A knowledge representation learning method for product indicator knowledge graph
By converting product indicator data into graph triplets and adopting joint training schemes, the problem of neglecting numerical attributes in the existing technology is solved, and the accuracy of knowledge graphs and the prediction ability of complex product indicator parameters are improved.
Patent Information
- Application Number
- CN202111007994.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-31
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-08-31
AI Technical Summary
The existing knowledge graph representation learning method only focuses on discrete correlations between entities-relationships, ignoring numerical properties, resulting in insufficient expression of complex product index parameters, affecting the model prediction accuracy.
A knowledge representation learning method for product indicator knowledge graph is proposed. By converting product indicator data into graph triplets, a discrete interval division strategy for numerical indicators is designed, and a relational triplet-indicator triplet is used to learn in sequence The joint training scheme uses cutting-edge knowledge representation models such as TransE or ComplEx to optimize the representation method of numerical indicators.
It improves the accuracy of the knowledge graph, enhances the prediction ability of product performance indicators, and improves the expression accuracy of complex product indicator parameters and model prediction accuracy.
Smart Images

Figure CN113705815B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of knowledge graphs and relates to a knowledge representation learning method oriented to a product indicator knowledge graph. Background Art
[0002] With the rapid development of science and technology, the number, variety and complexity of complex products in the field of industrial manufacturing are also increasing rapidly in terms of functions, configurations and performance indicators. Traditional complex product demand analysis methods are difficult to meet customers' requirements for rapid response, and it is difficult to maintain and improve the market competitiveness of complex product R&D companies. Constructing a knowledge graph of indicator parameters for complex products can assist enterprise product R&D personnel in conducting rapid statistical analysis of existing product data. At the same time, using cutting-edge graph representation reasoning algorithms can provide effective predictive reasoning for the indicator parameter requirements of new products.
[0003] Knowledge representation learning methods have been a hot topic in the field of artificial intelligence and knowledge graphs in recent years. By representing entities and relations in knowledge graphs as dense vectors in low-dimensional vector spaces, knowledge representation learning has been applied to knowledge graph completion, relationship link prediction, and a series of knowledge graph downstream tasks. However, the knowledge graphs used by existing representation learning methods only focus on discrete associations between entities and relations, without considering numerical attributes. For product indicator knowledge graphs, most product indicator parameters are numerical data. When the values of a category of product indicators are precise enough and the range of variation of different product indicators is large, similar values will be expressed as a large number of independent discrete entities. These entities not only consume a large number of training parameters, but also have a negative impact on the prediction accuracy of the model. Summary of the invention
[0004] The present invention aims to solve the problem that the existing knowledge graph representation learning method only focuses on the discrete associations between entities and relationships and ignores the numerical attributes, and provides a knowledge representation learning method for product indicator knowledge graph to improve the accuracy of the knowledge graph.
[0005] To solve the above problems, the present invention is achieved through the following technical solutions:
[0006] A knowledge representation learning method for product indicator knowledge graph includes the following steps:
[0007] Step 1: Extract product indicator graph data from product indicator data, organize it into editable documents, preprocess the data, and clean the data of a series of complex product indicators after verification. According to the pre-defined complex product ontology structure, "product-relationship-entity" is defined as a discrete relationship triple, and "product-indicator-value" is defined as a numerical indicator triple. In these two forms, the structured data is converted into graph triples as the data set of the knowledge representation learning algorithm.
[0008] Step 2: Design a discrete interval division strategy for product numerical indicators, and innovatively propose a distributed representation method for numerical indicator parameters. The key elements in the numerical indicator representation method include the number of intervals and the division method.
[0009] For the Number of Intervals factor, select five options, namely 5, 10, 20, 50 and 100.
[0010] As for the division method factor, the present invention selects two different methods, namely, uniform division based on numerical value and uniform division based on frequency.
