CN112818136A - Time convolution-based interactive knowledge representation learning model TCIM prediction method - Google Patents

Time convolution-based interactive knowledge representation learning model TCIM prediction method Download PDF

Info

Publication number
CN112818136A
CN112818136A CN202110220136.3A CN202110220136A CN112818136A CN 112818136 A CN112818136 A CN 112818136A CN 202110220136 A CN202110220136 A CN 202110220136A CN 112818136 A CN112818136 A CN 112818136A
Authority
CN
China
Prior art keywords
time
convolution
tcim
learning model
knowledge representation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110220136.3A
Other languages
Chinese (zh)
Inventor
汪璟玢
陆玉乾
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fuzhou University
Original Assignee
Fuzhou University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Fuzhou University filed Critical Fuzhou University
Priority to CN202110220136.3A priority Critical patent/CN112818136A/en
Publication of CN112818136A publication Critical patent/CN112818136A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Game Theory and Decision Science (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Development Economics (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Animal Behavior & Ethology (AREA)
  • Databases & Information Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a prediction method of an interactive knowledge representation learning model TCIM based on time convolution. The invention captures the characteristic information of time through the convolutional neural network, thereby completing the dynamic knowledge graph.

Description

Time convolution-based interactive knowledge representation learning model TCIM prediction method
Technical Field
The invention relates to the technical field of knowledge maps, in particular to a time convolution-based interactive knowledge representation learning model TCIM prediction method.
Background
Most of the existing models focus on researching the knowledge graph which does not contain time information, and the dynamic knowledge graph not only considers the information of the triples, but also considers the time information related to the triples. Relationships between entities in a knowledge graph may change over time. Such as the ICEWS 2014, ICEWS 2005-15 data set. Over time, the fact triplets contained may also change over time. For example (person wasBornIncity) this fact evolves into another fact (person diedIncity) as the relation evolves with time, and current models based on dynamic knowledge maps are the HyTE model proposed by Dasgupta in 2018, Garcia-Dur n proposes a time-aware representation model using recurrent neural networks to learn the type of relation, TA-DistMult, DE-SimplE, and so on. The main idea of the HyTE model is to project head and tail entities onto a time hyperplane respectively and translate on the time hyperplane. The HyTE model achieves a good effect on a time-aware model, but the evolution of roles of entities in relations is not considered in the evolution process of head and tail entities along with time, and the prediction efficiency on many-to-one and many-to-many relations is low. The TA-DistMult model processes time information into a time sequence, then uses a recurrent neural network to capture effective characteristics between the time sequence and the relationship, so that the relationship between entities can sense time, and finally carries out link prediction through a translation model. The nature of the two models is also a translation model, and the two models are different from the traditional translation model in that time information is added into the two models so as to enhance the expressive force of the models.
Most models today study static knowledge-graphs but ignore the important dimension of time, however, the relationships between entities in triples may change over time, and therefore it becomes especially important to develop time-fused knowledge-graph-embedded models. Existing time-aware models do not adequately consider the global features of extracting temporal information and the interactivity between temporal and triples.
Disclosure of Invention
In view of this, the present invention provides a time convolution-based prediction method for an interactive knowledge representation learning model TCIM, which captures characteristic information of time through a convolutional neural network, so as to complete a dynamic knowledge graph.
The invention is realized by adopting the following scheme: a prediction method of an interactive knowledge representation learning model TCIM based on time convolution is characterized in that feature information of time is captured through a convolution neural network, feature information of triplets is extracted through a circular convolution neural network, and features of the triplets and features of the time are fused and link prediction is carried out.
Further, the characteristic information of the time captured by the convolutional neural network is specifically:
the global features between times are extracted by using convolution nerve, and the year, month and day vector is expressed as (v)y,vm,vd) Treat it as a matrix
Figure BDA0002954463540000021
