Device and computer implemented data structure and method for explaining answer of similarity query, and method for training model for explaining answer of similarity query

Abstract

To provide a data structure, a method for explaining an answer of a similarity query and a method for training a model.SOLUTION: In a computer implemented method for explaining an answer of a similarity query, a knowledge graph includes nodes including first and second nodes, edges that represent relations between pairs of nodes, and attributes that are associated in the knowledge graph with an edge or a node. The method includes: providing embeddings of the first and second nodes of the similarity query, and providing an answer to the similarity query, the answer including a distance between the first node and the second node; and determining an output of a model, the model being configured for determining the output for explaining the answer to the similarity query depending on the embeddings of the first and second nodes, and the attributes, the output including one value that indicates the contribution of the attributes to the answer.SELECTED DRAWING: None

Description

[Technical Field]

[0001] The present invention relates to an apparatus, a computer-implemented data structure, a method for explaining answers to similarity queries, and a method for training a model for explaining answers to similarity queries. [Background technology]

[0002] The knowledge graph may be used to provide answers to similarity queries. Summary of the Invention [Problem to be solved by the invention]

[0003] DISCLOSURE OF THE INVENTION The present apparatus, the present computer-implemented data structures and the present methods describe answering similarity queries and provide advantageous properties.

[0004] First, models for explaining answers, and in particular the output of the graph neural networks they contain, rely on embeddings that capture various aspects of the entities in the knowledge graph used to answer the query and go beyond those aspects explicitly described in the knowledge graph.

[0005] This is particularly advantageous when the knowledge graph is incomplete or noisy. The knowledge graph may be incomplete due to missing links or entity attributes. The noise in the knowledge graph may be introduced during the construction of the knowledge graph.

[0006] The generalization capabilities of models, especially graph neural networks, allow them to infer missing links while remaining robust to noise in the knowledge graph. [Means for solving the problem]

[0007] A computer-implemented method for explaining an answer to a similarity query, wherein a knowledge graph includes nodes including a first node and a second node, wherein the knowledge graph includes edges representing relationships between pairs of nodes, wherein the knowledge graph includes attributes associated with at least one edge or at least one node in the knowledge graph, and wherein the similarity query includes the first node and the second node, the method comprising: providing an embedding of the first node for the similarity query; providing an embedding of the second node for the similarity query; providing an answer to the similarity query, wherein the answer includes a distance between the first node and the second node; and determining an output of a model, wherein the model is configured to determine the output for explaining the answer to the similarity query depending on the embedding of the first node, the embedding of the second node, and the attributes, and the output includes at least one value indicating a contribution of one of the attributes to the answer.

[0008] The method preferably includes determining an output indicating to what extent and / or in what direction at least one attribute contributes to the answer.

[0009] According to one example, the attributes include at least one attribute representing an edge of the knowledge graph or a value associated with an edge or node of the knowledge graph in the knowledge graph.

[0010] The embeddings of the first node and the second node may be predetermined or may be read from a storage device, wherein providing an embedding of the first node comprises determining, using an encoder, an embedding of the first node dependent on the first node and / or at least one other node in the vicinity of the first node and / or at least one edge in the vicinity and / or at least one attribute in the vicinity, and wherein providing an embedding of the second node comprises determining, using an encoder, an embedding of the second node dependent on the second node and / or at least one other node in the vicinity and / or at least one edge in the vicinity and / or at least one attribute in the vicinity.

[0011] According to one example, the nodes of the knowledge graph represent different production cells for manufacturing a workpiece or different workpieces, and / or the nodes of the knowledge graph represent different operations that can be performed to manufacture a workpiece or different workpieces, wherein the edges of the knowledge graph represent relationships between pairs of production cells, relationships between pairs of operations, or relationships between production cells and operations, in particular relationships indicating that production cells linked to each other by the relationship are interchangeable, or relationships indicating that operations linked to each other by the relationship are interchangeable, and the attributes represent relationships or properties associated with the relationships, production cells, or operations in the knowledge graph, in particular production time.

