Methods, system and computer programs for triggering an event and for training at least one machine learning model

Abstract

Various examples relate to computer-implemented methods, systems, and computer programs for triggering an event and for training at least one machine learning model. The proposed methods, systems and computer programs provide an artificial intelligence-based approach for helping decision-making in various fields, such as healthcare fields, which may be used to help healthcare professionals in diagnosing or treating a disease or condition, or by crime specialists to solve a crime case. A computer-implemented method for triggering an event, the method comprising inputting (130) a representation of a first sequence of knowledge graphs representing a plurality of stages of a case, such as a crime case or medical case, and a representation of a second sequence of actions having been performed in the plurality of stages of the case into at least one machine learning model, wherein the at least one machine learning model is trained to output a predicted next action to perform and a corresponding predicted next stage of the case in response to the representations of the first sequence and the second sequence being input into the machine learning model, and triggering (140) an event based on at least one of the predicted next action to perform and the predicted next stage of the case, wherein the event being triggered comprises at least one of processing, using an algorithm or machine learning model, sensor data, such as camera sensor data, being related to the predicted action, and controlling a device or system, such as a controllable sensor device, medical device, autonomous vehicle, traffic control system or wearable device, based on the predicted action.

Description

[0001] METHODS, SYSTEM AND COMPUTER PROGRAMS FOR TRIGGERING AN EVENT AND FOR TRAINING AT LEAST ONE MACHINE LEARNING MODEL

[0002] The present invention relates to a computer-implemented method, system, and computer program for triggering an event, and to a method, system, and computer program for training at least one machine learning model.

[0003] Investigation Management Unit (IMU) professionals are often burdened with a high case load. IMU professionals play a critical role in handling investigations, gathering evidence, and ensuring the smooth functioning of legal processes. However, the increasing complexity and volume of cases, coupled with limited resources, have resulted in an overwhelming workload for these professionals. This workload may negatively impact their efficiency, productivity, and overall performance.

[0004] Existing recommendation systems focus on simple or one-dimensional recommendations, i.e., they focus on preferences and properties to provide a recommendation such as a product or movie.

[0005] There may be a desire to ensure the effectiveness of the IMU and the timely resolution of investigations, by developing approaches to alleviate the burden on IMU professionals and improve their workflows.

[0006] It is therefore an objective of the present invention to improve and further develop a method to systematically support users, such as IMU professionals, to help them in their process.

[0007] This objective is addressed by the subject-matter of the independent claims.

[0008] In accordance with the invention, the aforementioned objectives are accomplished by a computer-implemented method for triggering an event comprising the features of claim 1. According to this claim, such a method comprises inputting a representation of a first sequence of knowledge graphs representing a plurality of stages of a case, such as a crime case or medical case, and a representation of a second sequence of actions having been performed in the plurality of stages of the case into at least one machine learning model. The at least one machine learning model is trained to output a predicted next action to perform and a corresponding predicted next stage of the case in response to the representations of the first sequence and the second sequence being input into the machine learning model. The method comprises triggering an event based on at least one of the predicted next action to perform and the predicted next stage of the case. Preferably, the event being triggered comprises at least one of processing, using an algorithm or machine learning model, sensor data, such as camera sensor data, being related to the predicted action, and controlling a device or system, such as a controllable sensor device, medical device, autonomous vehicle, traffic control system or wearable device, based on the predicted action.

[0009] Artificial Intelligence (Al) can play a pivotal role in alleviating the burden on overloaded Investigation Management Unit (IMU) professionals. By automating certain tasks and providing intelligent assistance, Al technology can significantly improve the workflow and streamline the investigation process. In particular, Al can be instrumental in assisting IMU professionals in creating action plans and providing recommendations for new cases, such as medical cases or crime cases. In the proposed concept, this is achieved through at least one machine learning model that is being tasked with predicting a next action to perform based on the case at hand, i.e. , based on a representation of the case at hand, and based on the actions that have previously been performed as part of the case. By predicting the next action to be performed, and then preferably automatically triggering an event that is associated with this action, the burden of the IMU professionals can be lowered, as tedious tasks have already been performed without involvement of the IMU professional. Thus, the method may be used to automatically assist users, such as IMU professionals, in an investigative process. This can lead to crime cases being solved faster and to medical cases being diagnosed or handled in an evidencebased manner, based on actions taken in previous cases.

[0010] Based on the data analysis, case similarity, and risk assessment, Al can generate an initial action plan for IMU professionals. For example, in the context of crime investigation, this plan may include recommended investigative techniques, evidence gathering strategies, interview or interrogation approaches, and potential leads to pursue. The Al system can continuously learn and improve its recommendations by incorporating feedback from IMU professionals and the outcomes of previous cases.

[0011] Depending on the specific context, an Action Plan Generation Tool can integrate with various real-world systems to trigger specific actions. For example, in emergency response scenarios, in triggering the event, the tool can connect with alarm systems, dispatch emergency services, or activate automated alerts. These integrations enable the tool to have a direct impact on the real world by initiating actions that mitigate risks or address critical situations.

[0012] Embodiments of the present disclosure address the challenge of analyzing / handling big data in a case management system to support specialists in making faster and consistent decisions, and by automatically performing actions to alleviate the burden of the specialists. Embodiments of the present disclosure provide action recommendations for case management systems by exploiting how the relationships between actions develop over time. In addition, the actions can automatically be performed, to further alleviate the burden of the specialists.

[0013] Other recommendation systems focus usually on simple or one-dimensional recommendations, i.e., they focus on preferences and properties to provide a recommendation such as a product or movie. This is different to the proposed concept, as in the proposed use case, the time aspect and the chronological sequence of actions is considered to make a suitable action recommendation. Compared to sequential recommendation or sequential path reasoning systems which focus on a single (knowledge) graph, the proposed concept faces the challenge of case dependent knowledge graphs. The proposed concept determines a suitable action recommendation by analyzing a set of knowledge graphs and temporal knowledge graphs.

[0014] For example, the at least one machine learning model may comprise at least one transformer model or transformer component. Transformer-based machine learning models have proven to be proficient at predicting a next item in a sequence of items, with the attention mechanism of the respective transformer being useful for determining, which aspects of the input are relevant for the prediction. Thus, the predictions made by the at least one machine learning model become better at identifying which aspects of the case at hand are most relevant for the purpose of predicting the next action to perform.

[0015] In the proposed concept, two inputs are used for the at least one machine learning model - the case at hand, and the actions that have already been performed, with the machine learning model being used to determine the action that is most useful to be performed next. As actions need to be seen in the context of a specific case, the prediction of the next action to be performed is also dependent on the respective case context. In the present disclosure, this can be handled by predicting both the next action to be performed and the next stage of the case. As the next stage of the case depends on the action to be performed, the prediction of the next stage of the case takes may take into account the action that is predicted to be performed next. For example, the at least one machine learning model may comprise an action prediction component trained to output the next action to perform based on the representation second sequence of actions, and a case prediction component trained to output the predicted next stage of the case based on the representation of the first sequence of actions and based on the output of the action prediction component. As the predicted next stage of the case can be evaluated, e.g., with respect to how likely it is that this stage is beneficial to resolving the case, the case prediction component can be used, in combination with a suitable evaluation mechanism, to determine how useful the predicted next action is. If the action is not deemed to be useful according to the evaluation mechanism, a different action can be picked.

[0016] The interrelationship between the predicted next action and the predicted next stage of the case can already be used during training of the respective component. For example, the action prediction component and the case prediction component may be trained such, that a loss function used for training the case prediction component impacts the training of the case prediction component, as the case prediction component is trained to output the predicted next stage of the case based on the output of the action prediction component. This way, backpropagation can be used to train the action prediction component to predict actions that are useful in the context of the respective case.

