Improved computer-implemented event prediction and information provision

Abstract

A computer-implemented method, computer system, and computer program product for event prediction and information provision are disclosed. For each entity of a group of entities, one or more events are obtained, and each event is associated with a category of a set of categories. A subset of model categories of the set of categories and a subset of target categories based on the subset of model categories are determined. For each category of the subset of target categories, an arrangement of categories in the subset of model categories is determined for the entity for which the event was obtained, and a corresponding probability is calculated. For the target entity, a target arrangement of categories in the subset of model categories is determined based on the events obtained for the target entity. The target category is determined based on the target arrangement and the calculated probability. Information is provided based on the target entity and the determined target category.

Description

[Technical field]

[0001] The present invention relates to computer-based methods for reasoning (G06N5 / 04) (in particular sequence pattern mining) based on the representation of knowledge about specific events (G06N5 / 02, G06N5 / 00). Run by The present invention relates to a method, a computer system, and a computer program product. Sensor-based fault tracking in a group of devices to predict targeted failures of the devices and provide information for maintaining the devices to avoid failures; Tracking the usage of features of the computer program product among a group of users and predicting the usage of the features by users and providing users with learning information about the usage of the features; and / or Tracking interactions with services within a group of users and predicting target interactions for the users and providing information about the target interactions to the users; can be used for [Background technology]

[0002] US Patent No. 5,465,321 relates to fault detection in dynamic systems. A system fault monitoring method and apparatus learns a mapping between symptoms and faults from training data. The state of the system is first estimated at discrete time intervals. A feature vector is estimated from a set of successive windows of sensor data. A pattern recognition construct then models an instantaneous estimate of the posterior fault class probability given the features. A hidden Markov model is used to estimate the class probabilities conditional on the recent past history, utilizing the temporal context.

[0003] US Patent No. 6,415,276 relates to Bayesian belief networks for industrial processes. A method and apparatus for diagnosing sensors and / or processes using a Bayesian network is disclosed. More specifically, the method and apparatus achieves fault detection, isolation, and adaptation of sensors and / or processes.

[0004] According to the above disclosure, since a hidden Markov model or a Bayesian network is utilized, the history is only partially taken into account, and therefore the above disclosure does not teach the technical features for taking into account a long history to a sufficient extent.

[0005] US Patent Application Publication No. 2016 / 0071126 relates to predicting and resolving customer journeys. A predictive model is disclosed in which each user is mapped to all available user journey information corresponding to a particular business. The predictive model is analyzed to understand the user's characteristics, preferences, and least-effort resolution associated with the services subscribed to by the user. The predictive model is analyzed to predict a service or collection of services for each user. The embodiment interacts with customers through channels and services, provides information to customers, receives information from customers, reacts to customers, and / or delivers actions to customers. All customer and system behavior, data, and actions are tracked, adjusted, and leveraged for continuous feedback and performance improvement.

[0006] US Patent Application Publication No. 2017 / 0372347 relates to a marketing attribution model based on sequences of customer journeys. The method includes extracting subsequences from a sequence of customer journeys that includes customer interactions on different channels at different times on different topics. The method further includes measuring the effectiveness of each subsequence based on journey success data by applying a statistical hypothesis testing approach. The method also includes determining the contribution of each of the customer interactions to a given one of the subsequences by applying a sequence-based journey attribution model.

[0007] US10067990 relates to a system, method and computer program for identifying important attributes of records. A plurality of records including a plurality of events having a plurality of attributes are stored. Furthermore, the attributes of the events are processed using machine learning. For this purpose, based on the processing, at least one of the attributes is identified as important. Then, the at least one attribute so identified can be displayed.

[0008] Gay, Lopez and Melendez, “Sequential learning for case-based pattern recognition in complex event domains”, Proceedings of the 16th UK Workshop on Case-Based Reasoning (UKCBR 2011), Cambridge, United Kingdom, 13 December 2011, published on 20 January 2012 in CEUR Workshop Proceedings (ISSN 1613-0073), Vol. 829, paper 6, http: / / ceur-ws.org / Vol-829 / paper6.pdf, hereinafter, Gay (2012) discloses sequence pattern mining.

[0009] At least some of the teachings disclosed in the cited documents above suffer from information overload. Typically, all or most of the information is treated on an equal footing. For example, Gay (2012) does not teach the actor to utilize the enumerated sequences (last sentence of section 3.1). Furthermore, in at least some of the cited documents above, motion prediction is not taught or is not properly taught. Proper motion prediction should provide timely intervention. Furthermore, the resulting sequences usually lack further temporal information, except for chronological information. For a first device whose next failure will occur soon, e.g. tomorrow, intervention is more urgent than for a second device whose next failure will occur in a more distant feature, e.g. within two years. In the latter case, intervention may no longer be required, e.g. considering the planned life of the device or upcoming scheduled maintenance. Furthermore, in an environment with limited resources, e.g. one maintenance man for the entire pool of devices, the limited resources must be optimally distributed. For example, for a pool of equipment, this may include minimizing the overall downtime across the pool. In a general formulation, this involves maximizing the overall retention (e.g., equipment uptime) across a group of entities (e.g., a pool of equipment) during at least an initial period of time (e.g., taking into account scheduled maintenance or scheduled lifespans of the equipment).

[0010] The present invention aims to solve at least some of the problems mentioned above. Summary of the Invention

[0011] In a first aspect, the present invention provides a computer for event prediction and information provision according to claim 1. Run by A method is provided.

[0012] In a second aspect, the present invention provides a computer system for event prediction and information provision as claimed in claim 12.