[0012] The method may include receiving a similarity query from a human-machine interface or machine-to-machine interface, where a first node represents a first production cell for execution on a workpiece; providing an answer; and transmitting the answer and an output indicative of the contribution of at least one attribute to the answer to the human-machine interface or machine-to-machine interface for operation, where a second node represents a second production cell for execution on the workpiece, where the at least one attribute is indicative of a characteristic of the first production cell and / or the second production cell.

[0013] The method may include, in particular, when the first production cell is out of order, for example, unavailable due to a defect in the first production cell or due to maintenance work, controlling the second production cell to perform an operation on the workpiece based on the answer to the similarity query and an output indicative of the contribution of at least one attribute to the answer, and based on the list, in particular the list specifying conditions, values ​​or ranges for the answer to the similarity query and the output to be satisfied.

[0014] A computer-implemented method for training a model to explain answers to similarity queries, wherein a knowledge graph includes nodes including a first node and a second node, wherein the knowledge graph includes edges representing relationships between pairs of nodes, wherein the knowledge graph includes attributes associated with at least one edge or at least one node in the knowledge graph, wherein the similarity query includes the first node and the second node, and wherein the answer includes a distance between the first node and the second node, the method comprising: providing a similarity query; providing an answer; providing an embedding of the first node; providing an embedding of the second node; and determining an output of the model; wherein the model is configured to determine an output for explaining an answer to the similarity query depending on the first node, the second node, the embedding of the first node, the embedding of the second node, and the attribute, wherein the output includes at least one value indicating a contribution of one of the attributes to the answer, and wherein providing the answer includes determining a similarity between the first node and the second node depending on the first node, the second node, the embedding of the first node, the embedding of the second node, and the attribute, and wherein the method includes determining an estimated similarity between the first node and the second node depending on the output for explaining the answer to the similarity query, and training the model depending on a difference between the similarity and the estimated similarity.

[0015] The step of determining the estimated similarity may comprise determining a sum, in particular a normalized sum, of the output values.

[0016] The apparatus includes at least one processor and at least one memory, where the at least one memory includes instructions that, when executed by the at least one processor, cause the apparatus to perform a method for explaining or training answers to similarity queries.

[0017] The computer-implemented data structure may include at least one data field for describing an answer to a similarity query, at least one data field for the similarity query, where the similarity query includes at least one data field containing a first node of the knowledge graph, at least one data field for the answer, where the answer includes at least one data field containing a second node of the knowledge graph, at least one data field for an embedding of the first node, at least one data field for an embedding of the second node, at least one data field for an attribute in the knowledge graph associated with at least one edge or at least one node of the knowledge graph, and at least one data field for an output of the model, in particular an output of a graph neural network that the model includes, where the output is at least one data field for describing the answer to the similarity query depending on the first node, the second node, the embedding of the first node, the embedding of the second node, and the attributes, where the output includes at least one value indicating a contribution of one of the attributes to the answer.

[0018] The computer-implemented data structure may include at least one data field for a similarity between the first node and the second node determined depending on the first node, the second node, the embedding of the first node, the embedding of the second node, and the attributes; at least one data field for an estimated similarity between the first node and the second node determined depending on the output to explain an answer to the similarity query; and at least one data field for a difference between the similarity and the estimated similarity.

[0019] Further exemplary embodiments can be derived from the following description and drawings. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is a schematic diagram outlining an approach for determining similarity and explanation for a particular node pair u and v in a knowledge graph. [Figure 2] FIG. 1 is a schematic diagram outlining an iterative algorithm for optimizing a model, and in particular the model's encoder and explanation modules. [Figure 3] FIG. 1 shows a flowchart including steps of a method for explaining answers to similarity queries. [Figure 4] FIG. 1 shows a flowchart including steps of a method for training a model to explain answers to similarity queries. [Figure 5] 1 is a schematic diagram illustrating an apparatus for carrying out a method for explaining an answer or a method for training a model. [Figure 6] FIG. 1 is a schematic diagram illustrating a particular computer-implemented data structure for carrying out a method for explaining an answer or a method for training a model. DETAILED DESCRIPTION OF THE INVENTION

[0021] FIG. 1 shows a schematic overview of an approach for determining similarity and explanation for a particular node pair u and v in a knowledge graph.

[0022] A knowledge graph is a tuple G=(V;R;E;X), where V is a set of n nodes v, R is a set of relations, E is a set of edges of the form (h;r;t), where h;t∈V and r∈R, and X is a matrix of node features.

