Prediction method, system and storage medium

Abstract

This specification provides a prediction method, system, and storage medium. In the prediction method, the prediction system obtains an interaction sequence of a target user. This interaction sequence includes representation information of interaction events corresponding to multiple known time steps. Each interaction event includes an interaction item and an interaction behavior. The interaction behavior corresponding to each interaction event belongs to a set of interaction behaviors, and the interaction behaviors in the set of interaction behaviors correspond to multiple intensity levels. The prediction system performs the following actions through a prediction model: determining the attention weight for each known time step based on the intensity levels and / or transition probabilities of the interaction behaviors corresponding to multiple interaction events and the representation information corresponding to multiple interaction events; and predicting the target user's target interaction event at an unknown time step based on the attention weight corresponding to each known time step and the representation information of the interaction events.

Description

Technical Field

[0001] This specification relates to the Internet field, and in particular to a prediction method, system, and storage medium. Background Technology

[0002] With the rapid development of internet technology, e-commerce, social media, content distribution, and other service platforms have emerged in large numbers. To provide users with a wide variety of services, the amount of information each platform can offer has exploded. In this context, recommendation systems have become a key tool connecting users with the massive amounts of information within these platforms. Recommendation systems can analyze users' historical interactions and potential preferences to predict content that users might be interested in. The accuracy of these predictions determines the quality of the user experience and user stickiness, making it a core element of competition among major service platforms.

[0003] The information in the background section is merely information known only to the inventor and does not imply that such information had entered the public domain before the date of this application, nor does it imply that it can be considered prior art in this disclosure. Summary of the Invention

[0004] This specification provides a prediction method, system, and storage medium that assigns attention weights to each historical interaction based on the intensity level and / or transition probability of the user's historical interactions during the prediction process, thereby improving the accuracy of the prediction.

[0005] In a first aspect, this specification provides a prediction method, which includes: obtaining an interaction sequence of a target user, the interaction sequence including representation information of interaction events corresponding to N known time steps, each interaction event including an interaction item and an interaction behavior, N being an integer greater than 1, each interaction behavior corresponding to an interaction event belonging to an interaction behavior set, the interaction behaviors in the interaction behavior set corresponding to multiple intensity levels; and performing the following through a prediction model: determining an attention weight for each known time step based on the representation information corresponding to the N interaction events and at least one of the following: the intensity level of the interaction behavior corresponding to the N interaction events, or the transition probability between the interaction behaviors corresponding to the N interaction events; predicting a target interaction event of the target user at an unknown time step based on the attention weight corresponding to each known time step and the representation information of the interaction events, the target interaction event including a target interaction behavior and a target interaction item.

[0006] In some embodiments, for each known time step, the representation information of the corresponding interactive event is determined based on the following method: obtaining the item representation vector of the interactive item and the behavior representation vector of the interactive behavior corresponding to the known time step; and fusing the item representation vector and the behavior representation vector into a single vector to obtain the representation information of the interactive event corresponding to the known time step.

[0007] In some embodiments, determining the attention weight for each known time step based on the representation information corresponding to N interaction events and the intensity level of the interaction behavior includes: for the i-th known time step, 1≤i≤N, determining a first attention score of the i-th known time step for each historical time step based on the representation information of the interaction events corresponding to the i-th known time step; determining a second attention score of the i-th known time step for each historical time step based on the intensity level of the interaction behavior corresponding to the i-th known time step; and determining the attention weight of the i-th known time step for each historical time step based on the first attention score and the second attention score.

[0008] In some embodiments, the interactive behaviors corresponding to the N interactive events belong to at least one intensity level. Determining the second attention score of the i-th known time step for each historical time step based on the intensity level of the interactive behavior corresponding to each historical time step and the i-th known time step includes: aggregating the representation information of the interactive events corresponding to each historical time step from the at least one intensity level to obtain the context information of the i-th known time step; and for each historical time step, obtaining the item representation vector of the interactive item and the behavior representation vector of the interactive behavior corresponding to the historical time step, and determining the second attention score of the i-th known time step for the historical time step based on the representation information of the interactive events corresponding to the i-th known time step, the item representation vector, the behavior representation vector, and the context information.

[0009] In some embodiments, the context information of the interaction events corresponding to each historical time step is aggregated from the at least one intensity level to obtain the context information of the i-th known time step includes: for each intensity level, determining that the interaction behavior belongs to at least one target historical time step of the intensity level, and aggregating the representation information of the interaction events corresponding to the at least one target historical time step to obtain the context component of the intensity level corresponding to the i-th known time step; and fusing the context component of the at least one intensity level based on the target intensity level to which the interaction behavior corresponding to the i-th known time step belongs to obtain the context information.

[0010] In some embodiments, fusing the context components of at least one intensity level based on the target intensity level of the interaction behavior corresponding to the i-th known time step to obtain the context information includes: dividing the context components of each of the at least one intensity level into a main context component and a secondary context component based on the target intensity level; determining the fusion weights of the main context component and the secondary context component; and determining the context information based on the main context component, the secondary context component and their respective fusion weights.

[0011] In some embodiments, determining the attention weight for each known time step based on the representation information corresponding to N interaction events and the transition probability between interaction behaviors includes: for the i-th known time step, 1≤i≤N, determining a first attention score of the i-th known time step to each historical time step based on the representation information of each historical time step and the interaction events corresponding to the i-th known time step; determining a third attention score of the i-th known time step to each historical time step based on the transition probability between each historical time step and the interaction behaviors corresponding to the i-th known time step; and determining the attention weight of the i-th known time step to each historical time step based on the first attention score and the third attention score.

[0012] In some embodiments, determining the third attention score of the i-th known time step for each historical time step based on the transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step includes: determining the target similarity between each historical time step and the interaction item corresponding to the i-th known time step; determining the target transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step; and for each historical time step, determining the third attention score of the i-th known time step for the historical time step based on the target similarity and the target transition probability.

[0013] In some embodiments, the target similarity is obtained based on at least one of the following: the semantic similarity between the historical time step and the interactive items corresponding to the i-th known time step; the item type similarity between the historical time step and the interactive items corresponding to the i-th known time step; or the similarity of the item representation vectors between the historical time step and the interactive items corresponding to the i-th known time step.

[0014] In some embodiments, determining the target transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step includes: obtaining a preset transition probability matrix for the set of interaction behaviors; and determining the target transition probability based on the interaction behavior corresponding to the historical time step, the interaction behavior corresponding to the i-th known time step, and the transition probability matrix.

[0015] In some embodiments, the third attention score is further determined based on at least one of the following: contextual statistics obtained from multiple statistical dimensions for each historical time step; or the time interval between the occurrence of the interaction event corresponding to each historical time step and the i-th known time step.

[0016] In some embodiments, for the i-th known time step, the plurality of statistical dimensions include at least one of the following: the number of historical time steps; the number of historical time steps corresponding to the same interactive item as the i-th known time step; the number of historical time steps corresponding to the same item type as the i-th known time step; or the number of historical time steps corresponding to the same intensity level as the i-th known time step.

[0017] In some embodiments, predicting the target user's target interaction event at an unknown time step based on the attention weights and representation information of interaction events corresponding to each known time step includes: determining the comprehensive context of the Nth known time step based on the attention weights and representation information of interaction events at each historical time step; and predicting the target interaction event based on the comprehensive context.

[0018] Secondly, this specification also provides a prediction system, comprising: at least one storage medium storing at least one instruction set; and at least one processor communicatively connected to the at least one storage medium, wherein the at least one processor reads the at least one instruction set during operation and executes the method as described in the first aspect according to the instructions of the at least one instruction set.

[0019] Thirdly, this specification provides a computer-readable non-transitory storage medium, wherein the computer-readable non-transitory storage medium stores at least one instruction set, which, when executed by at least one processor, implements the method as described in the first aspect.

[0020] Other features of the prediction methods, systems, and storage media provided in this specification are partially listed in the following description. The inventive aspects of the prediction methods, systems, and storage media provided in this specification can be fully understood through practice or by using the methods, systems, and combinations described in the detailed examples below. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A schematic diagram illustrating a predicted application scenario provided according to embodiments of this specification is shown; Figure 2 A hardware structure diagram of a computing system provided according to an embodiment of this specification is shown; Figure 3 A flowchart is shown of a training method for a prediction model provided according to an embodiment of this specification; Figure 4 A schematic diagram of the structure of a prediction model provided according to an embodiment of this specification is shown; Figure 5 Another structural schematic diagram of the prediction model provided according to an embodiment of this specification is shown; Figure 6 A further structural schematic diagram of the prediction model provided according to embodiments of this specification is shown; and Figure 7 A flowchart of a prediction method provided according to an embodiment of this specification is shown. Detailed Implementation

[0023] The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to make and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0024] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. Unless otherwise stated, the term “a plurality” means two or more, and “at least one” means one or more. Terms such as “first,” “second,” etc., may be used in this specification to describe various information, but such information should not be limited to these terms. These terms are used to distinguish information of the same type from one another and do not necessarily imply a specific order or sequence. For example, “first” may also be referred to as “second” without departing from the scope of embodiments described herein, and similarly, “second” may also be referred to as “first.”