Where D is the dimension of the embedding, vy、vm、vdRespectively representing year, month and day; using convolution kernels on convolutional layers
Figure BDA0002954463540000022
Extracting an embedded time triplet (v)y,vm,vd) The global relationship among the items with the same size is obtained, and the mutual information among the years, the months and the days is obtained at the same time; ω repeats the convolution operation on each line of T to ultimately generate a feature map vector f ═ f1,f2,…fi,...,fn](ii) a Wherein, the following:
Figure BDA0002954463540000023
wherein, b is an offset,
Figure BDA0002954463540000024
is a convolution operation, g is the ReLU activation function;
and splicing feature map vectors generated by different convolution kernels to form a final feature information representation of time information.
Further, the extracting of the feature information of the triplet through the circular convolution neural network specifically includes:
carrying out characteristic arrangement: randomly generating t head entities esAnd relation erIs shown as
Figure BDA0002954463540000025
Performing characteristic remodeling: by a reshaping function
Figure BDA0002954463540000026
Reshaping the feature arrangement, and capturing the maximum interactive features between the entity and the relationship features;
and (3) performing circular filling on the feature map, wherein the feature map after the circular filling is C ═ C1,c2,…,cn]；
Finally, the result obtained by convolution is as follows:
Figure BDA0002954463540000027
in the formula, bhIs a bias that is a function of the bias,
Figure BDA0002954463540000028
is a convolution operation, g is the ReLU activation function, hiRepresenting characteristic information of the triples.
Further, the arrangement of the features is reshaped bysAnd erIs arranged so that esAnd erThe middle elements are not adjacent to each other.
Further, the fusing the characteristics of the triples and the characteristics of the time and performing the link prediction specifically includes:
fusing the triple characteristic graphs and the time characteristic graphs, splicing the generated characteristic graphs, multiplying the characteristic graphs by the weight matrix, and performing dot product on the characteristic graphs and the target entity to obtain the final score of the quadruple (h, r, T, T); the final score is formulated as follows:
Figure BDA0002954463540000031
where x denotes that circular convolution is used,
Figure BDA0002954463540000032
is a convolution operation, T represents a time matrix,
Figure BDA0002954463540000033
representing the merging operation of a time vector and a triplet vector, vec representing the joining operation of the vectors, vtVector matrix representing target entity, W is a learnable parameter matrix, f is activation function, h represents head entity, r represents relation, t represents tail entity, omega1And Ω2Two different convolution kernels are represented.
Further, the tcam model is trained using Adam optimizer minimization loss function, and the score function is processed using sigmoid function δ (·), so that P ═ δ (S (h, r, t)), the loss function is as follows:
Figure BDA0002954463540000034
in the formula, t is a tag vector,
Figure BDA0002954463540000035
is the number of training data, PiIs a predicted value, tiRepresents a group of formulae and PiA corresponding label.
Further, the fusion method comprises feature addition, feature multiplication or gating fusion.
The invention also provides a prediction system for a time convolution based interactive knowledge representation learning model TCIM, comprising a memory, a processor and computer program instructions stored on the memory and executable by the processor, which when executed by the processor, are capable of implementing the method steps as described above.
The present invention also provides a computer readable storage medium having stored thereon computer program instructions executable by a processor, the computer program instructions when executed by the processor being capable of performing the method steps as described above.
Compared with the prior art, the invention has the following beneficial effects:
1. most of the traditional models are used for performing link prediction on a static knowledge graph, the important dimension of time is ignored, and the invention provides a method for capturing characteristic information of time through a convolutional neural network, so that a dynamic knowledge graph can be completed.
2. Most of the existing neural network models do not effectively fuse time characteristic information and triple characteristic information, and the three fusion methods provided by the invention can effectively fuse time and triple characteristics.
Drawings
FIG. 1 is a schematic diagram of a method according to an embodiment of the present invention.
FIG. 2 is a gating diagram according to an embodiment of the present invention.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
As shown in fig. 1, the embodiment provides a prediction method of a time convolution-based interactive knowledge representation learning model TCIM, which captures feature information of time through a convolution neural network, extracts feature information of a triplet through a circular convolution neural network, fuses features of the triplet and features of the time, and performs link prediction.