[0023] This means that an edge in the knowledge graph represents a relationship between a pair of nodes in the knowledge graph.

[0024] The nodes of the knowledge graph G can represent different production cells for manufacturing one workpiece or different workpieces.

[0025] The nodes of the knowledge graph G can represent different operations that can be performed to manufacture a workpiece or different workpieces.

[0026] The edges of the knowledge graph G can represent relationships between pairs of production cells, between pairs of operations, or between production cells and operations. A relationship can indicate that production cells linked to each other by the relationship are interchangeable. A relationship can also indicate that operations linked to each other by the relationship are interchangeable.

[0027] The k-hop neighborhood of a node u is defined as the set (h;r;t)∈E of edges (h;r;t) containing nodes whose shortest path distance to u is less than or equal to k. Formally, the k-hop neighborhood of u is the set {(h;r;t)∈Esp(u;h)≦k∧sp(u;t)≦k}, where sp is the shortest path distance.

[0028] Graph embedding is the space

number

[0029] The problem of similarity search on G involves ranking the nodes in V according to their similarity to a query node u∈V.

[0030] A similarity query in this context includes query nodes, which means that the similarity query includes nodes of the knowledge graph. A similarity query may be a text string that includes query nodes.

[0031] A distance function between nodes is a bivariate function of the nodes in a graph that returns a scalar value representing how close the nodes are according to some criterion.

[0032] This means that the answer to a similarity query includes the distance between a node in the knowledge graph and the query node, e.g., a scalar value that describes how close a node in the knowledge graph is to the query node. The answer may also be a text string containing the answer.

[0033] Formally, the distance function between nodes gives the following mapping:

number

[0034] A simple example is the function d(u;v) = |deg(u) - deg(v)|, where deg(·) returns the degree of the node. According to this function, the distance between nodes with similar degrees is small.

[0035] According to an exemplary mechanism, the embedding of nodes in a knowledge graph captures the similarity between nodes and provides an explanation for this similarity.

[0036] This mechanism is implemented, for example, via a model 100 that includes an interface 102 to a knowledge graph G and a user input 104. The model 100 is configured to provide an explanation for the similarity between nodes u and v from the knowledge graph G.

[0037] The interface 102, in this example, is configured to provide the nodes u and v. The interface 102 may be configured to provide the size k of the node neighborhood to be considered for determining the similarity and / or description, which may be a predetermined constant.

[0038] Model 100 is configured to select a k-hop neighborhood 108 for node u, step 106. Model 100 is also configured to select a k-hop neighborhood 112 for node v, step 110.

[0039] According to one example, model 100 includes a selector 114 configured to select, for each node, a respective k-hop neighbor.

[0040] The model 100 is configured, in step 116, to encode the node u and the k-hop neighborhood 108 of node u into an embedding 118 of node u.

[0041] The model 100 is configured, at step 120, to encode a node v and its k-hop neighborhood 112 into an embedding 122 of the node v.

[0042] According to one example, the model 100 includes an encoder 124. The encoder 124 is configured to process input data. The encoder 124 is configured to encode the input data.

[0043] The input data includes, for example, nodes u and v, as well as the k-hop neighborhood of node u and the k-hop neighborhood of node v.

[0044] The input data may additionally include at least one attribute that the knowledge graph G assigns to node u, node v, or any member of the k-hop neighborhood of node u or the k-hop neighborhood of node v. A member of the k-hop neighborhood in this context refers to an edge or a node of the knowledge graph G.

[0045] Attributes may represent relationships or properties associated with relations in the knowledge graph G, production cells, or operations, particularly production times.

[0046] The input data is a list of attributes (a1;a2;…;a A ) can be included. For example, the attribute a i represents a single edge in the k-hop neighborhood of node u or node v. Furthermore, the knowledge graph G may also include numerical attributes associated with the nodes or edges of the knowledge graph G. Attribute a i may be a numerical attribute assigned to node u, node v, or any member of the k-hop neighborhood of node u or the k-hop neighborhood of node v in the knowledge graph.

