Method, system, and non-transitory computer-readable medium for constructing and applying synaptic networks

• US 10,789,526 B2
• Filed: 01/30/2017
• Issued: 09/29/2020
• Est. Priority Date: 03/09/2012
• Status: Active Grant

• Current Assignee: Nara Logics, Inc.
• Original Assignee: Nara Logics, Inc.
• Inventors: Wilson, Nathan R., Zubair, Sahil, Ichinco, Denise, Plante, Raymond J., Eggers, Jana B.
• Primary Examiner(s): Miller, Brittney N
• Assistant Examiner(s): Weiner, Arielle E
• Application Number:
  US15/419,517
• Publication Number:
  US 20170140262A1
• Time in Patent Office: 1,338 Days
• Field of Search: 705 267
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: G06N 3/042 Knowledge-based neural networks; Logical representations of neural networks
• CPC Class Codes: G06N 3/082 modifying the architecture, e.g. adding, deleting or silencing nodes or connections
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising
• CPC Class Codes: G06Q 30/0269 based on user profile or attribute
• CPC Class Codes: G06Q 30/0282 Rating or review of business operators or products
• CPC Class Codes: H04L 67/01 Protocols
• CPC Class Codes: H04L 67/04 specially adapted for terminals or networks with limited capabilities; specially adapted for terminal portability
• CPC Class Codes: H04L 67/125 involving control of end-device applications over a network
• CPC Class Codes: H04L 67/306 User profiles
• CPC Class Codes: H04L 67/53 using third party service providers
• CPC Class Codes: H04L 67/535 Tracking the activity of the user network monitoring arrangements H04L43/00; recording of computer activity G06F11/34
• CPC Class Codes: H04L 67/568 Storing data temporarily at an intermediate stage, e.g. caching

Abstract

In selected embodiments a recommendation generator builds a network of interrelationships between venues, reviewers and users based on attributes and reviewer and user reviews of the venues. Each interrelationship or link may be positive or negative and may accumulate with other links (or anti-links) to provide nodal links the strength of which are based on commonality of attributes among the linked nodes and/or common preferences that one node, such as a reviewer, expresses for other nodes, such as venues. The links may be first order (based on a direct relationship between, for instance, a reviewer and a venue) or higher order (based on, for instance, the fact that two venue are both liked by a given reviewer). The recommendation engine in certain embodiments determines recommended venues based on user attributes and venue preferences by aggregating the link matrices and determining the venues which are most strongly coupled to the user.

Claims

1. A method comprising:

receiving, at one or more servers, item data for a plurality of items, the item data relating to a plurality of attributes of the items;

receiving, at the one or more servers, attribute data for the plurality of items, the attribute data relating to a plurality of attributes of at least a first item of the plurality of items for which at least one user of a plurality of users has an affinity;

receiving, at the one or more servers, reviewer data for the plurality of items, the reviewer data reflecting the affinity of a plurality of reviewers for the plurality of items;

generating, by processing circuitry, a synaptic data network having a plurality of nodes corresponding at least to each item of the plurality of items, each reviewer of the plurality of reviewers, and each attribute of the plurality of attributes and further including a plurality of connections between the plurality of nodes, each connection of the plurality of connections reflecting a strength of an interrelationship between at least two nodes of the plurality of nodes, wherein generating the synaptic data network comprises

organizing the plurality of items into one or more categories, each category comprising a plurality of sub-categories,

initializing the plurality of connections between the plurality of nodes based on relationships between pairs of the plurality of the nodes as identified from the item data, the attribute data, and the reviewer data, wherein the plurality of connections comprises a number of connections within each category of the one or more categories and/or the plurality of sub-categories of each category of the one or more categories,

accessing a framework comprising a plurality of synaptic learning rules for increasing accuracy of relationships between the plurality of nodes, and

applying, to the plurality of connections, each rule of at least a portion of the plurality of synaptic learning rules to modify the strength of the interrelationship between at least a portion of the plurality of nodes,

wherein applying at least the portion of the synaptic learning rules comprises constructing a plurality of inhibition nodes that manage relative numbers of nodal activations among the one or more categories and/or the plurality of sub-categories of each category of the one or more categories, thereby dynamically normalizing nodal activations to within a predetermined range; and

determining, by the processing circuitry based on the strength of interrelationship between a user of the plurality of users and each node of at least a portion of the plurality of item nodes, a plurality of recommended items having the strongest connections to the user, wherein determining the plurality of recommended items comprises

applying at least one of one or more synaptic retrieval rules to the synaptic data network to identify affinities of the user for at least a portion of the plurality of recommended items, and

applying a learning model incorporating at least one of the one or more synaptic retrieval rules to estimate an affinity of the user to at least one node, wherein applying the learning model comprises

identifying, among the portion of nodes, at least one node connected to a set of nodes representing a plurality of attributes and lacking connections representing reviewer data, and

applying the learning model to the respective plurality of attributes to estimate the affinity of the user to the respective node.

