Causal discovery using hyper-relational knowledge graph link prediction

Abstract

Causal discovery is performed using knowledge graph link prediction. [Solution] Information from a causal network is transformed into a causal knowledge graph according to a mapping, and the causal knowledge graph contains multiple causal links, each causal link containing causal entities, causal relationships, and effect entities with possibilities for mediation. The causal knowledge graph is transformed into an embedding that contains a latent vector space representation of the causal knowledge graph. The embedding is trained using a subset of the causal links in the causal knowledge graph. The embedding is used for causal discovery to predict additional causal links in the causal knowledge graph.

Description

[Technical Field]

[0001] Technical field The aspects of this disclosure generally relate to causal link prediction using hyperrelational knowledge graphs. [Background technology]

[0002] background A knowledge graph is a graphical data model that captures the semantic relationships between entities, which can be events, objects, or concepts. Knowledge graphs can be used to capture causality in terms of cause and effect. Such entity-based representation models enable a broader search space by linking causal entities to related result entities or concepts within the knowledge graph. [Overview of the Initiative] [Means for solving the problem]

[0003] overview In one or more exemplary examples, a method for causal discovery using knowledge graph link prediction comprises translating information from a causal network into a causal knowledge graph, the causal knowledge graph comprising a plurality of causal links, each of which comprises a causal entity, a causal relationship, and an effect entity, and at least one subset of the causal links comprising a hyperrelational causal link, the hyperrelational causal link further connecting the causal entity and the effect entity using a related modifier entity as a mediator, which is an intermediary in a serial causal connection between the causal entity and the effect entity; transforming the causal knowledge graph into an embedding which comprises a latent vector space representation of the causal knowledge graph; training the embedding using the causal links and hyperrelational causal links of the causal knowledge graph; and using the embedding to predict additional causal links in the causal knowledge graph for causal discovery.

[0004] In one or more exemplary examples, transforming a causal knowledge graph into an embedding involves using an approach based on a StarE graph neural network, in which a variable number of modifier pairs can be configured to be associated with each of the causal links.

[0005] In one or more exemplary examples, translation is performed in accordance with a causal ontology, which defines the concepts for structuring causal knowledge graphs.

[0006] In one or more exemplary examples, translation further includes mapping nodes in a causal network to causal entities in a causal knowledge graph, and mapping edges in a causal network to causal links in a causal knowledge graph.

[0007] In one or more exemplary examples, the modifier entity includes an intermediary entity in serial between the causative entity and the resulting entity such that the causative entity causes the resulting entity and / or the causative entity describes the resulting entity.

[0008] In one or more exemplary examples, a hyperrelational causal link includes a modifier relationship, which includes one or more hasMediator modifier relationships indicating the presence or absence of mediation, or hasMediatorType modifier relationships indicating the type of mediation.

[0009] In one or more exemplary examples, causal discovery includes a causal explanation for predicting the type of causal entity given an outcome entity.

[0010] In one or more exemplary examples, causal discovery includes causal prediction for predicting the type of result entity given a causal entity.

[0011] In one or more exemplary examples, a system for causal discovery using knowledge graph link prediction includes one or more hardware computing devices configured to translate information from a causal network into a causal knowledge graph, the causal knowledge graph comprising a plurality of causal links, each of which comprises a causal entity, a causal relationship, and an effect entity, and at least one subset of the causal links comprising a hyperrelational causal link, the hyperrelational causal link further connecting the causal entity and the effect entity using a related modifier entity as a mediator, which is an intermediary in a serial causal connection between the causal entity and the effect entity; transforming the causal knowledge graph into an embedding which comprises a latent vector space representation of the causal knowledge graph; training the embedding using the causal links and hyperrelational causal links of the causal knowledge graph; and using the embedding to predict additional causal links in the causal knowledge graph for causal discovery.

[0012] In one or more exemplary examples, transforming a causal knowledge graph into an embedding involves using an approach based on a StarE graph neural network, in which a variable number of modifier pairs can be configured to be associated with each of the causal links.

[0013] In one or more exemplary examples, the translation of information is carried out in accordance with a causal ontology, which defines the concepts for structuring a causal knowledge graph.

[0014] In one or more exemplary examples, translating information further includes mapping nodes in a causal network to causal entities in a causal knowledge graph, and mapping edges in a causal network to causal links in a causal knowledge graph.

[0015] In one or more exemplary examples, the modifier entity, in the case of a causal explanation, includes, serially between the result entity and the cause entity, an intermediary entity that explains the cause of the result entity.

[0016] In one or more exemplary examples, the modifier entity, in the case of causal prediction, includes an intermediary entity that brings about the result entity, in a serial relationship between the cause entity and the result entity.

[0017] In one or more exemplary examples, a hyperrelational causal link includes a modifier relationship, which includes one or more hasMediator modifier relationships indicating the presence or absence of mediation, and hasMediatorType modifier relationships indicating the type of mediation.

[0018] In one or more exemplary examples, causal discovery includes a causal explanation for predicting the type of causal entity given an outcome entity.

[0019] In one or more exemplary examples, causal discovery includes causal prediction for predicting the type of result entity given a causal entity.

[0020] In one or more exemplary examples, a non-temporal computer-readable medium causes one or more computing devices to perform an operation which includes: an instruction for causal discovery using knowledge graph link prediction, the instruction, when executed by one or more computing devices, to translate information from a causal network into a causal knowledge graph, the causal knowledge graph comprising a plurality of causal links, each of which comprises a causal entity, a causal relationship, and an effect entity, and at least one subset of the causal links comprising a hyperrelational causal link, the hyperrelational causal link further connecting the causal entity and the effect entity using a related modifier entity as a mediator, which is an intermediary in a serial causal connection between the causal entity and the effect entity; transforming the causal knowledge graph into an embedding which comprises a latent vector space representation of the causal knowledge graph; training the embedding using the causal links and hyperrelational causal links of the causal knowledge graph; and using the embedding for causal discovery to predict additional causal links in the causal knowledge graph.

[0021] In one or more exemplary examples, transforming a causal knowledge graph into an embedding involves using an approach based on a StarE graph neural network, in which a variable number of modifier pairs can be configured to be associated with each of the causal links.

[0022] In one or more exemplary examples, the translation of information is carried out in accordance with a causal ontology, which defines the concepts for structuring a causal knowledge graph.

[0023] In one or more exemplary examples, a hyperrelational causal link includes a modifier relationship, which includes one or more hasMediator modifier relationships indicating the presence or absence of mediation, and / or hasMediatorType modifier relationships indicating the type of mediation.