[0047] The task of the encoder 124 is to compute node embeddings that are useful for answering similarity search queries. According to one example, the encoder 124 computes the embeddings of nodes using the function

number

[0048] According to one example, f Θ takes as input a node u and its k-hop neighbors in the knowledge graph and computes the following embedding

number

[0049] This means that a GNN is a neural network that is configured to map nodes of a knowledge graph and the neighborhoods of the nodes in the knowledge graph to embeddings.

[0050] In step 126, model 100 is configured to determine a similarity 128 between node u and node v depending on the embedding 118 of node u and the k-hop neighborhood 108 of node u, and depending on the embedding 122 of node v and the k-hop neighborhood 112 of node v.

[0051] According to one example, the encoder 124 provides embeddings u and v. According to one example, the similarity between nodes u and v is determined depending on the embeddings u and v. According to one example, the similarity 128 is the dot product u of the embeddings u and v: T It is calculated as v.

[0052] The model 100 is configured, in step 130, to determine an explanation 132 for the similarity 128 between node u and node v.

[0053] According to one example, the model 100 includes an explainer 134 configured to output an explanation 132 for the similarity 128 between node u and node v.

[0054] According to one example, the explainer 134 receives as input nodes u and v, the k-hop neighborhood of node u and the k-hop neighborhood of node v, and an embedding u of node u and an embedding v of node v.

[0055] According to one example, the explainer 134 outputs at least one label, which indicates the contribution of the attribute associated with the at least one label to the answer. The output may include values, where each value indicates a respective label.

[0056] This means that the output of the explainer 134 includes at least one value that indicates the contribution of one of the attributes to the answer.

[0057] At least one label may be associated with, for example, node u, node v, or a member of node u's k-hop neighborhood or a member of node v's k-hop neighborhood.

[0058] The explainer 134 may, for example, provide a list of explanatory labels g Φ (u,v)=(e1,…,e A ) is output by a neural network g Φ where each e i ∈(-1,1) is the value of each attribute a i is the descriptive label assigned to the value g Φ (u,v)=e i attribute a i is the dot product u of the embedding u and v T It indicates how much and in what direction (positive or negative) it contributes to v.

[0059] Model 100 can be used to answer similarity queries. Given a query node u, the ranking for a similarity query depends on the input data, e.g., the dot product u of the embedding u and v for v∈V: T It is generated by computing v.

[0060] The model 100 can be used to provide explanations for answers to similarity queries. For example, the explainer 134 can be used to calculate the dot product u of the embeddings u and v. T It is performed to assign descriptive labels to parts of the input data that are responsible for an increase or decrease in v.

[0061] The dot product u of the embedding u and v T v is an example of an operation that determines a similarity score for the similarity between node u or node v. Other operations on the embeddings u and v may be used to determine a similarity score for the similarity between node u or node v.

[0062] 1 refers to a pair of nodes u and v. Model 100 may be configured to calculate similarities and explanations between node u and one or more other nodes in the knowledge graph G.

[0063] 2 shows a schematic overview of the iterative algorithm for optimizing the model 100, and in particular the encoder 124 and explainer 134 of the model 100. A selector 114 selects, in step 106, a k-hop neighborhood 108 for node u, and, in step 110, a k-hop neighborhood 108 for each node v,...,v t The encoder 124 is configured to select a plurality of k-hop neighborhoods 112-1, ..., 112-t for each node v. The encoder 124 is configured to determine an embedding 118 for the node u in step 116 depending on the k-hop neighborhood 108 of the node u. The encoder 124 selects a plurality of k-hop neighborhoods 112-1, ..., 112-t for each node v in step 120. t Depending on each k-hop neighborhood 112-1,...,112-t of each node v1,...,v t , 122-t for the plurality of embeddings 122-1, . . . , 122-t.

[0064] The model 100 then generates a k-hop neighborhood 108 of node u and a k-hop neighborhood 118 of node u, at step 126, for each node v. i For each embedding 122-i and each node v i Depending on the k-hop neighborhood 112-i of node u and each node v, i and a respective similarity 128 between the

[0065] The model 100, in step 130, calculates the k-hop neighborhood of node u, the k-hop neighborhood of node u, and each node v i Each embedding 122-i of each node v i Depending on the k-hop neighborhood 112-i of node u and each node v i and determining a respective explanation 132 for the similarity 128 between the

[0066] In this example model 100, neural networks are included in the encoder 124 and the explainer 134. The parameters Θ of the encoder 124 and / or the parameters Φ of the explainer 134 may be specifically optimized prior to deployment.