[0013] In a third aspect, the present invention provides a computer program product for event prediction and information provision as claimed in claim 13.

[0014] The present invention is advantageous for several reasons.

[0015] One reason, in particular, concerns the combination of (a) determining a model category subset of categories from the entire set of categories and (b) calculating the probability that a tail category follows a base based on the entire history within the subset of model categories. By reducing the entire set of categories to the subset of model categories, the entire history within the subset of model categories can be modeled via the sequence and taken into account in determining the probability. This should be contrasted with Hidden Markov Models or Bayesian Networks, where the time scale is intrinsic to the model and only recent history is considered.

[0016] Another reason is related to the specific hierarchy of the determination of the subset. If a subset of the target categories of categories is selected first, and then a subset of the model categories is selected, the accuracy mainly depends on the inclusion or exclusion of certain categories in the subset of the model categories, and often ends up insufficiently modeling certain transitions related to the categories of the subset of the target categories by operating in the reduced subset of the model categories. By first selecting a subset of the model categories and then selecting a subset of the target categories based on the subset of the model categories, the applicant surprisingly found that better transitions can be obtained in the reduced subset of the model categories, since the constraint of the subset of the target categories does not yet exist. Furthermore, although it is a priori clear that the subset of the target categories based on the subset of the model categories is inherently incomplete, the applicant surprisingly found that the accuracy does not decrease. This can be a posteriori attributed to the insufficient modeling of certain transitions related to the categories of the subset of the target categories when the subset of the target categories is first determined. In other words, both when initially selecting a subset of target categories and when initially selecting a subset of model categories, not all transitions can be captured in the reduced subset of model categories, whereas the initial determination of the subset of model categories and the subsequent determination of the subset of target categories based on the subset of model categories provides better modeling of transitions overall.

[0017] Gay (2012) fails to disclose at least the following salient features: ·Active subgroups of an entity -Determining active subgroups based on average event frequency A subset of the target category The first occurrence of a category in a subset of model categories Up to Determining a subset of target categories from a subset of model categories based on time period

[0018] Referring to the last sentence of section 3.1 of Gay (2012), Gay (2012) specifically does not teach specialization for subjects.

[0019] These salient features for Gay (2012) result in maximizing overall retention across the group of entities, at least during the initial period. ·Individual entities can behave differently and have inherent differences that are not known a priori. For example, in the case of a device, if a first fault occurs in the device, a second fault can have a very high probability of occurrence. Similarly, if a third fault occurs in the device, a fourth fault can have a very high probability of occurrence. On the other hand, the first and third faults may be very unlikely to occur in the same device. Without specialization for the entities, such differences between entities cannot be inferred. Therefore, the history of a particular entity is paramount in predicting the future of that particular entity, which can be based on similar history of other particular entities. By converting events associated with timestamps into an array, only the chronological information is typically preserved, but not absolute or relative time information. Up to Creating a subset of target categories based on time period introduces temporal information in the model space in addition to the temporal ordering of entity events without having to explicitly utilize absolute or relative timestamps to predict the target entity's category. Instead, the target category subset indirectly incorporates the absolute or relative timestamp aspect via the determination of which categories to include in the target category subset of the model category subset. The active subgroup of entities provides entities on which substantial entity sequence mining can be performed. There is also the added benefit that entities with high event frequency are favored over entities with the same amount of events but with a lower event frequency. Two entities with identical histories will have equal probability distribution of the next event category, except for event frequency, but the entity with a higher event frequency will have a greater urgency to process it since it occurs sooner. Retention across groups of entities is most beneficial by inferring from entities with high event frequency, i.e. entities with a large number of events occurring within a particular time period. [Brief description of the drawings]

[0020] [Figure 1] 1 shows a schematic diagram of an algorithm according to an embodiment of the present invention. [Figure 2A] 2 shows a schematic diagram of a set of categories (201), a subset of model categories of the category (202), and a subset of target categories of the category (203), according to an embodiment of the present invention. [Figure 2B] 2 shows a schematic diagram of a set of categories (201), a subset of model categories of the category (202), and a subset of target categories of the category (203), according to an embodiment of the present invention. [Diagram 3] 2 shows a schematic diagram of communication channels and data processing according to an embodiment of the present invention; [Figure 4] 13 illustrates a chart showing average time to first occurrence for categories of a subset of model categories, in accordance with an embodiment of the present invention. [Diagram 5] 1 shows a schematic diagram of an algorithm according to an embodiment of the present invention. [Figure 6] 1 shows a schematic diagram of an arrangement of categories of a subset of model categories based on an entity's events, according to an embodiment of the present invention; [Figure 7] 1 shows a schematic diagram of an array according to an embodiment of the present invention. [Figure 8] 1 shows a schematic diagram of an algorithm according to an embodiment of the present invention. [Figure 9] 1 shows a schematic diagram of an algorithm according to an embodiment of the present invention. [Figure 10] 13 shows a chart illustrating success rates of stimulated groups of entities versus control groups of entities, according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] The present invention relates to a computer-based method for predicting events and providing information. Run by The present invention relates to a method (CIM), a computer system, and a computer program product (CPP). The present invention is summarized in the corresponding parts above. In the following, the present invention is described in detail, preferred embodiments are described, and the present invention is illustrated by non-limiting examples.

[0022] Unless otherwise defined, all terms used in disclosing the present invention, including technical and scientific terms, have the meanings as commonly understood by one of ordinary skill in the art to which the present invention belongs. By way of further guidance, definitions of terms are included to better understand the teachings of the present invention.

[0023] As used herein, the following terms have the following meanings.