[0025] The term "at least one of A, B, or C" includes seven cases: A only, B only, C only, both A and B, both A and C, both B and C, and both A, B, and C. Similarly, the statement "at least one of multiple items" refers to all possible combinations based on these items. The term "and / or" refers to any or all possible combinations of one or more related listed items. For example, "A and / or B" includes three cases: A only, B only, and both A and B. "A, B, and / or C" is equivalent to "at least one of A, B, or C". The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0026] The term "comprising" is an open-ended description and should be understood as "including but not limited to," potentially including other content beyond what has been described. When used in this specification, the terms "comprising," "including," and / or "containing" mean the presence of the associated integers, steps, operations, elements, and / or components, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or the possibility of adding other features, integers, steps, operations, elements, components, and / or groups to the system / method.

[0027] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0028] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0029] Figure 1 A schematic diagram illustrating a predicted application scenario provided according to embodiments of this specification is shown. For example... Figure 1 As shown, application scenario 100 may include training system 10 and prediction system 20.

[0030] During the training phase, the training system 10 can train a prediction model based on multiple sample sequences. Taking sample sequence 1 as an example, sample sequence 1 can represent multiple interaction events of a user on a certain service platform. Each interaction event includes an interaction behavior and an interaction item. In e-commerce platforms, interaction behaviors include page browsing, clicking, adding to cart (hereinafter referred to as adding to cart), and purchasing; interaction items include different types of goods or services. In the insurance field, interaction behaviors include browsing, clicking, favorites, consultation, underwriting, and purchasing; interaction items include critical illness insurance and medical insurance.

[0031] It is understandable that interactive behaviors and interactive items may differ in different service areas. The following text will mainly focus on the insurance service area.

[0032] After training is complete, the prediction model can be deployed on prediction system 20. During the prediction phase, prediction system 20 (which can be understood as a recommendation system or a subsystem within a recommendation system) detects the prediction task, obtains the target user's interaction sequence, and inputs this sequence into the prediction model. The interaction sequence includes interaction events at N known time steps, and the prediction model can output predicted interaction events after the Nth time step (i.e., at unknown time steps). Therefore, the recommendation system can recommend page content related to the predicted interaction events to the user based on the output of the prediction model.

[0033] In some embodiments, the training system 10 may store data and instructions for implementing training methods of the prediction model, and may execute or be used to execute the data and instructions. In some embodiments, the training system 10 may include hardware devices with data processing capabilities and the necessary programs required to drive the hardware devices.

[0034] In some embodiments, the prediction system 20 may store data and instructions for implementing the prediction method, and may execute or be used to execute the data and instructions. In some embodiments, the prediction system 20 may include hardware devices with data processing capabilities and the necessary programs required to drive the hardware devices.

[0035] It is understood that training system 10 and prediction system 20 may correspond to the same system or different systems, and this specification does not impose any restrictions on this.

[0036] It should be noted that the training system 10 or the prediction system 20 can correspond to a single device or a cluster of devices; this specification does not impose any restrictions. When the training system 10 corresponds to a single device, the training method for the prediction model can be executed entirely on that device. When the training system 10 corresponds to a cluster of devices, the training method for the prediction model can be executed collaboratively on multiple devices corresponding to the device cluster. When the prediction system 20 corresponds to a single device, the prediction method can be executed entirely on that single device. When the prediction system 20 corresponds to a cluster of devices, the prediction method can be executed collaboratively on multiple devices corresponding to the device cluster.

[0037] Figure 2 A hardware structure diagram of a computing system 200 provided according to an embodiment of this specification is shown. The computing system 200 may be... Figure 1 The training system 10 or the prediction system 20 in the system.

[0038] like Figure 2 As shown, the computing system 200 may include at least one storage medium 230 and at least one processor 220. In some embodiments, the computing system 200 may also include a communication port 250 and an internal communication bus 210. Furthermore, the computing system 200 may also include I / O components 260.

[0039] The internal communication bus 210 can connect to different system components. For example, the internal communication bus 210 can connect to storage medium 230, processor 220, communication port 250, and I / O component 260.

[0040] I / O component 260 supports input / output between computing system 200 and other components.

[0041] Communication port 250 is used for data communication between computing system 200 and the outside world. For example, communication port 250 can be used for data communication between computing system 200 and a network. Communication port 250 can be a wired communication port or a wireless communication port.

[0042] In some embodiments, the network can be any type of wired or wireless network, or a combination thereof. For example, the network may include a cable network, a wired network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network™, a ZigBee™ short-range wireless network, a near field communication (NFC) network, or a similar network.

[0043] In some embodiments, the network may include one or more network access points. For example, the network may include wired or wireless network access points, such as base stations or internet switching points. Through these access points, one or more components of various devices corresponding to the computing system 200 can connect to the network to exchange data or information.

[0044] Storage medium 230 may include a data storage device. The data storage device may be a non-transitory storage medium or a temporary storage medium. For example, the data storage device may include one or more of a disk 232, a read-only storage medium (ROM) 234, or a random access storage medium (RAM) 236. Storage medium 230 also includes at least one instruction set stored in the data storage device. The instruction set may include computer program code, which may include programs, routines, objects, components, data structures, procedures, modules, etc., that execute the prediction methods provided in this specification.

[0045] Processor 220 can be communicatively connected to storage medium 230. Processor 220 is used to execute at least one of the above-described instruction sets. When computing system 200 is running, processor 220 reads the at least one instruction set and executes the prediction method provided in this specification according to the instructions of the at least one instruction set.

[0046] Processor 220 may be in the form of one or more processors. In some embodiments, processor 220 may include one or more hardware processors, such as microcontrollers, microprocessors, reduced instruction set computers (RISC), application-specific integrated circuits (ASICs), application-specific instruction set processors (ASIPs), central processing units (CPUs), graphics processing units (GPUs), physical processing units (PPUs), microcontroller units, digital signal processors (DSPs), field-programmable gate arrays (FPGAs), advanced RISC machines (ARMs), programmable logic devices (PLDs), any circuit or processor capable of performing one or more functions, or any combination thereof.

[0047] For the purpose of illustrating the point only, in the appendix Figure 2 Only one processor 220 is shown in the computing system 200. However, it should be noted that the computing system 200 may also include multiple processors. Therefore, the operation and / or method steps disclosed in this specification may be executed by one processor as described in this specification, or they may be executed jointly by multiple processors. For example, if processor 220 of computing system 200 in this specification executes steps A and B, it should be understood that steps A and B may also be executed jointly or separately by two different processors 220 (e.g., the first processor executes step A, the second processor executes step B, or the first and second processors jointly execute steps A and B).

[0048] Figure 3 A flowchart of a training method P300 for a prediction model according to an embodiment of this specification is shown. The training system 10 can execute the training method P300.

[0049] like Figure 3 As shown, the training method P300 includes the following steps.

[0050] S310: Obtain multiple sample sequences, each sample sequence including representation information of interaction events corresponding to multiple known time steps, and each interaction event including interaction items and interaction behaviors.

[0051] The training system 10 can obtain a set of interaction events from multiple users on the service platform. Taking a single user as an example, the user's set of interaction events can include multiple interaction events of that user on the service platform. Multiple interaction events in a set can be arranged according to the time of their occurrence.

[0052] An interaction event consists of an interaction item and an interaction action. Taking interaction event 1 as an example, it can indicate that a user interacted with interaction item 1 through interaction action 1 at time 1. For example, in the insurance service field: interaction event 1 could indicate that a user inquired about critical illness insurance at a certain time, or that a user browsed medical insurance at a certain time. An interaction event can also simply indicate that a user interacted with interaction item 1 through interaction action 1, such as interaction event 1 indicating that a user inquired about critical illness insurance; this manual does not impose such limitations.

[0053] Training system 10 can obtain one or more sample sequences based on each user's set of interaction events. (Set of interaction events) For example:

[0054] Among them, triplet Indicates user In time Through interactive behavior Interactive projects Interaction took place.

[0055] As an example, a user's interaction event set includes 200 interaction events arranged in chronological order. Training system 10 can obtain sample sequence 1 based on interaction events 1 to 199. In sample sequence 1, interaction events 1 to 199 are interaction events with known time steps, and interaction event 200 is an interaction event with an unknown time step (or a time step to be predicted), serving as the training label for sample sequence 1. Training system 10 can also obtain sample sequence 2 based on interaction events 1 to 99, where interaction event 100 is an interaction event with an unknown time step, serving as the training label for sample sequence 2.

[0056] This specification does not limit the number of interaction events included in each sample sequence. Different sample sequences may include the same or different numbers of interaction events. In the following text, interaction events corresponding to unknown time steps during training will be denoted as the true interaction events of the sample sequence, while interaction events corresponding to unknown time steps output by the prediction model during training will be denoted as the predicted interaction events of the sample sequence.