[0024] In one or more exemplary examples, the modifier entity includes an intermediary entity in serial between the causative entity and the resulting entity such that the causative entity causes the resulting entity and / or the causative entity describes the resulting entity. [Brief explanation of the drawing]

[0025] [Figure 1A] This diagram illustrates a typical series of causal connections where A causes B, and ultimately B causes C. [Figure 1B] This diagram shows a series of causal links encoded as knowledge graph links using Resource Description Format (RDF). [Figure 1C] This diagram illustrates causal links as hyperrelational links, where mediating entities are represented as hyperrelations using the RDF-Star format with the hyperrelational predicate hasMediator. [Figure 2A] This figure shows a flowchart of the four phases of the approach of this disclosure for discovering missing causal links in an incomplete causal network. [Figure 2B] This figure shows an example of causal link prediction including modifier relationships. [Figure 2C] This figure shows an example of causal link prediction without using modifier relationships. [Figure 3] This figure shows an example of a concrete relationship within the context of CausalKG. [Figure 4] This is a diagram illustrating an exemplary StarE encoder. [Figure 5] This figure shows an exemplary StarE architecture for a link prediction model. [Figure 6A] This figure shows an exemplary CausalKG structure, including subgraph C which has a causal link consisting only of causal relationships. [Figure 6B] This figure shows an exemplary CausalKG structure, including a subgraph CT with causal links containing information about causal relationships and entity types. [Figure 6C] This figure shows an exemplary CausalKG structure, including a subgraph CTP containing information about causal relationships, entity type relationships, and objects participating in causal events. [Figure 7A] This figure shows snapshots of the collision events shown in the video at times t-1, t, and t+1. [Figure 7B] This figure shows the causal event graph of the snapshot in Figure 7A. [Figure 7C] This figure shows the representation of causal and mediating links between causal events in two different CausalKGs. [Figure 8A] This figure shows exemplary experimental results in which causal explanation was performed using subgraph C. [Figure 8B] This figure shows exemplary experimental results in which causal explanation was performed using subgraph CT. [Figure 8C] This figure shows exemplary experimental results in which causal explanation was performed using subgraph CTP. [Figure 9A] This figure shows exemplary experimental results in which causal prediction was performed using subgraph C. [Figure 9B] This figure shows exemplary experimental results of causal prediction using subgraph CT. [Figure 9C] This figure shows exemplary experimental results of causal prediction using subgraph CTP. [Figure 10] This diagram illustrates an exemplary process for causal discovery using hyperrelationships. [Figure 11] This is a schematic diagram of a control system configured to control a robot assistant based on causal discovery. [Figure 12] This figure shows an exemplary manufacturing system for use in anomaly detection and / or generation of synthesized anomaly data. [Modes for carrying out the invention]

[0026] Detailed explanation While detailed embodiments of the invention are disclosed herein as necessary, it should be understood that the disclosed embodiments are merely illustrative examples of the invention, which may be realized in various alternative forms. The drawings are not necessarily to scale, and some features may be exaggerated or minimized to illustrate the details of certain components. Accordingly, certain structural and functional details disclosed herein should not be construed as limiting, but rather as merely representative grounds to teach those skilled in the art how to employ the invention in various ways.

[0027] Causality has traditionally been represented using Causal Bayesian Networks (CBNs), where nodes represent events and edges or links represent the causal relationship between two events. Obtaining a complete network with all causal relationships is important for many downstream applications. However, in practice, such causal networks are often incomplete due to missing causal links. Recent approaches have successfully addressed this problem by encoding causal networks within triple-based knowledge graphs such as resource description format (RDF), and then using knowledge graph link prediction techniques to discover missing causal links. While this approach works well for simple binary causal links, it does not support more complex links such as mediating causal links.

[0028] Figures 1A through 1C illustrate various examples of causal connection structures. Figure 1A shows an exemplary serial causal connection where A causes B, and ultimately B causes C. Figure 1B shows a serial causal link encoded as a knowledge graph link using RDF format. Figure 1C shows a causal link as a hyperrelational link, where the mediating entity is represented as a hyperrelation using the hyperrelational predicate hasMediator (e.g., total causal effect, natural direct effect, natural indirect effect). Causal links are encoded as knowledge graph links using RDF-Star format.

[0029] Referring to Figure 1A, we can define a simple binary causal link such that A causes B. In this example, A can be called the cause and B can be called the effect. Such causal links can also be chained together such that A causes B, and then B causes C. In this more complex example, there is a causal link between A and C with mediation by B. Nodes A and C can be called the cause and effect, respectively, and node B can be called the mediator.

[0030] Existing approaches using knowledge graph (KG) link prediction can predict direct, binary causal links, such as A causing B, but they cannot predict more complex causal links involving mediation, such as A causing C through mediation by B.

[0031] Figure 2A shows a flowchart of the four phases of the present disclosure's approach for discovering missing causal links 308 in an incomplete causal network. This approach can be called HyperCausalLP, which supports and leverages mediating links. This is achieved by using a hyperrelational knowledge graph to represent complex causal relationships. These four main phases are causal network construction 202, causal knowledge graph creation 204, embedding generation 206, and causal discovery 208.

[0032] Causal network construction 202 may include discovering known causal relationships and encoding them into a causal network 210. Examples in Figures 1A and 1C show how mediating causal links 308 can be encoded as hyperrelationships. RDF-Star1 can be used to encode these causal links 308. This causal network construction 202 may be carried out using observational data 212 and / or domain knowledge 214. Causal knowledge graph creation 204 may include translating the causal network 210 into a CausalKG216 conforming to a causal ontology 218.

[0033] Embedding generation 206 may include training KG embedding models 220A and 220B for CausalKG 216. Subsequently, a knowledge graph embedding (KGE) model is trained using a hyperrelational KG. This may be achieved using the StarE algorithm, which employs a neural network-based message delivery framework. Finally, a new causal link 308 is predicted using the KGE model. In this approach, embeddings include CausalKG-Base, which does not use mediating relations, and CausalKG-M, which is a hyperrelational graph-based embedding that uses mediation as a hyperrelation.

[0034] Causal discovery 208 may involve using knowledge graph embeddings for causal discovery 208 tasks. One example of such a task is predicting new causal links 308 within CausalKG216. More specifically, the following two causal tasks are performed: (1) causal explanation, where an outcome event is given and its cause is predicted, and (2) causal prediction, where a causal event is given and its outcome is predicted.

[0035] This approach to discovering missing causal links 308 using mediation can be evaluated using CLEVRER-Humans, a benchmark dataset for causal inference. This dataset provides a set of causal networks 210 representing collision events in simulated videos. A hyperrelational CausalKG is generated from this set of causal networks, a KG embedding model is trained, and two link prediction operations are performed in the form of causal explanation and causal prediction. The results of this evaluation demonstrate that predicting causal links 308 using a hyperrelational KG embedding model to encode mediation is superior to the baseline approach using a standard triple-based KG embedding model without mediation.

[0036] The causal link prediction 308 can be formulated as the KG link prediction problem. The following definition defines the main concepts, including causal relationships 304, causal links 308, causal entities 302, modifier entities 312, hyperrelational causal links 316, and causal knowledge graphs.

[0037] The causal knowledge graph CausalKG is a hyperrelational KG that contains causal knowledge in the form of causal relationships 304 and causal entities 302. Let CausalKG = (N, R, E, Ec), where, N: A set of nodes representing entities, R: A set of labels that represent relationships, E ⊆ N × R × N: A set of edges representing links between pairs of entities. Each link is a triple <h, r, t>, where h is the head entity, r is the relation, and t is the tail entity. N c ⊆ N: A set of nodes representing causal entities 302, R c ⊆ R: A set of labels representing causal relationships 304, P denotes the power set, E c ⊆ N c × R c × N c × P(R m × N m ): A set of edges representing hyper-relational causal links 316 connecting pairs of causal entities 302.