[0017] According to an example, the method may comprise encoding, using a second machine learning model, the first sequence of knowledge graphs representing the plurality of stages of the case into a sequence of embeddings representing the sequence of knowledge graphs. For example, the second machine learning model may be trained to encode each knowledge graph of the sequence of knowledge graphs into a corresponding embedding representing the knowledge graph. For example, the first sequence of knowledge graphs may be input as sequence of embeddings into the at least one machine learning model. By encoding the first sequence of knowledge graphs into a corresponding sequence of embeddings, a sequence of embeddings is created that is suitable for predicting an embedding that represents the case with a subsequent stage, and thus an additional knowledge graph in the sequence of knowledge graphs. This sequence of embeddings can be used, by the case prediction component, to predict the next stage of the case.

[0018] For example, the sequence of knowledge graphs may represent a temporal knowledge graph. This way, the development of a case through the stages can be represented in a way that allows a prediction of the next stage of the case.

[0019] A similar encoding process may be performed on the sequence of actions. For example, the method may comprise encoding, using a third machine learning model, the second sequence of actions having been performed in the plurality of stages of the case into a sequence of embeddings representing the actions having been performed in the plurality of stages of the case. The second sequence of actions may be input into the at least one machine learning model as sequence of embeddings. By encoding the second sequence of actions into a corresponding sequence of embeddings, a sequence of embeddings is created that is suitable for predicting an embedding that is likely to result from the sequence of embeddings, and thus next action to be taken. This sequence of embeddings can be used, by the action prediction component, to predict the next action. In machine learning, ensuring a wide range of different training samples in the training data is beneficial to the quality of the trained machine learning model. Therefore, the training data might not only contain positive samples, but also negative samples. For example, the third machine learning model may be trained using a first set of positive training samples representing sequences of actions having been performed in previous cases and using a second set of negative training samples representing illogical sequences of actions. Such negative samples are often not readily available, as illogical sequences of actions are generally not performed in real life. In some examples of the present disclosures, these negative samples may be generated using a data augmentation technique. In particular, the second set of negative training samples may be derived from the first set of positive training samples, for example by substituting actions of positive samples in a stepwise and / or backward manner using a language model, such as a Large Language Model (LLM). By substituting actions in the sequence of actions contained in positive examples with the help of an LLM, suitable negative samples can be generated. For example, by using a language model to generate the negative samples, the bias embodied within the existing training data can be avoided. By operating backwards, the latest action of the sequence of actions can be kept, and only actions occurring before the final action might be replaced. By using a stepwise approach, consistency can be maintained within the sequence of actions.

[0020] In general, transformers are trained to provide an output token in response to being fed input tokens. As there is a wide variety of case stages and actions to be performed, with a non-obvious relationship between each other, the cases and actions may be represented as embeddings, fed as input into the respective transformers, resulting in embeddings being output by the respective transformers. These embeddings may then be re-converted into the respective case and action representations. In other words, the at least one machine learning model may be trained to output at least one of a first embedding representing the predicted next stage of the case and a second embedding representing the predicted next action to perform. The method may comprise at least one of retrieving the predicted next stage of the case from a first data storage using the first embedding and retrieving the predicted next action to perform from a second data storage using the second embedding. Thus, after predicting the respective action to be taken and next stage of the case as respective embeddings, the respective action to be taken and next stage of the case are looked up so they can be evaluated with respect to their solvability score and / or or used to trigger an event.

[0021] These data storages may be populated from the results of the training of the at least one machine learning model. For example, information stored in the first data storage may be based on a training output of a training of a second machine learning model trained to output an embedding for a given knowledge graph and / or information stored in the second data storage may be based on a training output of a training of a third machine learning model trained to output an embedding for a given sequence of actions. This way, the lookup of the actions or next stages is in line with the training of the encoders being used to encode the respective sequences of actions and knowledge graphs.

[0022] The proposed concept is based on machine learning. Machine learning is a branch of artificial intelligence that involves the development of algorithms and models that allow computers to learn and make predictions or decisions without being explicitly programmed. It focuses on creating systems that can improve their performance over time by learning from data.

[0023] Training a machine learning model refers to the process of teaching the model to make accurate predictions or decisions. During training, the model is exposed to a large amount of data, which is used to adjust the model's internal parameters or weights. The model learns patterns, relationships, or rules from the training data, allowing it to generalize and make predictions on new, unseen data.

[0024] Training data is the set of examples or instances that is used to teach a machine learning model. It is often labeled data, meaning that each example is associated with a known outcome or target value. The training data consists of both input features and the corresponding output or target variable. The model learns from this data by analyzing the patterns and relationships between the input features and the target variable. Training algorithms, such as supervised learning, semi-supervised learning, unsupervised learning, or reinforcement learning may be used for training the machine learning model. Machine learning models, such as the machine learning model being trained in the present disclosure, are often implemented as Artificial Neural Networks (ANNs), and in particular Deep Neural Networks, Support Vector Machines, Decision Tree models, or Random Forest models.

[0025] In the following, a computer-implemented method is provided for training the at least one machine learning model being used by the method for triggering the event. The method comprises obtaining or generating training data comprising a plurality of training samples. Each training sample comprises a representation of a first sequence of knowledge graphs representing a plurality of stages of a case and a representation of a second sequence of actions having been performed in the plurality of stages of the case. The method comprises training, e.g., using at least one cross-entropy loss function, the at least one machine learning model, using the training data, to output a predicted next action to perform and a corresponding predicted next stage of the case in response to the first sequence and the second sequence being input into the at least one machine learning model. By training the machine learning model to predict the next action to be performed (with the prediction being used to automatically trigger an event that is associated with this action), a machine learning model (or models) can be created that can be used to lower the burden of the IMU professionals, as tedious tasks can be performed, according to the prediction provided by the machine learning model, without involvement of the IMU professional. This can lead to crime cases being solved faster and to medical cases being diagnosed or handled in an evidence-based manner, based on actions taken in previous cases.

[0026] In some examples, the at least one machine learning model may comprise at least one transformer model or transformer component. As outlined above, transformerbased machine learning models have proven to be proficient as predicting a next item in a sequence of items, with the attention mechanism of the respective transformer being useful for determining, which aspects of the input are relevant for the prediction. Thus, the predictions made by the at least one machine learning model become better at identifying which aspects of the case at hand are most relevant for the purpose of predicting the next action to perform. As outlined above, the at least one machine learning model may comprise an action prediction component and a case prediction component, which are linked together as the case prediction component uses the output of the action prediction component. Therefore, also the training of the two components may be linked together. For example, training the at least one machine learning model may comprise training, in a joint manner and using the training data, the action prediction component to output the next action to perform based on the representation second sequence of actions, and (training) the case prediction component to output the predicted next stage of the case based on the representation of the first sequence of actions and based on the output of the action prediction component. In this context, the joint training refers the action prediction component and the case prediction component being trained together, with the output of the action prediction component being input into the case prediction component during training. In particular, the action prediction component and the case prediction component may be trained such, that a loss function used for training the case prediction component impacts the training of the case prediction component, as the case prediction component is trained to output the predicted next stage of the case based on the output of the action prediction component. Thus, not only is the action prediction component trained to output an action that is likely to occur, but it is also trained to output an action that is likely to occur and that is beneficial to the solvability of the case at hand.

[0027] In general, the transformers may be provided with the cases and actions as sequences of embeddings, for the reasons listed above. Therefore, as part of the training pipeline, the sequence of knowledge graphs and the sequence of actions may be encoded using encoder models. For example, generating the training data may comprise encoding, using a second machine learning model, the first sequence of knowledge graphs representing the plurality of stages of the case into a sequence of embeddings representing the sequence of knowledge graphs. In addition to the training of the transformer components, the encoder model(s) may also be trained, e.g., in an end-to-end manner together with the transformer components. For example, the method may comprise training the second machine learning model to output an embedding of a knowledge graph for a knowledge graph being input into the second machine learning model. For example, the second machine learning model may be trained together with the at least one machine learning model in an end-to-end manner. This way, the second machine learning model may be trained to output a sequence of embeddings that enable the case prediction component to predict the next stage of the case.