The present embodiment first gives the following definitions:
definition 1 (quadruplet, S): let S ═ h, r, T denote a quadruple, where h denotes the head entity, r denotes the relationship, T denotes the tail entity, and T denotes time. A quadruple may also be referred to as a knowledge or fact.
Definition 2 (entity set, E): let E ═ E1, E2.., en, denote the set of all entities in the knowledge base.
Definition 3 (set of relationships, R): let the set of relationships R { R1, R2., rn }, denote the set of all relationships in the knowledge base.
In this embodiment, the characteristic information of capturing time by the convolutional neural network specifically includes:
existing temporal knowledge mapping models such as HyTE project entities and relationships onto a hyperplane of time, and TA-DistMult transforms time into time series, features between which are learned using LSTM. But these models ignore the capture of global features within time, so the present embodiment proposes to use a convolutional neural network to extract a model of global features between times.
The set of valid times in the knowledge base G is in the form of years, months, days, denoted as y, m, d. The global features between times are extracted by using convolution nerve, and the year, month and day vector is expressed as (v)y,vm,vd) Treat it as a matrix
Figure BDA0002954463540000051
Where D is the dimension of the embedding, vy、vm、vdRespectively representing year, month and day; using convolution kernels on convolutional layers
Figure BDA0002954463540000052
Extracting an embedded time triplet (v)y,vm,vd) The global relationship among the items with the same size is obtained, and the mutual information among the years, the months and the days is obtained at the same time; ω repeats the convolution operation on each line of T to ultimately generate a feature map vector f ═ f1,f2,…fi,...,fn](ii) a Wherein, as follows:
Figure BDA0002954463540000053
wherein, b is an offset,
Figure BDA0002954463540000054
is a convolution operation, g is the ReLU activation function;
and splicing feature map vectors generated by different convolution kernels to form a final feature information representation of time information.
In this embodiment, the extracting, by using the circular convolutional neural network, the feature information of the triplet specifically includes:
because the prior embedded model based on convolution ignores the interactivity of the entity and the relation, the expressive ability of the model can be strengthened by fully capturing the interactivity between the entity and the relation, and therefore the model has better prediction accuracy. Therefore, in order to enhance the interaction between the entity and the relationship, the embodiment performs three steps, 1) feature arrangement; 2) characteristic remodeling; 3) circular filling, specifically as follows:
carrying out characteristic arrangement: randomly generating t head entities esAnd relation erIs shown as
Figure BDA0002954463540000055
For t different permutations, it is expected that the total number of interactions is approximately t times the number of interactions for one permutation.
Performing characteristic remodeling: by a reshaping function
Figure BDA0002954463540000056
Reshaping the feature arrangement, and capturing the maximum interactive features between the entity and the relationship features;
performing circle filling on the feature map, wherein in the horizontal axis, the embodiment fills the right side of the feature map to the leftmost side of the image and vice versa; on the vertical axis, the top right of the image is filled to the top left, and vice versa. Circular fill was applied to each convolutional layer, since the fill in feature space resulted in better performance in the experiment, with the feature map after circular fill being C ═ C1,c2,…,cn]；
Finally, the result obtained by convolution is as follows:
Figure BDA0002954463540000061
in the formula, bhIs a bias that is a function of the bias,
Figure BDA0002954463540000062
is a convolution operation, g is the ReLU activation function, hiRepresenting characteristic information of the triples.
In this embodiment, the arrangement of features is reshaped by reshaping esAnd erIs arranged so that esAnd erAre not adjacent to each other.
In this embodiment, the fusing the characteristics of the triples and the characteristics of the time and performing the link prediction specifically includes:
fusing the triple characteristic graphs and the time characteristic graphs, splicing the generated characteristic graphs, multiplying the characteristic graphs by the weight matrix, and performing dot product on the characteristic graphs and the target entity to obtain the final score of the quadruple (h, r, T, T); the final score is formulated as follows:
Figure BDA0002954463540000063
where x denotes that circular convolution is used,
Figure BDA0002954463540000064
is a convolution operation, T represents a time matrix,
Figure BDA0002954463540000065