[0038] Causal entity 302, n c ∈ N c is an entity that is the head or tail of a causal link 308. There are two types of causal entities 302, namely, a causal entity (n cause ) and a resultant entity (n effect ) such that the causal entity causes the resultant entity. However, in the case of hyper-relational links, the causal entity 302 can also be a modifier entity 312 (n m ∈ N m ).

[0039] Causal relationship 304, r c ∈ R c is a relationship representing the causal relatedness between entities. There are four types of causal relationships 304, namely, causes(r causes ∈ R c ) is a causal relationship 304 from a causal entity to a resultant entity, causedBy(r causedBy ∈ R c) is a causal relationship from the result entity to the cause entity, 304, that is, the inverse of causes, ReceiveType(r causesType ∈R c ) is a causal relationship 304 from a cause entity to a type of effect entity, causedByType(r causedByType ∈R c This is a causal relationship 304 from the result entity to the type of cause entity.

[0040] Causal link 308,e c ∈E c This is an edge in causalKG216 that connects the pair of causal entities 302 to causal relationship 304. Causal link 308 is a triple <h c ,r c ,t c > and here, h c This is the head causal entity 302, and r c This is a causal relationship 304, and t c This is tail causal entity 302.

[0041] The modifier pair 310, q ∈ Q is a hyperrelation within CausalKG216, and this hyperrelation connects the causal link 308 to the "relation-entity" pair of that hyperrelation. Q is a modifier relation r m And, modifier entity 312,n m Modifier pair 310(r m ,n m It is a set of ).

[0042] Modifier entity 312,n m ∈N mis a causal entity 302, which is part of modifier pair 310. In a given serial causal connection, modifier entity 312 (i.e., mediator) is an entity between a sequentially connected cause entity and an effect entity, and is also known as mediator. In this disclosure, modifier entity 312 refers to mediator in a serial causal connection. In the context of this disclosure, the terms modifier entity 312 and mediator may be used interchangeably.

[0043] Modifier relationship 314,r m ∈R m This is a relationship that represents the association between the causal link 308 and the modifier entity 312 (e.g., the mediating entity). There are two types of modifier relationships 314, namely, hasMediator(r hasMediator ∈R m ) is a modifier relationship 314 from the causal link 308 to the mediating entity, hasMediatorType(r hasMediatorType ∈R m ) is a modifier relationship 314 from causal link 308 to type of mediating entity.

[0044] Hyper-relational causal links 316,e c ∈E c This is an edge within causalKG216 that connects a pair of causal entities 302 to a causal relationship 304 and their respective associated mediators (modifier entities 312). Each hyperrelational causal link 316 is a tuple <h c ,r c ,t c ,Q>, where h c This is the head causal entity 302, and r c This is a causal relationship 304, and t c Q is a tail causal entity 302, and Q is a modifier relation 314, r m And, modifier entity 312,n m Modifier pair 310(r m ,n m It is a set of ).

[0045] The task of extracting causal relationships 304 is to discover new causal links 308 within CausalKG216. Given CausalKG216,G, this task can be performed using knowledge graph link prediction. There are two types of causal relationship extraction: causal prediction and causal explanation. Causal prediction: causal entity (n cause ∈N c ) and causesType relationship (r causesType ∈R c Given ) and modifier pair 310(Q), <n cause ,r causesType Find the type (t) of related result entities such that t, Q>∈G. Causal explanation: Result entity (n effect ∈N c ) and the causeByType relationship (r causedByType ∈R c Given ) and modifier pair 310(Q), <n effect ,r causedByType Find the type (t) of related causal entities such that t, Q>∈G.

[0046] Referring again to Figure 2A, HyperCausalLP is structured into four main phases: causal network construction 202, in which known causal relationships 304 are discovered and encoded into a causal network 210; causalKG creation 204, in which the causal network 210 is translated into a CausalKG 216 conforming to a hyperrelational causal ontology 218 incorporating modifier pairs 310; embedding generation 206, in which hyperrelational KG embeddings 224 are learned for the CausalKG 216; and causal discovery 208, in which new causal links 308 within the CausalKG 216 are predicted.

[0047] The causal network 210 is a graphical model known as a causal Bayesian network structured as a directed acyclic graph. In this model, nodes symbolically represent events, and edges represent causal links 308 between these events. CN=(N cn ,E cn A network represented as ) has N nodes. cn and edge E cn This includes the direction of each edge in the network, which indicates the direction of causality. Given a causal network 210 with three nodes, the causal links 308 can have three different orientation structures: serial, branched, and collider. A serial structure is one in which causal relationships are traced in serial order, such that the first event causes the second event, and the second event causes the third event. In a branched structure, the first event causes both the second and third events. In a collider structure, two independent events together cause the third event. However, this disclosure focuses on a serial structure as shown in Figure 1A. The first node is considered the causal entity, the second node the mediating entity, and the third node the result entity.

[0048] The process of transforming data from a causal network 210 to a hyperrelational causal knowledge graph (CausalKG216) involves several direct transformations: N cn →N c Nodes within the causal network 210 become causal entities 302 within CausalKG216. Intermediary nodes within the causal network 210 become intermediary entities within CausalKG216, and intermediary entities are represented as modifier entities 312. E_(cn)→E_c: The edge in causal network 210 is, <n cause ,r causes ,n effect ,r m ,n mThis is causal link 308 within CausalKG216, in the format >.

[0049] CausalKG216 also incorporates other causal relationships 304, as well as modifier relationships 314 such as causedBy, causesType, causedByType, hasMediator, and hasMediatorType. CausalKG216 contains information from the causal network 210 and conforms to the causal ontology 218. The causal ontology 218 may be derived from concepts from causal Bayesian networks and causal artificial intelligence (AI) such as do-calculus. The causal ontology 218 may be used to define the semantics and structure of the causal relationships 304 and the nodes within the CBN. The ontology can define the main concepts used to structure CausalKG216, which includes the causal entities 302, the causal relationships 304, and the mediators.

[0050] CausalKG216 is used for causal link 308 prediction using KG link prediction. There are two causal link 308 prediction tasks: causal explanation and causal prediction. The goal of causal explanation is to predict the type of causal entity linked to the outcome entity. The goal of causal prediction is to predict the type of outcome entity linked to the causal entity. The goal for both tasks is not to predict instances of a specific causal entity (in the case of causal explanation) or outcome entity (in the case of causal prediction), but rather to predict the type of each of these entities.

[0051] The causal entity (in the case of a causal explanation) and the effect entity (in the case of a causal prediction) are not directly linked to the type of causal entity and the type of effect entity, respectively. They are two hops apart, that is, in the case of a causal explanation... <n effect ,r causedBy ,n cause >, <n cause, is of type, type>, and in the case of causal prediction, <n cause , r causes , n effect , <n effect , is of type, type>. The embedding model makes predictions about directly linked entities.