[0067] According to one example, the encoder 124 and / or the explainer 134 are optimized by iteratively adjusting the parameters of a neural network using gradient descent for a particular loss function.

[0068] To optimize the encoder 124, the loss function includes a ranking loss, which forces the pairwise distances between node embeddings to follow a distance relationship calculated using a distance function. An example of a distance function is the degree distance d(u,v).

[0069] According to one example, the model 100 includes a user input 202 configured to receive a distance function.

[0070] According to one example, the optimization may be performed in step 204 by: for a node v∈V in the knowledge graph G, selecting t target nodes u i The set T(u i ) i=1,…,t and determine their respective distances to node v, d(v,u1) <d(v,u2)<…<d(v,u t ) so that t target nodes are arranged as an ordered list T v (u1,…,u t )

[0071] According to one example, the optimization is performed in step 206 by v The i-th element of

number

number

number

number

[0072] This ranking loss L R (v) encourages the embedding distance in the embedding space to preserve the relationships contained in the distance function between nodes, such as degree distance.

[0073] According to one example, the optimization may be performed in step 208 by using the explainer loss L E The method may include determining:

[0074] In the case of the explainer 134, the explainer 134 may, for example, determine that the sum of the explanatory labels it generates is the dot product u of the embedding u and v between the pair of nodes u and v. T v is optimized to be the label e i The resulting estimated dot product from

number

number

[0075] Explainer 134 uses the following formula:

number

[0076] The optimization is performed in step 210 by computing the loss function L with respect to the parameters Θ of the encoder 124 and the parameters Φ of the explainer 134. R (v) and L E Determining the gradient of (u,v) and, for example, the loss function L R (v) and L E and updating the parameters Θ of the encoder 124 and the parameters Φ of the explainer 134 using gradient descent to minimize (u,v).

[0077] FIG. 3 shows a flowchart including steps of a computer-implemented method for explaining similarity query answering.

[0078] The method includes step 302 .

[0079] Step 302 includes providing a knowledge graph, for example, a knowledge graph G is provided.

[0080] The method includes step 304 .

[0081] In step 304, a similarity query is provided.

[0082] The similarity query includes a first node in the knowledge graph, for example, node u in the knowledge graph G. The similarity query includes a second node in the knowledge graph, for example, node v in the knowledge graph G.

[0083] The similarity query may be received from a human-machine interface, for example, a user interface or a machine-to-machine interface.

[0084] A first node may represent a first manufacturing cell for performing an operation on a workpiece, and a second node may represent a second manufacturing cell for performing an operation on a workpiece.

[0085] The method includes step 306 .

[0086] In step 306, attributes associated in the knowledge graph with at least one edge or at least one node of the knowledge graph are provided. The attributes may include at least one attribute representing an edge of the knowledge graph or a value associated in the knowledge graph with the edge or node of the knowledge graph.

[0087] For example, attributes of a k-hop neighborhood of a first node are provided. For example, attributes of a k-hop neighborhood of a second node are provided.

[0088] According to one example, the attributes (a1; a2; …; a A According to one example, at least one attribute a represents a single edge in the k-hop neighborhood of node u or in the k-hop neighborhood of node v. i is provided.

[0089] At least one attribute can be indicative of a characteristic of the first production cell. At least one attribute can be indicative of a characteristic of the second production cell. One attribute can be indicative of a characteristic of both the first production cell and the second production cell.

[0090] The method includes step 308 .

[0091] In step 308, an embedding of the first node is provided, for example, embedding 118 of node u.

[0092] Providing an embedding for the first node may include determining an embedding for the first node using the encoder 124 depending on the first node. For example, Θ (u) is determined.

[0093] Providing an embedding for the first node may include determining, using encoder 124, an embedding for the first node dependent on at least one other node in the vicinity of the first node.

[0094] Providing an embedding for the first node may include determining, using encoder 124, an embedding for the first node depending on the first node and at least one other node in the vicinity of the first node.