representing the merging operation of a time vector and a triplet vector, vec representing the joining operation of the vectors, vtVector matrix representing target entity, W represents a learnable parameter matrix, f is activation function (in this embodiment, f uses ReLU function), h represents head entity, r represents relation, t represents tail entity, and Ω1And Ω2Representing different convolution kernels. .
In this embodiment, the penalty function uses standard binary cross-entropy penalty in conjunction with tag smoothing, employs Adam optimizer to minimize the penalty function to train the TCIM model, and uses sigmoid function δ (·) to process the score function, so P ═ δ (S (h, r, t)), the penalty function is as follows:
Figure BDA0002954463540000066
in the formula, t is a tag vector,
Figure BDA0002954463540000071
is the number of training data, PiIs a predicted value, tiRepresents a group of formulae and PiA corresponding label.
In this embodiment, the fusion method includes feature addition, feature multiplication, or gated fusion, which is specifically as follows:
1) the features are added. And adding the time characteristic and the relation characteristic to obtain a final characteristic representation.
F＝Tf+Ef
2) The features are multiplied. And multiplying the time characteristic and the relation characteristic to obtain a final characteristic representation. The multiplication here is an element multiplication.
F＝Tf⊙Ef
3) Gated fusion. In the design of the embodiment, a core mechanism gate control mechanism of a long-time and short-time memory neural network (LSTM) is used for reference, effective information is screened by the gate control mechanism, parameter sharing is realized, and the risk of model overfitting is reduced. The calculated model diagram is shown in fig. 2. The formula is as follows:
F＝Ef⊙σ(concat(Ef，Tf)*wh+bh)
wherein wh，bhIs a parameter matrix to be learned,. is a hadamard product,. is a matrix multiplication,. concat is a join operation,. sigma.is an activation function, we use a sigmoid activation function herein.
The present embodiment also provides a prediction system for a time-convolution based interactive knowledge representation learning model TCIM, comprising a memory, a processor and computer program instructions stored on the memory and executable by the processor, which when executed by the processor, are capable of implementing the method steps as described above.
The present embodiments also provide a computer readable storage medium having stored thereon computer program instructions executable by a processor, the computer program instructions, when executed by the processor, being capable of performing the method steps as described above.
Preferably, when the method of the present embodiment is applied to the knowledge-graph complementation, the specific steps are as follows:
1. the triplets (h, r, T) in the knowledge graph are input as models.
2. Carrying out feature rearrangement and feature remodeling on the entity E (including the head entity h and the tail entity r) and the vector of the relation r, and extracting features by using circular convolution to obtain triple feature information Ef。
3. Analyzing the time T into information of (year, month, day), vectorizing the information, and performing convolution operation by using a convolution kernel of 1 x 3 to obtain characteristic information T of the timef。
4. The triple feature E is thenfAnd time characteristic information TfAnd performing fusion operation to obtain a final feature fusion vector representation F.
5. And finally, compressing the characteristics of the fused vector F through a full connection layer, multiplying the compressed characteristic representation by all entity vectors, and finally activating by using sigmoid to obtain the score of each triple.
The present embodiment uses the ICEWS dataset, which is the International crisis Warning System (ICEWS) project by Rockschid Martin. The data set for ICEWS14 is shown in the following table:
Figure BDA0002954463540000081
the evaluation index of this example is Hits @ N: and in the test triple set, the entity ranking is less than or equal to the proportion of N. And (4) ranking the head entity or the tail entity of the original test triple in the top N, adding 1 to the hit times, and adding 0 to the hit times if not. All Hits are then summed and averaged to obtain the value of Hits @ N. A larger Hits @ N indicates better performance.
Finally, the embodiment is verified to obtain the effectiveness of the algorithm of the embodiment, and in the ICEWS14 data set, the model of the embodiment is improved by 0.9% at Hits @10 and 20% at Hits @1 in comparison with a time-lapse embedding (DE-simplex) model for time-aware graph completion in testing. It is concluded from the above results that TCIM using convolutional neural networks can find the characteristics of entities, relationships and time well and achieve better performance.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The foregoing is directed to preferred embodiments of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow. However, any simple modification, equivalent change and modification of the above embodiments according to the technical essence of the present invention are within the protection scope of the technical solution of the present invention.