[0052] Figure 3 shows an example 300 of an embodied relationship in the context of CausalKG216. Example 300 shows the embodied causal relationships 304, namely, causesType and causedByType. causedByType is an embodied relationship from an instance of the result entity to the type of the cause entity. causesType is an embodied relationship from an instance of the cause entity to the type of the result entity. Example 300 also shows two modifier relationships 314 associated with the causal relationship, namely, hasMediator and hasMediatorType. The modifier relationship 314 is also associated with causedBy, which is the inverse of the causal relationship. Such an example 300 may be used for causal prediction to add a link connecting the cause entity to the type of the result entity, that is, for causesType(r causesType ∈R c ). In other examples, such an example 300 is used for causal explanation to add a link connecting the result entity to the type of the cause entity, that is, for causedByType(r causedByType ∈R c ). In addition to all the above knowledge, CausalKG216 can also incorporate additional domain knowledge 214 related to entities not explicitly mentioned into the causal network 210.

[0053] Referring back to Figure 2A, CausalKG216 can be transformed into a low-dimensional continuous latent vector space representation which can be called KGE224. KGE224 may be used for downstream tasks such as link prediction, entity classification, triple classification, etc. The Hyper-CausalLP approach can use the KG embedding algorithm to generate embeddings which can be used for causal link 308 prediction.

[0054] This approach trains two types of KGE224 for CausalKG216: 1) CausalKGE-Base224B embeddings without mediation (no hyperrelations), and 2) CausalKGE-M224A embeddings with mediation as hyperrelations (represented using modifier pairs 310). The CausalKGE-Base224B embeddings may be trained using causal links 308, ignoring the mediations associated with each link. On the other hand, the CausalKGE-M224A embeddings may be trained using causal links 308 with mediations. The CausalKGE-Base224B and CausalKGE-M224A embeddings may be evaluated against a causal link 308 prediction task using KG link prediction techniques. The CausalKG216 embedding for CausalKGE-Base224B may, in one example, be generated using the KG embedding algorithm available in the Ampligraph library 2.

[0055] Figure 2B shows an example of causal link 308 prediction including modifier relationship 314. This may be performed, for example, using the CausalKGE-M224A embedding along with mediation as a hyperrelation. Figure 2C shows an example of causal link 308 prediction without modifier relationship 314. This may be performed, for example, using the CausalKGE-Base224B embedding.

[0056] CausalKGE-M224A may be generated as a hyperrelational KGE224 model based on a graph neural network, such as StarE. StarE is a graph neural network-based approach. StarE allows various numbers of modifier pairs 310 to be associated with causal links 308. StarE combines causal relationship embeddings 304 with fixed-length vectors representing the associated modifier pairs 310. StarE incorporates modifier pairs with causal links 308 into a message delivery process. The StarE model includes two parts: a StarE encoder, shown in Figure 4, and a StarE architecture, which includes a transformer-based decoder, shown in Figure 5. The StarE encoder and the transformer-based decoder may be trained together.

[0057] Figure 4 shows an exemplary StarE encoder 400. The StarE encoder 400 encodes hyperrelations for causality. The modifier pairs (or mediation pairs) of the hyperrelations pass through the composite function φq, are summed together, and transformed by the weights Wq. The transformed vector is merged with γ and φr. The final node, i.e., the causal entity, combines the messages from all the hyperrelations. In StarE, this is specified as follows: φ is a composite function that combines nodes and their respective relationships. W λ(r) This is a shared parameter of directionality for outgoing relationships, incoming relationships, and self-looping relationships. γ is the principal relation (r c This is a function that combines the expression with the expression of its modifier (Q).

[0058] The StarE encoder 400 may be used in a KGE224 model or relational inference network. For example, r c (Cause) is r m1 ,r m2, interacts with a mediating entity that includes m1 and m2, and these relationships are processed by various different transformations (Φq, Φr) and aggregations (Σ). The model calculates a weighted sum or transformation of these intermediate relationships and outputs an encoded representation (h c ) of the causal entity, and this representation (h c ) captures how the causal entity is related to the result entity (t c ) and its mediation. As shown in the illustration, r c represents a set of causes, r m represents a set of mediators, and m represents a set of mediating entities. The Φ (phi) function refers to the transformation of an input vector or embedding, and the Φq, Φr nodes represent transformation functions or mappings, such as activation functions of a neural network or parameterized functions that encode relationships between entities. Φq can refer to a transformation specific to the mediator, while Φr can refer to a transformation applied after a summation step that aggregates intermediate information. The summation node (represented by Σ) can aggregate inputs from different mediators, combine information from multiple different mediator paths, and integrate different relationship signals before further processing. W q and W λ(r) refer to weights in a neural network that correspond to learnable parameters that adjust the influence of the aggregated information. W q may be applied to the aggregation of mediators, while W λ(r) can affect the causal entity path. γ represents a gating mechanism or activation function that modulates the encoded information before it is transmitted to the final summation step and prediction step.

[0059] Figure 5 shows an exemplary StarE architecture 500 for a link prediction model. As shown, the StarE architecture 500 updates N,R matrices, which are then used to encode relationships in a given query before passing through transformer, pooling, and fully connected layers. The fixed-dimensional output is then compared to N, and the result passes through a sigmoid function to generate a probability distribution across entities.

[0060] The StarE approach initializes two embedding matrices, R (relationships) and E (entities). StarE iteratively updates the embeddings by message transfer over edges in the training set. For a link prediction task, the query is first linearized, and then the relations and entities are encoded using the updated embeddings. The data then passes through a transformer. The transformer's output is averaged to obtain a fixed-dimensional vector representation of the query. The vector passes through a fully connected layer, is multiplied by the entities, and passes through a sigmoid function to obtain a probability distribution across all entities. The top n candidate entities for the link prediction query are obtained.

[0061] The HyperCausalLP approach described herein formulates the problem of predicting causal links 308 as a KG link prediction task. Trained CausalKG216 embedding models, namely CausalKGE-Base224B and CausalKGE-M, are used to predict missing causal links 308 between causal entities 302 within the KG. More specifically, HyperCausalLP is used for causal explanation and causal prediction tasks. Causal explanation aims to predict the cause of an outcome, while causal prediction aims to predict the outcome of a cause. For a given causal link 308, the causal explanation is: <n effect ,r causedByType Predicting links in the format ,?,Q>, causal prediction is, <n cause ,r causesType Predicts links in the format ,?,Q>.

[0062] For a given data set having causal entities 302 related to causal links 308 between entities, causal relationships 304, and mediators, HyperCausalLP can be used to create CausalKG216 and generate and train KGE224. The generated KGE224 can then be used for predicting causal links 308 in the form of causal explanations and causal predictions.

[0063] HyperCausalLP may be evaluated using CLEVRER-Humans, i.e., a benchmark data set for causal inference. More specifically, HyperCausalLP's hyperrelational graph-based causal link 308 prediction approach is 1) given outcome entities, <n effect ,r causedByType ,?,Q> A causal explanation that predicts the type of causal entity for causal link 308, and 2) given a causal entity, <n cause ,r causesType The causal prediction and KG link prediction task may be used to evaluate the type of result entity of a causal triple of the form ,?,Q> (see Figure 3). The above evaluation can be demonstrated using a benchmark data set for causal inference, namely CLEVRER-Humans. This section details the CLEVRER-Humans data set, data preprocessing steps, creation of CausalKG216 from the data set, experimental setup, evaluation metrics, and evaluation for various variations of CausalKG216.