[0095] Providing an embedding for the first node may include determining, using encoder 124, an embedding for the first node dependent on at least one edge in the neighborhood.

[0096] Providing an embedding for the first node may include determining, with the encoder 124, an embedding for the first node dependent on at least one attribute in a neighborhood of the first node.

[0097] The method includes step 310 .

[0098] In step 310, an embedding of a second node is provided. For example, an embedding 122 of node v is provided. For example, an embedding f Θ (u) is determined.

[0099] Providing an embedding of the second node may include determining, with the encoder 124, an embedding of the second node depending on the second node.

[0100] Providing an embedding for the second node may include determining, using encoder 124, an embedding for the second node dependent on at least one other node in the vicinity of the second node.

[0101] Providing an embedding for the second node may include determining, using encoder 124, an embedding for the second node depending on the second node and at least one other node in the vicinity of the second node.

[0102] Providing an embedding for the second node may include determining, with encoder 124, an embedding for the second node dependent on at least one edge in the neighborhood of the second node.

[0103] Providing an embedding for the second node may include determining, with the encoder 124, an embedding for the second node dependent on at least one attribute of the neighborhood.

[0104] The method may include determining embeddings for a plurality of nodes, including the first node and the second node, and storing the embeddings.

[0105] The method includes step 312 .

[0106] In step 312, an answer to the similarity query is provided.

[0107] The answer includes the similarity between the first node and the second node.

[0108] According to one example, the similarity between a first node and a second node is determined depending on the first node, the second node, the embedding of the first node, the embedding of the second node, and the attributes. For example, the dot product u of the embeddings u and v is T v is determined to indicate the similarity between a first node u and a second node v.

[0109] The response is sent, for example, via a machine-to-machine interface.

[0110] The answer is output, for example, via a human-machine interface.

[0111] The method includes step 314 .

[0112] In step 314, the output of the model 100 is determined depending on the first node, the second node, the embedding of the first node, the embedding of the second node, and the attribute.

[0113] According to one example, the explainer 134, e.g., a neural network g Φ (u,v)=(e1,…,e A ) output is determined.

[0114] The output is an explanation to explain the answer to the similarity query, e.g., explanatory labels e1,…,e A Includes.

[0115] The output includes at least one value indicating the contribution of one of the attributes to the answer.

[0116] Step 314 may include determining an output indicating to what extent and / or in what direction at least one attribute contributes to the answer.

[0117] The output is transmitted, for example, via a machine-to-machine interface.

[0118] The output is output, for example, via a human-machine interface.

[0119] The method may include determining the output of model 100 for the first node and the second node depending on the previously determined and stored embeddings, respectively, which means that it is not necessary to determine the similarity between the nodes in order to determine an explanation from the previously determined embeddings of the first node and the second node.

[0120] In this case, the output for explaining the answer to the similarity query is determined depending on the first node and the second node, in the sense that the embedding of the first node and the embedding of the second node are selected or identified depending on the first node and the second node.

[0121] The method optionally includes controlling a second production cell instead of the first production cell to perform the operation on the workpiece.

[0122] According to one example, a second production cell is controlled to perform an operation on a workpiece when a first production cell is unavailable, for example, due to a defect in the first production cell or due to maintenance work.

[0123] For example, the second production cell meets the following criteria: - Similarity query answers, - an output indicating the contribution of at least one of the attributes to the answer, - controlled to perform operations on workpieces based on lists, in particular lists specifying similarity query answers and output conditions, values ​​or ranges to be satisfied.

[0124] FIG. 4 shows a flowchart including steps of a computer-implemented method for training a model 100 for explaining answers to similarity queries.

[0125] Methods for training the model 100 include, for example, training the encoder 124 and / or training the explainer 134.

[0126] The method for training the model 100 includes steps 302 through 314 of a method for explaining answers to similarity queries.

[0127] The method for training additionally includes step 316 .

[0128] Step 316 includes determining an estimated similarity between the first node and the second node depending on the output to explain the answer to the similarity query.

[0129] The estimated similarity may depend on the sum of the model 100 output values, particularly the normalized sum.

[0130] For example, for a first node u and a second node v, the estimated similarity is:

number

[0131] The method for training includes step 318 .

