WO2023036450A1 - A method for operating a risk assessment system and a corresponding data processing system - Google Patents

Abstract

For providing an efficient risk assessment by simple means a method for operating a risk assessment system by means of a data processing system comprises the following steps: providing or collecting information in a domain of interest; transforming the information into a bipartite graph reflecting a relation between at least one unit of interest and the information; learning a representation for each of the at least one unit based on the bipartite graph related to the corresponding unit; clustering of representations and identifying a no risk cluster and a risk cluster based on historical data; repeating the providing or collecting step, the transforming step, the learning step and the clustering and identifying step one or more times and observing how the representations move and/or change; extracting for at least one unit of interest the part from the representation that caused the move and/or change over time; forecasting, based on historical data, a development of the part from the representation that caused the move and/or change over time; injecting at least one value of the forecasting of the development into the representation; and analyzing how the representation will change in a definable future time period. Further, a corresponding data processing system is provided.

Description

A METHOD FOR OPERATING A RISK ASSESSMENT SYSTEM AND A CORRESPONDING DATA PROCESSING SYSTEM

The present invention relates to a method for operating a risk assessment system by means of a data processing system.

Further, the present invention relates to a data processing system for risk assessment.

Corresponding prior art documents are listed as follows:

[1] Enrico Zio. The future of risk assessment. Reliability Engineering and System Safety, Elsevier, 2018,177, pp.176-190.

Further, WO 2021 034932 A1 discloses a method to make changes to a credit risk score from a present value to a target value using a tree based risk assessment model. The tree based decision models using a neural network are used for analyzing the behavior of nodes from a time period to a future time period in the cluster space. A risk mitigation recommendation is generated and information about re-training the model is also disclosed.

Further, US 10 734 101 B2 discloses a prediction of patients health based on sensor data of current patient and historical data of the patient having the same medical condition in the past. The method generates a bipartite graph and a machine learning, ML, model to analyze and train the model based on the similarity scores generated.

Further, US 10 453 015 B2 discloses a method to predict workplace injury events based on sensor data collected. Random forest models are used which are inferred as ensemble learning and regression models for classification of the features. Further, time series-based clustering is disclosed.

Risk assessment helps to understand and to control the risk of accident events. The outcome of risk assessment can be the input for decision making. Risk as a numeric quantity can be useful for making decisions such as on risk prevention and mitigation measures, prioritizing measures on different sources of risk, regulating an accepting risk, see [1],

There is the problem, that risk assessment has to include the following properties in order to be a useful basis particularly for safety-related rational decision making:

1. Risk assessment for decision making has to describe risk quantitatively, and with a proper representation of uncertainties, see [1],

2. Risk assessment for decision support needs to be able to incorporate all available knowledge, information and data, where it is not known which data is more or less important. The knowledge, information and data available for analyzing and characterizing hazards, modeling and computing risk are substantially grown and continue to do so, see [1],

3. Increased complexity of the systems, nowadays more and more made of heterogeneous elements - hardware, human, digital - organized in highly interconnected structures, leads to behaviors that are difficult to anticipate or predict, driven by unexpected events and corresponding emerging unknown systems responses. A risk assessment tool for decision support should be able to incorporate these heterogeneous elements and complex systems, as well as the interconnection between components, see [1],

4. To manage risk in a systematic and effective way, it is necessary that the risk assessment tool considers all phases of the potential risk scenarios that may occur, including prevention, mitigation, emergency crisis management and restoration, see [1],

5. Risk assessment must account for the time-dependent variations of components and systems. Changes can be physical - resulting from plant modifications, etc. -, operational - resulting from enhanced procedures, etc. -, organizational, but also changes in knowledge due to the acquisition of operational experience, field data, etc., see [1], The risk assessment tool has to be capable of capturing the time dependent behavior of the system risk profile, see [1], and has to be capable of updating its assessment as well as predicting future risk for prevention action recommendation. When trying to solve the task of risk assessment with given resources not all of the above mentioned properties can be found in combination.