[0011] First, the Uniform method gives the maximum value m of the indicator data m in the previous statistics. max and the minimum value m min Then, the maximum interval is directly divided into N parts (where N is the preset number of intervals). The length of each divided interval n = (m max -m min ) / N, the numerical intervals of each interval are basically equal. After partitioning using the Uniform partitioning method, the set of partitioned intervals Z_U can be expressed as:
[0012] Z_U={(m min +n*i,m min +n*(i+1))|0≤i<N}
[0013] Secondly, the Scale method considers the distribution of known indicator parameters to ensure that the number of indicator parameters in each interval is the same. The number of values in each partition interval is l, that is, l = L / N, where L represents the total number of samples. After partitioning using the Scale partitioning method, the set of partition intervals Z_S can be expressed as:
[0014] Z_S={(m l*i ,m l*(i+1) )|0≤i<N}
[0015] Step 3: Select a cutting-edge knowledge representation model and use negative sampling loss as the training target, so that the score of each correct triplet t is smaller than the randomly generated negative sample triplet. T represents the set of positive sample triplets, and T′ represents the set of negative sample triplets. Taking the maximum margin loss as an example, the loss function is defined as:
[0016]
[0017] where φ(t) is the score function of the model and γ>0 is the boundary value parameter.
[0018] Among them, the knowledge representation model between entity vectors and relationship vectors can be TransE or RotatE.
[0019] When the TransE knowledge representation model is used, the score function φ1(t) is:
[0020] φ1(t)=||e h +re t ||
[0021] When the ComplEx knowledge representation model is used, the score function φ2(t) is:
[0022]
[0023] Among them, (h, r, t) represents the triple of the knowledge graph, that is, the positive example triple, h represents the head entity, t represents the tail entity, r represents the relationship between the head entity h and the tail entity t, e c represents a complex vector.
[0024] Step 4: In view of the semantic differences between relation triples and indicator triples, a joint training scheme of learning relation triples and indicator triples in sequence is adopted.
[0025] The present invention proposes a sequential learning training scheme. For the same knowledge representation learning model, discrete relational triples are first used for training for several rounds and then numerical indicator triples are input for training. Relational triples are used to assist in training numerical triples to improve the accuracy of knowledge graph link prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 This is an example graph of entity and relationship triples in the numerical indicator knowledge graph.
[0027] Figure 2 This is a structural diagram of a knowledge representation learning method for product indicator knowledge graph according to the present invention.
[0028] Figure 3 It is a graph of the joint learning method (sequential learning) of relation-indicator triplets. DETAILED DESCRIPTION
[0029] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0030] Step 1: Extract product index graph data from product index data. Store the associations between products and product index parameters in the form of triples as data sets for the knowledge representation learning algorithm.
[0031] Step 1-1, select the data source. Select "Vehicle Information Detection Dataset" as the data source.
[0032] Step 1-2, data extraction. Since the "Vehicle Information Detection Dataset" is a non-editable PDF scan file, from the perspective of time cost and labor cost, an automated method of extracting PDF scan file information is adopted. Through optical character recognition (OCR) technology, complex product indicator data is identified from the PDF scan file and saved in the initial Excel spreadsheet file. However, there is obvious noise in the extracted data. Typical errors include the inability to recognize certain mathematical symbols, missing decimal points in numerical data, spaces in numerical data, and failure to recognize product names involving special characters.
[0033] Step 1-3: Data verification: To ensure that the extracted indicator data is consistent with the original data, we manually verify the Excel data content and modify the error information.
[0034] Steps 1-4, a series of data cleaning work on complex product indicators was carried out, mainly including splitting complex headers, adding complex product category information (primary classification information, secondary classification information), unifying the data format of numerical structure indicators, classifying text structure indicator data, generating complex product ontology structure files, etc. After the above pre-processing, the data are all structured data. According to the pre-defined complex product ontology structure, we define "product-relationship-entity" as discrete relationship triples and "product-indicator-value" as numerical indicator triples. In these two forms, the structured data is converted into graph triples, which are used as data sets for subsequent knowledge representation learning models.
[0035] Step 2: Design a discrete interval division strategy for product numerical indicators. The key elements in the numerical indicator representation method include the number of intervals and the division method.
[0036] Step 2-1: For the number of intervals, we selected five options, namely 5, 10, 20, 50 and 100. Intuitively, too many or too few interval divisions will affect the accuracy of the knowledge representation learning model.