[0064] The first step in generating the CausalKG216 in CLEVRER-Humans involves preprocessing the causal event graph (CEG). The CEG acts as a substitute for the causal network 210, and preprocessing the CEG is crucial to ensure that the CEG conforms to the definition of the causal network 210. Within the causal network 210, edges represent causal links 308 between nodes. The first step in this process is to remove edges with a score of 1, indicating that there is no causal responsibility between two nodes. Next, edges that create cycles within the CEG are removed to maintain the structure of a directed acyclic graph. Finally, a CEG is excluded if it has no remaining causal links 308 or if the depth from the root node to the leaf nodes is less than 2. After preprocessing, 764 CEGs remain in the CLEVRER-Humans data set.

[0065] Regarding event extraction, the CLEVRER-Humans data set features 27 distinct events such as collide, enter, exit, halt, and go. These events can be classified into two categories: binary events and singular events. Binary events involve two participating objects and include actions such as collide, bump, hit, bounce, and sideswipe. Singular events involve only one object and include actions such as enter, exit, and stop. Information regarding the event type and participating objects may be extracted from node descriptions within the CEG by parsing the CEG JavaScript Object Notation (JSON) file provided by the data set. Instead of using the conjugated forms of event labels (e.g., collided, hits, pushed), you can use the Berkeley Neural Semantic Parser and the Natural Language Toolkit (NLTK) Stem Lens Matterizer to capture the base forms of those verbs (e.g., collide, hit, push) (for example). Nodes describing multiple events, such as "The red ball collides with the blue sphere and hits the yellow cylinder," can be removed from the CEG because they describe more than one event. Instead, you can focus on nodes describing a single event.

[0066] In addition to extracting event types, the extraction of objects and object properties allows for the collection of information about participating objects and their characteristics, including color, shape, and material. Some object characteristics within the data set may be mislabeled, such as an object being labeled as gold instead of yellow. These mislabeling issues can be identified, and terminology can be normalized accordingly.

[0067] CausalKG216 can be created from CLEVRER-Humans by encoding causal information within the CEG in RDF format according to the causal ontology 218. The approach of this disclosure creates two different KGs, namely CausalKG-Base220B and CausalKG-M220A. CausalKG-Base220B may be a simple KG having causal links 308, while CausalKG-M220A is a hyperrelational KG consisting of mediation as hyperrelations (modifiers). A hyperrelation having mediation information between two given nodes in the CEG may be encoded using RDF-star format as described herein.

[0068] The KG may contain causal relationships and, in some examples, may include details about events (such as hit, collide, push, etc.), the objects involved, and their attributes. The CEG can function as a graphical representation of events in the video. Three ontologs can be used to represent information from the CEG: the causal ontology 218, the scene ontology (prefixed "so:"), and the semantic sensor network ontology (prefixed "ssn:"). The causal ontology 218 may be employed for events (as causal entities 302), causal relationships 304, and their corresponding causal mediations (i.e., modifier pairs 310). The scene ontology and sensor ontology represent additional video information such as scenes, objects, and object properties. Each video can be represented as a scene (so: Scene) using the concept of the scene ontology. This includes representing and connecting events within a scene (using the so:includes relationship), involved objects (using the so:hasParticipant relationship), and object properties (using the ssn:hasProperty relationship). In total, CausalKG216 from CLEVRER-Humans contains over 48K links, 5664 entities, 31 entity types, and 10 relationships.

[0069] As shown in Figures 6A to 6C, CausalKGE-Base224B embeddings and CausalKGE-M224A embeddings may be generated and evaluated against various CLEVRER-Humans CausalKG216 subgraph structures for causal explanation and causal prediction tasks. For CausalKG-M and a given subgraph, hyperrelations (modifier pairs 310) may be associated with causal relationships 304 of causes and causedBy, as shown in Figure 3. When different types of information are available in CausalKG216, different graph structures can be used to evaluate the performance of HyperCausalLP. For example, three distinct subgraph structures can be defined, increasing the level of expressiveness (as shown in Figures 6A to 6C).

[0070] Figure 6A shows an exemplary CausalKG216 structure, including subgraph C which has a causal link 308 with only causal relationships 304. These causal relationships 304 may include, for example, causes, causedBy, causesType, and causedByType. Figure 6B shows an exemplary CausalKG216 structure, including subgraph CT which has a causal link 308 with causal relationships 304 and information about entity types. These entity types may include, for example, rdf:type. Figure 6C shows an exemplary CausalKG216 structure, including subgraph CTP which has causal relationships 304, relationships of entity types, and information about objects participating in a causal event (e.g., hasParticipant). In the case of CausalKGE-M224A, a hyperrelation (modifier pair 310) is associated with the causal relationships 304 of causes and causedBy.

[0071] The hyperparameters for each of these graph structures can be optimized for both causal explanation and causal prediction tasks. The CausalKGE-Base224B model for each graph structure can be trained based on its respective optimized hyperparameters. The CausalKGE-M224A model can be trained based on the StarE hyperparameters. The trained CausalKGE can then be employed for the causal link 308 prediction task using a link prediction method.

[0072] HyperCausalLP can be evaluated using an experimental design setup for KG link prediction. Causal links 308,E within CausalKG216 c Given a set of , the set of causal links 308 <h c ,r c ,t c ,Q> tail t c or head h c By modifying using another causal entity 302 within KG, a set of tampered links T' is generated. For example, the head h' c ≠h c If you replace it with, <h’ c ,r c ,t c ,Q> and the tail is t' c ≠t c If you replace it with, <h c ,r c ,t' c ,Q>. The model is correct link <h c ,r c ,t c Q> and the tampered link <h’ c ,r c ,t c ,Q>, <h c ,r c ,t' cA score is assigned to ,Q>∈T'. The scores can be sorted to obtain the correct link rankings. A filtered evaluation setting and a filtered tampered link T' can be used to exclude links that exist in the training and validation sets. The performance of HyperCausalLP can be evaluated using two metrics: Mean reciprocal rank (MRR) and Hits@K (Hits@K for K=1,3,10). MRR is the average of the reciprocals of the individual ranks of the test links. Hits@k is the proportion of test links that are among the top k ranked links. The higher the values ​​of both metrics, the better the model performs.

[0073] Figure 7C shows exemplary snapshots of the representations of CausalKG-Base and CausalKG-M. As shown in Figure 7A, snapshots of collision events are shown in the video from CLEVRER-Humans at times t-1, t, and t+1. Three consecutive collision events occur: A: a red cube collides with a yellow ball, B: the yellow ball hits a blue cylinder, and C: the blue cylinder moves. A, B, and C are causal entities 302. A.Type is Collide, B.Type is Hit, and C.Type is Move. Figure 7B shows the causal event graph of the snapshot in Figure 7A. Figure 7C shows representations of causal links and mediating (modifier pair 310) links within two different CausalKG216.

[0074] Figures 8A to 8C summarize exemplary experimental results for CausalKGE-M224A and CausalKGE-Base224B for causal explanation. Figures 9A to 9C summarize exemplary experimental results for CausalKGE-M224A and CausalKGE-Base224B for causal prediction. More specifically, Figures 8A to 8C and 9A to 9C show the MRR and Hit@K (k=1,3,10) for five KGE224 models evaluated against various CausalKG216 subgraphs. These CausalKG216 subgraphs include the C subgraph shown in Figure 6A, the CT subgraph shown in Figure 6B, and the CTP subgraph shown in Figure 6C.

