METHODS OF INFERRING USER PREFERENCES USING ONTOLOGIES

• US 20080010272A1
• Filed: 07/09/2007
• Published: 01/10/2008
• Est. Priority Date: 07/07/2006
• Status: Active Grant

• Current Assignee: Ecole polytechnique fãdãrale de lausanne epfl (Schweizerische Eidgenossenschaft)
• Original Assignee: Ecole polytechnique fãdãrale de lausanne epfl (Schweizerische Eidgenossenschaft)
• Inventors: Schickel-Zuber, Vincent, Faltings, Boi
• Granted Patent: US 7,873,616 B2
• Time in Patent Office: Days
• US Class Current: 707/005
• CPC Class Codes: G06F 16/367 Ontology

Abstract

Methods of inferring user preferences using ontologies use ontology structure based similarity and an inference mechanism in defining ontology filtering (OF). Items are recommended to users by identifying items that the users may prefer, but may not have been aware of. This is performed by computing personalized scores for concepts and selecting an instance having a highest score.

Claims

1. A method for recommending an object to a user comprising:

defining an ontology as nodes in a directed acyclic graph each node representing a concept wherein general concepts are parent nodes and specific concepts are child nodes of the parent nodes;

using a propagating function to infer missing scores in the ontology;

calculating personalized scores of the concepts at the nodes of the ontology relative to a known user preference by propagating a score for the known user preference based on concepts of the known user preference, the propagating function, and the directed acyclic graph;

selecting a product because it is instance of a concept that has a high personalized score in the directed acyclic graph; and

recommending the product to the user.

2. The method of claim 1 wherein the scores depend on features of an item wherein all nodes representing instances of a concept will have the same score.

3. The method of claim 1 wherein features that are unknown contribute zero additional value to scores at nodes in the directed acyclic graph.

4. The method of claim 1 wherein concepts that contain identical features have an identical score.

5. The method of claim 1 wherein a formula for calculating an a-priori score is 1/(n_c+2).

6. The method of claim 1 wherein the novelty of the selection of a product is measured to determine the correctness of the product selection.

7. The method of claim 1 wherein a novelty metric is used to measure the correctness of a product selection by the use of the formula Novelty (r_a|r_b)=(|cr_a|−

|cr_a∩

cr_b|)/N.

8. The method of claim 1 wherein the distance between a node a and a node b is calculated.

9. The method of claim 1 wherein a distance between a node a and a node b is calculated by using a formula

10. The method of claim 1 wherein a similarity between two related concepts is calculated by determining the amount of score transferred between them.

11. The method of claim 1 wherein a transfer of a score from a general concept node to a specific concept node is calculated using the formula T(a,b)=S(b|a)/S(a).

12. A method for recommending an object to a user comprising:

defining an ontology as nodes in a directed acyclic graph each node representing a concept wherein general concepts are parent nodes and specific concepts are child nodes of the parent nodes;

wherein a propagating function can be computed from a ratio, or combination of a-priori scores at each node of the directed acyclic graph;

wherein an a-priori (APS) score can be calculated from the topology of concepts in the ontology;

calculating personalized scores of the concept at the nodes of the ontology relative to a known user preference by propagating a score for the known user preference based on concepts of the known user preference, the APSs, and the directed acyclic graph;

the propagating function; and

selecting a product because it contains a feature which is an instance of a concept that has a high personalized score at a node in the ontology.

13. The method of claim 12 wherein calculating an a-priori score is calculated using the formula 1/(n_c+2).

14. The method of claim 12 wherein calculating an inference for an upward link entails multiplying the score of the first node by a value, α

.

15. The method of claim 12 wherein propagating the at least one user specific preference requires an upward inference that is calculated by S(y|x)=α

S(x);

wherein α

is a coefficient of generalization.

16. The method of claim 12 wherein propagating downward requires the addition of a constant value, β

, to the score of the first node.

17. The method of claim 12 wherein propagating the at least one user specific preference requires a downward inference that is calculated by S(y|x)=S(x)+β

.

18. The method of claim 12 wherein an inference upward and then downward requires the multiplication of the score of the first node by a value, α

and then the addition of a separate value, β

.

19. The method of claim 12 wherein propagating the at least one user specific preference requires a upward inference followed by a downward inference that is calculated by S(y|x)=α

S(x)+β

.

20. The method of claim 12 wherein the link connecting the first node and the second node requires going upward and downward via a lowest common ancestor (LCA).

21. The method of claim 12 wherein a lowest common ancestor is determined as a node having a highest a-priori score.

22. The method of claim 12 wherein the novelty of the selection of a product is measured to determine the correctness of the product selection.

23. The method of claim 12 wherein a novelty metric is used to measure the correctness of a product selection by the use of the formula Novelty (r_a|r_b)=(|cr_a|−

|cr_a∩

cr_b|)/N.

24. A method of deriving a personal score of a concept by propagating a known personal score through an ontology comprising:

defining an a-priori score for a first node of an ontology;

defining an a-priori score for a second node of the ontology;

determining a lowest common ancestor node in the ontology connecting the first node and the second node;

defining an a-priori score for the lowest common ancestor node;

calculating an inference from the first node to the second node using the a-priori score for the first node, the a-priori score for the second node, and the a-priori score for the lowest common ancestor node; and

calculating a personalized score for the second node by using the inference and the known personal score for the first node.

25. The method of claim 24 wherein calculating a personalized score is accomplished by using a closest concept that is liked by the user.

26. The method of claim 25 further comprising:

recommending an item using the personalized score.

The method of claim 24 wherein calculating a personalized score is accomplished using a neighborhood concept.

27. The method of claim 26 further comprising:

selecting a product that is an instance of a concept with a highest inferred score.

29. The program of claim 24 wherein propagating the at least one user specific preference requires an upward inference.

30. The program of claim 24 wherein propagating the at least one user specific preference requires a downward inference.

31. The program of claim 24 wherein propagating the at least one user specific preference requires a upward inference followed by a downward inference.

32. The program of claim 24 wherein the link connecting the first node and the second node requires going upward and downward via a lowest common ancestor (LCA).

33. The program of claim 24 wherein the novelty of the selection of a product is measured to determine the correctness of the product selection.

28. A program embodied in a computer readable medium that when executed causes a computing device to:

define an a-priori score for a first node of an ontology;

define an a-priori score for a second node of the ontology;

determine a lowest common ancestor node in the ontology connecting the first node and the second node;

defining an a-priori score for the lowest common ancestor node;

calculate an inference from the first node to the second node using the a-priori score for the first node, the a-priori score for the second node, and the a-priori score for the lowest common ancestor node; and

calculate a personalized score for the second node by using the inference and the known personal score for the first node.