[0057] The training system 10 can pre-map each interactive behavior and each interactive item in the service platform to its corresponding representation vector. For example, the set of interactive items The training system 10 can map each interactive item to a corresponding item representation vector, thus obtaining a set of item representation vectors. In the insurance service field, Indicates browsing, Indicates click, Indicates collection, Indicates consultation, Indicates underwriting, Indicates a purchase. A collection of interactive behaviors. The training system 10 can map each interactive behavior to obtain a behavioral representation vector of the interactive behavior, and then obtain a set of behavioral representation vectors. In the field of insurance services, Indicates critical illness insurance. This refers to medical insurance, and so on. The interactive behaviors in a set of interactive behaviors can correspond to multiple intensity levels.

[0058] In some embodiments, the training system 10 may pre-classify all interactive behaviors in the set of interactive behaviors based on behavior intensity. Behavior intensity reflects the probability that a certain interactive behavior will lead to a user completing a transaction, or the degree of influence / contribution of the interactive behavior on the user's completion of a transaction. For example, in the field of insurance services, the behavior intensity of consultation represents the probability that consultation will lead to a user purchasing insurance, or the degree of influence or contribution of consultation on the user's purchase of insurance.

[0059] For example, in the field of insurance services, training system 10 can classify all interactive behaviors in the set of interactive behaviors according to Table 1.

[0060] Table 1

[0061] Taking a user's purchase of a transaction as an example, low-intensity interactions such as browsing and clicking have a lower impact or contribution to the user's purchase than high-intensity interactions such as adding to favorites, inquiring, underwriting, and purchasing. In other words, low-intensity interactions such as browsing and clicking are less likely to lead to a user's purchase than high-intensity interactions such as adding to favorites, inquiring, underwriting, and purchasing. The training system 10 can further categorize interactions into more intensity levels, for example, classifying underwriting and purchasing as the highest intensity level; this specification does not impose any limitations on this.

[0062] In some embodiments, taking a sample sequence H as an example, it includes representation information of interactive events corresponding to n known time steps, with each known time step corresponding to the representation information of one interactive event. For each known time step, the representation information of the corresponding interactive event is determined based on the following method: the training system 10 obtains the item representation vector of the interactive item corresponding to the known time step and the behavior representation vector of the interactive behavior corresponding to the known time step, and then fuses the item representation vector and the behavior representation vector into a single vector to obtain the representation information of the interactive event corresponding to the known time step.

[0063] The interactive items corresponding to a known time step can be understood as the interactive items included in the interactive events corresponding to that known time step. The interactive behaviors corresponding to a known time step can be understood as the interactive behaviors included in the interactive events corresponding to that known time step. For example, if the interactive event corresponding to the first known time step is "collect critical illness insurance", then the interactive behavior corresponding to the first known time step 1 is "collect", and the interactive item corresponding to the first known time step 1 is "critical illness insurance".

[0064] For example, the training system 10 can fuse the item representation vector and the behavior representation vector using the following formula to obtain the representation information of the interaction event.

[0065] (1) in, This represents the information (or vector) of the interaction event corresponding to the i-th known time step. Let i be the project representation vector of the interactive project corresponding to the i-th known time step. Let be the behavioral representation vector of the interaction behavior corresponding to the i-th known time step. The value of each element in can be and The sum of corresponding elements. In other words, , , The number of rows and columns are the same, that is, the dimensions are the same.

[0066] The training system 10 can obtain the representation information of each of the n+1 interaction events in a similar manner, including the representation information of the interaction events corresponding to the n known time steps and the representation information of the interaction event corresponding to the unknown time step. That is, the training system 10 obtains the sample sequence H and the representation information of the real interaction events at the unknown time step corresponding to the sample sequence H.

[0067] It should be noted that the above example uses the prediction of a single future time step, but the prediction model is not limited to this and can also predict multiple future time steps. The following text mainly uses the prediction of a single future time step as an example. When predicting multiple future time steps, the previously predicted unknown time steps can be treated as known time steps. Other content is similar and will not be repeated here.

[0068] S320: A prediction model is trained based on multiple sample sequences. For each sample sequence, the prediction model can execute S3201 and S3203 to obtain the predicted interaction event corresponding to that sample sequence. The parameters of the prediction model are updated with the goal of minimizing the difference between the real interaction event and the predicted interaction event corresponding to each sample sequence.

[0069] S3201: Determine the attention weight for each known time step based on the representation information corresponding to multiple interaction events involved in the sample sequence and at least one of the following: the intensity level of the interaction behavior corresponding to the multiple interaction events, and the transition probability between the interaction behaviors corresponding to the multiple interaction events.

[0070] For example, sample sequences , This represents the information about the interaction event corresponding to the first known time step, and so on. The prediction model can obtain the probability distribution of different interaction events corresponding to unknown time steps using the following formula, and thus select the interaction event with the highest probability as the predicted interaction event.

[0071] (2) in Represents the sample sequence. This represents all time steps prior to the i-th known time step. This represents the parameters of the prediction model.

[0072] The following uses sample sequence H as an example to illustrate the prediction process of the prediction model based on sample sequence H. In some embodiments, such as Figure 4-6 As shown, the prediction model includes an encoder and a decoder. The encoder consists of multiple identical modules, each containing a self-attention layer. The self-attention layer is used to calculate the correlation between each interaction event and other events in the input sequence. It is a crucial component of the prediction model, allowing it to dynamically assign different attention weights to each known time step (or position) when processing the input sequence, thereby capturing the dependencies between known time steps (or positions) in the sequence.

[0073] During the prediction process, the prediction model can assign attention weights to each known time step in the sample sequence H in a variety of ways. The following mainly uses three of these methods as examples.

[0074] Method 1 In some embodiments, the prediction model determines the attention weight for each known time step based on the representation information corresponding to the n interaction events involved in the sample sequence H and the intensity level of the interaction behavior corresponding to the n interaction events.

[0075] For example, such as Figure 4 As shown, the self-attention layer of the prediction model includes a multi-head self-attention network and a hierarchical behavior aggregation (HBA) network. During computation, the multi-head self-attention network can be based on HW...Q Generate matrix Q (query), based on HW K Generate matrix K (keys) based on HW V Generate matrix V (values). Where W... Q W K W V This is the linear transformation matrix (or linear projection matrix) that the prediction model can learn.

[0076] The sample sequence H includes representations of n interaction events. For the i-th (1≤i≤n) known time step, the multi-head self-attention network can determine the first attention score of the i-th known time step for each historical time step based on the representations of the interaction events corresponding to each historical time step and the representations of the interaction events corresponding to the i-th known time step. Alternatively, the multi-head self-attention network can determine the first attention score of the i-th known time step for each historical time step based on the representations of the interaction events corresponding to each historical time step and the representations of the interaction events corresponding to the i-th known time step, as well as the time step distance (or time interval) between each historical time step and the interaction events corresponding to the i-th known time step.

[0077] A historical time step is relative to the i-th known time step. It can be understood as any time step before the i-th known time step, such as any time step from the 1st known time step to the (i-1)th known time step. It can also include the i-th known time step, which is not a limitation in this specification. The following text mainly uses the example of including the i-th known time step.

[0078] In the self-attention mechanism, each current time step (position i) can generate a query vector Q. i The historical time step (position j) can generate the key vector K. j The attention score of the i-th known time step to each historical time step can be understood as follows: when the i-th known time step is used as the query time step, the attention score assigned to each historical time step represents the degree of dependence of the i-th known time step on each historical time step, or the importance, influence, or contribution of each historical time step to the i-th known time step. The higher the attention score (or attention weight) assigned by the prediction model to a certain historical time step, the stronger the correlation between that historical time step and the i-th known time step.

[0079] For example, a multi-head self-attention network determines the first attention score based on the following formula. .

[0080] (3) Among them, Q i K represents the vector in matrix Q corresponding to the i-th known time step.j Let d represent the vector in matrix K corresponding to historical time step j (i.e., the j-th known time step), and let d represent the number of columns in matrix Q or K, i.e., the vector dimension.

[0081] In some embodiments, while the training system 10 inputs the sample sequence H into the prediction model, it can also input the intensity level of each interaction behavior corresponding to the sample sequence H (i.e., the interaction behavior corresponding to all known time steps in the sample sequence H) into the prediction model. The prediction model sequentially performs query prediction for each known time step in H. For the i-th known time step, the prediction model can determine the second attention score of the i-th known time step to each historical time step based on the intensity level of the interaction behavior corresponding to each historical time step and the intensity level of the interaction behavior corresponding to the i-th known time step. Then, it can determine the attention weight of the i-th known time step to each historical time step based on the first attention score and the second attention score.

[0082] For example, the interaction events involved in the sample sequence H are: {Browse critical illness insurance A, Click critical illness insurance A, Browse medical insurance B, Add medical insurance B, Browse critical illness insurance C, Consult critical illness insurance A}. Then, the intensity level of the interaction behavior corresponding to all known time steps in the sample sequence H can be represented by the vector [0,0,0,1,0,1], where "1" represents a high intensity level and "0" represents a low intensity level. This specification does not restrict the representation of the intensity level of the input prediction model.