[0075] In these results, HyperCausalLP was evaluated for CausalKG216 generated from the CLEVRER-Humans dataset for causal link 308 prediction. This approach was evaluated for CausalKG-M using StarE in three different CausalKG216 subgraphs C, CT, and CTP, each with different hyperrelations: CausalKG-HasMediator: Modifier relationship 314 is hasMediator CausalKG-HasMediatorType: Modifier relationship 314 is hasMediatorType CausalKG-HasMediatorInstanceType: Modifier relationship 314: both hasMediator and hasMediatorType

[0076] The results (i.e., MRR, HitK) show a significant performance improvement in CausalKGE-M224A compared to CausalKGE-Base224B, a baseline model that only has links and no hyperrelations (or mediating information). CausalKGE-Base224B was evaluated using four KGE224 models: TransE, DistMult, HolE, and ComplEx. By incorporating additional knowledge (i.e., CT, CTP) into CausalKGE-M224A, performance improvements are observed compared to the simpler C subgraph.

[0077] By adding more knowledge, KG-link prediction performance for both causal explanation and causal prediction tasks is improved. The MRR score of CausalKGE-M224A with hasMediator, when enhanced with additional knowledge for causal prediction (i.e., CTP), is 13.77% better than C. The MRR score of CausalKGE-M224A with hasMediatorType, when enhanced with additional knowledge for causal prediction (i.e., CTP), is 12.55% better than C. The MRR score of CausalKGE-M224A with both hasMediator and hasMediatorType, when enhanced with additional knowledge for causal prediction (i.e., CTP), is 8.36% better than C. The MRR score of CausalKGE-M224A with hasMediator, when enhanced with additional knowledge for causal explanation (i.e., CTP), is 7.04% better than C. The MRR score of CausalKGE-M224A using hasMediatorType, when enhanced by additional knowledge for causal explanation (i.e., CTP), is 6.28% better than C. The MRR score of CausalKGE-M224A using both hasMediator and hasMediatorType, when enhanced by additional knowledge for causal explanation (i.e., CTP), is 10.13% better than C. Incorporating mediation into causal link 308 provides crucial additional knowledge for the causal link 308 prediction task. The hyperrelation hasMediator performs best compared to other hyperrelations hasMediatorType, and also compared to considering both hasMediator and hasMediatorType together. Hyperrelation-based KG performs better with a larger number of modifiers. This invention successfully demonstrates that knowledge incorporated into hyperrelations (modifiers) significantly improves causal link 308 prediction.

[0078] Therefore, an approach is disclosed for discovering missing causal links 308 within an incomplete causal network 210. The approach of this disclosure incorporates mediating information from the CBN into the KG as hyperrelationships. A KGE224 model trained with modifiers (mediating or hyperrelationships) outperforms all baseline KGE224 metrics without modifiers. The results demonstrate that the effective fusion of causal links 308 within the KG with modifiers (mediating or hyperrelationships) can facilitate the completion of the incomplete causal network 210.

[0079] Figure 10 shows an exemplary process 1000 for causal discovery using hyperrelationships. Process 1000 can implement the approach of this disclosure for causal discovery using knowledge graph link prediction, addressing a critical unresolved issue in the prior art by considering mediating information along with causal links 308. By using process 1000, it can be confirmed that a KGE model trained with mediating information outperforms a baseline KGE metric without mediating information. The results demonstrate that knowledge embedded in hyperrelationships significantly improves causal link 308 prediction.

[0080] In operation 1002, causal network construction 202 is performed. Causal network construction 202 may include discovering known causal relationships and encoding them into a causal network 210. This causal network construction 202 may be performed using observational data 212 and / or domain knowledge 214.

[0081] In operation 1004, causal knowledge graph creation 204 is performed. Causal knowledge graph creation 204 may include translating the causal network 210 into CausalKG216 conforming to the causal ontology 218. CausalKG216 contains multiple causal links, each of which contains a causal entity, a causal relationship, and an effect entity. In one example, information from the causal network 210 can be translated into CausalKG216 according to a mapping. The mapping may include mapping nodes in the causal network 210 to causal entities 302 in CausalKG216 and mapping edges in the causal network 210 to causal links 308 in CausalKG216.

[0082] In operation 1006, embedding learning 206 is performed. Embedding learning 206 may include training KG embedding models 220A and 220B for CausalKG216 using a training set and evaluating the training using a test set. This may be performed in two different approaches. In the first approach, CausalKGE-Base224B is generated using embeddings without mediation (no hyperrelations). The CausalKGE-Base224B embedding may be trained using causal links 308, ignoring the mediations associated with each link. In the second approach, CausalKGE-M224A is generated using embeddings with mediations as hyperrelations (represented using modifier pairs 310). The CausalKGE-M224A embedding may be trained using causal links 308 along with the mediations. The CausalKGE-Base224B and CausalKGE-M224A embeddings may be evaluated against a causal link 308 prediction task using KG link prediction techniques. The CausalKG216 embedding for CausalKGE-Base224B may, in one example, be generated using a KG embedding algorithm available in the Ampligraph library 2. The CausalKGE-M224A may be generated as a hyperrelational KGE224 model based on a graph neural network such as StarE. In many examples, applying modifier pairs 310 from the causal network 210 is superior to baseline KGE metrics that have not been trained on hyperrelationships.

[0083] In operation 1008, causal discovery 208 is performed. Causal discovery 208 may include using knowledge graph embeddings 224A,224B for the causal discovery task. One example of such a task is to predict a new causal link 308 in CausalKG216. In some examples, causal discovery 208 includes a causal description to predict the type of causal entity given an outcome entity. In some examples, causal discovery 208 includes a causal prediction to predict the type of outcome entity given a causal entity. After operation 1008, process 1000 terminates.

[0084] Figure 11 shows a schematic diagram of the interaction between a computer-controlled machine 1102 and a control system 1112. The computer-controlled machine 1102 can implement a causal discovery 208 and the use of predicted causal information. Referring to Figures 10 and 1A to 9C, the approaches described herein can be implemented in connection with such a computer-controlled machine 1102 and control system 1112. The computer-controlled machine 1102 includes an actuator 1114 and a sensor 1116. The actuator 1114 may include one or more actuators, and the sensor 1116 may include one or more sensors. The sensor 1116 is configured to detect the state of the computer-controlled machine 1102. The sensor 1116 may be configured to encode the detected state into a sensor signal 1118 and transmit the sensor signal 1118 to the control system 1112. Non-limiting examples of sensor 1116 include video sensors, radar sensors, LiDAR sensors, ultrasonic sensors, and motion sensors. In one embodiment, sensor 1116 is an optical sensor configured to detect an optical image of the environment near a computer-controlled machine 1102.

[0085] The control system 1112 is configured to receive a sensor signal 1118 from the computer-controlled machine 1102. As described below, the control system 1112 may be further configured to calculate an actuator control command 1120 depending on the sensor signal 1118 and to transmit the actuator control command 1120 to the actuator 1114 of the computer-controlled machine 1102.