[0028] Similar pre-processing may be applied to the sequence of actions. For example, generating the training data may comprise encoding, using a third machine learning model, the second sequence of actions having been performed in the plurality of stages of the case into a sequence of embeddings representing the actions having been performed in the plurality of stages of the case. In addition to the training of the transformer components, this encoder model may also be trained, e.g., in an end-to-end manner together with the action prediction component. In other words, the method may comprise training the third machine learning model to output an embedding representing an action in the context of preceding actions for a sequence of actions being input into the third machine learning model.

[0029] In machine learning, ensuring a wide range of different training samples in the training data is beneficial to the quality of the trained machine learning model. Therefore, the training data might not only contain positive samples, but also negative samples. Therefore, the third machine learning model may be trained, e.g., trained together with the at least one machine learning model in an end-to-end manner, using a first set of positive training samples representing sequences of actions having been performed in previous cases, and using a second set of negative training samples representing illogical sequences of actions. Such negative samples are often not readily available, as illogical sequences of actions are generally not performed in real life. In some examples of the present disclosures, these negative samples may be generated using a data augmentation technique. Accordingly, the method may comprise deriving the second set of negative training samples from the first set of positive training samples by substituting actions of positive samples in a stepwise and / or backward manner using a language model. For example, this may be done by assuming the last action of the respective sequence of actions as fixed. Starting from the penultimate action, a language model may be used to propose an alternate action to precede the last action and replace the penultimate action. Subsequently, the action preceding the penultimate action may be replaced by an alternative action also proposed by the language model. Thus, while retaining the last action, the language model may be used to propose alternative actions to replace the preceding actions, by traversing the sequence actions backwards (i.e., starting from the last action) and in a stepwise manner and replacing the actions preceding the last action based on the proposals of the language model. This way, consistency within the sequence of actions may be retained, while removing the bias inherent in the positive training samples.

[0030] Another aspect of the present disclosure relates to a system comprising one or more processors and one or more storage devices. For example, the system may be configured to perform both or either of the above methods.

[0031] Another aspect of the present disclosure relates to a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out at least one of the above methods.

[0032] Another aspect of the present disclosure relates to a non-transitory, computer- readable medium comprising a program code that, when the program code is executed on a processor, a computer, or a programmable hardware component, causes the processor, computer, or programmable hardware component to perform at least one of the above methods.

[0033] There are several ways how to design and further develop the teaching of the present invention in an advantageous way. To this end it is to be referred to the dependent claims on the one hand and to the following explanation of preferred embodiments of the invention by way of example, illustrated by the figure on the other hand. In connection with the explanation of the preferred embodiments of the invention by the aid of the figure, generally preferred embodiments and further developments of the teaching will be explained. In the drawing

[0034] Fig. 1 a shows a flow chart of an example of a computer-implemented method for triggering an event; Fig. 1 b shows a block diagram of an example of a system for triggering an event;

[0035] Fig. 2a shows a flow chart of an example of a computer-implemented method for training at least one machine learning model;

[0036] Fig. 2b shows a block diagram of an example of a system for training a machine learning model;

[0037] Fig. 3 shows a schematic diagram of an example of possible components of a system for triggering an event;

[0038] Fig. 4 shows a schematic diagram of examples of inputs and outputs of an encoder Al model;

[0039] Fig. 5 shows a schematic diagram of an example of a negative sampling Al agent; and

[0040] Fig. 6 shows a schematic diagram of an example of an action recommendation model.

[0041] In the following, two methods (and corresponding systems and computer programs) are introduced. The first of the two methods, shown in Fig. 1a, can be used to predict an action to be performed next in a case, such as a crime case, or a health management case (e.g., a case related to diagnosing a disease, or a case related to mid- or long-term treatment of a disease or management of a condition). Based on the prediction action, an event is triggered that is associated with the predicted action, e.g., to perform the action on behalf of an IMU specialist, or to perform a preparatory task that helps the IMU specialist in performing the action. The method of Fig. 1a uses at least one machine learning model to predict the action to be performed next. For example, the method of Fig. 2a may be used to train this at least one machine learning model. Fig. 1a shows a flow chart of an example of a computer-implemented method for triggering an event. The method comprises inputting 130 a representation of a first sequence of knowledge graphs representing a plurality of stages of a case, such as a crime case or medical case, and a representation of a second sequence of actions having been performed in the plurality of stages of the case into at least one machine learning model. The at least one machine learning model is trained, e.g., using the method of Fig. 2a, to output a predicted next action to perform and a corresponding predicted next stage of the case in response to the representations of the first sequence and the second sequence being input into the machine learning model. The method comprises triggering 140 an event based on at least one of the predicted next action to perform and the predicted next stage of the case. In the context of the present disclosure, the event is related to the predicted next action - in some cases, it may include the action (if the action can be performed automatically). In some other cases, the event may comprise one or more preparatory operations that aim to prepare an action being performed by a human operator.

[0042] This method may be used in a variety of different use cases, such as for automatically triggering actions in a crime case or a medical case. For example, the event being triggered comprises at least one of processing, using an algorithm or machine learning model, sensor data, such as camera sensor data, being related to the predicted action. Accordingly, the method may comprise processing, using an algorithm or machine learning model, sensor data being related to the predicted action. For example, in case the case is a crime case, the sensor data may be sensor data of a surveillance camera or traffic camera, with the sensor data being processed to automatically identify a vehicle registration number (e.g., license plate) or person (e.g., using a face recognition model or a face recognition database). For example, in case the case is a health case, the sensor data may be sensor data of a medical device, such as a microscope, an electrocardiogram (ECG), a computer tomography machine, an x-ray machine etc., which may be processed, using a trained machine learning model, to detect indications of a disease or of a treatment being (un-)successful. In the proposed concept, the output of the trained machine learning model is used to trigger an event, and in particular an action associated with the event, in the real world. Similarly, the knowledge graphs being used as input for the trained machine learning model represents the real world. For example, if the case is a medical case, the knowledge graph may represent the physical properties of the patient, e.g., the diagnostic tests that have been performed in the past, known pre-conditions, physical characteristics (age, weight, allergies etc.). Similarly, if the case is a crime case, the knowledge graph may also represent one or more physical properties of the case at hand, e.g., a blood sample, a piece of evidence, the scene of the crime etc.

[0043] Additionally, or alternatively, the event being triggered may comprise at least one of controlling a device or system, such as a controllable sensor device, medical device, autonomous vehicle, traffic control system or wearable device, based on the predicted action. For example, if the case is a crime case, if the action being predicted is surveillance of an area near a crime scene, a Pan-Tilt-Zoom camera may be controlled to adjust its field of view to the desired area. If the action is to search for evidence, weapons or suspect in an area, an autonomous vehicle, such as an autonomous aerial vehicle (i.e., an autonomous drone) or a robot may be instructed to fly, walk, or drive through the desired area and record image or video data of the area. For example, if the case is a health case, if the action is to adjust a dose of a medication, a medical device (such as a dispenser) may be controlled to adjust the dose given to the patient. If the action is to adjust the O2 concentration of the gas being used to ventilate a patient (e.g., for weaning the patient of a ventilator), a medical ventilator may be controlled to adjust the O2 concentration. For example, if the medical data available on a patient is insufficient for making a diagnosis, the action may be to increase the amount of medical data, and a wearable device, such as blood pressure monitor, glucose monitor, heart rate monitor etc. may be controlled to increase the number of measurements per unit of time. A lookup table or database may be used to determine how to implement the respective action. In other words, triggering the event may comprise performing a look-up in a look-up table or database to determine how to implement the respective predicted action, and performing or triggering performance of the actions according to the look-up. Fig. 1 b shows a block diagram of an example of a system 100, such as a computer system, for triggering the event. The system 100 comprises an optional interface 102 and one or more processors 104 to perform the method of Fig. 1 a. In some examples, the system 100 may further perform the method of Fig. 2a. Features introduced in connection with the methods of Fig. 1 a and / or 2a, and with respect to Figs. 3 to 6, may likewise be performed by the system 100 of Fig. 1 b. For example, the system 100 may comprise machine-readable instructions for providing the functionality of the method, with the one or more processors 104 executing the machine-readable instructions to perform the method of Fig. 1 a. The one or more processors 104 are coupled with the interface 102 and with memory and / or storage, such as one or more storage devices 106. For example, the one or more processors 104 may be configured to provide the functionality of the system 100, e.g., in conjunction with the interface 102 (for exchanging information) and / or with the memory / storage 106 (for storing information).