[0083] In some embodiments, the interactive behaviors corresponding to the n interactive events involved in the sample sequence H belong to M (integers greater than or equal to 1) intensity levels. The prediction model aggregates the representation information of the interactive events corresponding to each historical time step from the M intensity levels to obtain the context information of the i-th known time step. For each historical time step, the prediction model obtains the item representation vector of the interactive item and the behavior representation vector of the interactive behavior corresponding to that historical time step, and determines the second attention score of the i-th known time step to the historical time step based on the representation information of the interactive events corresponding to the i-th known time step, the item representation vector, the behavior representation vector, and the context information.

[0084] The M intensity levels can be all intensity levels of the interaction behavior set, or only a portion of them; this specification does not impose any restrictions on this. The following explanation will primarily use the example of these M intensity levels including low and high intensity levels.

[0085] For example, for the i-th time step, the HBA network determines the second attention score based on the following formula. .

[0086] (4) in, Represents a multilayer sensing function. This represents the representation information of the interaction event corresponding to the i-th known time step. This indicates the concatenation of vectors. This represents the behavior representation vector of the interaction behavior corresponding to historical time step j. This represents the project representation vector of the interactive project corresponding to historical time step j. This represents the context information for the i-th known time step. This indicates that for the i-th known time step, the prediction model only focuses on... Using historical time steps can reduce the impact of time steps j > i on prediction, improve prediction accuracy, and also improve computational efficiency.

[0087] In some embodiments, for each of the M intensity levels, the prediction model determines at least one target historical time step belonging to that intensity level for the interaction behavior, and aggregates the representation information of the interaction events corresponding to the at least one target historical time step to obtain the context component corresponding to the intensity level for the i-th known time step. Furthermore, the prediction model can fuse the context components of the M intensity levels based on the target intensity level to which the interaction behavior corresponding to the i-th known time step belongs to obtain the context information for the i-th known time step.

[0088] For example, the HBA network includes two parallel aggregation channels: a high-intensity aggregation channel and a low-intensity aggregation channel. The low-intensity aggregation channel aggregates the representation information of interaction events belonging to the low-intensity level in each historical time step, thereby obtaining the low-intensity context component corresponding to the i-th known time step. The high-intensity aggregation channel aggregates the representation information of interaction events belonging to the high-intensity level in each historical time step, thereby obtaining the high-intensity level context component corresponding to the i-th known time step. Low-intensity interaction events are those that contain low-intensity interactive behaviors. High-intensity interaction events are those that contain high-intensity interactive behaviors.

[0089] In some embodiments, for each intensity level, the prediction model masks other time steps besides the target historical time step, and then aggregates the representation information of the interaction events corresponding to the unmasked time steps in the n known time steps of the sample sequence H.

[0090] For example, the low-intensity aggregation channel generates a mask matrix corresponding to the low-intensity level according to formula (5). And determine the low-intensity context component corresponding to the i-th known time step according to formula (6). . It aggregates low-intensity purchase signals from users.

[0091] (5) Here, the time steps where j is greater than i, and the time steps where j is less than or equal to i and the corresponding interaction behavior belongs to the high intensity level, are the mask positions (the mask position takes values ​​of...). ), This indicates that the interaction behavior corresponding to historical time step j belongs to the low intensity level.

[0092] (6) For example, the high-intensity aggregation channel generates a mask matrix corresponding to the high-intensity level according to formula (7). And determine the high-intensity context component corresponding to the i-th time step according to formula (8). . It aggregates high-intensity purchase signals from users.

[0093] (7) Here, the time steps where j is greater than i, and the time steps where j is less than or equal to i and the corresponding interaction behavior belongs to the low intensity level, are the mask positions (the mask positions take values ​​of...). ), This indicates that the interaction behavior corresponding to historical time step j belongs to the high-intensity level.

[0094] (8) In some embodiments, the prediction model divides the context components corresponding to each intensity level into primary context components and secondary context components based on the target intensity level corresponding to the i-th known time step. The target intensity level is the intensity level to which the interactive behavior included in the interactive event corresponding to the i-th known time step belongs.

[0095] For example, the HBA network uses the gating mechanism given in formula (9) to divide the context components corresponding to different intensity levels into main context components. and auxiliary context components : (9) in, express The vector corresponding to the i-th known time step. express The vector corresponding to the i-th known time step. When the target intensity level corresponding to the i-th known time step is low intensity level, for , for ,or for Main, As a secondary concatenation of vectors or matrices, for Main, A concatenated vector or matrix, used as a secondary element. When the target intensity level corresponding to the i-th known time step is a high intensity level... for , for ,or, for Main, As a secondary concatenation of vectors or matrices, for Main, The concatenation of vectors or matrices as secondary components is not limited in this specification.

[0096] After obtaining the primary and secondary context components, the prediction model determines their respective fusion weights. As an example, the prediction model determines the fusion weights of the primary context component based on formula (10). Therefore, the fusion weights of the auxiliary context components are determined as follows: . ∈[0,1] d .

[0097] (10) in, and These are the parameters that the prediction model can learn during training.

[0098] Furthermore, the prediction model determines the context information for the i-th time step based on the main context component, the auxiliary context component, and their respective fusion weights. As an example, the prediction model determines the context information for the i-th known time step based on formula (11). .

[0099] (11) in, This indicates a product operation.

[0100] For the i-th known time step, the prediction model determines the attention weights of the i-th known time step to each historical time step based on the first attention score and the second attention score.

[0101] For example, the self-attention layer assigns a target attention score to historical time step j based on the following formula (12). Attention weights are assigned to historical time step j based on formula (13). .

[0102] (12) (13) in, These are the parameters that the prediction model can learn during training.

[0103] Method 2 In some embodiments, the prediction model determines the attention weight for each known time step based on the representation information corresponding to n interaction events in the sample sequence H and the transition probability between the interaction behaviors corresponding to the n interaction events.

[0104] For example, such as Figure 5 As shown, the self-attention layer includes a multi-head self-attention network and a transition relation encoding (TRE) network. The TRE network can learn historical time steps. The transition correlation between the i-th known time step and the i-th time step. The transition probability from interaction behavior 1 to interaction behavior 2 can represent: the likelihood of a user transitioning from interaction behavior 1 to interaction behavior 2, or the probability or frequency of a user transitioning from interaction behavior 1 to interaction behavior 2.

[0105] The sample sequence H includes representation information of n interaction events. For the i-th known time step, the multi-head self-attention network calculates the first attention score of the i-th known time step to the historical time step j based on formula (3). For the detailed calculation process, please refer to Method 1, which will not be repeated here.

[0106] In some embodiments, for the i-th known time step, the prediction model determines the third attention score of the i-th known time step to the historical time step j based on the transition probability between the interaction behavior corresponding to each historical time step and the interaction behavior corresponding to the i-th known time step. Therefore, the attention weight of the i-th known time step for each historical time step can be determined based on the first attention score and the third attention score.

[0107] For example, the TRE network determines the target similarity between each historical time step and the corresponding interactive item at the i-th known time step, and determines the target transition probability between each historical time step and the corresponding interactive behavior at the i-th known time step. For each historical time step, the TRE network determines the third attention score of the i-th known time step to the historical time step based on the target similarity and the target transition probability.

[0108] For example, the target similarity between each historical time step j (j≤i) and the interactive item corresponding to the i-th known time step is obtained based on at least one of the following: the semantic similarity between historical time step j and the interactive item corresponding to the i-th known time step; the item type similarity between historical time step j and the interactive item corresponding to the i-th known time step; or the similarity of the item representation vectors between historical time step j and the interactive item corresponding to the i-th known time step.

[0109] The training system 10 can pre-classify different interactive items. Taking the insurance service field as an example, critical illness insurance A and critical illness insurance B can be classified as interactive items of the same category.

[0110] Specifically, for each historical time step j, the TRE network can determine the target similarity between historical time step j and the interactive item corresponding to the i-th known time step using formula (14). . It can also represent the semantic proximity between historical time step j and the interactive item corresponding to the i-th known time step.

[0111] (14) in, This represents the semantic similarity between historical time step j and the interactive item corresponding to the i-th known time step. This represents the semantic similarity between the historical time step j and the project type of the interactive item corresponding to the i-th known time step. Let represent the cosine distance between the historical time step j and the item representation vector corresponding to the i-th known time step.

[0112] If the interactive item corresponding to the historical time step j is the same as that corresponding to the i-th known time step, then The maximum value is achieved if the historical time step j and the i-th known time step correspond to different interactive items, but belong to the same item type. The value is in the middle. If the historical time step j and the i-th known time step correspond to different interactive items and belong to different item types, then... The value is the smallest.

[0113] In some embodiments, for a sample sequence H, the training system 10 can obtain semantic representation information of the interactive items corresponding to each known time step, and semantic representation information of the item type to which the interactive items belong at each known time step. The aforementioned semantic representation information can be represented by vectors and may differ from the item representation vector obtained by the training system 10 mapping each interactive item; this specification does not impose any limitations on this. While inputting the sample sequence H into the prediction model, the training system 10 can also input the semantic representation information of the interactive items and their respective item types corresponding to each known time step into the prediction model.

