CN112396184A - Relationship mining method and system based on graph structure data - Google Patents
Relationship mining method and system based on graph structure data Download PDFInfo
- Publication number
- CN112396184A CN112396184A CN202011380368.7A CN202011380368A CN112396184A CN 112396184 A CN112396184 A CN 112396184A CN 202011380368 A CN202011380368 A CN 202011380368A CN 112396184 A CN112396184 A CN 112396184A
- Authority
- CN
- China
- Prior art keywords
- graph structure
- path
- structure data
- relationship
- attention weight
- 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.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
- G06N5/022—Knowledge engineering; Knowledge acquisition
- G06N5/025—Extracting rules from data
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
- G06N5/046—Forward inferencing; Production systems
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a relationship mining method and a system based on graph structure data, wherein the method comprises the following steps: s1, acquiring and analyzing the image to obtain image structure data; s2, processing the graph structure data based on the task layer, and performing relation reasoning on the graph structure to obtain subtasks; s3, completing the interaction with the environment according to the subtasks to obtain corresponding rewards; s4, feeding back the corresponding reward to the task layer; and S5, circulating the steps S2-S4 until the subtask with the maximum reward is completed. The system comprises: the system comprises an object and relation detection module, a task layer module, an action layer module, a feedback module and a circulation module. The invention directly models and utilizes the relation between objects, and can provide explanation performance when reaching the same performance. The method and the system for relationship mining based on the graph structure data can be widely applied to the field of reinforcement learning.
Description
Technical Field
The invention belongs to the field of reinforcement learning, and particularly relates to a relation mining method and system based on graph structure data.
Background
Most of the existing deep reinforcement learning uses unstructured data such as scene similar images and the like, a strategy from an observation state to an action is learned by means of the nonlinear fitting capability of a neural network, however, the application of the neural network in fields with higher safety requirements such as automatic driving and the like is difficult due to the inexplicability of an intermediate reasoning process, and in addition, the implicit relation between objects learned by the neural network cannot be intuitively understood by human beings.
Disclosure of Invention
In order to solve the above technical problems, an object of the present invention is to provide a method and a system for relationship mining based on graph structure data, which utilize a structural expression to model the relationship of an object, and use deep learning to propagate and infer the relationship in the structural relationship, so as to learn the strategy from state to action in a complex task scene.
The first technical scheme adopted by the invention is as follows: a relationship mining method based on graph structure data comprises the following steps:
s1, acquiring and analyzing the image to obtain image structure data;
s2, processing the graph structure data based on the task layer, and performing relation reasoning on the graph structure to obtain subtasks;
s3, completing the interaction with the environment according to the subtasks to obtain corresponding rewards;
s4, feeding back the corresponding reward to the task layer;
and S5, circulating the steps S2-S4 until the subtask with the maximum reward is completed.
Further, the step of obtaining and analyzing the image to obtain the graph structure data specifically includes:
s101, acquiring an image and detecting the image to obtain a detected object and corresponding attributes of the object;
s102, representing the detected object as a point of a graph structure;
s103, embedding prior knowledge into the detected object and the corresponding attribute of the object to generate an edge of a graph structure;
and S104, combining the points and the edges of the graph structure to obtain graph structure data.
Further, the step of processing the graph structure data based on the task layer and performing relation reasoning on the graph structure to obtain the subtasks specifically includes:
s201, obtaining a relation edge attention weight according to graph structure data;
s202, obtaining a path attention weight according to the graph structure data;
s203, carrying out weighted aggregation on the relationship paths according to the relationship edge attention weight and the path attention weight to obtain path aggregation information;
and S204, generating an action strategy according to the path aggregation information to obtain a subtask.
Further, the step of obtaining the attention weight of the relationship edge according to the graph structure data specifically includes:
s2011, expanding the graph structure data to obtain a two-dimensional matrix;
s2012, inputting the two-dimensional matrix into a transform frame to obtain the attention weight and the hidden vector of the relation edge in the first step;
s2013, inputting the hidden vectors and the two-dimensional matrix into a transform frame again to obtain the attention weight of the next step and the corresponding hidden vectors;
s2014, the step S2013 is circulated until the repetition times reach the maximum relation path length;
s2015, outputting the attention weight and the corresponding hidden vector of the relation edge of each step.