It is an object of the present invention to improve and further develop a method for operating a risk assessment system by means of a data processing system and a corresponding data processing system for risk assessment for providing an efficient risk assessment by simple means.

In accordance with the invention, the aforementioned object is accomplished by a method for operating a risk assessment system by means of a data processing system, comprising the following steps:

- providing or collecting information in a domain of interest;

- transforming the information into a bipartite graph reflecting a relation between at least one unit of interest and the information;

- learning a representation for each of the at least one unit based on the current bipartite graph related to the corresponding unit and for the historic data of bipartite graphs;

- clustering of representations and identifying a no risk cluster and a risk cluster based on historical data;

- repeating the providing or collecting step, the transforming step, the learning step and the clustering and identifying step one or more times and observing how the representations move and/or change;

- extracting for at least one unit of interest the part from the representation that caused the move and/or change over time;

- forecasting, based on historical data, a development of the part from the representation that caused the move and/or change over time;

- injecting at least one value of the forecasting of the development into the representation; and

- analyzing how the representation will change in a definable future time period. Further, the aforementioned object is accomplished by a data processing system for risk assessment, comprising:

- providing or collecting means for providing or collecting information in a domain of interest;

- transforming means for transforming the information into a bipartite graph reflecting a relation between at least one unit of interest and the information;

- learning means for learning a representation for each of the at least one unit based on the current bipartite graph related to the corresponding unit and for the historic data of bipartite graphs;

- clustering means for clustering of representations and identifying means for identifying a no risk cluster and a risk cluster based on historical data;

- repeating means for repeating the providing or collecting step, the transforming step, the learning step and the clustering and identifying step one or more times and observing means for observing how the representations move and/or change;

- extracting means for extracting for at least one unit of interest the part from the representation that caused the move and/or change over time;

- forecasting means for forecasting, based on historical data, a development of the part from the representation that caused the move and/or change over time;

- injecting means for injecting at least one value of the forecasting of the development into the representation; and

- analyzing means for analyzing how the representation will change in a definable future time period.

According to the invention it has been recognized that it is possible to provide a very efficient method and data processing system by combining a suitable provision or collection of information in a domain of interest with the provision of a representation of a unit of interest based on a bipartite graph reflecting a relation between the unit and the provided or collected information. Generally a unit of interest can be a user or a system, for example. It has been further recognized that clustering of representations and identifying a no risk cluster and a risk cluster based on historical data provides a suitable tool within the method and system according to the invention. The bipartite graph can provide a differentiation between units and their respective information or their attributes, activities and/or related events. Further, a development of the part from the representation that caused a move and/or change over time is provided and at least one value of the forecasting of the development is injected into the representation. Then, an analyzing step follows which analyzes how the representation will change in a definable future time.

Thus, on the basis of the invention an efficient risk assessment is provided by simple means.

According to an embodiment of the invention the information can comprise events, activities and/or attributes, preferably related to the at least one unit or preferably related to at least one unit of the bipartite graph. Depending on the individual situation attributes can, for example, comprise time, location and/or dependencies to other events or activities. Information can comprise the daily log of a unit of interest.

Within another embodiment the providing or collecting step can comprise providing or collecting the information through observation by one or more sensors or observation units. Sensors or observation units can be arranged within a sensor network or observation unit network for effectively providing or collecting according to the present method.

According to a further embodiment the learning step can include one or more units from the past. This will result in a very flexible and effective method and system.

Within another embodiment the method can further comprise determining the respective centers of the clusters. Such centers or centroids can simply be determined by a clustering module which can be provided for taking the output of a representation discovery module for creating clusters.

According to a further embodiment the identifying step can be performed through label propagation. Labels of historical data can be propagated, wherein current units can simply be labeled in a semi-supervised way. Within a further embodiment the forecasting step can be performed by a regression process, a regression analysis or a time series forecasting. A regression process, a regression analysis or a time series forecasting helps understanding how crucial attributes will develop. Such a regression process, a regression analysis or a time series forecasting can simply be performed within a forecast module.