[0132] Step 318 includes training the model 100 depending on the difference between the similarity and the estimated similarity. According to one example, the loss function L R (v) and L E (u,v) is used for training.

[0133] The step of training the model 100 is to use the loss function L R (v) and L E This may include minimizing (u,v).

[0134] This training may involve repeatedly performing steps 304 to 318. The parameters Θ and Φ are, for example, determined by the loss function L R (v) and L E It may be determined by gradient descent based on (u,v). The parameters Θ and Φ may be updated between two iterations.

[0135] 5 shows a schematic diagram of an apparatus 500 for performing a method for explaining an answer or a method for training a model 100. The apparatus 500 includes an interface 506. The interface 506 may include a human-machine interface or a machine-to-machine interface.

[0136] The device 500 includes at least one processor 502 and at least one memory 504. The at least one memory 504, in this example, includes non-transitory, volatile memory.

[0137] According to one example, the at least one memory 504 includes instructions that, when executed by the at least one processor 502, cause the device 500 to perform a method for explaining answers to similarity queries. According to one example, the at least one memory 504 includes instructions that, when executed by the at least one processor 502, cause the device 500 to perform a method for training.

[0138] 6 shows a schematic diagram of a particular computer-implemented data structure 600 for carrying out a method for explaining an answer or a method for training a model. The data structure 600 includes a number of data fields 602.

[0139] The data structure 600 includes at least one data field 602 for describing the answer to the similarity query.

[0140] The data structure 600 includes at least one data field 602 for a similarity query, where the similarity query includes a first node of a knowledge graph.

[0141] The data structure 600 includes at least one data field 602 for an answer, where the answer comprises a second node in the knowledge graph.

[0142] The data structure 600 includes at least one data field 602 for an embedding of a first node, at least one data field 602 for an embedding of a second node, and at least one data field 602 for an attribute associated with at least one edge or at least one node in the knowledge graph.

[0143] The data structure 600 includes at least one data field 602 for the output of the model 100, particularly the output of the graph neural network that the model 100 includes. The output includes at least one value that indicates the contribution of one of the attributes to the answer.

[0144] The data structure 600 may additionally include at least one data field 602 for the first node, the second node, the embedding of the first node, the embedding of the second node, and a similarity between the first node and the second node determined depending on the attributes.

[0145] The data structure 600 may additionally include at least one data field 602 for an estimated similarity between the first node and the second node determined dependent on the output to explain the answer to the similarity query.

[0146] The data structure 600 may additionally include at least one data field 602 for the difference between the similarity and the estimated similarity.

Claims

1. 1. A computer-implemented method for explaining answers to similarity queries, comprising: the knowledge graph includes a node including a first node and a second node; the knowledge graph includes edges representing relationships between pairs of nodes; the knowledge graph includes an attribute associated in the knowledge graph with at least one edge or at least one node of the knowledge graph; the similarity query includes a first node and a second node; The method comprises: providing an embedding of the first node of the similarity query (308); providing an embedding of the second node of the similarity query (310); providing (312) an answer to the similarity query, the answer including a distance between the first node and the second node; determining (314) the output of the model (100); Including, the model (100) is configured to determine an output to explain an answer to the similarity query depending on the embedding of the first node, the embedding of the second node, and the attributes; the output includes at least one value indicating the contribution of one of the attributes to the answer; A method characterized by:

2. The method comprises: determining (314) an output indicating to what extent and / or in what direction the at least one attribute contributes to the answer; The method of claim 1.

3. the attributes include at least one attribute representing an edge of the knowledge graph or a value associated within the knowledge graph with an edge or node of the knowledge graph; 3. The method according to claim 1 or 2.

4. The step of providing an embedding of the first node (308) comprises: determining an embedding of the first node using an encoder (124) depending on the first node and / or at least one other node in the vicinity of the first node and / or at least one edge in the vicinity and / or at least one attribute in the vicinity; The step of providing an embedding of the second node (310) comprises: determining an embedding of the second node using the encoder (124) depending on the second node and / or at least one other node in its vicinity and / or at least one edge in its vicinity and / or at least one attribute in its vicinity; 4. The method according to any one of claims 1 to 3.

