Application number: 17/452,818 - Filed: 2021-10-29 - Published: 2023-05-04 - Granted: 2025-02-11 - Earliest claimed: 2021-10-29 - Assignees: Microsoft Technology Licensing, LLC - Applicants: Microsoft Technology Licensing, LLC - Inventors: Oren Barkan; Avi Caciularu; Idan Rejwan; Yonathan Weill; Noam Koenigstein; Ori Katz; Itzik Malkiel; Nir Nice - Web page: https://patentgrep.com/patent/US12223274B2 ## CPC - G06F 16/35: Clustering; Classification - G06F 40/295: Named entity recognition - G06F 16/282: {Hierarchical databases, e.g. IMS, LDAP data stores or Lotus Notes} - G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks - G06N 20/00: Machine learning - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract A relational similarity determination engine receives as input a dataset including a set of entities and co-occurrence data that defines co-occurrence relations for pairs of the entities. The relational similarity determination engine also receives as input side information defining explicit relations between the entities. The relational similarity determination engine jointly models the co-occurrence relations and the explicit relations for the entities to compute a similarity metric for each different pair of entities within the dataset. Based on the computed similarity metrics, the relational similarity determination engine identifies a most similar replacement entity from the dataset for each of the entities within the dataset. For a select entity received as an input, the relational similarity determination engine outputs the identified most similar replacement entity. ## Claims ### Claim 1 A method comprising: training a machine learning (ML) model with inputs including: a dataset defining entities and including co-occurrence data, the co-occurrence data including co-occurrence relations for pairs of the entities within the dataset; and side information defining explicit relations between the pairs of the entities, the side information including an explicitly-provided relation type identifier that identifies a relationship type for multiple of the pairs of the entities in the dataset; and jointly modeling, with the ML model, the co-occurrence relations and the explicit relations for the pairs of the entities to compute similarity metrics for different pairs of the entities within the dataset, the similarity metrics for a given pair of the entities including: a first probability distribution representing a relational similarity between two entities of the given pair based on the co-occurrence relations for the two entities within the dataset; a second probability distribution representing a second relational similarity between the two entities of the given pair based on the explicitly-provided relation type identifier for the given pair; and a joint probability distribution that is based on the first probability distribution and the second probability distribution; based on the similarity metrics, identifying a most similar replacement entity from the dataset for each of the entities within the dataset; and for a select entity received as an input, outputting the most similar replacement entity. ### Claim 2 (depends on 1) The method of claim 1, wherein the side information further includes hierarchical relations in addition to the explicit relations, each of the hierarchical relations providing a network of informative priors for an individual entity within the dataset. ### Claim 3 (depends on 2) The method of claim 2, wherein jointly modeling further comprises: jointly modeling the hierarchical relations, co-occurrence relations, and explicit relations to compute at least one of the similarity metrics for each different pair of the entities within the dataset. ### Claim 4 (depends on 2) The method of claim 2, wherein jointly modeling the hierarchical relations, co-occurrence relations, and the explicit relations further comprises: learning a single representation for each of the entities within the dataset that is informed by the hierarchical relations for the entity, the explicit relations for the entity, and the co-occurrence relations for the entity. ### Claim 5 (depends on 1) The method of claim 1, wherein joint modeling the co-occurrence relations and explicit relations includes mapping the entities onto probability densities computed by a Bayesian network. ### Claim 6 (depends on 5) The method of claim 5, wherein the first probability distribution is a first posterior predictive distribution based on the co-occurrence relations and the second probability distribution is a second posterior predictive distribution based on the explicit relations. ### Claim 7 (depends on 6) The method of claim 6, wherein the the joint probability distribution is based on the first posterior predictive distribution and the second posterior predictive distribution. ### Claim 8 (depends on 1) The method of claim 1, wherein the method is implemented by a recommendation system and the dataset is a collaborative filtering dataset. ### Claim 9 A system comprising: a relational similarity determination engine stored in memory and configured to: receive as a first input a dataset defining entities and including co-occurrence data, the co-occurrence data including co-occurrence relations for pairs of the entities; and receive as a second input side information defining explicit relations between the pairs of the entities, the side information including an explicitly-provided relation type identifier that identifies