According to a further embodiment the representation can be modified to analyze a future development of the corresponding at least one unit, preferably in a cluster space. For performing this step forecast feature trajectories can be used in a simple and effective way.

Within a further embodiment, if the representation moves closer to a risk cluster or does not move away, the method can further comprise extracting information from at least one historical data sample, which has a similar trajectory for crucial features of the part until a time t or the actual time, but different for the time which was forecast.

According to a further embodiment, at least one non-crucial feature, which developed different to the ones of the at least one unit of interest, can be extracted from the at least one historical data sample.

Within a further embodiment and for providing a very effective and helpful method the method can comprise action recommendation, preferably for prevention and counteraction of unwanted situations.

According to a further embodiment at least one recommended action can be - preferably automatically - derived from at least one non-crucial feature, which has a different trajectory comparing the at least one unit of interest and at least one historical data sample. Particularly an automatic proceeding provides a very effective method and system. Generally, action recommendation can be performed within an action recommender module which can forward its output to other units or systems to simply adapt them accordingly. Within a further embodiment at least one recommended action can consist of at least one instruction to adjust at least one non-crucial feature of the at least one unit of interest to those of at least one selected historical data sample. Depending on the individual application the recommended action can comprise other instructions.

According to a further embodiment the method can start or can be repeated, if a defined feature or part of the information related to at least one definable unit changes. A corresponding automatic process can be implemented for providing a very effective and simple method and corresponding system.

Advantages and aspects of embodiments of the present invention are summarized as follows:

1 ) The way of forecasting based on given representations and historic data: Observe and forecast the development of a unit over time, where focusing on the crucial features of a unit representation. a. Isolating the features which cause a change over time in the cluster space - crucial features - and forecast the development of those through a regression analysis based on historical data which has similar trajectories until present/today. b. Use the forecast feature trajectories to modify the original unit representation to analyze the future development of that unit - risk vs. no risk - in the cluster space.

2) The way of counter action computation-. Find historical samples where the trajectory of the crucial features is initially similar but then develops in opposite directions. Use the non-crucial features of those to compute the counter actions. a. Compile a list of counter actions by extracting samples from the historical data which have a similar trajectory pattern for the crucial features in the cluster space until present/today but then develop as desired - in contrast to the forecast trajectory. b. The identified samples are used to automatically analyze how the non-crucial features developed until present/today. The features trajectories which differ to the observed unit are the ones which encode the required actions. Hence, the set of recommendations consists of these features as these need to be adapted/influenced accordingly.

Further advantages and aspects of embodiments of the present invention are summarized as follows:

1) Collect information in the domain of interest through observation, e.g., with a sensor network.

2) Transform the information into a bipartite graph which reflects the relation between the units of interest and the logged events/activities/attributes.

3) Learn a representation for each unit based on the directly connected nodes, i.e. , events/activities/attributes. This includes also units from the past where the outcome is known - historical data.

4) Cluster the representations - e.g. embeddings - and identify the “no risk” and “risk” cluster based on historical data through label propagation.

5) Repeat Step 1-4 n-times and observe how the representations - e.g. embeddings - move/change.

6) Extract for each unit of interest the part from the representation that caused the changes of the representation over time - crucial features.

7) Forecast the development of the crucial features based on historical data - regression.

8) Inject the forecast values into the representation and analyze how the representation will change in the future.

9) If the representation moves closer to a “risk” cluster or does not move away, extract information from the historical data samples which have a similar trajectory for the crucial features until time t but different for the time which was forecast. Extract from these samples the non-crucial features which developed different to the ones of the unit of interest.

10)The recommended action is automatically derived from the non-crucial features which have a different trajectory - comparing the unit of interest and samples from the historical data. The recommendation consists of instructions to adjust the non-critical features of the unit of interest to those of the selected historical samples. Further advantages and aspects of embodiments of the present invention, particularly in contrast to current state-of-the-art, are summarized as follows:

When trying to solve the task of risk assessment with given resources - not using embodiments of the present invention - not all of the mentioned properties - problems - can be found in combination. By using embodiments of the present invention, it is possible to fulfill all those properties, which have been described by [1], This has the following advantages:

1) By being able to describe risk quantitatively and with a representation of uncertainties, the system has the advantage of higher acceptance and trustworthiness, as compared to systems without this property. This can be covered by our representation, clustering and action recommender module.