2. The method of claim 1, wherein generating the synaptic network comprises generating a recurrent structure configured to indicate the connection strengths between the nodes across at least one of i) at least two categories of the one or more categories or ii) at least two sub-categories of a given category of the one or more categories.

3. The method of claim 1, wherein applying each rule of the plurality of synaptic learning rules further includes iteratively applying the plurality of synaptic learning rules to the connections between the nodes to increase or decrease the strength of the interrelationship between at least a portion of the plurality of nodes.

4. The method of claim 1, wherein applying each rule of the plurality of synaptic learning rules further includes applying a bottom-up approach to updating the strengths of the interrelationships between at least a portion of the plurality of nodes, wherein

a first set of connections of the plurality of connections between a set of local nodes of the plurality of nodes are updated before further sets of connections of the plurality of connections between one or more sets of higher level nodes.

5. The method of claim 1, wherein applying each rule comprises forming or strengthening each connection of one or more connections between two or more nodes responsive to the two or more nodes being activated within a predetermined period of time of one another.

6. The method of claim 5, wherein forming or strengthening each connection between the two or more nodes comprises increasing the strength of connection of each connection between the two or more nodes as a number of activations of the two or more nodes within the predetermined period of time increases.

7. The method of claim 5, wherein forming or strengthening the connection between the two or more nodes comprises increasing the strength of connection of each connection between the two or more nodes as an amount of time between activations of the two or more nodes decreases.

8. The method of claim 1, wherein applying each rule further includes weakening the strength of connection of each connection between the two or more nodes as a number of connections to a particular node of the two or more nodes increases.

9. The method of claim 1, wherein applying each rule further includes weakening a strength of connection of a first connection of the plurality of connections between a first node and one or more additional nodes in response to increasing the strength of connection of a second connection of the plurality of connections between the first node and a second node.

10. The method of claim 1, wherein applying each rule further includes increasing a strength of connection of a first connection of the plurality of connections between a first node and one or more additional nodes in response to weakening a strength of connection of a second connection of the plurality of connections between the first node and a second node.

11. The method of claim 1, wherein applying each rule further includes consolidating multiple paths of indirect connection between a given pair of nodes by adding, to the plurality of connections, a new direct connection between the given pair of nodes.

12. The method of claim 1, wherein applying each rule further includes uniformly applying a scaling factor to strengths of connection of the plurality of connections between the plurality of nodes based on an amount of activity across the synaptic data network.

13. The method of claim 1, wherein applying each rule further comprises:

identifying two or more converging input nodes each converging as an affinity of the user to a given item node; and

increasing the connection strengths of the two or more converging input nodes.

14. The method of claim 1, wherein applying the at least one of the one or more synaptic retrieval rules comprises suppressing one or more competing or contradictory connections to the at least one recommended item.

15. The method of claim 1, wherein applying the at least one of the one or more synaptic retrieval rules further includes applying a lateral inhibition/competition synaptic retrieval rule configured to generate two or more nodal sub-networks, each nodal sub-network comprising two or more connected nodes, wherein each nodal sub-network of the two or more nodal sub-networks, upon activation of one or more nodes within the respective nodal sub-network, cross-suppresses each other nodal subnetwork of the two or more nodal sub-networks.

16. The method of claim 1, wherein applying the two or more synaptic retrieval rules further includes dynamically normalizing a number of activations to one or more nodes to within a predetermined range based on a connection from each of the one or more nodes to a respective inhibition node of the plurality of inhibition nodes.

17. The method of claim 1, wherein:

determining the plurality of recommended items comprises activating a subset of the plurality of nodes responsive to identifications of likes and dislikes from the user, each like or dislike corresponding to an item or attribute; and

applying the at least one synaptic retrieval rule comprises propagating the likes and dislikes through the synaptic data network to surface at least the portion of the plurality of recommended items.

18. The method of claim 1, wherein determining the plurality of recommended items comprises determining, for each item of the plurality of recommended items, a reason for recommending the respective item, wherein the reason is derived from information attached to the node corresponding to the respective item and at least one of a path of connection leading to the node or a strength of connection leading to the node.