[0024] As used herein, "a" refers to both singular and plural referents unless the context clearly dictates otherwise. By way of example, "a compartment" refers to one or more compartments.

[0025] As used herein, "including," "comprising," and "having" are synonymous and are inclusive or open-ended terms that specify the presence of what follows (e.g., a component) and exclude or do not exclude the presence of additional, non-recited components, features, ingredients, elements, steps that are known in the art or disclosed herein.

[0026] As used herein, "based on" is synonymous with "based at least in part on" and is an inclusive or open-ended term that specifies the presence of what follows and does not exclude or exclude the presence of additional, unrecited components, features, ingredients, elements, steps, etc.

[0027] As used herein, an "array" is an ordered table of elements. An array includes at least two elements. An array includes a "base" and a "tail element." An array specifically includes a base followed by a tail element. A base includes one or more elements. If the base includes at least two elements, the base is an array, and the base itself includes a base and a tail element. In a non-limiting example, a first array can include elements E1, E4, E7, E8, and E9 in the order E1→E9→E7→E4→E8. The first array includes a second array, base E1→E9→E7→E4, and tail element E8. The second array includes a third array, base E1→E9→E7, and tail element E4.

[0028] A "group," "set," "collection," "table," "array," etc., including elements as used herein, refers to a computer-readable, preferably digital, representation of the group, set, collection, table, array, etc., including pointers to the elements, such as identifiers of the elements. A set of categories thus refers to a set of computer-readable representations including pointers to the categories, such as identifiers of the categories. Each category is then represented by a category identifier. A group of devices thus refers to a group of computer-readable representations including device pointers, such as device identifiers. Each device is then represented by a device identifier. A group of users thus refers to a group of computer-readable representations including pointers to the users, such as identifiers of the users. Each user is then represented by a user identifier.

[0029] In a first aspect, the present invention provides a CIM for event prediction and information provision. In a second aspect, the present invention provides a computer system for event prediction and information provision, the computer system comprising one or more processors, the computer system configured to execute the CIM according to the first aspect of the present invention. In a third aspect, the present invention provides a CPP for event prediction and information provision, the CPP comprising instructions that, when executed by a computer such as a computer system according to the second aspect, cause the computer to execute the CIM according to the first aspect. The third aspect may provide a tangible, non-transitory, computer-readable data communication comprising the CPP.

[0030] Therefore, those skilled in the art will understand that the three aspects of the present invention are interrelated. Therefore, in the following, explicit reference to a specific aspect may be omitted. Furthermore, each feature described above or below may relate to each of the aspects of the present invention, even if it is disclosed in relation to a specific aspect of the present invention.

[0031] The CIM comprises several steps, see figures 1 and 2. For each entity of the group of entities, one or more events are obtained (101). Preferably, for each entity of the group of entities, a plurality of events are obtained. Each event is associated (201) with a category of the set of categories. The event can further be associated with a timestamp. Furthermore, the event can be associated with a communication channel from the set of communication channels. The event can in particular comprise a pointer to a category of the set of categories. The event can further comprise a timestamp. Also, the event can comprise a pointer to a communication channel of the set of communication channels. Preferably, for each entity of the group, an average event frequency is calculated. Preferably, an active subgroup of entities is determined based on the calculated average event frequency. A subset (202) of model categories of the set of categories (201) is determined in (102). Thus, the set of categories (201) comprises the subset (202) of model categories. Preferably, the subset of model categories is a subset of the categories of the set of categories in the phenomenon obtained for the entities in the active subgroup. occurrence Degree (prevalence) A subset of target categories (203) is determined (102) based on the subset of model categories (202). In particular, the subset of model categories (202) includes the subset of target categories (203). Preferably, the subset of target categories is determined based on a first occurrence of a category of the subset of model categories in an event that is obtained for an entity in the active subgroup. Up toFrom the subset of model categories, an arrangement of categories of the subset of model categories (202) is determined (103) based on the events obtained for the entity, preferably based on the chronological occurrence of the categories of the subset of model categories in the events obtained for the entity, for each entity of the group, where at least one event associated with the category has been obtained for each category of the subset of target categories (203), based on the events obtained for the entity, preferably based on the chronological occurrence of the categories of the subset of model categories in the events obtained for the entity. The arrangement consists of a base, which includes one or more categories, and a tail category after the base. Two or more categories may be obtained. 3 or more A base containing a category is itself a subarray consisting of a base containing one or more categories and a tail category following the base. 3 or more For each sequence of the determined sequence and its recursive bases including the categories of the determined sequence, a probability is calculated (104) that the tail category of the sequence follows the base of the sequence. For a target entity of a group of entities for which, for at least one category of the subset of target categories, no event associated with at least one category has been obtained, i.e. the events obtained for the entity do not include an event associated with the category for at least one category of the subset of target categories, a target sequence of the categories of the subset of model categories is determined (105) based on the events obtained for the target entity, preferably based on the chronological occurrence of the categories of the subset of model categories in the events obtained for the target entity. A target category of the at least one category is determined based on the target sequence and the calculated probability (105). Information based on the target entity and the determined target category is provided (106).