2) By being able to incorporate all available knowledge, information and data, the system is able to capture and exploit all aspects of the situation. This in turn results in a more accurate risk assessment. By automatically selecting the data or features with higher importance, the system is able to improve accuracy even further. The automatic selection of data with higher importance is related to advantages and aspects 1a) and 2b).

3) By being able to incorporate heterogeneous elements as well as interconnections, the system has the advantage of linking and combining information. As a consequence, it is able to grasp structural dependencies between components, and thus to more accurate risk assessment as compared to systems without this property. Embodiments of our invention are able to do this as they rely on graph structured data.

4) By being able to consider all phases of the potential risk scenarios that may occur - namely prevention (future risk), timely reaction to emergencies (due to its dynamic updates) and restoration (improving bad risk scores from past) - the system is able to manage risk in a more systematic and effective way as compared to systems which are not able to consider all phases. This is covered in embodiments of our invention by the combination of our representation, clustering, and event forecasting module. 5) By being able to capture the time-dependent behavior and variations, the system has the advantage of more timely reactions to changes. Further, it enables a better prevention by more realistic and thus accurate risk assessment. This is related to advantages and aspects 1).

The advantages map to the problems as defined at the beginning of this document.

Embodiments of the invention describe a system for dynamic risk assessment and action recommendation for prevention and counteraction of unwanted situations. This risk assessment bases on a bipartite graph, which differentiates between units, e.g., humans, machines, and their attributes or related events. By computing each unit’s affiliation to certain clusters based on the graph’s node representation, and by analyzing and predicting the units’ movement in the cluster space, embodiments of the invention are able to assign a dynamic risk score for each unit through label propagation and, finally to compute counter actions if needed.

Embodiments of the invention solve the task of action recommendation for prevention and counter action of unwanted situations based on a dynamic risk assessment.

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 following explanation of examples of embodiments of the invention, illustrated by the drawing. In the drawing

Fig. 1 shows in a diagram a system architecture according to an embodiment of the present invention.

As depicted in Fig. 1 , an embodiment of a data processing system according to the invention consist of several modules, in particular a digital log module, a graph generation module, a representation discovery module, a clustering module, an event forecasting module, and an action recommender. In the following, we describe the scope and functionality of these modules and explain how they are connected. A daily log of the unit of interest, e.g., a user or system, records the daily routine/procedure, i.e. , through sensors it logs each event or activity along related attributes such as time, location, and the dependencies to other events or activities. The daily log has the ability to log only selected events/activities and can be considered as a digital diary. The daily log provides the input information to our system. It is assumed that several - similar or identical - units are observed simultaneously.

The graph generation module takes the output of the daily log, e.g., on a daily basis, along historical data. The historical data is a dataset which was earlier provided by the daily log, but the respective outcome is already known, i.e., the historical data is labeled so that it can be used as a reference to assess the new incoming data. The graph generation module takes the new and the historical data and converts them into a bipartite graph. Formulating the problem as a graph, allows to incorporate heterogeneous elements as well as interconnections - Problem 3. Further, it makes a clear distinction between the observed units and the recorded events/activities and the related attributes. In particular, the graph covers only relations between the observed units and the events/activities/attributes. There are no direct connections between units or between features, e.g., events, activities, attributes. This is an important and required characteristic in embodiments of our invention. The graph generation module does not consider or rely on the labels or outcome of the historical data for generating the graph.

The representation discovery module takes the generated bipartite graph as input and learns a representation for each observed unit based on the linked attributes/events/activities - features. The model allows to weight or prioritize different features differently, which has an influence on the similarity. This allows to incorporate all available knowledge, and to assess which data is more or less important - Problem 2. The idea is to learn a representation which is similar for units that share a similar set of features. This is achieved with a neural network based approach. The actual label/outcome of the historical data is not considered, i.e., this is an unsupervised process. The learned representation is the output of this module. Please note that later this module is also triggered by the event forecasting module, for this reason, the state of the module is always saved for later purposes.

