KR-20260063725-A - APPARATUS AND METHOD FOR ANALYZING COUPLED TENSOR WITH KNOWLEDGE GRAPH

Abstract

The embodiments disclosed in this specification relate to a combined tensor analysis apparatus and a method. A combined tensor analysis apparatus according to one embodiment includes a memory storing a program and data for performing combined tensor decomposition and one or more processors, and operates by executing the program. It performs tensor decomposition of a temporal tensor composed of a plurality of slice matrices and containing dynamic information and a knowledge graph tensor composed of a plurality of slice matrices and containing static information, and includes a control unit that performs combined tensor decomposition by sharing some of the plurality of factor matrices obtained as a result of the tensor decomposition of the temporal tensor as factor matrices of the knowledge graph tensor.

Inventors

• 강유
• 이승주
• 박용찬

Assignees

• 서울대학교산학협력단

Dates

Publication Date

20260507

Application Date

20241031

Claims (11)

1. Memory storing a program and data for performing combined tensor decomposition; and A combined tensor analysis device comprising one or more processors and operating by executing the program, and including a control unit that performs tensor decomposition of a temporal tensor composed of a plurality of slice matrices containing dynamic information and a knowledge graph tensor composed of a plurality of slice matrices containing static information, wherein the control unit performs combined tensor decomposition by sharing some of the plurality of factor matrices obtained as a result of the tensor decomposition of the temporal tensor as factor matrices of the knowledge graph tensor.
2. In paragraph 1, The above control unit is, When performing tensor decomposition of the above temporal tensor, factor matrices that approximate the above temporal tensor are obtained, wherein the factor matrices are updated one by one alternately while keeping the factor matrices other than the factor matrix being updated fixed, A combined tensor analysis device that updates the factor matrix to be updated by additionally reflecting, at a preset ratio, a change amount calculated by comparing the update of the previous round and the update of the current round for the factor matrix to be updated.
3. In paragraph 1, A combined tensor analysis device characterized in that each of the factor matrices obtained as a result of decomposing the knowledge graph tensor corresponds to an entity, an item, and a relationship between the entity and the item, which are components of the knowledge graph.
4. In paragraph 3, The above control unit is, A combined tensor analysis device that generates an entity factor matrix, which is a factor matrix corresponding to the above entity, corresponding to each slice matrix of the above knowledge graph tensor, and initializes an entity factor matrix corresponding to the specific slice matrix based on the previously obtained entity factor matrix when decomposing a specific slice matrix.
5. In paragraph 1 The above control unit is, A combined tensor analysis device that performs combined tensor decomposition based on a loss function including a normalization term to reflect the relationship between the reconstruction loss of the temporal tensor and the knowledge graph tensor and the knowledge graph.
6. In a combined tensor analysis method performed by a combined tensor analysis device, A step of obtaining a temporal tensor composed of a plurality of slice matrices and containing dynamic information, and a knowledge graph tensor composed of a plurality of slice matrices and containing static information; and A method for combined tensor analysis, comprising the step of performing tensor decomposition of the temporal tensor and the knowledge graph tensor, wherein some of the factor matrices obtained as a result of the tensor decomposition of the temporal tensor are shared as factor matrices of the knowledge graph tensor to perform combined tensor decomposition.
7. In paragraph 6, The step of performing the above combined tensor decomposition is, When performing tensor decomposition of the above temporal tensor, factor matrices that approximate the above temporal tensor are obtained, wherein the factor matrices are updated one by one alternately while keeping the factor matrices other than the factor matrix being updated fixed, A combined tensor analysis method comprising the step of updating the factor matrix to be updated by additionally reflecting, at a preset ratio, a change amount calculated by comparing the update of the previous round with the update of the current round for the factor matrix to be updated.
8. In paragraph 6, A method for analyzing combined tensors, characterized in that each of the factor matrices obtained as a result of decomposing the knowledge graph tensor corresponds to an entity, an item, and a relationship between the entity and the item, which are components of the knowledge graph.
9. In paragraph 8, The step of performing the above combined tensor decomposition is, A method for analyzing combined tensors, comprising the step of generating an entity factor matrix, which is a factor matrix corresponding to the entity, corresponding to each slice matrix of the knowledge graph tensor, and, when decomposing a specific slice matrix, initializing an entity factor matrix corresponding to the specific slice matrix based on the previously obtained entity factor matrix.
10. A computer program stored on a computer-readable recording medium to perform the method described in claim 6, which is performed by a combined tensor analysis device.
11. A computer-readable recording medium on which a computer program for performing the method described in paragraph 6 is recorded.