[0037] Step 2-2: For the division method factor, the present invention selects two different methods, namely, uniform division based on numerical value and scale division based on frequency. First, the uniform method takes the maximum value m of the indicator data m in the previous statistics as max and the minimum value m min Then, the maximum interval is directly divided into N parts (where N is the preset number of intervals). The length of each divided interval n = (m max -m min ) / N, the numerical intervals of each interval are basically equal. After partitioning using the Uniform partitioning method, the set of partitioned intervals Z_U can be expressed as: Z_U={(m min +n*i,m min +n*(i+1))|0≤i<N};
[0038] Secondly, the Scale method considers the distribution of known indicator parameters to ensure that the number of indicator parameters in each interval is the same. When N = 4, the Scale method is equivalent to statistical quartiles, using quartiles and mean to divide the entire value space into 4 parts. Then the number of values in each partition interval is l, that is, l = L / N, where L represents the total number of samples. After partitioning using the Scale partitioning method, the set of partition intervals Z_S can be expressed as: Z_S = {(m l*i ,m l*(i+1) )|0≤i<N};
[0039] Step 3: Select a cutting-edge knowledge representation model and use negative sampling loss as the training target, so that the score of each correct triplet t is smaller than the randomly generated negative sample triplet. T represents the set of positive sample triplets, and T′ represents the set of negative sample triplets. Taking the maximum margin loss as an example, the loss function is defined as:
[0040]
[0041] where φ(t) is the score function of the model and γ>0 is the boundary value parameter.
[0042] Step 4: In view of the semantic differences between relation triples and indicator triples, a joint training scheme of relation triples and indicator triples is adopted. A training scheme of sequential learning is proposed. For the same knowledge representation learning model, relation triples are first trained for several rounds and then indicator triples are input for training. For example, Figure 2 As shown, relational triples are used to assist in training numerical triples to improve the accuracy of knowledge graph link prediction.
[0043] The present invention constructs a knowledge representation model, projects entities, relationships and indicator parameters in the product indicator map into a low-dimensional vector space, realizes the semantic information representation of entity relationships, mines the potential associations between indicator parameters, and realizes the effective prediction of product performance indicators. An innovative vector representation method for numerical indicator parameters is proposed, and the product-indicator-attribute data contained in the knowledge graph is represented in a distributed manner. A joint training scheme of learning relationship triples-indicator triples in sequence is adopted, which improves the accuracy of knowledge graph link prediction.
[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A knowledge representation learning method for product indicator knowledge graph, whose characteristics include Follow these steps: Step 1: Extract product indicator graph data from product indicator data, organize it into editable documents, pre-process the data, and clean the data of a series of complex product indicators after verification; according to the pre-defined complex product ontology structure, define "product-relationship-entity" as discrete relationship triples, and define "product-indicator-value" as numerical indicator triples. In these two forms, convert structured data into graph triples as data sets for knowledge representation learning algorithms; Step 2: The key elements in the numerical indicator representation method include the number of intervals and the division method; For the factor of number of intervals, five options were selected, namely 5, 10, 20, 50 and 100; For the division method factor, two different methods were selected, namely, equal division based on numerical value (Uniform) and equal division based on frequency (Scale); First, the Uniform method gives the maximum value m of the indicator data m in the previous statistics. max and the minimum value m min After that, the maximum interval is directly divided into N parts, where N is the preset number of intervals; the length of each divided interval n = (m max -m min ) / N, the numerical intervals of each interval are equal; after partitioning using the Uniform partitioning method, the set of partitioned intervals Z_U can be expressed as: Z_U={(m min +n*i,m min +n*(i+1))|0≤i<N} Secondly, the Scale method considers the distribution of known indicator parameters to ensure that the number of indicator parameters in each interval is the same; the number of values in each partition interval is l, that is, l = L / N, where L represents the total number of samples; after partitioning using the Scale partitioning method, the set Z_S of partition intervals is expressed as: Z_S={(m l*i ,m l*(i+1) )|0≤i<N} Step 3: Select a cutting-edge knowledge representation model and use negative sampling loss as the training target, so that the score of each correct triplet t is less than the randomly generated negative sample triplet; T represents the set of positive sample triplets, and T′ represents the set of negative sample triplets; taking the maximum margin loss as an example, the loss function is defined as: Where φ(t) is the score function of the model, and γ>0 is the boundary value parameter; Among them, the knowledge representation model between entity vectors and relationship vectors is selected from TransE or RotatE; When the TransE knowledge representation model is used, the score function φ1(t) is: φ1(t) = ||e h +r - e t || When the ComplEx knowledge representation model is used, the score function φ2(t) is: Among them, (h, r, t) represents the triple of the knowledge graph, that is, the positive example triple, h represents the head entity, t represents the tail entity, r represents the relationship between the head entity h and the tail entity t, e c represents a complex vector; Step 4: In order to better learn the vector representation of the product indicator graph, a joint training scheme of learning relation triples and indicator triples in sequence is adopted to target the semantic differences between relation triples and indicator triples.