The clustering module takes the output of the representation discovery module and creates clusters. In this context, the nodes - units - with labels, i.e. , the ones which result from the historical data, are used to create meaningful clusters. Meaningful clusters mean that the clusters actually represent the groups of interest. Based on these clusters, the module determines centroids which represent the center of the respective clusters. The distance of the nodes without labels to the closest cluster center reflects the tendencies to which group they belong. This allows to propagate the labels of the historical data, i.e., to label the current units in a semi-supervised way. The distance metric is not specified but can be selected based on the use case. A potential instantiation is the Euclidean distance.

In this context, the system starts essentially to observe over time how the nodes without labels move over time. Hence, if parameters for a certain unit change, e.g., a new activity/event occurred, the entire procedure which is described above is repeated. This makes the system capable of capturing the time dependent behavior - Problem 5.

As it is ensured that the representation of units with labels are constant, the system can essentially track how nodes without labels move over time, which in turn indicates if a unit moves closer to a certain cluster or not. After a predefined time, the logged movement information is, along the change in representation of these nodes, forwarded to the event forecast module. It is notable that our invention does not simply rely on the movement of the nodes to assess a risk or outcome, but analyzes with the event forecasting module how the node might move in the future to finally compute a risk or outcome.

The event forecasting module analyzes the movement of the individual nodes, i.e., the nodes without labels. In particular, the module extracts from the representations of a node, i.e., from the numerical vectors, those elements - features - which cause the position changes of the node in the cluster space over time. As a result, we have for each extracted feature a time series. The time series reflects how the respective feature changed over time in the cluster space - clustering module. This allows to find similar node - and also feature - movement patterns in the historical data. This in turn allows to forecast how the crucial features might develop. In particular, the event forecast module performs a regression analysis to understand how the crucial attributes will develop.

As an example, let’s assume, we observed an unlabeled node n for 10 time-steps in the clustering module, wherein n is an integer. Then, the event forecasting module analyzes the corresponding representations - numerical vectors - of node n. As we have 10 time-steps, we also have 10 representations. The event forecasting module extracts from these n-dimensional representation the elements which cause the movement of the node in the cluster space. The extracted elements represent our features. As a result, for each extracted feature, we have a time series, as we have observed node n for 10 time steps. Those are used to identify similar/related patterns/trajectories in the historical data. The related historical data is used to train a model to subsequently forecast the future trend of the extracted features of node n.

Finally, these results are fed back to the representation discovery module, i.e. , the features - attributes/events/activities - of the observed unit - node - are stepwise modified/extended based on the forecast result and the representation of the node is stepwise re-computed from the model which was initially trained. Finally, the new representations - based on the forecast features - are passed to the clustering module to analyze how the observed node is expected to move. The predicted movement of the node - unit - allows to understand and extract the potential risk, e.g., whether a unit will fail in the near future. The outcome is a risk score reflected by the movement of the node and the distance, e.g., Euclidean distance, to the closest centroid, the actual classification of the node based on the closest centroid, and an explanation of the classification result through the extracted and forecast features. The risk score describes the assessed risk quantitatively - Problem 1 .

Overall, the forecasting module for risk score computation allows to consider all phases of the potential risk scenarios that may occur. It can be used for prevention - future risk but also for timely reaction to emergencies - due to its dynamic updates - and for restoration - improving bad risk scores from past - Problem 4.

Finally, the action recommender takes these outcomes, to compute, if necessary, counter actions. These counter actions should ensure that the predicted movement/develop is at least delayed or changed. The counter actions aim to influence the crucial features which were extracted by the event forecasting module and are derived from historical data. In particular, the action recommender extracts from the historical data units where the crucial features evolve as desired. Subsequently, the remaining features - non-crucial - are compared with the ones from the observed unit to understand - e.g. by correlation - which features need to be treated to deserve the desired outcome. The output of the action recommender module, is forwarded to other units or systems to adapt them accordingly.