Further, the step of performing weighted aggregation on the relationship path according to the relationship edge attention weight and the path attention weight to obtain path aggregation information specifically includes:
s2031, selecting a current reasoning step and weighting all relationship side information according to the relationship side attention weight of the current reasoning step to obtain the relationship side information of the current reasoning step;
s2032, selecting the length of the current path, and obtaining the path information of the current length according to the relationship side information of each inference step of the current path;
s2033, carrying out weighted average on all path information according to the path attention weight to obtain path aggregation information.
Further, the generating of the action policy according to the path aggregation information specifically includes deriving a state-action function value of the action atom according to the path aggregation information.
The second technical scheme adopted by the invention is as follows: a relationship mining system based on graph structure data comprises the following modules:
the object and relation detection module is used for acquiring and analyzing an image to obtain image structure data;
the task layer module is used for processing the graph structure data based on the task layer and performing relation reasoning on the graph structure to obtain subtasks;
the action layer module is used for finishing the interaction with the environment according to the subtasks and obtaining corresponding rewards;
the feedback module is used for feeding back the corresponding reward to the task layer;
and the loop module is used for executing the loop instruction.
Further, the task layer module further comprises the following sub-modules:
the relation edge attention submodule is used for obtaining the attention weight of the relation edge according to the graph structure data;
the path attention submodule is used for obtaining a path attention weight according to the graph structure data;
the reasoning submodule is used for carrying out weighted aggregation on the relationship path according to the attention weight and the path attention weight of the relationship edge to obtain path aggregation information;
and the strategy sub-module is used for generating an action strategy according to the path aggregation information to obtain the subtask.
The method and the system have the beneficial effects that: according to the method, high-dimensional picture data are converted into graph structure data, and then a deep learning model is used for carrying out relationship reasoning on the graph structure data to obtain a series of relationship paths, so that certain interpretability is provided for an algorithm, direct modeling can be realized, relationships among objects can be utilized, and interpretation performance can be provided when the same performance is achieved.
Drawings
FIG. 1 is a flowchart illustrating steps of a method for relationship mining based on graph structure data according to an embodiment of the present invention;
FIG. 2 is a block diagram of a relationship mining system based on graph structure data according to an embodiment of the present invention;
FIG. 3 is a diagram illustrating image parsing to obtain graph structure data according to an embodiment of the present invention;
FIG. 4 is a block diagram illustrating the overall architecture and data flow of an embodiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments. The step numbers in the following embodiments are provided only for convenience of illustration, the order between the steps is not limited at all, and the execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.
As shown in FIG. 1, the present invention provides a relationship mining method based on graph structure data, which comprises the following steps:
s1, acquiring and analyzing the image to obtain image structure data;
s2, processing the graph structure data based on the task layer, and performing relation reasoning on the graph structure to obtain subtasks;
s3, completing the interaction with the environment according to the subtasks to obtain corresponding rewards;
s4, feeding back the corresponding reward to the task layer;
and S5, circulating the steps S2-S4 until the subtask with the maximum reward is completed.
Further, as a preferred embodiment of the method, the step of obtaining and analyzing the image to obtain the graph structure data specifically includes:
s101, acquiring an image and detecting the image to obtain a detected object and corresponding attributes of the object;
s102, representing the detected object as a point of a graph structure;
in particular, meaningful objects are abstracted from unstructured state data, represented as points in a graph structure, whose input is unstructured data such as pictures.
S103, embedding prior knowledge into the detected object and the corresponding attribute of the object to generate an edge of a graph structure;
specifically, human priori knowledge is embedded into detected objects and attributes thereof, such as coordinates, sizes and the like, for describing the upper, lower, left and right, the distance and the like, and the priori knowledge is used for constructing the mutual relation between every two objects, so that the edges of the graph structure are generated.
And S104, combining the points and the edges of the graph structure to obtain graph structure data.