[0086] As shown in Figure 11, the control system 1112 includes a receiving unit 1122. The receiving unit 1122 may be configured to receive a sensor signal 1118 from the sensor 1116 and convert the sensor signal 1118 into an input signal X. In an alternative embodiment, the sensor signal 1118 is received directly as an input signal X without using the receiving unit 1122. Each input signal x may be a part of each sensor signal 1118. The receiving unit 1122 may be configured to process each sensor signal 1118 to generate each input signal x. The input signals x may include data corresponding to an image recorded by the sensor 1116.

[0087] The control system 1112 includes a machine learning (ML) processing unit 1124. The ML processing unit 1124 may be configured to perform learning, classification, inference, generation, etc., using one or more models as detailed above. In one example, the ML processing unit 1124 is configured to determine an output signal Y from an input signal X. Each output signal y contains information to assign one or more labels to each input signal X. The ML processing unit 1124 can transmit the output signal Y to a conversion unit 1128. The conversion unit 1128 is configured to convert the output signal Y into an actuator control command 1120. The control system 1112 is configured to transmit the actuator control command 1120 to an actuator 1114, which is configured to operate a computer-controlled machine 1102 in response to the actuator control command 1120. In another embodiment, the actuator 1114 is configured to operate a computer-controlled machine 1102 directly based on the output signal Y.

[0088] When actuator 1114 receives an actuator control command 1120, actuator 1114 is configured to perform an action corresponding to the associated actuator control command 1120. Actuator 1114 may include control logic configured to translate actuator control command 1120 into a second actuator control command 1120 used to control actuator 1114. In one or more embodiments, the actuator control command 1120 can be used to control a display instead of or in addition to actuator 1114.

[0089] In other embodiments, the control system 1112 includes sensor 1116 in place of or in addition to computer-controlled machine 1102 which includes sensor 1116. The control system 1112 may also include actuator 1114 in place of or in addition to computer-controlled machine 1102 which includes actuator 1114.

[0090] As shown in Figure 11, the control system 1112 also includes a processor 1130 and a memory 1132. The processor 1130 may include one or more processors. The memory 1132 may include one or more memory devices. A hyperrelational causal link, as determined by one or more embodiments, may be implemented by the control system 1112, which includes non-volatile storage 1126, the processor 1130, and the memory 1132.

[0091] Non-volatile storage 1126 may include one or more persistent data storage devices, such as hard drives, optical drives, tape drives, non-volatile solid-state devices, cloud storage, or any other devices capable of permanently storing information. Processor 1130 may include one or more devices selected from the following, which process (analog or digital) signals based on computer executable instructions residing in memory 1132: namely, high-performance computing (HPC) systems including high-performance cores, microprocessors, microcontrollers, digital signal processors, microcomputers, central processing units (CPUs), field-programmable gate arrays (FPGAs), programmable logic devices, state machines, logic circuits, analog circuits, digital circuits, or any other devices. Memory 1132 may include one or more memory devices, but are not limited to, random-access memory (RAM), volatile memory, non-volatile memory, static random-access memory (SRAM), dynamic random-access memory (DRAM), flash memory, cache memory, or any other devices capable of storing information.

[0092] The processor 1130 may reside in the non-volatile storage 1126 and be configured to load and execute computer executable instructions that embody one or more ML algorithms and / or methodologies of one or more embodiments into memory 1132. The non-volatile storage 1126 may include one or more operating systems and applications. The non-volatile storage 1126 can store compiled and / or interpreted computer programs written using a variety of programming languages ​​and / or technologies, including, but not limited to, Java, C, C++, C#, Objective-C, Fortran, Pascal, JavaScript, Python, Perl, and PL / SQL, either alone or in combination.

[0093] When executed by the processor 1130, the computer-executable instructions of the non-volatile storage 1126 can cause the control system 1112 to implement one or more of the ML algorithms and / or methodologies disclosed herein. The non-volatile storage 1126 may also contain ML data (including data parameters) that support the functions, features and processes of one or more embodiments described herein.

[0094] Program code embodying the algorithms and / or methodologies described herein may be distributed individually or collectively as various different forms of program products. The program code may be distributed using computer-readable storage media containing computer-readable program instructions for causing a processor to implement aspects of one or more embodiments. Computer-readable storage media that are essentially non-transient may include volatile and non-volatile removable and non-removable tangible media implemented in any way or technique for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media may further include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid-state memory technology, portable compact disc read-only memory (CD-ROM) or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other media that can be used to store and read information of a desired information. Computer-readable program instructions may be downloaded from a computer-readable storage medium to a computer, another type of programmable data processing device, or another device, or they may be downloaded via a network to an external computer or external storage device.

[0095] Computer-readable program instructions stored on a computer-readable medium may be used to cause a computer, other types of programmable data processing devices, or other devices to function in a particular manner so that the instructions stored on the computer-readable medium produce a product containing instructions that perform functions, behaviors, and / or actions specified in a flowchart or diagram. In certain alternative embodiments, multiple functions, behaviors, and / or actions specified in a flowchart and diagram can be rearranged, processed serially, and / or processed simultaneously according to one or more embodiments. Furthermore, both flowcharts and / or diagrams may contain more or fewer nodes or blocks than those illustrated according to one or more embodiments.

[0096] The process, method, or algorithm may be implemented, in whole or in part, using appropriate hardware components such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), state machines, controllers, or other hardware components or hardware devices, or combinations of hardware components, software components, and firmware components.

[0097] Figure 12 shows an exemplary manufacturing system 1200 for use in anomaly detection and / or generation of synthesized anomaly data. The system 1200 may be configured to control manufacturing machinery 1202 such as a punch cutter, cutter, or gun drill, as part of a production line.

[0098] System 1200 may be configured to control an actuator 1114 configured to control a manufacturing machine 1202. Sensor 1116 of System 1200 may be configured to capture one or more characteristics of a manufactured product 1204. ML processing unit 1124 may be configured to determine the state of the manufactured product 1204 from one or more captured characteristics. Actuator 1114 may be configured to control System 1200 (e.g., the manufacturing machine) depending on the determined state of the manufactured product 1204 for subsequent manufacturing steps of the manufactured product 1204. In particular, actuator 1114 may be configured to control the function of System 1200 (e.g., the manufacturing machine) for subsequent manufactured products 1206 depending on the determined state of the manufactured product 1204.

[0099] For example, system 1200 can use CausalKG216 to predict the reason for a problem in the manufacturing system 1200, such as what kind of problem it is and what caused it (causedBy) (e.g., causedByType). Alternatively, system 1200 can use CausalKG216 to predict the outcome that should be addressed, such as whether a detected input may cause a problem (e.g., causesType).

[0100] The processes, methods, or algorithms disclosed herein may be deliverable to, or implemented by, an apparatus, controller, or computer, and the apparatus, controller, or computer may include any existing programmable electronic control unit or a dedicated electronic control unit. Similarly, the processes, methods, or algorithms may be stored as data and instructions executable by a controller or computer in various forms, including, but not limited to, information permanently stored in a non-writable storage medium such as a read-only memory (ROM) device, and information variably stored in a writable storage medium such as a flexible disk, magnetic tape, compact disk (CD), RAM device, and other magnetic and optical media. The processes, methods, or algorithms may be implemented in a software-executable object. Alternatively, the processes, methods, or algorithms may be embodied in whole or in part using appropriate hardware components, such as an ASIC, FPGA, state machine, controller, or other hardware component or hardware device, or a combination of hardware components, software components, and firmware components.