a relationship type for of the pairs of the entities in the dataset; and jointly model the co-occurrence relations and the explicit relations for the entities to compute similarity metrics for each different pair of the entities within the dataset, the similarity metrics for a given pair of the entities including: a first probability distribution representing a relational similarity between two entities of the given pair based on the co-occurrence relations for the two entities within the dataset: a second probability distribution representing a second relational similarity between the two entities of the given pair based on the explicitly-provided relation type identifier for the given pair; and a joint probability distribution that is based on the first probability distribution and the second probability distribution; identify, based on the similarity metrics, a most similar replacement entity from the dataset for each of the entities within the dataset; and for a select entity received as an input, output the most similar replacement entity. ### Claim 10 (depends on 9) The system of claim 9, wherein the side information further includes hierarchical relations in addition to the explicit relations, each of the hierarchical relations providing a network of informative priors for an individual entity within the dataset. ### Claim 11 (depends on 10) The system of claim 10, wherein the relational similarity determination engine jointly models the hierarchical relations, co-occurrence relations, and the explicit relations to compute a similarity metric for each different pair of the entities within the dataset. ### Claim 12 (depends on 10) The system of claim 10, wherein the joint modeling of the hierarchical relations, co-occurrence relations, and the explicit relations further comprises: learning a single representation for each of the entities within the dataset that is informed by the hierarchical relations for the entity, the explicit relations for the entity, and the co-occurrence relations for the entity. ### Claim 13 (depends on 9) The system of claim 9, wherein jointly modeling the co-occurrence relations and the explicit relations includes mapping the entities onto nodes of a Bayesian network. ### Claim 14 (depends on 9) The system of claim 9, wherein the relational similarity determination engine maps the entities onto probability densities computed by a Bayesian network. ### Claim 15 (depends on 9) The system of claim 9, wherein computing the similarity metric for each pair of the entities further comprises: computing a first posterior predictive distribution based on the co-occurrence relations and computing a second posterior predictive distribution based on the explicit relations. ### Claim 16 (depends on 9) The system of claim 9, wherein the system is a recommendation system and the co-occurrence relations are included within a collaborative filtering dataset. ### Claim 17 One or more non-transitory computer-readable storage media encoding processor-executable instructions for executing a computer process, the computer process comprising: training a machine learning (ML) model with inputs including: a dataset defining entities and including co-occurrence data, the co-occurrence data including co-occurrence relations for pairs of the entities; and side information defining explicit relations between the pairs of the entities, the side information including an explicitly-provided relation type identifier that identifies a relationship type for each of the pairs of the entities in the dataset; and jointly modeling, with the ML model, the co-occurrence relations and the explicit relations for the entities to compute similarity metrics for each different pair of the entities within the dataset, the similarity metrics for a given pair of the entities including: a first probability distribution representing a relational similarity between two entities of the given pair based on the co-occurrence relations for the two entities within the dataset; a second probability distribution representing a second relational similarity between the two entities of the given pair based on the explicitly-provided relation type identifier for the given pair; and a joint probability distribution that is based on the first probability distribution and the second probability distribution; based on the joint probability distribution, identifying a most similar replacement entity from the dataset for each of the entities within the dataset; and for a select entity received as an input, outputting the most similar replacement entity. ### Claim 18 (depends on 17) The one or more computer-readable storage media of claim 17, wherein the side information further includes hierarchical relations in addition to the explicit relations, each of the hierarchical relations providing a network of informative priors for an individual entity within the dataset. ### Claim 19 (depends on 18) The one or more computer-readable storage media of claim 18, wherein jointly modeling further comprises: jointly modeling the hierarchical relations, co-occurrence relations, and the explicit relations to compute a similarity metric for each of the different pairs of the entities within the dataset. ### Claim 20 (depends on 17) The one or more computer-readable storage media of claim 17, wherein jointly modeling the co-occurrence relations and the explicit relations includes mapping the entities onto probability densities computed by a Bayesian network.