Further Embodiments

1. Public Safety: Crime Prevention

Use Case: The police observes the different areas/district in a city and records the crimes.

Data Source: Database of the police force. This includes information of the respective district, including characteristics of the social life, availability of, e.g., schools and police forces, and ethnicities.

Our Method: Embodiments of our invention observe and predict whether certain districts slip into crime, e.g., the number of crimes increases.

Output: A - preferably ranked - list of districts which become more criminal or might become more criminal in the future. Further, the reason why the crime level increases and a list of suggestions how to prevent this development - Measures against criminalization.

Physical Change - Technicity: Add/Adapt monitoring units, e.g., drones, continuous monitoring is usually not ok, sent cleaning androids/units/machines - might increase the moral -, sent autonomous and mobile library - give people the chance to educate - , adapt advertising on digital advertising panel - screens - to fight disinformation, enable/disable automatically the body cams of a police officer, preferably depending on his/her region.

2. Smart City: Predictive Maintenance

Use Case: Buildings like bridges, towers, factories but also urban vehicles like buses and trams need - preferably regular - maintenance. These buildings can be owned by the public domain but also a private company.

Data Source: Embodiments of the invention provide a sensor network which monitors the - preferably smart - buildings and vehicles.

Our Method: Embodiments of our invention observe if buildings/vehicles drift - possibly too fast - to the “need maintenance” cluster.

Output: A - preferably ranked - list of units which need or might need maintenance soon. Further, the reason why the maintenance is necessary and a list of suggestions how to prevent or delay the maintenance.

Physical Change - Technicity: Reroute traffic or reduce useable lanes to reduce the load on a bridge, lock/block elevators in a building or adapt the permitted weight, lock/block trams/buses automatically from usage to avoid accidents, evacuate a building - preferably enable alarm, e.g. after an earthquake.

3. Digital Government: Infection Protection

Use Case: The government runs a service for the society, e.g. an app like for Covid19, to inform people but also to gather information, e.g., whether a person is infected or not.

Data Source: A - preferably central - database which is part of the back-end of the service/app.

Our Method: Embodiments of our invention can analyze/predict, e.g., on a regionlevel how the infection will develop, i.e. , whether it becomes worse or better.

Output: A list of endangered regions along the reasons why it becomes worse - preferably derived from comparison with regions where it is better. Further, a list of potential counter actions.

Physical Change - Technicity: close public buildings automatically, e.g., adapted opening hours, automatically adapted control of how many people are in a public building, adapt advertising on digital advertising panel - preferably screens - to fight disinformation, adapt the frequency of the public transport, sent autonomous and mobile testing station.

4. Health Care: Assisted Living

Use Case: After a surgery people - especially elderly - need support during their daily routine. Uncertainty and insecurity are the biggest problems, especially when people live alone.

Data Source: Smart Home Sensor Network which records/logs the daily activities of the resident.

Our Method: Embodiments of our invention observe the condition of the patient and forecast the expected development in terms of recovery, i.e. , whether it becomes better or worse. This is derived from other patients which had a similar surgery and living situation.

Output: A list of observed patients which need more support as the expected recovery takes longer than expected or does not set in at all.

Physical Change - Technicity: automatically adaption of the sport/training equipment - level of difficulty -, establish connection to a suitable therapist/doctor - smart routing -, adapting the camera system - preferably to give the user the feeling of more security -, automatic adaption of the furniture/chairs, e.g., height - standing table.

5. Public Safety: Offender Recidivism Prediction

Use Case: Depending on the crime, offenders might be placed under house arrest, or have a day pass. In such cases, these people are often monitored through an electronic ankle tag.

Data Source: The - preferably smart - electric ankle tag which transmits locations and movements. In addition, a database with historical cases.

Our Method: Embodiments of our invention observe the locations and movements of an individual and derive from that if the risk of reoffending increases. This information is derived from the association between certain places and certain crimes, e.g., some areas are well known for drug dealing. Output: A risk score which reflects whether the risk of becoming a reoffender increases or decreases.