5. the nodes of the knowledge graph represent different production cells for manufacturing a workpiece or different workpieces, and / or the nodes of the knowledge graph represent different operations that can be performed to manufacture a workpiece or different workpieces; the edges of the knowledge graph represent relationships between pairs of production cells, relationships between pairs of operations, or relationships between production cells and operations, in particular relationships indicating that production cells linked to each other by the relationship are interchangeable, or relationships indicating that operations linked to each other by the relationship are interchangeable; the attributes represent relationships within the knowledge graph or relationships or properties associated with a production cell or operation, in particular production time; 5. The method according to any one of claims 1 to 4.

6. The method comprises: receiving (304) a similarity query from a human-machine interface or a machine-to-machine interface, the first node representing a first production cell for performing an operation on a workpiece; providing said answer; sending the answer and an output indicating the contribution of at least one attribute to the answer to a human-machine or machine-to-machine interface for operation, the second node representing a second production cell for performing an operation on the workpiece; Including, the at least one attribute indicates a characteristic of the first production cell and / or the second production cell. The method of claim 5.

7. The method further comprises: in particular when the first production cell is unavailable, for example due to a defect in the first production cell or due to maintenance work, controlling the second production cell to perform an operation on the workpiece based on the answer to the similarity query and the output indicating the contribution of at least one of the attributes to the answer, and based on a list, in particular a list specifying conditions, values ​​or ranges of the answer to the similarity query and of the output to be fulfilled, The method of claim 6.

8. A computer-implemented method for training a model (100) for explaining answers to similarity queries, comprising: the knowledge graph includes a node including a first node and a second node; the knowledge graph includes edges representing relationships between pairs of nodes; the knowledge graph includes an attribute associated in the knowledge graph with at least one edge or at least one node of the knowledge graph; the similarity query includes a first node and a second node; the answer includes a distance between the first node and the second node; The method comprises: providing the similarity query (304); Providing the answer (312); providing an embedding of the first node (308); providing an embedding of the second node (310); determining (314) an output of the model (100); Including, the model is configured to determine an output to explain the answer to the similarity query depending on the first node, the second node, an embedding of the first node, an embedding of the second node, and the attribute; the output includes at least one value indicating a contribution of one of the attributes to the answer; providing the answer (312) includes determining a similarity between the first node and the second node depending on the first node, the second node, an embedding of the first node, an embedding of the second node, and the attributes; The method comprises: determining (316) an estimated similarity between the first node and the second node depending on the output to explain the answer to the similarity query; training (318) the model (100) depending on the difference between the similarity and the estimated similarity; Including, A method characterized by:

9. - determining (316) the estimated similarity comprises determining a sum, in particular a normalized sum, of the values ​​of the outputs; The method of claim 8.

10. In the apparatus (500), The apparatus comprises at least one processor (502) and at least one memory (504); The at least one memory (504) comprises instructions that, when executed by the at least one processor (502), cause the device (500) to perform a method for explaining similarity query answers according to any one of claims 1 to 7 or a method for training according to claim 8 or 9.

1. An apparatus (500) comprising:

11. A computer-implemented data structure (600), comprising: The data structure (600) At least one data field (602) for describing the answer to the similarity query; At least one data field (602) for the similarity query, the similarity query including a first node of a knowledge graph; at least one data field (602) for the answer, the answer including a second node in the knowledge graph; at least one data field (602) for embedding said first node; at least one data field (602) for embedding said second node; In the knowledge graph, at least one data field (602) for an attribute associated with at least one edge or at least one node of the knowledge graph; at least one data field (602) for the output of the model, in particular the output of the graph neural network that said model comprises, for explaining said answer to said similarity query depending on said first node, said second node, the embedding of said first node, the embedding of said second node, and said attributes; Including, the output includes at least one value indicating the contribution of one of the attributes to the answer. A computer-implemented data structure (600).

12. The data structure (600) at least one data field (602) for a similarity between the first node and the second node determined depending on the first node, the second node, the embedding of the first node, the embedding of the second node, and the attribute; at least one data field (602) for an estimated similarity between the first node and the second node determined in dependence on the output for explaining the answer to the similarity query; at least one data field (602) for the difference between said similarity and said estimated similarity; 11. The computer-implemented data structure (600) of claim 10, comprising:

Images