[0032] In a preferred embodiment, the CIM includes the following steps: · For each entity of the group of entities, obtaining one or more, preferably multiple, categories of the set of categories associated with each event, with an event,timestamp. · Calculating the average event frequency for each entity in the group. · Determining active subgroups of entities based on the calculated average event frequency. The set of categories in the events captured for entities in the active subgroup occurrence Determining a subset of model categories from the set of categories based on the degrees. The first occurrence of a category of a subset of model categories in an event that is obtained for an entity in the active subgroup Up to Determining a subset of target categories from the subset of model categories based on a time period. determining, for each entity of the group of the plurality of entities, an ordering of the categories of the subset of model categories based on a chronological occurrence of the categories of the subset of model categories in the events obtained for the entity, the events obtained for the entity including, for each category of the subset of target categories, at least one event associated with the category, the ordering comprising: A base that includes one or more categories; · A tail category after the base, It consists of two or 3 or more The base that contains the category of is itself A base that includes one or more categories; · A tail category after the base, It is an array consisting of. · For each of the determined sequences and their subsequences, calculating the probability that the tail category of the sequence follows the base of the sequence. For a target entity of the group of multiple entities, where the events obtained for the entity do not include an event associated with a category for at least one category of the subset of target categories, determining a target sequence of categories of the subset of model categories based on a chronological occurrence of the categories of the subset of model categories in the events obtained for the target entity, and determining a target category for the at least one category based on the target sequence and the calculated probability. · Providing information based on covered entities and determined covered categories.

[0033] For a discussion of some of the advantages, see the Summary of the Invention section above.

[0034] In a preferred embodiment, the invention is configured for failure prediction and information provision in devices. In this embodiment, the entity is a device, whereby the device preferably comprises one or more sensors. The device can be represented by an identifier of the device. Then, one or more events of each device are preferably obtained based on data from one or more sensors of the device. In this embodiment, the information is preferably information provided for failure avoidance. In the same or similar devices, certain symptoms can indicate the appearance of a failure. This is because the symptoms can cause the failure. This is alternatively or additionally because the symptoms and the upcoming failure can be caused by a common root cause.

[0035] In another embodiment, the invention is configured for prediction and information provision of feature usage by users in a CPP. In this embodiment, the entities are users. The users can be represented by their identifiers. In this embodiment, the categories are functions of computer program products. In this embodiment, the step of providing information based on the target entities and the target categories is preferably a step of providing learning information on the target features to the target users via visualization means. When learning to use the CPP, the amount of new features can be overwhelming for the user. In this embodiment, the invention provides learning information to the user about the target features, which are specifically selected based on the user's feature usage history and therefore can be of particular interest to the user. This in turn can result in a more time and resource efficient learning process, since the user is not presented with a steep learning curve based on long documents that are mostly irrelevant but mostly relevant tutorials. This can also result in selective introduction of new features to experienced users when releasing new editions of the CPP.

[0036] In yet another embodiment, the invention is configured for predicting and providing information on a type of interaction by a user with a service. In this embodiment, the entity is a user, and the user can be represented by a user identifier. In this embodiment, the category is a type of interaction. In this embodiment, providing information based on a target entity and a target category is providing information to a user on a type of interaction of interest.

[0037] In a preferred embodiment, the determination of the entity of the sequence of categories of the subset of model categories based on the events obtained for the entity is based on the chronological occurrence of the categories of the subset of model categories in the events obtained for the entity. Thereby, the determined sequence of the entity can include each category of the subset of model categories at most once, whereby the determined sequence is preferably based on the first chronological occurrence of the categories of the subset of model categories in the events obtained for the entity. Alternatively, the determined sequence of the entity can include the categories of the subset of model categories a number of times equal to the number of events obtained for the entity associated with the category, i.e., the sequence includes the categories of the subset of model categories as many times as they occur in the associated events obtained for the entity. The recurrence of a category may be important or may not be very relevant. As an example of the presentation of learning of CPP functions, repeated use of a function or a function sequence can indicate that a user would benefit from a particular other function, while a single use can indicate trial or error. As an example of predicting equipment failure, a positive correlation may exist between category and failure, such that the probability of failure is further independent of the number of occurrences of a particular category.

[0038] The step of determining the subset of target categories and the subset of model categories of the set of categories may comprise one or more of the following steps: Preferably, for each entity of the group, an average frequency is calculated. Preferably, an active subgroup of entities is determined based on the calculated average event frequency. The active subgroup of entities preferably comprises entities whose calculated average event frequency is greater than or not less than a predefined frequency cut-off value. Preferably, a subset of model categories of the set of categories is determined based on the occurrence of the categories, preferably within the events obtained for the entities in the active subgroup. Preferably, a subset of model categories of the set of categories is determined based on the occurrence of the categories, preferably within the events obtained for the entities in the active subgroup. occurrence Preferably, an average category frequency is calculated for each category of the set of categories, preferably based on the events obtained for the entities in the active subgroup. Preferably, a subset of the model categories is determined, preferably over the events obtained for the entities in the active subgroup, from the set of categories based on the frequency of the categories. Preferably, an average category frequency is calculated for each category of the set of categories, preferably based on the events obtained for the entities in the active subgroup. Top A subset of the target categories is then determined from the subset of the model categories, such that the subset of the model categories includes the subset of the target categories.

[0039] The step of determining the subset of target categories and the subset of model categories of the set of categories may comprise one or more of the following steps: Preferably, for each entity of the group, an average frequency is calculated. Preferably, an active subgroup of entities is determined based on the calculated average event frequency. The active subgroup of entities preferably comprises entities whose calculated average event frequency is greater than or not less than a predefined frequency cut-off value. The subset of model categories of the set of categories is determined, preferably as described above. Preferably, for each category of the subset of model categories, preferably across the active subgroup of entities, the first occurrence of an event associated with the category is determined. Up to An average duration is calculated. Preferably, a subset of target categories is determined from the subset of model categories based on the average durations determined for the categories of the subset of model categories. Preferably, the subset of target categories is determined based on the first occurrence of the categories of the subset of model categories in an event obtained for an entity in the active subgroup. Up to Based on the time period, a subset of the model categories is determined. Preferably, the subset of the target categories is , Mo Categories that are subsets of Dell categories Of these, the average period determined is the lowest A given number Categories or a specified percentage Categories is determined based on.