Specifically, referring to fig. 3, according to the human priori knowledge, we first determine objects such as doors, keys, stairs, ropes, conveyor belts, people, skulls, etc. to be extracted from the image, and then embed the human priori knowledge, such as on the left, on the top, near-far, accessibility, etc., to construct relationships between the objects, which are the edges of the image structure. The graph data is very intuitive and can be easily understood by people, and simultaneously, the relation between objects can be mined.
In addition, the algorithm is mainly applied to a task layer, and can carry out targeted relationship reasoning according to the relationship between objects in the graph structure, so that key tasks in the graph structure are mined. As shown in the environment of FIG. 3, in the figure, a small person needs to acquire the key first, then arrives at the lower right corner stair and finally can pass through the door at the upper right corner, so that the customs clearance is completed. The invention can collect rewards from the process of continuous trial and error to train a task layer, and finally excavate and finish the passing key tasks. Meanwhile, according to the weight of the attention module, the method can visualize the reason why the relationship reasoning path answers the selection of the key task, and compared with a general reinforcement learning algorithm of a black box, the method has better interpretability.
Further, as a preferred embodiment of the method, the step of processing the graph structure data based on the task layer and performing relationship reasoning on the graph structure to obtain the subtasks specifically includes:
s201, obtaining a relation edge attention weight according to graph structure data;
s202, obtaining a path attention weight according to the graph structure data;
s203, carrying out weighted aggregation on the relationship paths according to the relationship edge attention weight and the path attention weight to obtain path aggregation information;
and S204, generating an action strategy according to the path aggregation information to obtain a subtask.
Further, as a preferred embodiment of the present invention, the step of obtaining the attention weight of the relationship edge according to the graph structure data specifically includes:
s2011, unfolding the graph structure data M to obtain a two-dimensional matrix M';
s2012, inputting the two-dimensional matrix M' into a Transformer framework to obtain the attention weight of the relation edge of the first step
And hidden vector v1；
S2013, converting the hidden vector viAnd the two-dimensional matrix M' is input into the Transformer framework again to obtain the attention weight of the relation edge in the step i +1
And corresponding hidden vector vi+1；
S2014, the step S2013 is circulated until the repetition times reach the maximum relation path length;
specifically, repeat T times, where T is the maximum relationship path length.
S2015, outputting the relation edge attention weight S of each stepiAnd corresponding hidden vector vi。
As a further preferred embodiment of the present invention, the step of obtaining the path attention weight according to the graph structure data specifically includes:
s2021, splicing the corresponding hidden vectors of each step into an integrated vector;
specifically, the hidden vector v1，v2，……，vTAre spliced into a vector v.
S2022, inputting the integration vector v and the two-dimensional matrix M' into a Transformer frame to obtain a path attention weight Sφ。
As a further preferred embodiment of the present invention, the step of performing weighted aggregation on the relationship path according to the relationship edge attention weight and the path attention weight to obtain path aggregation information specifically includes:
s2031, selecting a current reasoning step and weighting all relationship side information according to the relationship side attention weight of the current reasoning step to obtain the relationship side information of the current reasoning step;
specifically, at the current inference step t, attention weight is given according to the kth relationship edge of the current step
Weighting all relationship side information:
wherein M iskA relationship matrix representing the kth relationship.
S2032, selecting the length of the current path, and obtaining the path information of the current length according to the relationship side information of each inference step of the current path;
specifically, the current path length is selected, the path information of the current length is obtained from the relationship side information of each inference step of the current path, wherein the relationship side information is represented by a matrix, the matrix of each inference step is continuously multiplied to obtain the path information of the current path length, and when the current path length is t ', the relationship side information of the previous t' step is aggregated:
s2033, carrying out weighted average on all path information according to the path attention weight to obtain path aggregation information.
Specifically, according to the path attention weight, information of paths of different lengths is aggregated:
where T is the maximum relationship path length.
As a further preferred embodiment of the method, the generating the action policy according to the path aggregation information specifically includes deriving a state-action function value of the action atom according to the path aggregation information.