Claims (9)

1. A prediction method of an interactive knowledge representation learning model TCIM based on time convolution is characterized in that feature information of time is captured through a convolution neural network, feature information of a triple is extracted through a circular convolution neural network, and features of the triple and features of the time are fused and subjected to link prediction.
2. The method for predicting the interactive knowledge representation learning model TCIM based on the time convolution as claimed in claim 1, wherein the characteristic information of the time captured by the convolutional neural network is specifically:
the global features between times are extracted by using convolution nerve, and the year, month and day vector is expressed as (v)y,vm,vd) Treat it as a matrix
Figure FDA0002954463530000011
Where D is the dimension of the embedding, vy、vm、vdRespectively representing year, month and day; using convolution kernels on convolutional layers
Figure FDA0002954463530000012
Extracting an embedded time triplet (v)y,vm,vd) The global relationship among the items with the same size is obtained, and the mutual information among the years, the months and the days is obtained at the same time; ω repeats the convolution operation on each line of T to ultimately generate a feature map vector f ═ f1,f2,…fi,...,fn](ii) a Wherein, the following:
Figure FDA0002954463530000013
wherein, b is an offset,
Figure FDA0002954463530000014
is a convolution operation, g is the ReLU activation function;
and splicing feature map vectors generated by different convolution kernels to form a final feature information representation of time information.
3. The prediction method of the interactive knowledge representation learning model TCIM based on time convolution as claimed in claim 1, wherein the extracting of the feature information of the triplet by the circular convolution neural network is specifically:
carrying out characteristic arrangement: randomly generating t head entities esAnd relation erIs shown as
Figure FDA0002954463530000015
Performing characteristic remodeling: by a reshaping function
Figure FDA0002954463530000016
Reshaping the feature arrangement, and capturing the maximum interactive features between the entity and the relationship features;
and (3) performing circular filling on the feature map, wherein the feature map after the circular filling is C ═ C1,c2,…,cn]；
Finally, the result obtained by convolution is as follows:
Figure FDA0002954463530000017
in the formula, bhIs a bias that is a function of the bias,
Figure FDA0002954463530000021
is a convolution operation, g is the ReLU activation function, hiRepresenting characteristic information of the triples.
4. The method for predicting the interactive knowledge representation learning model TCIM based on time convolution of claim 3, wherein the method for reshaping the arrangement of the features is to use esAnd erIs arranged such that esAnd erThe middle elements are not adjacent to each other.
5. The method for predicting the interactive knowledge representation learning model TCIM based on time convolution according to claim 1, wherein the fusing the characteristics of the triples and the characteristics of the time and performing the link prediction specifically includes:
fusing the triple characteristic graphs and the time characteristic graphs, splicing the generated characteristic graphs, multiplying the characteristic graphs by the weight matrix, and performing dot product on the characteristic graphs and the target entity to obtain the final score of the quadruple (h, r, T, T); the final score is formulated as follows:
Figure FDA0002954463530000022
where x denotes that circular convolution is used,
Figure FDA0002954463530000023
is a convolution operation, T represents a time matrix,
Figure FDA0002954463530000024
representing the merging operation of a time vector and a triplet vector, vec representing the joining operation of the vectors, vtVector matrix representing target entity, W is a learnable parameter matrix, f is activation function, h represents head entity, r represents relation, t represents tail entity, omega1And Ω2Two different convolution kernels are represented.
6. The prediction method for the interactive knowledge representation learning model TCIM based on time convolution is characterized in that an Adam optimizer is adopted to minimize a loss function to train the TCIM model, and a sigmoid function delta () is used to process a score function, so that P ═ delta (S (h, r, t)), the loss function is as follows:
Figure FDA0002954463530000025
in the formula, t is a tag vector,
Figure FDA0002954463530000026
is the number of training data, PiIs the predicted value, tiRepresents a group of formulae and PiA corresponding label.
7. The method for predicting the interactive knowledge representation learning model TCIM based on the time convolution is characterized in that the fusion method comprises feature addition, feature multiplication or gating fusion.
8. A prediction system for a time-convolution based interactive knowledge representation learning model TCIM, comprising a memory, a processor and computer program instructions stored on the memory and executable by the processor, the computer program instructions, when executed by the processor, being capable of carrying out the method steps of any one of claims 1 to 7.
9. A computer-readable storage medium, having stored thereon computer program instructions executable by a processor, the computer program instructions, when executed by the processor, being capable of carrying out the method steps according to any one of claims 1 to 7.
CN202110220136.3A 2021-02-26 2021-02-26 Time convolution-based interactive knowledge representation learning model TCIM prediction method Pending CN112818136A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110220136.3A CN112818136A (en) 2021-02-26 2021-02-26 Time convolution-based interactive knowledge representation learning model TCIM prediction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110220136.3A CN112818136A (en) 2021-02-26 2021-02-26 Time convolution-based interactive knowledge representation learning model TCIM prediction method

Publications (1)

Publication Number Publication Date
CN112818136A true CN112818136A (en) 2021-05-18

Family

ID=75864129

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110220136.3A Pending CN112818136A (en) 2021-02-26 2021-02-26 Time convolution-based interactive knowledge representation learning model TCIM prediction method

Country Status (1)

Country Link
CN (1) CN112818136A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113190691A (en) * 2021-05-28 2021-07-30 齐鲁工业大学 Link prediction method and system of knowledge graph
CN114117064A (en) * 2021-11-09 2022-03-01 西南交通大学 Knowledge dynamic evolution method based on multi-time granularity and application

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113190691A (en) * 2021-05-28 2021-07-30 齐鲁工业大学 Link prediction method and system of knowledge graph
CN113190691B (en) * 2021-05-28 2022-09-23 齐鲁工业大学 Link prediction method and system of knowledge graph
CN114117064A (en) * 2021-11-09 2022-03-01 西南交通大学 Knowledge dynamic evolution method based on multi-time granularity and application

Similar Documents

Publication Publication Date Title
Liu et al. Weighing counts: Sequential crowd counting by reinforcement learning
US8756183B1 (en) System for representing, storing, and reconstructing an input signal
CN111581966B (en) Context feature-fused aspect-level emotion classification method and device
TWI655587B (en) Neural network and method of neural network training
CN112818136A (en) Time convolution-based interactive knowledge representation learning model TCIM prediction method
CN115147891A (en) System, method, and storage medium for generating synthesized depth data
KR20210081769A (en) Attack-less Adversarial Training for a Robust Adversarial Defense
Fang et al. Multiscale CNNs ensemble based self-learning for hyperspectral image classification
Kim et al. Exploring temporal information dynamics in spiking neural networks
CN109242029A (en) Identify disaggregated model training method and system
Kamilaris et al. Training deep learning models via synthetic data: Application in unmanned aerial vehicles
Pan et al. Fully convolutional neural networks with full-scale-features for semantic segmentation
CN114723784A (en) Pedestrian motion trajectory prediction method based on domain adaptation technology
Zhang et al. The performance research of the data augmentation method for image Classification
Qu et al. Improving the reliability for confidence estimation
Xu et al. Cross-domain few-shot classification via inter-source stylization
Du et al. Structure tuning method on deep convolutional generative adversarial network with nondominated sorting genetic algorithm II
Wang et al. Lightweight bilateral network for real-time semantic segmentation
CN116824334A (en) Model back door attack countermeasure method based on frequency domain feature fusion reconstruction
CN116051320A (en) Multitasking attention knowledge tracking method and system for online learning platform
Lambertenghi et al. Assessing Quality Metrics for Neural Reality Gap Input Mitigation in Autonomous Driving Testing
Xu et al. MMT: Mixed-Mask Transformer for Remote Sensing Image Semantic Segmentation
Ramachandra Causal inference for climate change events from satellite image time series using computer vision and deep learning
Shah et al. Reasoning over history: Context aware visual dialog
CN111898708A (en) Transfer learning method and electronic equipment

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20210518

RJ01 Rejection of invention patent application after publication