[0044] As outlined above, the method and system of Figs. 1 a and 1 b rely on at least one machine learning model. In general, training of this machine learning model may be performed according to the method of Fig. 2a, and / or using the system of Fig. 2b. Fig. 2a shows a flow chart of an example of a computer-implemented method for training the at least one machine learning model. The method comprises obtaining I generating 210 training data comprising a plurality of training sample. Each training sample comprises a representation of a first sequence of knowledge graphs representing a plurality of stages of a case and a representation of a second sequence of actions having been performed in the plurality of stages of the case. The method comprises training 220 the at least one machine learning model, using the training data, to output a predicted next action to perform and a corresponding predicted next stage of the case in response to the first sequence and the second sequence being input into the at least one machine learning model.

[0045] Fig. 2b shows a block diagram of an example of a system 200, such as a computer system, for training the at least one machine learning model. The system 200 comprises an optional interface 202 and one or more processors 204 to perform the method of Fig. 2a. In some examples, the system 200 may further perform the method of Fig. 1 a. Features introduced in connection with the methods of Fig. 2a and / or 1a, and with respect to Figs. 3 to 6, may likewise be performed by the system 200 of Fig. 2b. For example, the system 200 may comprise machine-readable instructions for providing the functionality of the method, with the one or more processors 204 executing the machine-readable instructions to perform the method of Fig. 2a. The one or more processors 204 are coupled with the interface 202 and with memory and / or storage, such as one or more storage devices 206. For example, the one or more processors 204 may be configured to provide the functionality of the system 200, e.g., in conjunction with the interface 202 (for exchanging information) and / or with the memory / storage 206 (for storing information).

[0046] For example, the interface 102 or interface 202 of Figs. 1 b and / or 2b may include or correspond to a network interface circuitry and / or a device interface circuitry configured to be communicatively coupled to one or more other devices, such as the one or more processors 104 or 204 of Figs. 1 b and / or 2b. For example, the interface 102 or 202 of Figs. 1 b and / or 2b may include a transmitter, a receiver, or a combination thereof (e.g., a transceiver), and may enable wired communication, wireless communication, or a combination thereof. For example, the one or more processors 104 or 204 of Figs. 1 b and / or 2b may include or correspond to a digital signal processor circuitry (DSP), a graphical processing unit (GPU), and / or a central processing unit (CPU). The one or more processors 104 or 204 of Figs. 1 b and / or 2b may be coupled to the memory / storage circuitry(s) 106 or 206 of Figs. 1 b and / or 2b. The memory / storage circuitry 106 or 206 of Figs. 1 b and / or 2b may include instructions (e.g., executable instructions), such as computer-readable instructions or processor circuitry-readable instructions. The instructions may include one or more instructions that are executable by a computer, such as by the one or more processors 104 or 204 of Figs. 1 b and / or 2b. For example, the memory / storage circuitry 106 or 206 of Figs. 1 b and / or 2b may include or correspond to volatile or nonvolatile storage circuitry, such as Random Access Memory (RAM), magnetic disks, optical disks, or flash memory devices. The memory / storage circuitry 106 or 206 of Figs. 1 b and / or 2b may include both removable and non-removable memory devices. The methods of Figs. 1a and 2a, as well as the corresponding systems 100, 200 and corresponding computer programs, will now be introduced in more detail with respect to an application of the proposed methods on crime cases. However, the more general methods, systems and computer programs may likewise be applied to other applications or use cases, such as medical use cases. Features introduced in connection with the “crime case” embodiment shown in connection with Figs. 3 to 6 may likewise be applied, in a more general manner, on arbitrary types of cases.

[0047] Embodiments of the present disclosure relate to an Al-based approach to support professionals during their daily routine when assessing new (crime) cases. The scope and focus of the proposed concept are to support human decision makers in processing a huge amount of information to make consistent and faster decisions. In particular, it focuses on the aspect of action plan generation. Based on the action plan, an event is triggered that is associated with an action predicted for the action plan. Optionally, the approach may include triggering actions automatically as part of triggering the event.

[0048] In the following, as a running example, the focus is on crime cases. Investigators have to assign iteratively new actions to a crime case, such as “review CCTV (Closed Circuit Television) footage” and depending on the outcome, e.g., “search car number plate XY”. Similar scenarios exist in other domains such as in the health care domain (e.g., assigning treatments to patients) or in the information technology domain (e.g., recommend actions to customers to handle / solve their issues).

[0049] Embodiments of the present disclosure connect to the (Crime) Case Management System, the Solvability Assessment System (for determining a probability of having a positive outcome at the current stage), and, if available, additional databases (e.g., publicly available information). Based on this information, the system disclosed herein may provide iteratively the next best fitting action. The proposed concept can be embedded in or connected to, e.g., loT (Internet of Things) devices such as (wearable) cameras, sensors, and traffic signal control systems, or drones. In the following, the individual components of the proposed system are described, including its insights and technical details. According to an example, schematically illustrated in Fig. 3, which shows a schematic diagram of an example of possible components of a system for triggering an event, the system comprises or consists of several different components which can be grouped into three processes: The first process learns representations of the potential actions (1 -4), starting from a generated action knowledge graph 1 , an encoder with an Al model 2 (e.g., a second machine learning model) and a negative sampling Al agent 3, resulting in an action embedding 4. The second process learns a representation of the crime cases at different stages (5-7, 13), starting from the crime case temporal knowledge graph(s), e.g., a set, 5, using an encoder with an Al model 6 (e.g., a third machine learning model), to obtain a list of sequences 7, which may be evaluated using a solvability score engine 13. The last process takes the input of the first two processes to learn the mapping from the current state of a case to the most valuable action (8-12), by using at least one transformer 8 to generate inputs for an action decoder 9 and a crime case decoder 10, with the action decoder 9 outputting an action 11 and the crime case decoder 10 outputting a score 12.

[0050] These components are introduced in detail in the following, starting with block 1 : Action Knowledge Graph Generation. In an Action Knowledge Graph, entities are actions like "analyze CCTV footage", "look for witnesses", or other actions an Investigation Management Unit professional could assign to a crime case. Two nodes are connected if there is the chance that those two actions can be executed in sequence. For example, if it is possible that after "analyze CCTV footage" the action "look for witnesses" is performed, then there is a connection. The connection may have a probability weight, reflecting how likely it is that the action is a follow-up action. The probabilities may be derived from historical data.

[0051] Thus, if historical data suggests that after analyzing CCTV footage, the next logical action is usually to look for witnesses, an edge may connect the nodes "analyze CCTV footage" and "look for witnesses" with a probability weight of, for example, 0.8. This weight indicates an 80% chance of the action "look for witnesses" being performed after "analyze CCTV footage."

[0052] For example, the action knowledge graph may be a representation of the second sequence of actions having been performed in the plurality of stages of the case. The Action Knowledge Graph may be constructed based on the historical action protocols of a crime case, with an action knowledge graph representing one crime case and a crime case being represented by one (e.g., a single) action knowledge graph. In this case, the edges of the graph might have no weight and the relations indicate just the sequence of the actions. For each of these graphs, the nodes may be labeled with the result of the particular action (i.e. , whether there was a positive, negative, or neutral outcome). Then, the graphs that describe the same crime type (e.g., burglary) may be merged. This process may assign the mentioned probabilities to the edges. In the following, two examples are given:

[0053] For example, the following triple (analyze CCTV footage, look for witnesses, 0.3) indicates that there is an 30% probability that after analyzing the CCTV footage, the next action would be to look for witnesses. The triple (analyze CCTV footage, review forensic evidence, 0.4) indicates that there is a 40% probability that after analyzing the CCTV footage, the next action would be to review forensic evidence. In this context, a triple represents an edge between a first action (analyze CCTV footage) as first entry of the triple, a second action (look for witnesses or review forensic evidence) as second entry of the triple, and the probability / weight of the edge as third entry of the triple.