[0101] Causal discovery in one or more embodiments predicts causal links between physical objects or components. Such causal links between physical objects or components may be dynamic and / or static interactions. Such causal links between physical objects or components may be mechanical, electrical and / or chemical interactions. For example, without limiting the applicable physical objects or components, the physical objects or components may be the cube and ball described in relation to Figure 7A, and the causal link may be a dynamic mechanical interaction between the cube and the ball. In another non-limiting example, the physical objects or components may be a manufacturing machine 1202 and a manufactured product 1204, and the causal link may be a mechanical interaction between these components.

[0102] While exemplary embodiments have been described above, these embodiments are not intended to describe all possible forms that are covered by the claims. The terms used herein are descriptive, not limiting, and it should be understood that various modifications are possible without departing from the spirit and scope of this disclosure. As stated above, features of various embodiments can be combined to form further embodiments of the invention not expressly described or illustrated. While various embodiments may be described as offering advantages over or being more preferable to other embodiments or implementations of the prior art with respect to one or more desired characteristics, those skilled in the art will recognize that one or more features or characteristics may be compromised to achieve desired overall system attributes that depend on the particular application and implementation. These attributes may include, but are not limited to, strength, durability, lifespan, marketability, appearance, packaging, dimensions, maintainability, weight, manufacturability, ease of assembly, etc. Therefore, to the extent that any embodiment is described as being less desirable than other embodiments or implementations of the prior art with respect to one or more characteristics, these embodiments are not outside the scope of this disclosure and may even be desirable for a particular application.

Claims

1. A method for causal discovery that uses knowledge graph link prediction, The aforementioned method, Translating information from a causal network into a causal knowledge graph, wherein the causal knowledge graph includes a plurality of causal links, each of which includes a causal entity, a causal relationship, and an effect entity, and at least one subset of the causal links includes a hyperrelational causal link, the hyperrelational causal link further connects the causal entity and the effect entity using related modifier entities as mediators, which are mediators in the serial causal connection between the causal entity and the effect entity. The causal knowledge graph is transformed into an embedding that includes the latent vector space representation of the causal knowledge graph, Training the embedding using the causal links and hyperrelational causal links of the causal knowledge graph, To discover causality, the embedding is used to predict additional causal links in the causal knowledge graph, Methods that include...

2. Converting the aforementioned causal knowledge graph into an embedding involves using an approach based on a StarE graph neural network, in which a variable number of modifier pairs can be configured to be associated with each of the causal links. The method according to claim 1.

3. The aforementioned translation is carried out in accordance with a causal ontology, which defines a concept for structuring the causal knowledge graph. The method according to claim 1.

4. The above translation is, Mapping nodes in the causal network to causal entities in the causal knowledge graph, Mapping edges within the causal network to causal links within the causal knowledge graph, The method according to claim 1, further comprising:

5. The modifier entity includes an intermediary entity in series between the cause entity and the result entity such that the cause entity causes the result entity and / or the cause entity explains the result entity. The method according to claim 4.

6. The aforementioned hyper-relational causal link includes modifier relationships, the modifier relationships including one or more hasMediator modifier relationships indicating the presence or absence of mediation, and / or hasMediatorType modifier relationships indicating the type of mediation. The method according to claim 1.

7. The aforementioned causal discovery includes a causal explanation for predicting the type of causal entity given the result entity, The method according to claim 1.

8. The aforementioned causal discovery includes causal prediction for predicting the type of the resulting entity given a causal entity, The method according to claim 1.

9. A system for causal discovery that uses knowledge graph link prediction, The aforementioned system, Translating information from a causal network into a causal knowledge graph, wherein the causal knowledge graph includes a plurality of causal links, each of which includes a causal entity, a causal relationship, and an effect entity, and at least one subset of the causal links includes a hyperrelational causal link, the hyperrelational causal link further connects the causal entity and the effect entity using related modifier entities as mediators, which are mediators in the serial causal connection between the causal entity and the effect entity. The causal knowledge graph is transformed into an embedding that includes the latent vector space representation of the causal knowledge graph, Training the embedding using the causal links and hyperrelational causal links of the causal knowledge graph, To discover causality, the embedding is used to predict additional causal links in the causal knowledge graph, A system comprising one or more hardware computing devices configured to perform the following.

10. Converting the aforementioned causal knowledge graph into an embedding involves using an approach based on a StarE graph neural network, in which a variable number of modifier pairs can be configured to be associated with each of the causal links. The system according to claim 9.

11. The translation of information is carried out in accordance with a causal ontology, which defines a concept for structuring the causal knowledge graph. The system according to claim 9.

12. Translating information is Mapping nodes in the causal network to causal entities in the causal knowledge graph, Mapping edges within the causal network to causal links within the causal knowledge graph, The system according to claim 9, further comprising:

13. The modifier entity includes an intermediary entity in series between the cause entity and the result entity such that the cause entity causes the result entity and / or the cause entity explains the result entity. The system according to claim 12.

14. The aforementioned hyper-relational causal link includes modifier relationships, the modifier relationships including one or more hasMediator modifier relationships indicating the presence or absence of mediation, and / or hasMediatorType modifier relationships indicating the type of mediation. The system according to claim 9.

15. The modifier entity includes an intermediary entity in series between the cause entity and the result entity such that the cause entity causes the result entity and / or the cause entity explains the result entity. The system according to claim 9.

16. A non-temporary computer-readable medium, the non-temporary computer-readable medium includes instructions for causal discovery using knowledge graph link prediction, and the instructions, when executed by one or more computing devices, Translating information from a causal network into a causal knowledge graph, wherein the causal knowledge graph includes a plurality of causal links, each of which includes a causal entity, a causal relationship, and an effect entity, and at least one subset of the causal links includes a hyperrelational causal link, the hyperrelational causal link further connects the causal entity and the effect entity using related modifier entities as mediators, which are mediators in the serial causal connection between the causal entity and the effect entity. The causal knowledge graph is transformed into an embedding that includes the latent vector space representation of the causal knowledge graph, Training the embedding using the causal links and hyperrelational causal links of the causal knowledge graph, To discover causality, the embedding is used to predict additional causal links in the causal knowledge graph, A non-temporary computer-readable medium for causing one or more computing devices to perform operations that include the above.

17. Converting the aforementioned causal knowledge graph into an embedding involves using an approach based on a StarE graph neural network, in which a variable number of modifier pairs can be configured to be associated with each of the causal links. The medium according to claim 16.

18. The translation of information is carried out in accordance with a causal ontology, which defines a concept for structuring the causal knowledge graph. The medium according to claim 16.

19. The aforementioned hyper-relational causal link includes modifier relationships, the modifier relationships including one or more hasMediator modifier relationships indicating the presence or absence of mediation, and / or hasMediatorType modifier relationships indicating the type of mediation. The medium according to claim 16.

20. The modifier entity includes an intermediary entity in series between the cause entity and the result entity such that the cause entity causes the result entity and / or the cause entity explains the result entity. The medium according to claim 16.

Images