Description

Apparatus and Method for Analyzing Coupled Tensors with Knowledge Graph The embodiments disclosed in this specification relate to a knowledge graph-based combined tensor analysis apparatus and method, and more specifically, to a knowledge graph-based combined tensor analysis apparatus and method that integrally decomposes a temporal tensor and a knowledge graph tensor into a plurality of factor matrices. An irregular tensor refers to a tensor composed of multiple matrices with different row sizes. Each of the multiple matrices constituting the tensor is called a slice matrix, and each slice matrix can have the same column size and different row sizes. In particular, as data is generated in complex environments recently, much data, such as stock data and traffic volume data, can be represented as irregular tensors. In particular, since data irregularity often arises from temporal variations, many existing irregular tensor decomposition methods have primarily focused on capturing dynamic features that change over time, and have often failed to consider static features, such as knowledge information that remains constant over time. However, since static features are just as important as dynamic features for accurately modeling factor matrices, the need has arisen for a technique capable of accurately modeling static features by effectively integrating and analyzing static and dynamic features. Meanwhile, the aforementioned background technology is technical information that the inventor possessed for the derivation of the present invention or acquired during the process of deriving the present invention, and it cannot be considered as prior art disclosed to the general public prior to the filing of the present invention. FIG. 1 is a block diagram illustrating a combined tensor analysis device according to one embodiment. FIGS. 2 to 4 are drawings for explaining a combined tensor decomposition process according to one embodiment. FIG. 5 is a flowchart illustrating a combined tensor analysis method according to one embodiment. Various embodiments are described in detail below with reference to the attached drawings. The embodiments described below may be implemented in various different forms. In order to explain the features of the embodiments more clearly, detailed descriptions of matters widely known to those skilled in the art to which the following embodiments belong have been omitted. Additionally, parts of the drawings unrelated to the description of the embodiments have been omitted, and similar parts throughout the specification have been given similar reference numerals. Throughout the specification, when a configuration is described as being "connected" to another configuration, this includes not only cases where they are "directly connected," but also cases where they are "connected with another configuration in between." Furthermore, when a configuration is described as "including" another configuration, this means that, unless specifically stated otherwise, it does not exclude other configurations but may include additional configurations. The embodiments will be described in detail below with reference to the attached drawings. However, before explaining this, the meanings of the terms used below are first defined. An 'irregular tensor' refers to a tensor composed of multiple slice matrices with the same column size and irregular row size. Examples of irregular tensors include stock data, traffic data, and music data. For example, in the case of stock data, it can be composed of slice matrices corresponding to individual stocks, and the format of each slice matrix can be (date, feature); in other words, rows correspond to dates, and columns correspond to features such as trading volume, opening price, and closing price. Therefore, the elements of each slice matrix can represent, for example, the trading volume or the closing price of a specific date. As with stock data, data is stored along a time axis such as dates, so irregular tensors containing time information are called 'temporal tensors,' and since they are features that change over time, they can also be called dynamic features or temporal features. The 'Knowledge Graph' is information A large-scale graph-structured database that stores information in the form of a knowledge triplet, is the Subject(Head) Entity, means Object(Tail) Entity, It refers to the relationship between two entities. This structured knowledge representation can serve as auxiliary data by effectively integrating rich factual information and associations between items, and it can be called a static feature because it is not a feature that changes over time. A knowledge graph can be represented as a third-order binary tensor or as a random tensor. Each element of the knowledge graph corresponds to a triplet, where '1' represents a true fact and '0' represents an unknown fact, which can mean false or missing. In this specification, a knowledge graph may be a random tenso