[0054] It may be beneficial to consider that those probabilities are local as they were computed just based on the direct predecessor. It can be expected that the probabilities have a strong variation as the entire set of preceding actions influence the probability of the next action. For that reason, the Al model 2 may take both (the Action Knowledge Graphs with probabilities, and the list of crime case specific Action Knowledge Graphs) as input.

[0055] Block 2: Encoder: Al Model. In the context of the present disclosure, the Al model of block 2 may also be denoted “third machine learning model”. For example, the methods of Figs. 1 a and 2a may comprise encoding, using the third machine learning model, the second sequence of actions having been performed in the plurality of stages of the case into a sequence of embeddings representing the actions having been performed in the plurality of stages of the case. This second sequence of actions, i.e., the output of the Al model of the encoder, may be input into the at least one machine learning model as sequence of embeddings. The encoder model may thus take as input a list of Action Knowledge Graphs (one for each crime case, in the following denoted sample graph) and the merged Action Knowledge Graphs (with the probabilities; one for each crime type, in the following denoted static graph). In other words, in addition to an individual sequence of actions at hand, a merged sequence of actions may be input into the third machine learning model (e.g., as knowledge graph). This model may also be used to encode the training data in preparation of or during training. Thus, in the method of Fig. 2a, generating 210 the training data may comprise encoding 214, using the third machine learning model, the second sequence of actions having been performed in the plurality of stages of the case into a sequence of embeddings representing the actions having been performed in the plurality of stages of the case. The model (i.e., the transformer model, illustrated in more detail in Fig. 6) may be implemented as a sequence to vector model, as it predicts the next action (as vector representation) for a given sequence of actions (modeled as a graph).

[0056] Fig. 4 outlines the concept and architecture of the encoder model 2 according to an example. The encoder model takes as input a sequence of actions (e.g., the second sequence of actions) having been performed (e.g., in a crime case), and outputs an action embedding that represents the next action to be performed. The actions are input as part of a sequence actions, thereby providing not only the latest action having been performed, but also the action(s) having preceded the latest action (if available). For example, in Fig. 4, for time ti , the encoder model may output an action embedding representing the action “review CCTV”. For time t2, the encoder model may output an action embedding representing the sequence of actions “review CCTV” and “search car” etc.

[0057] To implement the functionality of the encoder Al model 2, the encoder Al model 2 (i.e., the third machine learning model) may be trained using machine learning. For example, the method of Fig. 2a may comprise training 224 the third machine learning model to output an embedding representing an action in the context of preceding actions for a sequence of actions being input into the third machine learning model. For example, the third machine learning model may be trained using a plurality of training samples, with each training sample comprising a sequence of actions (and optionally a merged sequence of actions with weighted edges) as desired input data. Goal of the training is to train the encoder Al model 2 (i.e. , the third machine learning model) such, that the embeddings generated by the encoder Al model 2 are a sequence- and context-dependent representation of the respective (final) action of the sequence of actions. This property becomes apparent when the embedding is being used by the transform er(s) (see block 8 in Fig. 3 and the “Action Transformer” and “Case Transformer” in Fig. 6). Thus, during training of the third machine learning model, the quality of the output of the transformers may be taken into account for calculating the loss function of the training of the third machine learning model. Thus, the third machine learning model may be trained together with the at least one machine learning model (comprising the transformers) in an end-to- end manner. In particular, a loss function being used for training the third machine learning model being based on the output of the action transformer and case transformer.

[0058] The architecture shows a novel aspect: By learning a sequence- and context- dependent representation for each potential action, the Al model is enabled to learn the consequences of actions. This ensures that a learned representation is always characterized by its preceding actions, i.e., the latest action is a consequence of preceding actions. In other words, the representation of a particular node may be derived from a sequence of graph representations, representing the different stages of a crime case.

[0059] Compared to other memory-based approaches (e.g., “conversational memory” as employed by the langchain project or “recurrent neural networks”), the main difference is that the proposed concept is specifically designed for learning stepwise sequence- and context-dependent embedding representations. The mentioned memory-based approaches provide in general a memory function (e.g., to help LLMs remember what they have already answered) but do not capture that the representation of current activity is essentially constructed from the representations of the preceding actions (i.e., embeddings).

[0060] To improve the output of the Al model, a Negative Sampling Al Agent 3 (which is only used during training) may be used to provide counter examples to counteract the bias in the data. The Negative Sampling Al Agent 3 may be configured to provide input to the previous component (the encoder Al model 2). In particular, the negative samples may be generated on-demand when the corresponding training batch is constructed. In the context of training an Al model, negative samples are examples or instances that do not belong to the target class or category that the model is being trained to recognize or classify. These samples represent the "negative" or "nontarget" class, contrasting with the "positive" or "target" class samples. By including negative samples during the training process, the model learns to differentiate between the target class and other potential classes or categories. This helps the model develop a more comprehensive understanding of the features and characteristics that define the target class, improving its ability to accurately classify and make predictions. Thus, the third machine learning model may be trained using a first set of positive training samples representing sequences of actions having been performed in previous cases and using a second set of negative training samples representing illogical sequences of actions. As will be shown in the following, the second set of negative training samples may be derived from the first set of positive training samples.

[0061] This means essentially for the running example that the triple may describe a transition between two activities which does not exist in our training data. To counteract this bias, a large language model (LLM) may be leveraged. The LLM may represent an additional and independent data source that may be used to ensure diversity. It may be considered that some models are bad at following negative instructions (or negations), i.e., the prompt to generate the negative samples may be of the same type and not a counter example. For example, when instructing the LLM with the following prompt “In your response, do not use words that start with the letter "a". How is the weather?”, a first LLM may respond with “The weather is quite pleasant today. The sun is shining, and ...” and a second LLM may respond with “The weather is sunny and ... ”.

[0062] Obviously, these answers do not adhere to the prompt, as the goal is to generate sequences excluding certain actions or subsequences. This leads to a further novel aspect: By substituting stepwise and backwards the performed actions with a LLM, the common knowledge of the LLM is leveraged to turn the positive sample into a reasonable, non-trivial negative sample. Thus, the method of Fig. 2a may comprise deriving the second set of negative training samples from the first set of positive training samples by substituting actions of positive samples in a stepwise and / or backward manner using a language model. This may ensure that negative samples are generated that are not biased by the historical data. This can be illustrated with the help of Fig. 5. In Fig. 5, the positive sample comprises the actions review CCTV (which has a positive label, as it leads to a license plate), search car (which as a positive label, as the car has been found) and interview (which has a negative label, as the wrong person has been found). Using stepwise and backward substitution, the last action (interview) is retained, and the penultimate action (search car) has been substituted with “search person” (see the first positive sample on the top right of Fig. 5). This can be done by prompting the language model to propose an action that would lead the police to interviewing a person. This operation is repeated for the action preceding “search car” (which is now “search person”). Again, the language model may be prompted to propose an action that would leads the police to search a person. Such an action is, for example, the action “analyze fingerprints”. This way, a language model, such as an LLM, can be leveraged to derive negative samples from positive samples.