19. A system comprising:

circuitry of one or more computing devices configured to

receive item data for a plurality of items, the item data relating to a plurality of attributes of the items,

receive attribute data for the plurality of items, the attribute data relating to a plurality of attributes of at least a first item of the plurality of items for which at least one user of a plurality of users has an affinity,

receive reviewer data for the plurality of items, the reviewer data reflecting the affinity of a plurality of reviewers for the plurality of items,

generate a synaptic data network having a plurality of nodes corresponding at least to each item of the plurality of items, each reviewer of the plurality of reviewers, and each attribute of the plurality of attributes and further including a plurality of connections between the plurality of nodes, each connection of the plurality of connections reflecting a strength of an interrelationship between at least two nodes of the plurality of nodes, wherein generating the synaptic data network comprises

organizing the plurality of items into one or more categories, each category comprising a plurality of sub-categories,

initializing the plurality of connections between the plurality of nodes based on relationships between pairs of the plurality of the nodes as identified from the item data, the attribute data, and the reviewer data, wherein the plurality of connections comprises a number of connections within each category of the one or more categories and/or the plurality of sub-categories of each category of the one or more categories,

accessing a framework comprising a plurality of synaptic learning rules for increasing accuracy of relationships between the plurality of nodes, and

applying, to the plurality of connections, each rule of at least a portion of the plurality of synaptic learning rules to modify the strength of the interrelationship between at least a portion of the plurality of nodes,

wherein applying at least the portion of the synaptic learning rules comprises constructing a plurality of inhibition nodes that manage relative numbers of nodal activations among the one or more categories and/or the plurality of sub-categories of each category of the one or more categories, thereby dynamically normalizing nodal activations to within a predetermined range; and

determine, based on the strength of interrelationship between a user of the plurality of users and each node of at least a portion of the plurality of item nodes, a plurality of recommended items having the strongest connections to the user, wherein determining the plurality of recommended items comprises

applying at least one of one or more synaptic retrieval rules to the synaptic data network to identify affinities of the user for at least a portion of the plurality of recommended items, and

applying a learning model incorporating at least one of the one or more synaptic retrieval rules to estimate an affinity of the user to at least one node, wherein applying the learning model comprises

identifying, among the portion of nodes, at least one node connected to a set of nodes representing a plurality of attributes and lacking connections representing reviewer data, and

applying the learning model to the respective plurality of attributes to estimate the affinity of the user to the respective node.

20. A non-transitory computer-readable medium having stored thereon computer-readable instructions which when executed by a processing circuitry cause the processing circuitry to perform a method comprising:

receiving item data for a plurality of items, the item data relating to a plurality of attributes of the items;

receiving attribute data for the plurality of items, the attribute data relating to a plurality of attributes of at least a first item of the plurality of items for which at least one user of a plurality of users has an affinity;

receiving reviewer data for the plurality of items, the reviewer data reflecting the affinity of a plurality of reviewers for the plurality of items;

generating a synaptic data network comprising a plurality of nodes corresponding at least to each item of the plurality of items, each reviewer of the plurality of reviewers, and each attribute of the plurality of attributes and further including a plurality of connections between the nodes, each connection of the plurality of connections reflecting a strength of an interrelationship between at least two nodes of the plurality of nodes, wherein generating the synaptic data network comprises

organizing the plurality of items into one or more categories, each category comprising a plurality of sub-categories,

initializing the plurality of connections between the plurality of nodes based on relationships between pairs of the plurality of the nodes as identified from the item data, the attribute data, and the reviewer data, wherein the plurality of connections comprises a number of connections within each category of the one or more categories and/or the plurality of sub-categories of each category of the one or more categories,

accessing a framework comprising a plurality of synaptic learning rules for increasing accuracy of relationships between the plurality of nodes, and

applying, to the plurality of connections, each rule of at least a portion of the plurality of synaptic learning rules to modify the strength of the interrelationship between at least a portion of the plurality of nodes,

wherein applying at least the portion of the synaptic learning rules comprises constructing a plurality of inhibition nodes that manage relative numbers of nodal activations among the one or more categories and/or the plurality of sub-categories of each category of the one or more categories, thereby dynamically normalizing nodal activations to within a predetermined range; and

determining, by the processing circuitry based on the strength of interrelationship between a user of the plurality of users and each node of at least a portion of the plurality of item nodes, a plurality of recommended items having the strongest connections to the user, wherein determining the plurality of recommended items comprises

applying at least one of one or more synaptic retrieval rules to the synaptic data network to identify affinities of the user for at least a portion of the plurality of recommended items, and

applying a learning model incorporating at least one of the one or more synaptic retrieval rules to estimate an affinity of the user to at least one node, wherein applying the learning model comprises

identifying, among the portion of nodes, at least one node connected to a set of nodes representing a plurality of attributes and lacking connections representing reviewer data, and

applying the learning model to the respective plurality of attributes to estimate the affinity of the user to the respective node.