[0114] For example, the sequence of interactive events corresponding to sample sequence H is: {Browse critical illness insurance A, Click critical illness insurance A, Browse medical insurance B, Add medical insurance B, Browse critical illness insurance C, Consult critical illness insurance A}. The semantic representation information of all interactive items represented by sample sequence H can be represented by the vector [1,1,0,0,2,1], which is not limited in this specification. The semantic representation information of the item category to which all interactive items represented by sample sequence H belong can also be represented by the vector [1,1,0,0,1,1], which is not limited in this specification.

[0115] In some embodiments, for each historical time step, the prediction model can obtain a preset transition probability matrix (which can be provided by the training system) for the set of interaction behaviors, and then determine the target transition probability between the historical time step and the interaction behavior corresponding to the i-th known time step based on the interaction behavior corresponding to the historical time step, the interaction behavior corresponding to the i-th known time step, and the transition probability matrix.

[0116] In the insurance service field, the interactive behaviors executed by users in their actual decision-making process regarding whether to purchase insurance generally follow a logical sequence. For example, the user's decision-making process follows a one-way funnel logic of browsing - clicking - saving - consulting - underwriting - purchasing. When calculating the third attention score, the TRE network in the predictive model can determine which historical time steps are more relevant for the i-th known time step. For instance, if the interaction corresponding to the i-th known time step is "underwriting," the TRE network assigns a higher third attention score to the historical time step corresponding to the same interaction item or category, and whose corresponding interaction behavior is "consulting," because "consulting" is a direct preceding behavior to "underwriting." Conversely, if the interaction corresponding to the i-th known time step is "underwriting," the TRE network assigns a lower third attention score to the historical time steps corresponding to the same interaction item or category, and whose corresponding interaction behaviors are "clicking," "browsing," etc., because "clicking," "browsing," etc., are indirect preceding behaviors to "underwriting." The TRE network assigns the lowest third attention score to historical time steps corresponding to different interaction items or different categories of interaction items.

[0117] It is understandable that when the number of transition layers between historical time step j and the interaction corresponding to the i-th known time step is positive, the larger the number of transition layers, the lower the third attention score assigned to historical time step j by the TRE network. When the number of transition layers between historical time step j and the interaction corresponding to the i-th known time step is negative, the larger the number of transition layers, the higher the third attention score assigned to historical time step j by the TRE network, or the same attention score can be assigned to historical time steps corresponding to different numbers of transition layers. This specification does not impose any restrictions on this. For example, if the i-th known time step corresponds to "favorite" and historical time step j corresponds to "click", then the transition layer between them is +1; if the i-th known time step corresponds to "favorite" and historical time step j corresponds to "browse", then the transition layer between them is +2; if the i-th known time step corresponds to "click" and historical time step j corresponds to "favorite", then the transition layer between them is -1, and so on.

[0118] The transition probability matrix can be pre-set by the training system 10, which can adjust the transition probabilities between different interaction behaviors in the transition probability matrix during the training of the prediction model. For example, the training system 10 pre-defines... In the insurance service field, the following can also be defined: Browse < Click < Favorite < Consult < Underwriting < Purchase.

[0119] (15) The prediction model can be adjusted during training to values ​​such as 0.8, 0.5, and 0.2, and the transition probability matrix... For example: Browse, click, save, consult, underwrite, purchase Views [0.50.20.2 0.20.20.2 ] Click [0.80.5 0.20.20.20.2] Add to favorites [0.8 0.8 0.5 0.2 0.2 0.2] Consultation [0.8 0.8 0.5 0.2 0.2] Underwriting [0.8 0.8 0.8 0.5 0.2] Purchase [0.8 0.8 0.8 0.8 0.5 ] The first row [0.5 0.2 0.2 0.2 0.2 0.2] represents the probability that, when the interaction behavior corresponding to the i-th known time step is "browse", the corresponding interaction behavior is [browse, click, favorite, consult, underwrite, purchase]. The second row [0.8 0.5 0.2 0.2 0.2 0.2] represents the probability that, when the interaction behavior corresponding to the i-th known time step is "click", the corresponding interaction behavior is [browse, click, favorite, consult, underwrite, purchase]. And so on.

[0120] For example, if the TRE network determines that the interaction behavior corresponding to the i-th known time step is "consultation" and the interaction behavior corresponding to the historical time step j is "favorite", then the transition probability between the historical time step j and the interaction behavior corresponding to the i-th known time step is determined to be 0.8. If the TRE network determines that the interaction behavior corresponding to the i-th time step is "consultation" and the interaction behavior corresponding to the historical time step j is "consultation", then the transition probability between the historical time step j and the interaction behavior corresponding to the i-th known time step is determined to be 0.5, and so on.

[0121] It should be noted that the probability transition matrix above is only an example, and will be used in the example. In all cases, the transition probability is set to a uniform 0.8. The training system 10 can also set different transition probabilities for interactions with different transition layers based on the number of transition layers between time step j and time step i. For example, if the number of transition layers between time step j and time step i is 1, then... The value is 0.9; the number of transition layers between time step j and time step i is 2; then... If the transition value is 0.8 and the number of transition layers between time step j and time step i is 3, then... It is 0.7, and so on.

[0122] In some embodiments, for the i-th known time step, the prediction model can determine the transfer correlation of the interaction behavior corresponding to the i-th known time step and the historical time step j based on formulas (16) and (17). , It can represent the attention allocated from the perspective of behavioral shifting, and is used to determine the third attention score.

[0123] (16) (17) in, These are the parameters that the prediction model can train during the training process.

[0124] In some embodiments (10), the third attention score may also be determined based on at least one of the following: ① contextual statistics obtained from multiple statistical dimensions for each historical time step, or ② the time interval between the occurrence of the interaction event corresponding to each historical time step and the i-th known time step.

[0125] It should be noted that the prediction model can also determine independent attention scores based on ① and / or ②, such as the fourth attention score, for the purpose of fusion calculation of attention weights. Alternatively, the attention allocated based on ① and / or ② can be fused into the calculation process of the second attention score. This specification does not impose any restrictions on this.

[0126] For example, the TRE network can also determine the transfer relevance based on at least one of ① and ②, and then determine the third attention score based on the transfer relevance. The following mainly introduces the example where the transfer relevance is also determined based on ① and ②. The way the transfer relevance is also determined based on ① or ② is similar, the difference being that one of the terms is missing in the relevant formula.

[0127] In some embodiments, the TRE network determines the time decay effect of the interaction event corresponding to the historical time step when determining the third attention score. For the i-th known time step, the TRE network can determine the transfer correlation between the interaction behavior corresponding to the i-th known time step and the historical time step j based on formulas (17), (18), and (19). : (18) (19) in, This represents the Sigmoid function. This represents the time interval between the interaction event at historical time step j and the interaction event at the i-th known time step. The logarithmic function allows the TRE network to learn the importance of smaller time intervals, such as a 1-hour time interval being more important than a 12-hour time interval.

[0128] It is understandable that, while inputting the sample sequence H into the prediction model, the training system 10 can also input the occurrence time of the interaction event corresponding to each known time step in the sample sequence H into the prediction model, so that the prediction model can calculate based on this. .

[0129] In some embodiments, the third attention score is also determined based on contextual statistics. For the i-th known time step, the plurality of statistical dimensions include: statistical dimension 1, the number of historical time steps (e.g., the number of time steps covered by the 1st to the i-th known time steps); statistical dimension 2, the number of historical time steps corresponding to the same interaction item as the i-th known time step; statistical dimension 3, the number of historical time steps corresponding to the same item type as the i-th known time step; or statistical dimension 4, the number of historical time steps corresponding to the same intensity level (referring to the intensity level of the interaction behavior) as the i-th known time step.

[0130] For the i-th known time step, the prediction model can obtain statistical information based on the multiple statistical dimensions. i Alternatively, the training system 10 may obtain statistical information based on the multiple statistical dimensions. i and will i Input the prediction model. The prediction model is then based on... i Determine the contextual statistics CX. The contextual statistics CX include statistical information obtained from all known time steps in the sample sequence H, which can reflect the user's preferences for interactive items and the intensity distribution of interactive behaviors. i It includes statistical values ​​obtained from each statistical dimension, representing the statistics of its historical time steps for the i-th known time step.

[0131] For example, the TRE network can also concatenate the behavior representation vector of the interaction behavior, the item representation vector of the interaction item, and statistical information corresponding to the i-th known time step to obtain contextual statistics CX. For instance, for the i-th known time step, the prediction model determines the contextual statistics CX based on the following formula. i This allows us to obtain the CX corresponding to all time steps in the sample sequence H. i CX is obtained by splicing.

[0132] (20) It is understandable that, while training system 10 inputs the sample sequence H into the prediction model, it can also input the behavior representation vector of the interaction behavior and the item representation vector of the interaction item corresponding to each known time step in the sample sequence H into the prediction model. It can also input the behavior representation vector of the interaction behavior and the item representation vector of the interaction item corresponding to each known time step into the prediction model. i Input the prediction model.