[0063] Fig. 5 illustrates an example of a concept to derive several negative samples from one positive sample. The given sequence of actions may be provided to the LLM along an action list with descriptions. The action list may define and provide the scope of the substitution. Using the LLM instead of just historical information allows to avoid bias in the historical data. The substation process may aim to replace some actions with alternative actions while ensuring the consistency of the sequence. Consistency refers to the probability and authenticity in respect of the preceding and subsequent actions. The substitution process may operate backwards as it may aim to reach the latest action of the sequence. A negative sample in our use case may be a sequence leading to the same final action as the positive sample but with a different sequence of actions. This is because the action is defined by the preceding actions, hence, as described, the same action can have different representations if the trigger is different. Block 4: Action Embedding may be a database storing the output of the Al Encoder Module 2. It may be used as a look up for the subsequent operation. For example, in the context of the present disclosure, the Action Embedding block 4 may also be referred to as second data storage, which is based on a training output of the training of a third machine learning model (the encoder Al model 2). For example, an action embedding (i.e., the output of the encoder Al model 2 or third machine learning model) may be a n-dimensional numerical vector.

[0064] Block 5: Crime Case TKG (Set). According to an example, to model a set of crime cases as temporal knowledge graphs, an ordered sequence of knowledge graphs (i.e., the first sequence of knowledge graphs representing the plurality of stages of the case) is created that capture the evolving nature of the cases over time. In particular, the nodes and edges (of the respective knowledge graphs) may have the following characteristics: Entities (Nodes): In the “crime case” embodiment, these entities can include suspects, victims, witnesses, locations, evidence, and any other relevant entities involved in the investigation. Relations (Edges): The relationships represent the connections or interactions between different entities. For example, relationships can include "committed," "witnessed," "located at," "related to," "collected as evidence," etc. These relationships may reflect the various aspects of the crime cases and the interactions between the entities. Similar definitions may be implemented in medical cases. For example, in the “health case” embodiment, the entities may include diagnostic equipment, medical equipment for treating a patient, medicine for treating a patient etc., while the relations may be “has performed”, “has taken” etc.

[0065] Returning to the “crime case” embodiment, the crime cases (as well as other cases) may be divided into distinct time intervals or stages (e.g., initial investigation, arrest, trial, and resolution). For each time interval, a knowledge graph is created that captures the entities and relationships relevant to that specific period. This temporal information allows to order the knowledge graphs in the temporal knowledge graph sequence. The sequence of knowledge graphs thus represents a temporal knowledge graph. Using an existing solution / system, for each Knowledge Graph, the solvability score (e.g., the probability, that the respective stage leads to a resolution of the case) may be determined using the solvability score engine 13. The score may be used at a later stage to assess the influence of the recommended action.

[0066] Block 6: Encoder: Al Model. In an example the encoder model may be configured to take as input a list of Temporal Knowledge Graph sequences and to learn a suitable representation. Suitable means that the model can predict or forecast the next Knowledge Graph. In other words, the methods of Figs. 1a and 1 b may comprise encoding 120, using a second machine learning model (the encoder Al model 6), the first sequence of knowledge graphs representing the plurality of stages of the case into a sequence of embeddings representing the sequence of knowledge graphs. Similarly, generating 210 the training data may comprise encoding 212, using the second machine learning model, the first sequence of knowledge graphs representing the plurality of stages of the case into the sequence of embeddings representing the sequence of knowledge graphs. As a result, the first sequence of knowledge graphs may be input as sequence of embeddings into the at least one machine learning model. This module may embed an existing TKG Forecasting algorithm. Training may be performed similar to the training of the encoder Al model 2 being used for the sequences of actions. In particular, the second machine learning model may be trained 222, e.g., using end-to-end training together with the at least one machine learning model / the transformers, to encode each knowledge graph of the sequence of knowledge graphs into a corresponding embedding representing the knowledge graph. After the TKG model was trained, the learned embedding representations may be output. Those are the output of this component.

[0067] Block 7: List of Sequences. Similar to block 4, this component may be a database storing the output of the Al Encoder Model 6. It may be used as a look up for the subsequent operation. A sequence may be implemented as an ordered list of n- dimensional numerical vectors or matrices, each representing a knowledge graph of the respective knowledge graph sequence. For example, the database storing the output of the Al encoder model 6 may also be denoted the “first data storage”, which is used to store the training output of the training of a second machine learning model trained to output an embedding for a given knowledge graph. Block 8: Transformer (Action Recommendation Model). The Action Recommendation Model uses as input the “List of Sequences” (one sequence per case) and “Action Embeddings” (the corresponding sequence of actions for the given case).

[0068] As shown in Fig. 6, the architecture of this Al model may comprise or consist of two sequence-to-vector transformer models. Thus, the at least one machine learning model being referred to by the methods of Figs. 1a and 2a may comprise at least one transformer model or transformer component, namely an action transformer and a case transformer (shown in Fig. 6).

[0069] The transformers are in series, i.e. , the output of the action transformer may be part of the input to the case transformer. In other words, the at least one machine learning model may comprise an action prediction component (i.e., the action transformer) trained to output the next action to perform based on the representation second sequence of actions, and a case prediction component (i.e., the case transformer) trained to output the predicted next stage of the case based on the representation of the first sequence of actions and based on the output of the action prediction component. The subsequent crime decoder model may map the n- dimensional numerical vector (output of the case transformer model) to the actual to-be-performed action label. Likewise, based on the output of the action transformer, the action decoder model may return an updated Crime Solvability Score, reflecting the influence of the provided / recommended action.

[0070] Two loss functions assessing the output of the crime case decoder and action decoder respectively may be used. For example, the loss functions may be calculated based on the output of the crime case decoder and of the action decoder. The action transformer may learn to predict the next element in the sequence of actions. As the predicted action is considered by the subsequent Case Transformer Model, the loss function of the crime case decoder model may also influence the action transformer (due to back-propagation). This leads to a third novel aspect: The learning process of the first transformer model may be led with a second subsequent transformer model where the output of the first transformer model (predicted action) is used in combination with a KG sequence (TKG) to predict how the predicted action influences the crime case. The loss function of the Crime Case Decoder may be reduced or minimized, i.e., the Solvability Score may be increased or maximized (i.e., the action is most effective). Thus, training the at least one machine learning model may comprise training 226, 228, in a joint manner (i.e., the action prediction component and the case prediction component together, optionally together with the second and third machine learning model) and using the training data, a) the action prediction component to output the next action to perform based on the representation of the second sequence of actions and b) the case prediction component to output the predicted next stage of the case based on the representation of the first sequence of actions and based on the output of the action prediction component. In particular, the action prediction component and the case prediction component may be trained such, that a loss function used for training the case prediction component impacts the training of the case prediction component, as the case prediction component is trained to output the predicted next stage of the case based on the output of the action prediction component. The loss function may use the existing ground truth as reference value to assess the output. The loss function of the action decoder may compare the output to the ground truth (i.e., any valid action at this stage, independent of the current crime case). The described components (e.g., action transformer and crime transformer) may belong to a single model.

[0071] Block 9: Action Decoder. The Action Decoder is part of the Action Recommendation Model. The decoder model may be configured to map the n-dimensional numerical vector (output of the transformer model) to the actual to-be-performed action label. The associated loss function may, for example, be a cross-entropy loss function. The loss function may assess whether the predicted action is in the expected range (set of values).

[0072] Block 10: Crime Case Decoder. The Crime Case Decoder may also be part of the Action Recommendation Model. The decoder model may be configured to return an updated Crime Solvability Score, reflecting the influence of the provided / recommended action. The associated loss function may, for example, be a (binary) cross-entropy loss function. In other words, the at least one machine learning model may be trained using at least one cross-entropy loss function. The loss function may assess whether the predicted score increased (i.e., there is a higher chance to solve / close the case).

[0073] In summary, the at least one machine learning model (i.e., the transformers) may be trained to output at least one of a first embedding representing the predicted next stage of the case (denoted vector at the output of the case transformer) and a second embedding representing the predicted next action to perform (denoted vector at the output of the action transformer). The actual next stage of the case and the predicted action may be retrieved, using the action decoder and action transformer, e.g., from the first and second data storage. Accordingly, the method of Fig. 1a may comprise retrieving the predicted next stage of the case from a first data storage using the first embedding (e.g., the vector output by the case decoder) and retrieving the predicted next action to perform from a second data storage using the second embedding (e.g., the vector output by the action decoder). For example, the first and second data storage may be implemented as vector database supporting lookup of entries according to the similarity between an input vector (e.g., the embedding) and the stored vectors (the stored embeddings being associated with the sequences of actions and sequences of knowledge graphs representing stages of a case).