Physical Change - Technicity: Embodiments of the invention predict a significantly higher risk score due to an event related to the offender. In reaction to this, a system for electronic monitoring, see https://en.wikipedia.org/wiki/Electronic_monitoring_in_the_United_States, could be triggered to start monitoring the location and to restrict the offenders from specific geographical areas or even to enforce curfew hours - given that this is within probation regulations decided by the court.

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.

Claims

C l a i m s

1. A method for operating a risk assessment system by means of a data processing system, comprising the following steps:

- providing or collecting information in a domain of interest;

- transforming the information into a bipartite graph reflecting a relation between at least one unit of interest and the information;

- learning a representation for each of the at least one unit based on the bipartite graph related to the corresponding unit;

- clustering of representations and identifying a no risk cluster and a risk cluster based on historical data;

- repeating the providing or collecting step, the transforming step, the learning step and the clustering and identifying step one or more times and observing how the representations move and/or change;

- extracting for at least one unit of interest the part from the representation that caused the move and/or change over time;

- forecasting, based on historical data, a development of the part from the representation that caused the move and/or change over time;

- injecting at least one value of the forecasting of the development into the representation; and

- analyzing how the representation will change in a definable future time period.

2. A method according to claim 1 , wherein the information comprises events, activities and/or attributes, preferably related to at least one unit of the bipartite graph.

3. A method according to claim 1 or 2, wherein the providing or collecting step comprises providing or collecting the information through observation by one or more sensors or observation units.

4. A method according to any one of claims 1 to 3, wherein the learning step includes one or more units from the past.

5. A method according to any one of claims 1 to 4, further comprising determining the respective centers of the clusters.

6. A method according to any one of claims 1 to 5, wherein the identifying step is performed through label propagation.

7. A method according to any one of claims 1 to 6, wherein the forecasting step is performed by a regression process, a regression analysis or a time series forecasting.

8. A method according to any one of claims 1 to 7, wherein the representation is modified to analyze a future development of the corresponding at least one unit, preferably in a cluster space.

9. A method according to any one of claims 1 to 8, wherein, if the representation moves closer to a risk cluster or does not move away, the method further comprises extracting information from at least one historical data sample, which has a similar trajectory for crucial features of the part until a time t or the actual time, but different for the time which was forecast.

10. A method according to claim 9, wherein at least one non-crucial feature, which developed different to the ones of the at least one unit of interest, is extracted from the at least one historical data sample.

11. A method according to any one of claims 1 to 10, wherein the method comprises action recommendation, preferably for prevention and counteraction of unwanted situations.

12. A method according to claim 11 , wherein at least one recommended action is - preferably automatically - derived from at least one non-crucial feature, which has a different trajectory comparing the at least one unit of interest and at least one historical data sample.

13. A method according to claim 11 or 12, wherein at least one recommended action consists of at least one instruction to adjust at least one non-crucial feature of the at least one unit of interest to those of at least one selected historical data sample.

14. A method according to any one of claims 1 to 13, wherein the method starts or is repeated, if a defined feature or part of the information related to at least one definable unit changes.

15. A data processing system for risk assessment, preferably for carrying out a method for operating a risk assessment system according to any one of claims 1 to 14, comprising:

- providing or collecting means for providing or collecting information in a domain of interest;

- transforming means for transforming the information into a bipartite graph reflecting a relation between at least one unit of interest and the information;

- learning means for learning a representation for each of the at least one unit based on the bipartite graph related to the corresponding unit;

- clustering means for clustering of representations and identifying means for identifying a no risk cluster and a risk cluster based on historical data;

- repeating means for repeating the providing or collecting step, the transforming step, the learning step and the clustering and identifying step one or more times and observing means for observing how the representations move and/or change;

- extracting means for extracting for at least one unit of interest the part from the representation that caused the move and/or change over time;

- forecasting means for forecasting, based on historical data, a development of the part from the representation that caused the move and/or change over time;

- injecting means for injecting at least one value of the forecasting of the development into the representation; and

- analyzing means for analyzing how the representation will change in a definable future time period.