[0133] After obtaining contextual statistics, the prediction model can spatially map these statistics to obtain a context matching score. For example, the TRE network determines the context matching score based on the following formula:

[0134]

[0135] (twenty one) in, express The vector corresponding to the i-th known time step; express The vector corresponding to historical time step j; The context matching score between the i-th known time step and the historical time step j represents the availability of the context statistics of the historical time step j for the predictive model to query the i-th known time step.

[0136] For the i-th known time step, the TRE network can determine the transfer correlation of the interaction behavior corresponding to the i-th known time step and the historical time step j based on formula (22). : (twenty two) in, τ represents the parameters that the model can learn and adjust during training, while τ represents the preset parameters in the TRE network that do not need to be adjusted during training.

[0137] TRE network determines transfer correlation Then, the third attention score for the interaction event corresponding to historical time step j at the i-th known time step can be obtained by performing the following formula. : (twenty three) in, These are the parameters that the prediction model can learn and adjust during training.

[0138] For the i-th known time step, the prediction model determines the attention weights corresponding to each historical time step for the i-th known time step based on the first attention score and the third attention score.

[0139] For example, the predictive model (such as a self-attention layer) assigns a target attention score to historical time step j based on the following formula (24). Attention weights are assigned to historical time step j based on formula (13). .

[0140] (twenty four) in, These are the parameters that the prediction model can learn during training.

[0141] Method 3 In some embodiments, such as Figure 6 As shown, the self-attention layer includes a multi-head self-attention network, an HBA network, and a TRE network. For the i-th known time step, the multi-head self-attention network determines its first attention score for each historical time step, the HBA network determines its second attention score for each historical time step, and the TRE network determines its third attention score for each historical time step. Further, the self-attention layer determines the attention weights of the i-th time step for each historical time step based on the first, second, and third attention scores.

[0142] The methods for determining the first attention score, second attention score, and third attention score are as described above and will not be repeated here.

[0143] The prediction model can assign a target attention score to historical time step j based on the following formula (25). Attention weights are assigned to historical time step j based on formula (13). .

[0144] (25) The softmax function mentioned in the formula above can also be replaced with the sigmoid function, and this manual does not restrict this.

[0145] In summary, the prediction model can obtain the attention weights for each historical time step when predicting unknown time steps based on methods 1, 2, or 3.

[0146] S3203: Based on the attention weights and representation information of interaction events corresponding to each known time step, predict the user's predicted interaction events at unknown time steps corresponding to the sample sequence. The predicted interaction events include predicted interaction behaviors and predicted interaction items.

[0147] In some embodiments, the prediction model can obtain the comprehensive context of the nth known time step based on the attention weights and representation information of the interaction events corresponding to each historical time step, and then obtain the predicted interaction events based on the comprehensive context.

[0148] For example, the prediction model can obtain the comprehensive context of the nth known time step based on the following formula. : (26) in, Let V be the vector corresponding to historical time step j.

[0149] For example, the prediction model can adopt a transform architecture. Unlike the traditional transform architecture, the self-attention layer of the prediction model can include not only a multi-head self-attention network but also an HBA network or a TRE network. The prediction model obtains the comprehensive context at the nth known time step. Subsequently, other modules within the transform architecture can be used to determine the predicted interaction events. For example, based on the similarity between the integrated context and each item representation vector in the item representation vector set, and the similarity between the integrated context and each behavior representation vector in the behavior representation vector set, the prediction model determines the probability of the combination of the interaction item corresponding to any item representation vector and the interaction behavior corresponding to any behavior representation vector as the predicted interaction event at an unknown time step. Thus, the combination of the interaction behavior and interaction item with the highest probability is selected as the predicted interaction event. Using a similar approach, the prediction model can obtain the predicted interaction events corresponding to each sample sequence.

[0150] In some embodiments, the prediction model updates its parameters with the goal of minimizing the difference between the real and predicted interaction items for each sample sequence, as well as the difference between the real and predicted interaction behaviors for each sample sequence.

[0151] For example, the prediction model updates its parameters using the following loss function L.

[0152] (27) in, The loss term corresponding to the interaction behavior represents the difference between the actual interaction item and the predicted interaction item; The loss term for the interaction item represents the difference between the actual interaction item and the predicted interaction item. A fixed empirical value can be used, such as 0.3, but this manual does not impose any restrictions on this.

[0153] Figure 7 A flowchart of a prediction method P400 provided according to an embodiment of this specification is shown. The prediction system can deploy a prediction model trained based on training method P300, and then call the prediction model to execute prediction method P400.

[0154] like Figure 7 As shown, the prediction method P400 includes the following steps.

[0155] S410: Obtain the interaction sequence of the target user. The interaction sequence includes the representation information of the interaction events corresponding to N known time steps. Each interaction event includes an interaction item and an interaction behavior. N is an integer greater than 1. The interaction behavior corresponding to each interaction event belongs to the interaction behavior set. The interaction behaviors in the interaction behavior set correspond to multiple intensity levels.

[0156] The prediction system can be a recommendation system within a service platform or a subsystem within a recommendation system. After a target user performs a target action (such as the various interactive behaviors listed above), the prediction system obtains the target user's interaction sequence and calls the prediction model to trigger a prediction. Alternatively, the service platform may pre-prepare multiple recommendation items for each user. Once all the platform-prepared recommendation items have been viewed by the target user, the prediction system can also obtain the target user's interaction sequence and call the prediction model to trigger a prediction. This specification does not impose any restrictions on the timing of prediction triggering.

[0157] For example, the prediction system can obtain a set of user interaction events, which includes multiple interaction events of the target user on the service platform prior to the prediction trigger time. An interaction event includes an interaction item and an interaction behavior.

[0158] In some embodiments, for each known time step, the prediction system obtains the item representation vector of the interactive item corresponding to the known time step and the behavior representation vector of the interactive behavior corresponding to the known time step, and merges the item representation vector and the behavior representation vector into a single vector to obtain the representation information of the interactive event corresponding to the known time step.

[0159] In this embodiment, the prediction system fuses item representation vectors and behavior representation vectors, effectively addressing the challenges posed by sequence inflation. Traditional methods, which alternate items and behaviors as independent tokens, not only significantly increase computational complexity but also waste storage resources. Vector fusion avoids these problems. Furthermore, in long sequences, if items and behaviors are arranged independently, the prediction model often struggles to accurately distinguish between items and behaviors belonging to the same interaction event, easily leading to incorrect associations of behaviors with unrelated items. This not only reduces the accuracy of item and behavior representations but also negatively impacts prediction results. The fusion of item and behavior representation vectors improves representation accuracy, enabling the prediction model to more precisely capture items and behaviors of the same event, thereby effectively enhancing prediction accuracy.

[0160] S420: Perform S4201 and S4203 using the prediction model.

[0161] S4201: Determine the attention weight for each known time step based on the representation information corresponding to N interaction events and at least one of the following: the intensity level of the interaction behavior corresponding to the N interaction events, or the transition probability between the interaction behaviors corresponding to the N interaction events.

[0162] In some embodiments, for the i-th known time step, 1≤i≤N, the prediction model determines the first attention score of the i-th known time step to each historical time step based on the representation information of the interaction event corresponding to each historical time step and the representation information of the interaction event corresponding to the i-th known time step; it determines the second attention score of the i-th known time step to each historical time step based on the intensity level of the interaction behavior corresponding to each historical time step and the intensity level of the interaction behavior corresponding to the i-th known time step; and then determines the attention weight of the i-th known time step to each historical time step based on the first attention score and the second attention score.

[0163] In this embodiment, by distinguishing the intensity levels of different interactive behaviors, the prediction model can significantly enhance its ability to perceive the user's true intentions (such as the shallow exploration stage or the deep decision-making stage). Specifically, for the i-th time step, the prediction model, based on the attention weights calculated based on content similarity, further introduces an intensity-aware attention bias, thereby strengthening the attention to interactions with the same intensity level in historical time steps. This mechanism achieves selective activation dependent on behavior intensity, enabling the model to assign more precise attention weights to each historical interaction, thus effectively improving the accuracy of prediction.

[0164] In some embodiments, the interactive behaviors corresponding to the N interactive events in the interaction sequence belong to at least one intensity level. For the i-th (1≤i≤N) known time step, the prediction model aggregates the representation information of the interactive events corresponding to each historical time step from the at least one intensity level to obtain the context information of the i-th known time step. Then, for each historical time step j, the prediction model obtains the item representation vector of the interactive item and the behavior representation vector of the interactive behavior corresponding to historical time step j, and determines the second attention score of the i-th known time step to historical time step j based on the representation information of the interactive events corresponding to the i-th known time step, the item representation vector, the behavior representation vector, and the context information.