[0074] Block 11 : Action may be a string label part of a predefined list of actions. The list may be case-dependent.

[0075] Block 12: Score may be a probability score value ranging from 0 to 1. The higher the score, the higher is the probability that the case can be solved / closed.

[0076] Block 13: Solvability Score Engine. The Solvability Score Engine is an existing module or algorithm and might not be a novel aspect of the proposed concept. The scope of the Solvability Score Engine is that it provides Solvability Score for a given case. The score may reflect the likelihood a case is being solved at the current stage. There is no assumption or prediction about the next (potential) steps. The concepts disclosed herein may be applied in various technical areas, for instance:

[0077] For example, the proposed concept may be used in the realm of Public Safety, for the purpose of crime investigation, by supporting Investigators to create Action Plans. After an initial investigation phase, crime investigators must review and assign actions to a crime case. This is an iterative process, i.e., the investigator assigns an action, and the outcome of the action influences the follow-up action. This is a time-consuming process, as the investigators have to review much information. In addition, a consistent decision-making across crime cases is difficult. Embodiments of the concepts disclosed herein support the decision-maker to make faster and more consistent decisions. As data source, a crime case management system (including fingerprints, analyzed blood samples, and other physical evidence), publicly available databases (including crime articles, academic literature, weather forecast, etc.), and / or smart sensor networks (e.g., cameras, temperature, presence, accelerometer, microphone) may be used.

[0078] The proposed concept may be applied to this use cases by taking historical crime cases, and the crime case of interest as input to compute the most effective action for a given crime type. The method supports the investigator in their decision-making process. As output, the method may provide a recommended action along with a score reflecting the influence of the recommended action.

[0079] The systems / methods disclosed herein can be integrated with loT (Internet of Things) devices, such as surveillance cameras and video analytics systems. When an action plan is generated, the proposed method may include, as event, sending commands to the cameras, e.g., to adjust their field of view, focus on specific areas, or activate facial recognition algorithms to identify potential suspects in real-time.

[0080] The proposed concept may also be integrated with loT devices such as traffic cameras, sensors, and traffic signal control systems that allow to convert suggested actions directly into action. For example, the proposed method may include, as event being triggered, sending commands to adjust traffic signal timings, activate dynamic message signs, or reroute traffic in real-time to alleviate congestion or divert traffic away from accidents.

[0081] The systems / methods disclosed herein can also be integrated with drones equipped with cameras, sensors, or loudspeakers for public safety purposes. When an action plan is generated, the proposed concept can be used to send commands to the drones to perform tasks like aerial surveillance, provide situational awareness, or broadcast important announcements to the public during emergencies.

[0082] The systems / methods disclosed herein can also be integrated with loT-enabled wearables, such as body cameras or biometric sensors worn by first responders or law enforcement personnel, allowing the proposed concept to directly apply the recommended actions. For example, the methods / systems disclosed herein can analyze data from body cameras to identify potentially dangerous situations, generate alerts, or provide guidance to ensure the safety of the personnel involved.

[0083] The proposed concept may also be used in the realm of healthcare, e.g., for patient treatment and / or electronical healthcare records. When treating a patient, e.g., in a hospital, it is common to assign a treatment and then observe the effect. This is an iterative process, i.e. , the medical doctor assigns treatments, observes the outcome, and then plans the follow-up treatment. Treatments may comprise, e.g., administering medications, scheduling therapy sessions, or preparing for a surgical procedure. Embodiments of the concepts disclosed herein support the decisionmaker to make faster and more consistent decisions by analyzing the treatments of related patients (e.g., having the same disease). As data source, electronic healthcare records, a patient case management system, publicly available databases, and / or smart sensor networks may be used. According to embodiments of the present disclosure, historical patient records, and the patient of interest may be taken as input to compute the most effective treatment for a given disease. The method supports the medical staff in their decision-making process when assigning treatments. According to embodiments of the present disclosure, the method may provide a recommended treatment (aka. action) along a score reflecting the influence of the recommended treatment. To automatically apply the recommended action, the proposed concept can be integrated with loT devices such as wearable health trackers, patient monitoring systems, and medication management systems (e.g., pill dispensers). Alternatively, or additionally, the systems / methods disclosed herein can be connected with alarm systems, dispatch emergency services, or activate automated alerts.

[0084] Various examples of the present disclosure are based on learning a stepwise sequence- and context-dependent embedding representation for each action to capture the characteristics of the preceding actions (novel embedding representation). The representation of a particular node may be derived from a sequence of graph representations, presenting the different stages of a case. This may be enabled by a sequence to vector model, as it predicts the next action for a given sequence of actions.

[0085] Various examples of the present disclosure are based on, for training an Al model, deriving negative samples from a positive sample by substituting stepwise and backwards the performed action with the help of a large language model. This may be used to overcome biased historical data by injecting the knowledge of a large language model.

[0086] Various examples of the present disclosure are based on controlling the learning process of a sequence to vector transformer model with a second subsequent sequence to vector transformer model. The output of a first transformer model (predicted action) is combined with a sequence of KGs (TKGs) as the input to a second transformer model. The goal of the second transformer model is to predict how the output of the first transformer influences the next step of the case. For example, the loss function of the Crime Case Decoder should be reduced or minimized, i.e. , the Solvability Score may be increased or maximized (i.e. , the action is most effective). The loss function may use the existing ground truth as reference value to assess the output.

[0087] Embodiments of the present disclosure provide computer-implemented methods and systems for Knowledge Graph based Sequential Decision Support for Action Plan Generation in Case Management Systems. The methods / systems may comprise one or more of the following operations / components. For example, the methods / systems may comprise Action Representation Learning, which may be used to generate Action Knowledge Graphs. An Al model (Sequence-dependent Action Representation) (related to the first inventive step) may be trained with help of a Negative Sampling Al Agent (related to the second novel aspect step), see Fig. 3, Blocks 2 and 3. This may be used to compute Action Embeddings. For example, the methods / systems may comprise Case Representation Learning, which may be used to generate the Case Temporal Knowledge Graphs. An Al model may be trained (with a TKG algorithm), and a list of sequences (Embedding representation of TKGs) may be generated. For example, the methods / systems may comprise training an Al Action Recommendation Model. This may include training a sequence to vector Al model (with a transformer-based approach) to learn a mapping from the learned Action Representation and List of Sequences to the recommended action and solvability score (related to the third novel aspect), see Fig. 3, Blocks 8, 9 and 10. The action decoder may be applied to derive the recommended action. The crime case decoder may be applied to derive the updated case solvability score. The output (recommended action) may be used to control an arbitrary loT device (e.g., surveillance cameras or traffic signal control).

[0088] Existing recommendation systems usually focus on simple or one-dimensional recommendations, i.e. , they focus on static preferences or properties to provide a recommendation such as a product or movie (see Guo, Qingyu, et al. 'A survey on knowledge graph-based recommender systems." or Wu, Shiwen, et al. "Graph neural networks in recommender systems: a survey."). This is very different to the methods / systems according to the concepts disclosed herein, as in the presently discussed case, the time aspect and the chronological sequence of actions needs to be considered to make a suitable action recommendation. Compared to sequential recommendation or sequential path reasoning systems which focus on a single (knowledge) graph, the methods / systems according to the concepts disclosed herein face the challenge of case dependent knowledge graphs, i.e., a suitable action recommendation is to be determined by analyzing a set of knowledge graphs and temporal knowledge.