Specifically, using action atoms to represent a single action, the input is the inference module result:
the output is the Q value of the action atom. Here, for each action relationship, we use the fully-connected neural network to simulate, outputting:
in the above formula, action _ a (x, x') represents an atom acting as a, Q represents a state-action value at which the atom is currently acting in the current state, S represents the current state, v represents the current statexOne-hot coding representing corresponding entities using a fully-connected network MLPaA relationship matrix representing the action predicate a.
Referring to fig. 4, the present invention is based on the premise that the objects and the relationship detector can abstract the objects well and model the relationship between the objects. The environment first inputs unstructured data, the image, into the object and relationship detector, then the detector inputs the structured data into the task layer, the task layer assigns the task to the action layer for execution, the action layer interacts with the environment through the controller and collects rewards, and then feeds back the rewards to the task layer. The task layer will use the rewards to train attention modules, inference modules, and policy modules therein.
The invention provides a novel deep reinforcement learning framework, which can abstract the complex game environment of objects in real life, construct graph structure data describing the relationship between the objects by embedding human priori knowledge, can dig out important key tasks through an attention module, an inference module and a strategy module, and can visualize relationship paths so as to provide certain interpretability.
As shown in fig. 2, a relationship mining system based on graph structure data includes the following modules:
the object and relation detection module is used for acquiring and analyzing an image to obtain image structure data;
specifically, the object and relationship detection module realizes the expression from unstructured data to a graph structured state based on a supervised learning paradigm, so that human priori knowledge can be embedded in the process. Therefore, at each time step, the unstructured data is extracted by using the module, and the structured expression of the state of each step can be obtained to provide input for the subsequent modules.
The task layer module is used for processing the graph structure data based on the task layer and performing relation reasoning on the graph structure to obtain subtasks;
the action layer module is used for finishing the interaction with the environment according to the subtasks and obtaining corresponding rewards;
the feedback module is used for feeding back the corresponding reward to the task layer;
and the loop module is used for executing the loop instruction.
As a further preferred embodiment of the present system, the task layer module further includes the following sub-modules:
the relation edge attention submodule is used for obtaining the attention weight of the relation edge according to the graph structure data;
specifically, the attention module of the relationship edge has the input of the graph structure data M output by the detection module and the output of the attention module of the relationship edge
The path attention submodule is used for obtaining a path attention weight according to the graph structure data;
in particular, the path attention module whose input is the hidden vector output v of the relational edge attention module1，v2，……，vTAnd a two-dimensional matrix M' whose output is the path attention weight sφ. In which the hidden vector embeds the hidden information of different length paths.
The reasoning submodule is used for carrying out weighted aggregation on the relationship path according to the attention weight and the path attention weight of the relationship edge to obtain path aggregation information;
specifically, a plurality of relationship paths exist in the graph structure, and the inference module is used for performing weighted aggregation on the relationship paths according to the attention weight generated by the attention module, so that the information of the whole graph structure path is embedded into the matrix. And in each inference step, weighting all the relation side information, then carrying out matrix multiplication on the weighted relation side information and the path information obtained in the previous step to obtain the path information of the current length, and then carrying out weighted average on all the path information according to the path attention weight to obtain an inference result.
And the strategy sub-module is used for generating an action strategy according to the path aggregation information to obtain the subtask.
Specifically, reinforcement learning finally requires learning a strategy capable of maximizing the external environment reward sum, so that the strategy module has the function of deducing a state-action function value of an action atom according to the path aggregation information obtained by the reasoning module, and outputting an action strategy.
The contents in the system embodiments are all applicable to the method embodiments, the functions specifically realized by the method embodiments are the same as the system embodiments, and the beneficial effects achieved by the method embodiments are also the same as the beneficial effects achieved by the system embodiments.
While the preferred embodiments of the present invention have been illustrated and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (9)
1. A relationship mining method based on graph structure data is characterized by comprising the following steps:
s1, acquiring and analyzing the image to obtain image structure data;
s2, processing the graph structure data based on the task layer, and performing relation reasoning on the graph structure to obtain subtasks;
s3, completing the interaction with the environment according to the subtasks to obtain corresponding rewards;
s4, feeding back the corresponding reward to the task layer;
and S5, circulating the steps S2-S4 until the subtask with the maximum reward is completed.