[0040] Each sequence of the determined sequence and two or 3 or more The step of calculating the probability that the tail category of the sequence follows the base of the sequence for each of the determined sequences and their recursive bases including the category of the sequence may comprise one or more of the following steps: Preferably, an input table of input sequences is obtained. Preferably, the input table includes the determined sequences. Preferably, the input table includes, for each determined sequence, two or more 3 or moreThe input table further includes all corresponding recursive bases which include the categories of at once, i.e. the input table further includes, for each determined sequence, all corresponding subsequences at once. Preferably, an output table is obtained. Preferably, the output table includes all unique bases of the input sequences of the input table. The output table preferably includes all unique bases of the input sequences of the input table at once. Preferably, for each base of the output table, a number of predecessors is determined. Preferably, the number of predecessors of a base is the number of input sequences in the input table of which the input sequence is a base. Preferably, for an input sequence of the input table, more preferably for each input sequence of the input table, a probability is calculated that the tail category of the input sequence follows the base of the input sequence, preferably based on the number of occurrences of the input sequence in the input table and the corresponding number of predecessors of its base in the output table. Preferably, the probability that the tail category of the input sequence follows the base of the input sequence is calculated by dividing the number of occurrences of the input sequence in the input table by the corresponding number of predecessors of its base in the output table.

[0041] In a preferred embodiment, the event includes a communication channel of a collection of communication channels. A collection of target communication channels is determined for a target entity. Information based on the target entity and the determined target category is preferably provided via the target communication channel. In a first alternative embodiment, the category can include the communication channel and the target communication channel can be determined via the determination of the target category. In a second alternative alternative embodiment, an array of communication channels or an array of category and communication channel pairs can be used to determine the target communication channel.

[0042] The present invention is further described by the following non-limiting examples which further illustrate the invention and are not intended, and should not be construed, as limiting the scope of the invention. [Example] EXAMPLES

[0043] <Embodiments of the present invention> For each entity of a group of entities, one or more events are captured. Each event is associated with a timestamp and a category from the set of categories. A category "occurs" for an entity if an event associated with the category has been captured for the entity. A category has not "yet occurred" for the entity if an event associated with the category has not been captured for the entity.

[0044] For each entity in the group, an average event frequency is calculated. For example, the average event frequency of an entity can be the average number of events per month of the entity over the last year. Also, if the entity has not yet been active for a full year, the average event frequency of the entity can be the average number of events per month of the entity during the active period. An active subgroup of active entities is determined. An "active entity" is an entity whose calculated average event frequency is greater than or not less than a predefined frequency cutoff value. The other entities are referred to as "inactive entities."

[0045] The events captured for active entities (in the active subgroups) are filtered out. For each category in the set of categories, the average frequency of the active category is calculated based on the filtered events, i.e., the events of the active entities in the active subgroups. For example, the average frequency of the active category can be the average number of occurrences of the category per active entity per month. The subset of model categories is ,mosquito A collection of categories Included in category Among them, the calculated average frequency of active categories is the highest. A given number Categories or a specified percentage CategoriesBased on the calculated average frequency of the active categories, the categories of the set of categories can be determined from the set of categories based on the calculated average frequency of the active ...

[0046] For each model category, the first occurrence (of the event associated with the model category) Up to The average period is determined over the active subgroups of active entities. Preferably, for active entities where the model category has not yet occurred, the average period is determined over the first occurrence. Up to A predefined default period value is used to determine the averaging period or the first occurrence Up to Either these active entities are not considered for determining the average period, but most preferably the latter. A subset of target categories is determined from the subset of model categories based on the determined average period for the first occurrence of the model categories in the subset of model categories. Preferably, the subset of target categories is , Mo Dell Category Subset Included in Model Category Among these, the average period determined for the first occurrence is the lowest A given number Categories or a specified percentage Categories For example, the model categories of the subset of model categories can be ordered in the second selection table according to the determined average period for first occurrence. Depending on whether an increasing or decreasing ordering is utilized, the top or bottom of the model categories of the second selection table are retained as a subset of the target categories of the subset of model categories based on a predetermined number of predetermined frequencies. Additionally, the categories of the subset of the target categories are also referred to as target categories.

[0047] <Construction of predictive models> A full track subgroup of full track entities is determined. An entity is a full track entity when, for each object category of the subset of object categories, at least one event associated with the object category has been obtained, i.e., for an entity, each object category has occurred. An entity for which, for at least one object category of the subset of object categories, no event associated with at least one category has been obtained, i.e., at least one object category has not yet occurred, is called a partial track entity. Those skilled in the art will understand that the following four cases may occur: active full track entities, inactive full track entities, active partial track entities, and inactive partial track entities.

[0048] For each all-track entity of the group, an arrangement of model categories of the subset of model categories is determined based on the events obtained for all-track entities, whereby the arrangement is determined based on the chronological occurrence of the model categories in the events obtained for all-track entities, whereby the determined arrangement can include each model category at most once, whereby the model categories of the determined arrangement are ordered according to the chronological first occurrence of the model categories of the subset of model categories in the events obtained for all-track entities, in particular according to the timestamp of the event, i.e. only the first occurrence of the model category is represented in the determined arrangement, whereby the determined arrangement can alternatively include the model category a number of times equal to the number of events obtained for all-track entities associated with the model category, i.e. each occurrence of the model category for all-track entities is represented in the determined arrangement.