[0089] Zhao et al.: "Time-aware path reasoning on knowledge graph for recommendation." noticed the limitation of existing graph-based recommender systems (“current KG- based explainable recommendation methods unfortunately ignore the temporal information”). Zhao et al limit the scope on the temporal aspect, by focusing on path reasoning for providing recommendations. A difference between their approach and the present disclosure is that Zhao et al. ignore the order of actions or events of the same case. While they model the information as Temporal Knowledge Graph, the graph (and so their algorithm) does not describe which events belong to the same process (or case).

[0090] Multi-hop knowledge graph reasoning may also be considered a related field. Bai et al. "SQUIRE: a sequence-to-sequence framework for multi-hop knowledge graph reasoning." introduces a sequence-to-sequence transformed based approach to reason sequences (e.g., a sentence). However, Bai et al.’s approach does not consider the time-aspect, i.e. , it focuses on a single, static knowledge graph. As such, this approach cannot capture the characteristics of a case. Another difference is that that Bai et al. attempt to predict / reason sequences. In contrast, in the proposed concept, vectors are predicted. In addition, Bai et al. focus on a different use case I application.

[0091] With respect to Negative Sampling, the approaches of Wang et al.: "Reinforced negative sampling over knowledge graph for recommendation." and Rony et al.: "LEMON: LanguagE MOdel for Negative Sampling of Knowledge Graph Embeddings." may provide alternative approaches. In contrast to the present concept, Wang et al. uses reinforcement learning and is unrelated to Large Language Models. In addition, sequences are not taken into account. Rony et al. relies on language models to generate negative samples, by using the language model to form neighborhood clusters by utilizing the distances between entities to obtain representations of symbolic entities via their textual information.

[0092] Many modifications and other embodiments of the invention set forth herein will come to mind to the one skilled in the art to which the invention pertains having the benefit of the teachings presented in the foregoing description and the associated drawings. Therefore, it is to be understood that the invention is not to be limited to the specific embodiments disclosed and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.

[0093] List of reference signs

[0094] 1 Action Knowledge Graph Generation

[0095] 2 Encoder Al model (for action KG)

[0096] 3 Negative Sampling Al Agent

[0097] 4 Action Embedding

[0098] 5 Crime Case Temporal Knowledge Graph

[0099] 6 Encoder Al model (for crime case TKG)

[0100] 7 List of sequences

[0101] 8 Transformer

[0102] 9 Action Decoder

[0103] 10 Crime Case Decoder

[0104] 11 Action

[0105] 12 Score

[0106] 13 Solvability Score Engine

[0107] 100 System

[0108] 102 Interface

[0109] 104 Processor

[0110] 116 Storage device

[0111] 110 Encoding a first sequence of knowledge graphs into a sequence of embeddings

[0112] 120 Encoding a second sequence of actions into a sequence of embeddings

[0113] 130 Inputting a representation of the first sequence of embeddings and a representation of the second sequence of actions into at least one machine learning model Triggering an event System Interface

[0114] Processor

[0115] Storage device

[0116] Generating training data Encoding a first sequence of knowledge graphs into a sequence of embeddings

[0117] Encoding a second sequence of actions into a sequence of embeddings

[0118] Training at least one ML model Training a second ML model Training a third ML model Training an action prediction component of an ML model

[0119] Training a case prediction component of an ML model

Claims

C l a i m s1. A computer-implemented method for triggering an event, the method comprising: inputting (130) a representation of a first sequence of knowledge graphs representing a plurality of stages of a case, such as a crime case or medical case, and a representation of a second sequence of actions having been performed in the plurality of stages of the case into at least one machine learning model, wherein the at least one machine learning model is trained to output a predicted next action to perform and a corresponding predicted next stage of the case in response to the representations of the first sequence and the second sequence being input into the machine learning model; and triggering (140) an event based on at least one of the predicted next action to perform and the predicted next stage of the case, wherein the event being triggered comprises at least one of processing, using an algorithm or machine learning model, sensor data, such as camera sensor data, being related to the predicted action, and controlling a device or system, such as a controllable sensor device, medical device, autonomous vehicle, traffic control system or wearable device, based on the predicted action.

2. The computer-implemented method according to claim 1 , wherein the at least one machine learning model comprises an action prediction component trained to output the next action to perform based on the representation second sequence of actions, and a case prediction component trained to output the predicted next stage of the case based on the representation of the first sequence of actions and based on the output of the action prediction component.

3. The computer-implemented method according to claim 2, wherein the action prediction component and the case prediction component are trained such, that a loss function used for training the case prediction component impacts the training of the case prediction component, as the case predictioncomponent is trained to output the predicted next stage of the case based on the output of the action prediction component.

4. The computer-implemented method according to one of the claims 1 to 3, wherein the method comprises encoding (110), using a second machine learning model, the first sequence of knowledge graphs representing the plurality of stages of the case into a sequence of embeddings representing the sequence of knowledge graphs.

5. The computer-implemented method according to one of the claims 1 to 4, wherein the method comprises encoding (120), using a third machine learning model, the second sequence of actions having been performed in the plurality of stages of the case into a sequence of embeddings, the sequence of embeddings representing the actions having been performed in the plurality of stages of the case.

6. The computer-implemented method according to claim 5, wherein the third machine learning model is trained using a first set of positive training samples representing sequences of actions having been performed in previous cases, and using a second set of negative training samples representing illogical sequences of actions, wherein the second set of negative training samples is derived from the first set of positive training samples by substituting actions of positive samples in a stepwise and / or backward manner using a language model.

7. The computer-implemented method according to one of the claims 1 to 6, wherein the at least one machine learning model is trained to output at least one of a first embedding representing the predicted next stage of the case and a second embedding representing the predicted next action to perform, with the method comprising at least one of retrieving the predicted next stage of the case from a first data storage using the first embedding and retrieving the predicted next action to perform from a second data storage using the second embedding.

8. A computer-implemented method for training at least one machine learning model, comprising: generating (210) training data comprising a plurality of training samples, with each training sample comprising a representation of a first sequence of knowledge graphs representing a plurality of stages of a case and a representation of a second sequence of actions having been performed in the plurality of stages of the case; and training (220) the at least one machine learning model, using the training data, to output a predicted next action to perform and a corresponding predicted next stage of the case in response to the first sequence and the second sequence being input into the at least one machine learning model.

9. The computer-implemented method according to claim 8, wherein the at least one machine learning model comprises an action prediction component and a case prediction component, wherein training (220) the at least one machine learning model comprises training (226, 228), in a joint manner and using the training data, a) the action prediction component to output the next action to perform based on the representation second sequence of actions, and b) the case prediction component to output the predicted next stage of the case based on the representation of the first sequence of actions and based on the output of the action prediction component.

10. The computer-implemented method according to claim 9, wherein the action prediction component and the case prediction component are trained such, that a loss function used for training the case prediction component impacts the training of the case prediction component, as the case prediction component is trained to output the predicted next stage of the case based on the output of the action prediction component.

11. The computer-implemented method according to one of the claims 8 to 10, wherein the method comprises training (222) a second machine learning model to output an embedding of a knowledge graph for a knowledge graph being input into the second machine learning model, wherein the secondmachine learning model is trained together with the at least one machine learning model in an end-to-end manner.

12. The computer-implemented method according to one of the claims 8 to 11 , wherein the method comprises training (224) a third machine learning model to output an embedding representing an action in the context of preceding actions for a sequence of actions being input into the third machine learning model, wherein the third machine learning model is trained using a first set of positive training samples representing sequences of actions having been performed in previous cases, and using a second set of negative training samples representing illogical sequences of actions, and / or wherein the third machine learning model is trained together with the at least one machine learning model in an end-to-end manner.

13. The computer-implemented method according to claim 12, wherein the method comprises deriving the second set of negative training samples from the first set of positive training samples by substituting actions of positive samples in a stepwise and / or backward manner using a language model.

14. A system comprising one or more processors and one or more storage devices, wherein the system is configured to perform at least one of the method according to one of the claims 1 to 7 and the method according to one of the claims 8 to 13.

15. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out at least one of the method according to one of the claims 1 to 7 and the method according to one of the claims 8 to 13.

Images