2. The relationship mining method based on graph structure data according to claim 1, wherein the step of obtaining and analyzing the image to obtain the graph structure data specifically comprises:
s101, acquiring an image and detecting the image to obtain a detected object and corresponding attributes of the object;
s102, representing the detected object as a point of a graph structure;
s103, embedding prior knowledge into the detected object and the corresponding attribute of the object to generate an edge of a graph structure;
and S104, combining the points and the edges of the graph structure to obtain graph structure data.
3. The relationship mining method based on graph structure data according to claim 2, wherein the step of processing the graph structure data based on the task layer and performing relationship reasoning on the graph structure to obtain the subtasks specifically comprises:
s201, obtaining a relation edge attention weight according to graph structure data;
s202, obtaining a path attention weight according to the graph structure data;
s203, carrying out weighted aggregation on the relationship paths according to the relationship edge attention weight and the path attention weight to obtain path aggregation information;
and S204, generating an action strategy according to the path aggregation information to obtain a subtask.
4. The relationship mining method based on graph structure data according to claim 3, wherein the step of obtaining the attention weight of the relationship edge according to the graph structure data specifically comprises:
s2011, expanding the graph structure data to obtain a two-dimensional matrix;
s2012, inputting the two-dimensional matrix into a transform frame to obtain the attention weight and the hidden vector of the relation edge in the first step;
s2013, inputting the hidden vectors and the two-dimensional matrix into a transform frame again to obtain the attention weight of the next step and the corresponding hidden vectors;
s2014, the step S2013 is circulated until the repetition times reach the maximum relation path length;
s2015, outputting the attention weight and the corresponding hidden vector of the relation edge of each step.
5. The graph structure data-based relationship mining method according to claim 4, wherein the step of obtaining the path attention weight according to the graph structure data specifically comprises:
s2021, splicing the corresponding hidden vectors of each step into an integrated vector;
and S2022, inputting the integration vector and the two-dimensional matrix into a transform framework to obtain a path attention weight.
6. The method according to claim 5, wherein the step of performing weighted aggregation on the relationship paths according to the relationship edge attention weight and the path attention weight to obtain the path aggregation information specifically includes:
s2031, selecting a current reasoning step and weighting all relationship side information according to the relationship side attention weight of the current reasoning step to obtain the relationship side information of the current reasoning step;
s2032, selecting the length of the current path, and obtaining the path information of the current length according to the relationship side information of each inference step of the current path;
s2033, carrying out weighted average on all path information according to the path attention weight to obtain path aggregation information.
7. The graph structure data-based relationship mining method according to claim 6, wherein the generating of the action policy according to the path aggregation information is specifically to derive a state-action function value of an action atom according to the path aggregation information.
8. A relationship mining system based on graph structure data is characterized by comprising the following modules:
the object and relation detection module is used for acquiring and analyzing an image to obtain image structure data;
the task layer module is used for processing the graph structure data based on the task layer and performing relation reasoning on the graph structure to obtain subtasks;
the action layer module is used for finishing the interaction with the environment according to the subtasks and obtaining corresponding rewards;
the feedback module is used for feeding back the corresponding reward to the task layer;
and the loop module is used for executing the loop instruction.