[0165] In this embodiment, the prediction model aggregates interactions at different intensity levels, avoiding the problem of high-intensity interaction events being masked by low-intensity interaction events, or vice versa. The aggregated contextual information can provide macro-level guidance during the prediction process. Furthermore, when determining attention weights, the model also considers the independent representation vectors and fusion vectors of items and behaviors. This ensures that the model follows the macro-logic of behavior intensity while preserving the original characteristics of each behavior and item during attention allocation. This allows for further differentiation of subtle differences between different behaviors or items based on macro-level guidance, improving the accuracy of attention allocation.

[0166] For example, taking intensity level 1 as an example, the prediction model determines that the corresponding interactive behavior belongs to at least one target historical time step of intensity level 1, and aggregates the representation information of the interactive events corresponding to the at least one target historical time step to obtain the context component of intensity level 1 corresponding to the i-th known time step. After obtaining the context component corresponding to each intensity level of N interactive events in a similar manner, the prediction model fuses the context components of each of the at least one intensity level of the N interactive events based on the target intensity level to which the interactive behavior belongs to the i-th known time step, to obtain the context information of the i-th known time step.

[0167] In this example, the prediction model fuses contextual components based on the target intensity level corresponding to the i-th known time step. This allows it to adaptively capture contextual components that are more important to the i-th known time step, thereby fusing more accurate contextual information and improving the accuracy of the second attention score.

[0168] For example, after the prediction model obtains the context component corresponding to each intensity level, it divides each context component into a main context component and a secondary context component based on the target intensity level, determines the fusion weights of the main context component and the secondary context component, and then determines the context information of the i-th known time step based on the main context component, the secondary context component and their respective fusion weights.

[0169] In this example, the prediction model flexibly adjusts the dependence of the context components corresponding to different intensity levels according to the target intensity level corresponding to the i-th known time step. By fusing the main and auxiliary vectors with different fusion weights, the dominance of the main vector can be guaranteed, while useful information in the auxiliary vector can be used, making the context information clear and comprehensive.

[0170] In some embodiments, for the i-th known time step, 1≤i≤N, the prediction model determines the first attention score of the i-th known time step to each historical time step based on the representation information of the interaction event corresponding to each historical time step and the representation information of the interaction event corresponding to the i-th known time step; it determines the third attention score of the i-th known time step to each historical time step based on the transition probability between the interaction behavior corresponding to each historical time step and the interaction behavior corresponding to the i-th known time step; and then determines the attention weight of the i-th known time step to each historical time step based on the first attention score and the third attention score.

[0171] In this embodiment, by utilizing the transition probabilities between different interactive behaviors, the prediction model can accurately capture the causal dependencies and funnel logic in the user's decision-making path, thereby prioritizing historical interactions with logical precursor relationships in attention calculation. Specifically, for the i-th time step, the model, based on the attention weights calculated according to content similarity, further introduces an attention bias reflecting the behavior transition relationship, enhancing the focus on behaviors that constitute decision precursors in the past. This mechanism achieves selective dependency activation of behavior transition logic, enabling the model to assign attention weights that better align with the decision-making path logic to each historical interaction, thereby effectively improving the accuracy of prediction.

[0172] In some embodiments, for the i-th known time step, the prediction model determines the target similarity between each historical time step and the corresponding interactive item at the i-th known time step, and determines the target transition probability between each historical time step and the corresponding interactive behavior at the i-th known time step. Then, for each historical time step, the prediction model determines the third attention score of the i-th known time step on the historical time step based on the target similarity and the target transition probability.

[0173] The predictive model jointly considers target similarity and target transition probability, which can filter out historical time steps similar to the project at the i-th known time step, and identify historical time steps that conform to the user's decision-making logic under similar projects. It avoids noise interference from historical interactions of different projects or different project categories at the i-th known time step, thereby accurately capturing the behavioral evolution pattern of user decision-making paths for the same or similar projects in complex behavioral sequences and improving the accuracy of the third attention score.

[0174] In some embodiments, the target similarity is obtained based on at least one of the following: the semantic similarity between the historical time step and the interactive item corresponding to the i-th known time step; the item type similarity between the historical time step and the interactive item corresponding to the i-th known time step; or the similarity of the item representation vectors between the historical time step and the interactive item corresponding to the i-th known time step.

[0175] In this embodiment, determining the target similarity from multiple perspectives can more accurately and meticulously distinguish different interactive behaviors of the same project type, as well as the differences between interactive behaviors of different project types. The similarity of project representation vectors can capture a more nuanced degree of similarity between projects through semantic measurement in continuous space, making up for the shortcomings of discrete markers such as the project itself and project type, and achieving more accurate mining of relationships between projects.

[0176] In some embodiments, the prediction model can obtain a preset transition probability matrix for the set of interaction behaviors (i.e., the transition probability matrix optimized and adjusted by the prediction model during training), and then determine the target transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step based on the interaction behavior corresponding to each historical time step, the interaction behavior corresponding to the i-th known time step, and the transition probability matrix.

[0177] In some embodiments, the third attention score of the i-th known time step to the historical time step may also be determined based on at least one of the following: ① contextual statistics obtained from multiple statistical dimensions for each historical time step; ② the time interval between the occurrence of the interaction events corresponding to each historical time step and the i-th known time step.

[0178] It should be noted that the attention allocated by the prediction model based on ① and / or ② can be integrated into the calculation process of the second attention score, or into the calculation process of the third attention score, or an attention score can be calculated separately. This specification does not impose any restrictions on this.

[0179] In the above embodiments, incorporating statistical information into the attention score calculation imbues the model with a global behavioral pattern awareness capability. This allows attention allocation to move beyond the content similarity of a single interaction and dynamically adjust based on the cumulative state of the user's historical behavior. Consequently, the prediction model can distinguish between accidental clicks and sustained attention, and between casual browsing and strong intent. This complementary macro-perspective based on statistical features and micro-perspective based on content enables the attention mechanism to make predictions that better reflect the current user state, based on an understanding of the dynamic evolution of user behavior.

[0180] In the above embodiments, the introduction of event time intervals into the attention score calculation enables the prediction model to possess dynamic timeliness perception capabilities. On the one hand, time interval information can effectively distinguish between recent and long-term behaviors, thereby assigning higher weight to short-term explosive interests while appropriately attenuating long-term drifting interests to avoid outdated signals interfering with current decisions. On the other hand, through time interval information, the model can adaptively learn behavioral patterns across different time spans and capture the differentiated semantics that the same time interval may have for different items (for example, the decision-making cycle for purchasing medical insurance and critical illness insurance is 3 days, but this duration may have different meanings for the two types of insurance). Thus, in attention allocation, it both respects the temporal logic of behavior occurrence and flexibly responds to the time dependencies specific to each item, achieving more accurate predictions. Introducing event time intervals into the attention score calculation enables the prediction model to possess dynamic timeliness perception and capture capabilities.

[0181] For example, for the i-th known time step, the above-mentioned multiple statistical dimensions include at least one of the following: the number of historical time steps; the number of historical time steps corresponding to the same interactive item as the i-th known time step; the number of historical time steps corresponding to the same item type as the i-th known time step; or the number of historical time steps corresponding to the same intensity level as the i-th known time step.

[0182] In some embodiments, for the i-th known time step, the prediction model can determine the attention weights of the i-th known time step to each historical time step based on the first attention score, the second attention score, and the third attention score.

[0183] S4203: Based on the attention weights and representation information of interaction events corresponding to each known time step, predict the target user's target interaction events at unknown time steps. The target interaction events include target interaction behaviors and target interaction items.

[0184] In some embodiments, the prediction model determines the comprehensive context of the Nth known time step based on the attention weights and representation information of the interaction events at each historical time step, and then predicts the target interaction event corresponding to the unknown time step based on the comprehensive context.

[0185] It should be noted that the implementation of all embodiments in prediction method 400 can refer to the prediction process of the prediction model for sample sequence H in training method 300, including but not limited to: the way the training system determines the sample sequence, all the contents of the training system inputting the prediction model for each sample sequence, the structure of the prediction model, the calculation process of different networks in the prediction model, the explanation of terms, formulas (n can be replaced with N when referring to the formulas), etc.

[0186] After the predictive model obtains the target interaction event, the service platform can push relevant content matching the target interaction event to the target user. For example, if the target interaction event is browsing critical illness insurance, the service platform will push information about critical illness insurance to the target user. As another example, if the target interaction event is saving critical illness insurance, the service platform will push information about purchasing channels and methods for critical illness insurance to the target user.

[0187] In summary, the prediction method and system provided in this specification significantly enhance the ability to perceive users' true intentions (such as shallow exploration or deep decision-making stages) by distinguishing the intensity levels of different interactive behaviors. The prediction model accurately captures causal dependencies and funnel logic in the user's decision-making path through the transition probabilities between different interactive behaviors, thus prioritizing historical interactions with logical precursors in attention calculation. Furthermore, in determining attention weights, the prediction model introduces an intensity-perceived attention bias and / or a behavior transition relationship bias, in addition to calculating attention weights based on content similarity. This enhances attention to interactions with the same intensity level and / or constituting decision precursors in the past, achieving selective dependency activation of behavior intensity and / or behavior transition logic. This allows the model to assign more accurate attention weights to each historical interaction, thereby effectively improving prediction accuracy.

[0188] This specification, in another aspect, provides a computer-readable non-transitory storage medium storing at least one set of executable instructions for prediction. When the executable instructions are executed by a processor, they instruct the processor to perform the steps of the prediction method P400 described herein. In some possible embodiments, various aspects of this specification may also be implemented as a program product comprising program code. When the program product is run on a computing system 200, the program code causes the computing system 200 to perform the steps of the method P300 described herein. The program product for implementing the above method may employ a portable compact disc read-only memory (CD-ROM) containing program code and may run on the computing system 200. However, the program product of this specification is not limited thereto. In this specification, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system. The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can send, propagate, or transmit programs for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing the operations described herein can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on computing system 200, partially on computing system 200, as a standalone software package, partially on computing system 200 and partially on a remote electronic device, or entirely on a remote electronic device.

[0189] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0190] In summary, after reading this detailed disclosure, those skilled in the art will understand that the foregoing detailed disclosure is presented by way of example only and is not restrictive. Although not explicitly stated herein, those skilled in the art will understand that this specification requires various reasonable changes, improvements, and modifications to the embodiments. These changes, improvements, and modifications are intended to be made by this specification and are within the spirit and scope of the exemplary embodiments described herein.

[0191] Furthermore, certain terms in this specification have been used to describe embodiments of this specification. For example, "an embodiment," "an embodiment," and / or "some embodiments" mean that a particular feature, structure, or characteristic described in connection with that embodiment may be included in at least one embodiment of this specification. Therefore, it is to be emphasized and understood that two or more references to "an embodiment" or "an embodiment" or "alternative embodiment" in various parts of this specification do not necessarily refer to the same embodiment. Moreover, specific features, structures, or characteristics may be suitably combined in one or more embodiments of this specification.

[0192] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and aiding in the understanding of a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art will readily identify some of them as separate embodiments when reading this specification. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. It is also valid when each secondary embodiment contains fewer than all the features of a single foregoing disclosed embodiment.

[0193] Every patent, patent application, publication of a patent application, and other material cited herein, such as articles, books, specifications, publications, documents, and literature (excluding any related historical examination documents), is referenced for all purposes relevant to this document, including in the specification and claims herein. However, in the event of any inconsistency or conflict between the descriptions, definitions, and / or terms used in the foregoing and those used herein, the descriptions, definitions, and / or terms used herein shall prevail.

[0194] Finally, it should be understood that the embodiments disclosed herein are illustrative of the principles of the embodiments described in this specification. Other modified embodiments are also within the scope of this specification. Therefore, the embodiments disclosed in this specification are merely examples and not limitations. Those skilled in the art can implement the applications described in this specification using alternative configurations based on the embodiments in this specification. Therefore, the embodiments in this specification are not limited to the embodiments precisely described in the applications.

Claims

1. A prediction method, comprising: Obtain the interaction sequence of the target user, the interaction sequence including representation information of interaction events corresponding to N known time steps, each interaction event including an interaction item and an interaction behavior, where N is an integer greater than 1, and the interaction behavior corresponding to each interaction event belongs to an interaction behavior set, the interaction behaviors in the interaction behavior set corresponding to multiple intensity levels; and Executed via a predictive model: The attention weight for each known time step is determined based on the representation information corresponding to N interaction events and at least one of the following: the intensity level of the interaction behavior corresponding to the N interaction events, or the transition probability between the interaction behaviors corresponding to the N interaction events. Based on the attention weights and representation information of interaction events corresponding to each known time step, predict the target interaction events of the target user at unknown time steps. The target interaction events include target interaction behaviors and target interaction items.

2. The method according to claim 1, wherein, For each known time step, the representation information of the corresponding interaction event is determined based on the following method: Obtain the item representation vector and the behavior representation vector of the interaction behavior corresponding to the known time step; and The project representation vector and the behavior representation vector are fused into a single vector to obtain the representation information of the interaction event corresponding to the known time step.

3. The method according to claim 1, wherein, The attention weight for each known time step is determined based on the representation information corresponding to N interaction events and the intensity level of the interaction behavior, including: For the i-th known time step, 1≤i≤N, Based on the representation information of the interaction events corresponding to each historical time step and the i-th known time step, determine the first attention score of the i-th known time step for each historical time step; Based on the intensity level of the interaction behavior corresponding to each historical time step and the i-th known time step, determine the second attention score of the i-th known time step for each historical time step; and The attention weight of the i-th known time step for each historical time step is determined based on the first attention score and the second attention score.

4. The method according to claim 3, wherein, The interactive behaviors corresponding to the N interactive events belong to at least one intensity level. Determining the second attention score of the i-th known time step for each historical time step based on the intensity level of the interactive behavior corresponding to each historical time step and the i-th known time step includes: The representation information of the interaction events corresponding to each historical time step is aggregated from the at least one intensity level to obtain the context information of the i-th known time step; and For each historical time step, obtain the item representation vector of the interactive item and the behavior representation vector of the interactive behavior corresponding to the historical time step, and determine the second attention score of the i-th known time step to the historical time step based on the representation information of the interactive event corresponding to the i-th known time step, the item representation vector, the behavior representation vector and the context information.

5. The method according to claim 4, wherein, The representation information of the interaction events corresponding to each historical time step is aggregated from the at least one intensity level to obtain the context information of the i-th known time step, including: For each intensity level, at least one target historical time step belonging to that intensity level is determined, and the representation information of the interaction events corresponding to the at least one target historical time step is aggregated to obtain the context component of the intensity level corresponding to the i-th known time step; and The context information is obtained by fusing the context components of at least one intensity level based on the target intensity level of the interaction behavior corresponding to the i-th known time step.

6. The method according to claim 5, wherein, The context information is obtained by fusing the context components of at least one intensity level based on the target intensity level of the interaction behavior corresponding to the i-th known time step, including: Based on the target intensity level, the context components of each of the at least one intensity level are divided into a primary context component and a secondary context component; Determine the fusion weights of the primary context component and the secondary context component; and The context information is determined based on the primary context component, the secondary context component, and their respective fusion weights.

7. The method according to claim 1, wherein, The attention weights for each known time step are determined based on the representation information corresponding to N interaction events and the transition probabilities between interaction behaviors, including: For the i-th known time step, 1≤i≤N, Based on the representation information of the interaction events corresponding to each historical time step and the i-th known time step, determine the first attention score of the i-th known time step for each historical time step; Based on the transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step, determine the third attention score of the i-th known time step for each historical time step; and The attention weight of the i-th known time step for each historical time step is determined based on the first attention score and the third attention score.

8. The method according to claim 7, wherein, The determination of the third attention score of the i-th known time step for each historical time step based on the transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step includes: Determine the target similarity between each historical time step and the interactive item corresponding to the i-th known time step; Determine the target transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step; and For each historical time step, the third attention score of the i-th known time step to the historical time step is determined based on the target similarity and the target transition probability.

9. The method according to claim 8, wherein, The target similarity is obtained based on at least one of the following: The semantic similarity between the historical time step and the interactive item corresponding to the i-th known time step; The similarity of project types between the historical time step and the interactive project corresponding to the i-th known time step; or The similarity between the project representation vectors of the historical time step and the interactive project corresponding to the i-th known time step.

10. The method of claim 8, wherein, Determining the target transition probability between each historical time step and the interaction behavior corresponding to the i-th known time step includes: Obtain a preset transition probability matrix for the set of interaction behaviors; and The target transition probability is determined based on the interaction behavior corresponding to the historical time step, the interaction behavior corresponding to the i-th known time step, and the transition probability matrix.

11. The method of claim 8, wherein, The third attention score is also determined based on at least one of the following: Contextual statistical information obtained from statistical analysis of each historical time step across multiple statistical dimensions; or The time interval between the occurrence of the interactive event corresponding to each historical time step and the i-th known time step.

12. The method according to claim 11, wherein, For the i-th known time step, the plurality of statistical dimensions include at least one of the following: The number of historical time steps; The number of historical time steps corresponding to the same interactive item as the i-th known time step; The number of historical time steps of the same project type corresponding to the i-th known time step; or The number of historical time steps with the same intensity level as the i-th known time step.

13. The method according to claim 1, wherein, The prediction of the target user's target interaction events at unknown time steps based on the attention weights and representation information of interaction events corresponding to each known time step includes: The comprehensive context of the Nth known time step is determined based on the attention weights and representation information of interaction events at each historical time step; and The target interaction event is predicted based on the comprehensive context.

14. A prediction system, comprising: At least one storage medium storing at least one instruction set; as well as At least one processor is communicatively connected to the at least one storage medium, wherein the at least one processor reads the at least one instruction set during operation and performs the method as described in any one of claims 1-13 according to the instructions of the at least one instruction set.

15. A computer-readable non-transitory storage medium, wherein, The computer-readable non-transitory storage medium stores at least one set of instructions, which, when executed by at least one processor, implement the method as described in any one of claims 1-13.

Images