[0049] Each sequence of the determined sequence and two or 3 or moreFor those recursive bases that contain the category of (and therefore every base is a sequence), the probability that the tail category of the sequence follows the base of the sequence is calculated. Thus, the following steps are preferably utilized for the calculation of the probabilities: An input table of input sequences is obtained. The input table includes the determined sequences. The input table further includes two or 3 or more A recursive base is obtained that contains all unique bases of the input arrays of the input table at once. For each base in the output table, the number of precedents of the input arrays in the input table for which the input array is a base is determined. For each input array in the input table, the probability that the tail category of the input array follows the base of the input array is calculated by dividing the number of occurrences of the input array in the input table by the corresponding number of precedents of the base in the output table.

[0050] <Partial track entity prediction> For the entities of the partial track of interest, a target sequence of model categories of the subset of model categories is determined based on the events acquired for the entities of the partial track of interest, whereby the sequence is determined based on the chronological occurrence of the model categories in the events acquired for the entities of the partial track of interest. Thus, the determined sequence can include each model category once, whereby the model categories of the determined sequence are ordered according to the chronological first occurrence of the model categories of the subset of model categories in the events acquired for the entities of the partial track of interest, in particular according to the timestamp of the event, i.e. only the first occurrence of a model category is represented in the determined sequence. Thus, the determined sequence can alternatively include model categories a number of times equal to the number of events acquired for the entities of the partial track of interest associated with the model categories, i.e. each occurrence of a model category is represented in the determined sequence.

[0051] For the entities of the partial track of interest, an object category is determined based on the object sequence and the calculated probabilities. The determined object category is in particular a category of a subset of object categories for which no events associated with the object category have been obtained for the entities of the partial track of interest. Thus, the occurrence of the determined object category for the entities of the partial track of interest will bring the entities of the partial track of interest "closer" to being entities of the full track. Preferably, the determined object category includes the maximum probability according to the calculated probabilities following the object sequence.

[0052] <Information provision> Information based on the entities of the partial track of the object and the determined object category is provided. EXAMPLES

[0053] <User Involvement> This example further builds on example 1 and is related to user interaction with an organization. The entities are users. The categories are also called event types. The model categories are also called waypoints. The target categories are also called milestones. The sequences (of the model categories) are also called journeys (of the points). The bases are also called sub-journeys. The full track entities are also called onboarded users. The partial track entities are also called pre-joined users.

[0054] Figure 3 shows a schematic diagram of communication channels and data processing according to an embodiment of the present example. Users can interact with the organization through six communication channels: face-to-face interaction (301), telephone (302), mail (303), online and / or email via the organization's web platform (304), mobile via the organization's mobile or web platform (305), and debit and credit card transactions (306). Each user interaction can be recorded in one of several systems based on a unique user identifier associated with the user. Periodically, preferably at the start of each day, the recorded data from the previous period, preferably the previous day, can be extracted from storage, converted to the correct format, and loaded into a relational database (307).

[0055] New data in the relational database (307) can be converted (308), preferably daily, into events and added (309) to an event log. An event includes a user identifier, an event type (category) from a collection of event types, and a date and time (timestamp). The event log (309) can be queried for events for use in an algorithm (310) according to the present invention.

[0056] For each user identity, the average number of events per month (average event frequency) is determined. User identities associated with an average number of events at or above a predefined threshold (predefined frequency cutoff value) are flagged as active. All other user identities are flagged as inactive. The predefined threshold can be adjusted via a user input device.

[0057] To obtain a viable forecasting algorithm, the number of variables or degrees of freedom is reduced, especially on a computer system with a finite or limited amount of resources. A subset of model categories of points of interest (model categories) is selected. Points of interest are selected by calculating the average number of times each event type occurs per active user identifier, and flagging the most frequent ones as points of interest. This selection is motivated by the fact that the more frequent an event type is in a group of users, the more important that event type is to that group of users. Events of active user identifiers are filtered out, the average monthly frequency of each event type in active subgroups of active users is calculated, and a subset of model categories containing a predetermined number of points of interest is output. The points of interest are the event types with the highest average monthly frequency. The predetermined number of points of interest can be adjusted via a user input device.

[0058] The subset of target categories for waypoints is selected from the subset of model categories for points of interest. The date and time when a user joins an organization and the occurrence of the user's points of interest (first occurrence) are Up to The average amount of time between the first occurrence of a point of interest (average period) is determined. FIG. 4 shows an example chart showing the average period for the first occurrence of a point of interest in increasing order of average period. Point of interest 00 can be selected as the midpoint because it is the least number of days (on average) between the first occurrence of a point of interest and the first occurrence of a user. Decide Determined Lower The selection criterion for the averaging period reflects the timeliness of the information intervention. For other purposes, other criteria can be used.

[0059] A subscribing user identifier is associated with a user who has "achieved" all waypoints, i.e., the event log contains at least one event including the user's identifier and the waypoint for each waypoint. A pre-subscribing user identifier is associated with a user who has "not achieved" all waypoints, i.e., the event log lacks an event including the user's identifier and the waypoint for at least one waypoint. The following procedure is preferably utilized to determine which user's identifiers are subscribing and which user's identifiers are pre-subscribing. A data table is created that contains a column with the user's identifier, as well as a column for each waypoint. The data table entry that corresponds to the column and the user's identifier contains the date and time of the event where the corresponding user first reached the waypoint. If the waypoint has not yet been achieved by the user, the corresponding entry in the data table contains a null entry. For each row, the column associated with the waypoints is looped over, and if a null value is detected, the user's identifier is flagged as pre-subscribing. If no null value is detected, the user's identifier is flagged as subscribing. A flag is entered into an additional column of the data table. The data table is then split into two data tables: a subscription data table containing entries for subscribing user identifiers, and a pre-subscription data table containing entries for pre-subscription user identifiers.

[0060] The subscription data table is then used to build a predictive model for pre-subscription user identifiers. For each user identifier, the first occurrence of a point of interest is mapped in chronological order. For subscription user identifiers, in addition to the journey, two or more 3 or moreAll sub-paths that contain the point information of are also determined. This can be accomplished by isolating the point information of the path and incrementally adding them in order, one at a time, until the path is obtained again. An input table is obtained that contains all the paths and at least two lengths of sub-paths. An output table is determined that contains all the unique bases of the sequences in the input table. The (precedent) number of sequences (paths and at least two lengths of sub-paths) in the input table that correspond to each unique base is determined and added to a separate table. The number of occurrences of each unique sequence in the input table is then determined and divided by the predecessor number of the corresponding base. This results in a counting probability of each unique sequence in the input table across the subscription group of the subscription user. A probability table can be obtained that contains the calculated probability for each unique path and sub-path of the subscription user's identifier. The probability table can be visually represented in a Sankey diagram, an example of which is shown in FIG. 7.

[0061] Refer to FIG. 6. Given a user identifier x, if the data table entry of point information W1 includes the date 2015-09-05, the data table entry of point information W2 includes a null entry, the data table entry of point information W3 includes the date 2014-12-11, the data table entry of point information W4 includes the date 2018-11-09, and the data table entry of point information W5 includes the date 2016-05-02, then the journey of user identifier x can be determined as W3→W1→W5→W4. If user identifier x is a subscribing user identifier, sub-route W3→W1→W5 and W3→W1 are also determined. Given a user identifier y, the journey W3→W1→W4→W5 is determined, and similarly, sub-route W3→W1→W4 and W3→W1 are also determined. Then the input table can be "W3 → W1, W3 → W1 → W5, W3 → W1 → W5 → W4, W3 → W1, W3 → W1 → W4, W3 → W1 → W4 → W5". Then the output table can be "W3, W3 → W1, W3 → W1 → W5, W3 → W1 → W4" with the corresponding predecessor table "2, 2, 1, 1]". Then the probability of the sequence W3 → W1 → W4 is 1 / 2 = 0.5.

[0062] FIG. 5 shows a schematic diagram of an embodiment of an algorithm for calculating the probability of a sequence, which can be executed from start (501) to end (508). Journey data of a subscribed user can be provided (502). The point information of the journey of the subscribed user can be ordered by ascending date (503). All possible sub-journeys made by the subscribed user can be calculated, and the number of times each sequence occurs in the data can be calculated (504). If the journey a→b→c→d→e is obtained for the user's identifier, the sub-journeys are a, a→b, a→b→c, and a→b→c→d (509). The sub-journeys of at least two lengths are a→b, a→b→c, and a→b→c→d. For each sequence (journey or sub-journey of at least two lengths), a base is determined by removing the last point information of the sequence (505). The counting probability of each of the bases transitioning to point information is calculated by dividing the number of times each sequence occurs by the number of predecessors of its corresponding base (506). For example, if base a→b transitions to point information c 10 times and to point information d 5 times, the resulting transition probability is 0.6666... ​​for a→b→c and 0.3333... for a→b→d. The corresponding sequences and probabilities are output in a probability table (507).

[0063] For the pre-subscription user's identifier, a target sequence can be determined based on the pre-subscription data table. A transition probability for each waypoint that the corresponding user has not yet reached is determined based on the calculated probability, i.e., the transition probability is read from the probability table if the calculated probability exists, and is set to zero if the calculated probability does not exist. The waypoint corresponding to the maximum probability is selected. EXAMPLES

[0064] <Information provision> This example further builds on examples 1 and 2.

[0065] When a waypoint is predicted for a pre-joining group of pre-joining users, the user is contacted. The pre-joining group can include pre-joining users for whom a particular waypoint is predicted. The pre-joining group can include all pre-joining users. The pre-joining group can include pre-joining users who have not been contacted recently, i.e., within a predefined contact period. The user's contact data can include a first and second name, a home address, an email address, a telephone number, and / or a mobile phone number. Through an SQL query based on the user's identifier, the predicted waypoint can be combined with the corresponding contact data. The query can be specifically filtered for one or more specific parts of the contact information. The contact information can be associated with a specific event type. For example, if the waypoint is the use of a mobile platform, the associated contact information can be a telephone number for sending an SMS. In this case, the first name, the second name, and the telephone number are obtained.

[0066] Figure 8 shows a schematic diagram of an algorithm for user contact, which can be executed from start (801) to end (807). For each pre-joining user's identifier, the next most likely waypoint is predicted (802) as described in Example 2 above. Contact details of the user's identifier with a waypoint prediction are obtained (803). The obtained contact details are sent to a verification team (804). Upon verification, the contact details are updated in a contact log (805). Based on the contact log, the users are contacted. The contacted users are tracked periodically (806), especially with regard to the achievement of the predicted and / or other waypoints.

[0067] The contact information can be tested via one or more test functions. If all test functions do not pass, manual intervention can be required. If all test functions pass, the contact information can be output as an Excel file for sending to a contact center.

[0068] Figure 9 shows an algorithm including steps that may be involved in creating user contact details and sending an email to a contact center, which may run from start (901) to finish (905). A six character password may be randomly generated (902). An Excel file may be encrypted based on the password (903). An email containing the encrypted Excel file may be sent to the contact center (904). The password may be provided to the contact center via another communication channel via another email (904).

[0069] An event including the user's identifier and the event type can be added to a log of events whereby the event type is associated with the contact. EXAMPLES

[0070] <Evaluation of algorithm performance> This example further builds on examples 1, 2 and 3.

[0071] A control group of users that does not overlap with the (predicted) group of users is obtained. For the users of the control group, unattained waypoints are randomly assigned. Also, the users of the control group are contacted in relation to the randomly selected waypoints. The performance of the invention is evaluated by comparing the results of the group of users with the results of the control group of users.

[0072] Performance can be evaluated based on the achievement of predicted or randomly selected waypoints by the user, either within a predefined time window or as next point information. Preferably, in this embodiment, performance is evaluated based on the achievement of predicted or randomly selected waypoints within a predefined time window from contact, and particularly within the context of this embodiment of a predefined time window of three months from contact.

[0073] 10 shows a chart depicting the percentage of contacted users who meet the per-way achievement of the predicted waypoint (black bars) or the randomly selected waypoint (white bars) separately for the control group (white bars) and the predicted group (black bars). It can be observed that users are significantly more likely to achieve the predicted waypoint than the randomly selected waypoint.

Claims

1. 1. A computer-implemented method for event prediction and information provision, said method comprising: obtaining one or more events for each entity of a group of entities, each event being associated with a timestamp and a category from a set of categories; calculating an average event frequency for each entity in the group; determining an active subgroup of entities based on the calculated average event frequency, the active subgroup of entities including those entities for which the calculated average event frequency is greater than or not less than a predetermined frequency cutoff value; determining a subset of model categories from the set of categories based on the occurrence of the categories in the set of categories in the events obtained for the entities in the active subgroup; For each category in the set of categories, calculating an average frequency of the category based on the events obtained for the entities in the active subgroup; determining a subset of the model categories based on a predetermined number or a predetermined percentage of categories in the set of categories that have the highest average frequency of the calculated categories; determining a subset of target categories from the subset of model categories based on time periods to first occurrences of the subset of model categories in the events obtained for the entities in the active subgroup; For each category in the subset of model categories, calculating the average time to first occurrence of an event associated with that category across the active subgroups of entities; determining a subset of the target categories based on a predetermined number or a predetermined percentage of categories included in the subset of model categories that have the lowest determined average period; determining, for each entity of the group of entities, where the events obtained for the entity include, for each category of the subset of target categories, at least one event associated with the category, an ordering of categories of the subset of model categories based on chronological occurrence of the categories of the subset of model categories in the events obtained for the entity, wherein the ordering consists of a base including one or more categories and a tail category after the base, and a base including two or more categories is itself a sub-array consisting of a base including one or more categories and a tail category after the base; calculating, for each of the determined sequences and its subsequences, the probability that the tail category of the sequence follows the base category of the sequence; For a target entity of the group of entities, wherein the events acquired for the target entity do not include, for at least one category of the subset of target categories, an event associated with the at least one category, determining a target sequence of categories of the subset of model categories based on a chronological occurrence of the categories of the subset of model categories in the events acquired for the target entity, and determining a target category of the at least one category based on the target sequence and the calculated probability; providing information based on the target entities and the determined target categories; A method comprising:

2. 2. The computer-implemented method of claim 1, wherein the determined ordering for an entity includes each category of the subset of model categories at most once, and the determined ordering is based on a first occurrence in chronological order of the categories of the subset of model categories within the events obtained for the entity.

3. 2. The computer-implemented method of claim 1, wherein the determined sequence for an entity includes categories of the subset of model categories a number of times equal to the number of events obtained for the entity associated with that category.

4. 4. The computer-implemented method of claim 1, wherein the step of calculating, for each of the determined sequences and its subsequences, a probability that the tail category of the sequence follows the base of the sequence comprises: obtaining an input table of input sequences that includes the determined sequences and, for each determined sequence, includes all corresponding sub-sequences at once; obtaining an output table containing all unique bases of the input sequences of the input table; for each base of the output table, determining the number of predecessors of the input sequence in the input table for which the input sequence is the base; calculating, for each input sequence in the input table, the probability that the tail category of the input sequence follows the base of the input sequence by dividing the number of occurrences of the input sequence in the input table by the corresponding number of predecessors of its base in the output table; A method comprising:

5. 5. The computer-implemented method of claim 1, wherein the event includes a communication channel of a collection of communication channels, the method including determining, for the target entity, a target communication channel of the collection, and the information based on the target entity and the determined target category is provided via the target communication channel.

6. 6. The computer-implemented method for failure prediction and information provision in a device according to claim 1, wherein the entities are devices including one or more sensors, and the one or more events of each device are obtained based on data from the one or more sensors of the device, and the information is provided for failure avoidance.

7. 6. The computer-implemented method for event prediction and information provision of claim 1, wherein entities are users and categories are functions of a computer program product, and wherein the step of providing information based on the target entities and the target categories is a step of providing learning information about the function of a target to the user of the target via visualization means.

8. 6. A computer-implemented method for predicting a type of interaction by a user with a service and providing information as described in any one of claims 1 to 5, wherein the entity is a user and the category is a type of interaction, and the step of providing information based on the target entity and the target category is a step of providing the user with information about the type of interaction of interest.

9. A computer system for event prediction and information provision, the computer system configured to perform the computer-implemented method of any one of claims 1 to 8.

10. 9. A computer program for event prediction and information provision, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the computer-implemented method of any one of claims 1 to 8.

Images