9. The graph structure data based relationship mining system according to claim 8, wherein the task layer module further comprises the following sub-modules:
the relation edge attention submodule is used for obtaining the attention weight of the relation edge according to the graph structure data;
the path attention submodule is used for obtaining a path attention weight according to the graph structure data;
the reasoning submodule is used for carrying out weighted aggregation on the relationship path according to the attention weight and the path attention weight of the relationship edge to obtain path aggregation information;
and the strategy sub-module is used for generating an action strategy according to the path aggregation information to obtain the subtask.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011380368.7A CN112396184B (en) | 2020-12-01 | 2020-12-01 | Relation mining method and system based on graph structure data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011380368.7A CN112396184B (en) | 2020-12-01 | 2020-12-01 | Relation mining method and system based on graph structure data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112396184A true CN112396184A (en) | 2021-02-23 |
| CN112396184B CN112396184B (en) | 2023-09-05 |
Family
ID=74604875
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011380368.7A Active CN112396184B (en) | 2020-12-01 | 2020-12-01 | Relation mining method and system based on graph structure data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112396184B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014146463A1 (en) * | 2013-03-19 | 2014-09-25 | 中国科学院自动化研究所 | Behaviour recognition method based on hidden structure reasoning |
| WO2020056984A1 (en) * | 2018-09-19 | 2020-03-26 | 平安科技（深圳）有限公司 | Shortest path query method, system, computer device and storage medium |
| CN111581343A (en) * | 2020-04-24 | 2020-08-25 | 北京航空航天大学 | Reinforced learning knowledge graph reasoning method and device based on graph convolution neural network |
-
2020
- 2020-12-01 CN CN202011380368.7A patent/CN112396184B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014146463A1 (en) * | 2013-03-19 | 2014-09-25 | 中国科学院自动化研究所 | Behaviour recognition method based on hidden structure reasoning |
| WO2020056984A1 (en) * | 2018-09-19 | 2020-03-26 | 平安科技（深圳）有限公司 | Shortest path query method, system, computer device and storage medium |
| CN111581343A (en) * | 2020-04-24 | 2020-08-25 | 北京航空航天大学 | Reinforced learning knowledge graph reasoning method and device based on graph convolution neural network |
Non-Patent Citations (1)
| Title |
|---|
| 裴继升;叶晓俊;: "基于语法推导的溯源依赖关系路径模式挖掘算法", 清华大学学报(自然科学版), no. 06, pages 3 - 10 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112396184B (en) | 2023-09-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Peng et al. | Variational discriminator bottleneck: Improving imitation learning, inverse rl, and gans by constraining information flow | |
| Chen et al. | Fostering linguistic decision-making under uncertainty: a proportional interval type-2 hesitant fuzzy TOPSIS approach based on Hamacher aggregation operators and andness optimization models | |
| Puppe | Systematic introduction to expert systems: Knowledge representations and problem-solving methods | |
| Gero et al. | A function–behavior–structure ontology of processes | |
| US20190130266A1 (en) | System and method for improved neural network training | |
| Lee et al. | A genetic fuzzy agent using ontology model for meeting scheduling system | |
| Vlasselaer et al. | Exploiting local and repeated structure in dynamic Bayesian networks | |
| Tuan et al. | Dental diagnosis from X-ray images using fuzzy rule-based systems | |
| Bobrow et al. | Systematic understanding: synthesis, analysis, and contingent knowledge in specialized understanding systems | |
| Zhang et al. | Unifying generative models with GFlowNets and beyond | |
| Shindo et al. | Neuro-symbolic forward reasoning | |
| CN106022294A (en) | Intelligent robot-oriented man-machine interaction method and intelligent robot-oriented man-machine interaction device | |
| Papageorgiou et al. | A web-based tool for Fuzzy Cognitive Map Modeling | |
| Li et al. | Learning to solve optimization problems with hard linear constraints | |
| Shindo et al. | α ILP: thinking visual scenes as differentiable logic programs | |
| Lv et al. | Cognitive robotics on 5G networks | |
| Firt | The missing G | |
| Münch et al. | High-level situation recognition using fuzzy metric temporal logic, case studies in surveillance and smart environments | |
| CN112396184B (en) | Relation mining method and system based on graph structure data | |
| Kaur et al. | Non-parametric learning of lifted restricted boltzmann machines | |
| Gong et al. | Path-specific effects based on information accounts of causality | |
| CN112036546B (en) | Sequence processing method and related equipment | |
| Kumar | APTx: better activation function than MISH, SWISH, and ReLU's variants used in deep learning | |
| Ikram | A benchmark for evaluating Deep Learning based Image Analytics | |
| Les et al. | Shape understanding system: The visual reasoning process |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |