Target object state evaluation method and device, electronic equipment, medium and program product

Abstract

The application provides a target object state evaluation method and device, electronic equipment, medium and program product, which can be applied to the field of artificial intelligence technology. The method comprises the following steps: obtaining a time sequence feature sequence and a static feature vector authorized by a target user; performing time sequence coding on the time sequence feature sequence to obtain a time sequence state representation corresponding to each time step; performing static coding on the static feature vector to obtain a static state representation; generating a fusion gate vector based on the time sequence state representation; performing weighted processing on the static state representation based on the fusion gate vector, and performing fusion processing on the weighted static state representation and the time sequence state representation corresponding to the time step to obtain a fusion state representation corresponding to each time step; and arranging the fusion state representations corresponding to the time steps in a time step sequence to obtain a fusion state sequence, and generating a state evaluation result of the target user based on the fusion state sequence.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, to the application of large models in business processing, and more specifically to a method, apparatus, device, medium, and program product for evaluating the state of a target object. Background Technology

[0002] Currently, deep learning technology has been widely applied in scenarios such as marketing services, customer status assessment, and service fit analysis. Recurrent neural networks and their variants, with their ability to model temporal dependencies, are often used to process sequential data within continuous time windows; the introduction of attention mechanisms further enhances the model's ability to focus on key time steps and key features. Among existing methods, some schemes process customer information through predictive models and output account opening recommendations or information push results; others extract feature relationships at different levels in parallel using cross-networks, linear models, and deep neural networks to generate customer intention prediction results.

[0003] However, the above methods still have shortcomings: on the one hand, some methods treat customer information as a whole and input it into the prediction model at once, making it difficult to make full use of the dynamic changes generated by users within a continuous time window, resulting in limited ability of the model to characterize the evolution of user state; on the other hand, although some methods can extract feature relationships at different levels, they are still biased towards static modeling as a whole, and tend to treat user features as time-independent fixed values, making it difficult to capture the phased features of user state changes over time. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide a method, apparatus, device, medium, and program product for evaluating the state of a target object.

[0005] According to a first aspect of this application, a method for evaluating the state of a target object is provided. The method includes: acquiring a temporal feature sequence and a static feature vector authorized by a target user, wherein the temporal feature sequence is used to characterize the dynamic changes of the target user within a continuous time window, and the static feature vector is used to characterize the stable attribute information of the target user within the continuous time window; performing temporal encoding on the temporal feature sequence to obtain a temporal state representation corresponding to each time step; performing static encoding on the static feature vector to obtain a static state representation; generating a fusion gating vector based on the temporal state representation; performing weighted processing on the static state representation based on the fusion gating vector, and fusing the weighted static state representation with the temporal state representation of the corresponding time step to obtain a fused state representation corresponding to each time step; and arranging the fused state representations corresponding to each time step in the order of time steps to obtain a fused state sequence, and generating a state evaluation result of the target user based on the fused state sequence.

[0006] According to an embodiment of this application, obtaining the time-series feature sequence and static feature vector authorized by the target user includes: acquiring dynamic data authorized by the target user according to a preset time granularity and performing time alignment processing; generating time-series features corresponding to multiple time steps based on the aligned dynamic data to form the time-series feature sequence; and determining the stable attribute information from the attribute information based on the frequency or period of change of the attribute information authorized by the target user within the continuous time window, and generating the static feature vector based on the stable attribute information.

[0007] According to an embodiment of this application, the method further includes: establishing a time-series input channel corresponding to the time-series feature sequence and a static input channel corresponding to the static feature vector; establishing a time-step mapping relationship between the static feature vector and multiple time steps, wherein the time-step mapping relationship is used to determine the fusion position of the static state representation at each time step; and inputting the time-series feature sequence and the static feature vector into the corresponding time-series encoder and static encoder based on the time-series input channel, the static input channel and the time-step mapping relationship.

[0008] According to an embodiment of this application, generating a fusion gating vector based on the temporal state representation includes: acquiring multiple trained state prototype vectors, the state prototype vectors being used to characterize different state evolution stages of the target user within the continuous time window; determining the state attribution weight corresponding to the current time step based on the similarity between the temporal state representation of the current time step and each state prototype vector; performing weighted processing on the multiple state prototype vectors based on the state attribution weights to obtain the state basis vector corresponding to the current time step; and performing gating mapping processing on the state basis vector to obtain the fusion gating vector corresponding to the current time step.

[0009] According to an embodiment of this application, the step of performing gating mapping on the state basis vector to obtain the fusion gating vector corresponding to the current time step includes: constructing a conditional context representation corresponding to the current time step based on the temporal state representation of the current time step, the state change amount between adjacent time steps, and the local temporal context; inputting the conditional context representation into a gating network to obtain a scaling factor corresponding to the current time step; and adjusting the state basis vector element by element based on the scaling factor to obtain the fusion gating vector corresponding to the current time step.

[0010] According to an embodiment of this application, the static state representation is weighted based on the fusion gating vector, and the weighted static state representation is fused with the temporal state representation of the corresponding time step to obtain the fused state representation for each time step. This includes: determining the fusion position of the static state representation at the current time step based on the time step mapping relationship; projecting the static state representation to obtain a static representation to be fused; and weighting the static representation to be fused element-wise based on the fusion gating vector corresponding to the current time step, and fusing the weighted static representation to be fused with the temporal state representation of the current time step according to the fusion position to obtain the fused state representation corresponding to the current time step.

[0011] According to an embodiment of this application, fusing the weighted static representation to be fused with the temporal state representation of the current time step to obtain the fused state representation corresponding to the current time step includes: generating a feedback modulation term based on the temporal state representation of the current time step; weighting the feedback modulation term based on the fusion gating vector corresponding to the current time step to obtain a gated feedback representation; and fusing and normalizing the temporal state representation of the current time step, the weighted static representation to be fused, and the gated feedback representation to obtain the fused state representation corresponding to the current time step.

[0012] According to an embodiment of this application, generating a state assessment result for the target user based on the fused state sequence includes: inputting the fused state sequence into an aggregation layer in a state assessment model to obtain an overall state representation of the target user; inputting the overall state representation into at least one assessment head in the state assessment model to obtain a state assessment value; and generating a state assessment result for the target user based on the state assessment value, wherein the state assessment result includes at least one of a state level, a service adaptation result, or a resource configuration reference result.

[0013] A second aspect of this application provides a target object state assessment device, the device comprising: a data acquisition module, configured to: acquire a temporal feature sequence and a static feature vector authorized by a target user, wherein the temporal feature sequence is used to characterize the dynamic change information of the target user within a continuous time window, and the static feature vector is used to characterize the stable attribute information of the target user within the continuous time window; an encoding module, configured to: perform temporal encoding on the temporal feature sequence to obtain a temporal state representation corresponding to each time step; perform static encoding on the static feature vector to obtain a static state representation; a fusion module, configured to: generate a fusion gating vector based on the temporal state representation; perform weighted processing on the static state representation based on the fusion gating vector, and perform fusion processing on the weighted static state representation and the temporal state representation of the corresponding time step to obtain a fused state representation corresponding to each time step; and a result generation module, configured to: arrange the fused state representations corresponding to each time step in the order of time steps to obtain a fused state sequence, and generate a state assessment result of the target user based on the fused state sequence.

[0014] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0015] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0017] According to embodiments of this application, the dynamic change information and stable attribute information of a target user within a continuous time window can be encoded separately, and a fusion gating vector can be generated based on the temporal state representation at each time step to dynamically control the injection intensity of the static state representation at the corresponding time step. Compared to simply concatenating temporal features and static features or fusing them with fixed weights, embodiments of this application can adaptively adjust the degree of influence of stable attribute information according to different temporal states, allowing the static state representation to participate in the fusion process with different intensities at different time steps, thereby more accurately depicting the conditional dependency relationship between the user's dynamic change information and stable attribute information. The resulting fused state sequence retains the evolutionary characteristics of the user's state over time and introduces stable attribute constraints or supplementary information that match the current temporal context, thereby improving the accuracy of the target user's state representation and the stability of the state evaluation results, which helps to improve the reliability of application results such as marketing service adaptation, resource allocation reference, or state level determination. Attached Figure Description

[0018] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0019] Figure 1 The illustration schematically depicts an application scenario of the target object state assessment method, apparatus, device, medium, and program product according to embodiments of this application;

[0020] Figure 2 A flowchart illustrating a target object state evaluation method according to an embodiment of this application is shown schematically.

[0021] Figure 3 The flowchart illustrates a method for generating fusion gating vectors based on temporal state representations according to some exemplary embodiments of this application.

[0022] Figure 4 The flowchart illustrates a method for gating mapping of state basis vectors according to some exemplary embodiments of this application.

[0023] Figure 5 A schematic diagram illustrating the structure of a target object state evaluation apparatus according to an embodiment of this application is shown; and

[0024] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a target object state assessment method according to an embodiment of this application. Detailed Implementation

[0025] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0026] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0027] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0028] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by a person skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0029] First, the terminology used in this article will be described and explained.

[0030] Temporal feature sequence: refers to a feature sequence formed after time alignment and feature processing of dynamic data authorized by the target user according to a preset time granularity, used to characterize the dynamic changes of the target user within a continuous time window. The dynamic change information may include changes in access, transactions, interactions, service usage, or other data that changes over time.

[0031] Static feature vector: This refers to a feature vector generated based on attribute information authorized by the target user, used to characterize the stable attribute information of the target user within a continuous time window. The stable attribute information can be determined based on the frequency or period of change of the attribute information within the continuous time window, and is typically used to represent relatively stable user attributes.

[0032] Continuous time window: refers to a data observation interval consisting of multiple consecutive time steps. Within this interval, the system acquires dynamic data and attribute information of the target user, and generates time-series feature sequences and static feature vectors based on the data within this interval.

[0033] Time step: refers to a time unit obtained by dividing a continuous time window according to a preset time granularity. Each time step corresponds to a temporal feature and can correspond to a temporal state representation and a fused state representation.

[0034] State prototype vector: A vector representing the different state evolution stages of a target user within a continuous time window. Multiple state prototype vectors can be obtained through training and used to calculate the state assignment weights for the current time step.

[0035] State basis vector: refers to the vector obtained by weighting multiple state prototype vectors based on state assignment weights, and is used as the basic representation for generating fusion gating vectors.

[0036] It's important to note that the term "neural network" can refer to a machine learning network based on deep learning. A neural network processes input and provides corresponding output, typically consisting of an input layer, an output layer, and one or more hidden layers between them. Neural networks used in deep learning applications often include many hidden layers, increasing the network's depth. The layers of a neural network are connected sequentially, so that the output of the previous layer serves as the input to the next layer. The input layer receives the input to the neural network, while the output layer's output becomes the final output. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each processing the input from the layer above.

[0037] It should be understood that machine learning generally includes three phases: training, testing, and application (also known as inference). In the training phase, a given model is trained using a large amount of training data, iteratively updating parameter values ​​until the model can consistently generate inferences that meet the expected goals from the training data. Through training, the model can be considered to have learned the relationship between inputs and outputs (also known as the input-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing phase, test inputs are applied to the trained model to test whether it can provide the correct output, thus determining the model's performance. In the application phase, the model can be used to process actual inputs based on the trained parameter values ​​to determine the corresponding output.

[0038] In recent years, deep learning technology has been widely applied in scenarios such as marketing services, customer status assessment, service suitability analysis, and resource allocation decision support. In these scenarios, systems typically need to evaluate a user's current status or subsequent service suitability based on dynamic changes within a continuous time window and relatively stable user attribute information. Dynamic changes usually exhibit clear temporal order and evolutionary patterns, such as a user's access, transactions, income and expenditure, interactions, or service usage at different times. Stable attribute information, on the other hand, typically remains relatively stable over a longer period, such as a user's basic attributes, risk level, preference type, or account attributes. Effectively integrating these two types of heterogeneous features is a crucial factor affecting the accuracy and stability of user status assessment.

[0039] In existing solutions, recurrent neural networks (RNNs) and their variants have become common methods for processing sequential data due to their ability to model temporal dependencies. These methods can extract trends in dynamic data in chronological order and, to some extent, capture state dependencies over longer time spans. The introduction of attention mechanisms further enhances the model's ability to focus on key time steps and key features, allowing the model to assign different importance weights across multiple time steps, thereby improving the expressive power of sequential features. In addition, some technical solutions extract feature relationships at different levels in parallel using cross-networks, linear models, and deep neural networks to obtain user intentions, service needs, or state evaluation results.

[0040] However, the existing methods mentioned above still have certain shortcomings. Some methods treat customer or user information as a single, holistic input for prediction models, relying primarily on static snapshot-like features for evaluation. This makes it difficult to fully utilize the dynamic changes in user information generated within continuous time windows. For marketing service scenarios with significant temporal evolution characteristics, user states are not fixed but continuously change with recent interactions, transactions, visits, asset changes, or service response status. If the model cannot effectively characterize this dynamic evolution process, it will be difficult to accurately identify the changing trends of user states at different time stages, thus affecting the reliability of results such as service adaptation, resource allocation, or reach priority.

[0041] Other methods, while extracting feature relationships at different levels through multiple models or network structures, still tend towards static modeling, often treating user features as time-independent fixed values. This makes it difficult to reflect the phased changes and contextual differences in user state over time. Especially when both temporal features and stable attribute information exist, existing solutions typically lack differentiated processing mechanisms for these two types of heterogeneous features. A common approach is to simply concatenate temporal and static features and input them uniformly into the model, or to repeatedly introduce the same static attribute information at each time step. While this approach is simple to implement, it keeps the stable attribute information at a fixed strength throughout the temporal evaluation process, failing to adaptively adjust its influence based on the dynamic state of the current time step.

[0042] In reality, the role of stable attribute information in user state assessment is not entirely the same at different time steps or different stages of state evolution. When user dynamic changes exhibit different trends, abrupt changes, or local patterns, the same stable attribute information may need to play different degrees of constraint, supplementation, or modulation roles. If static attributes are always involved in the calculation with a fixed strength, they may be overemphasized at some time steps and fail to play a full role at others, making it difficult for the model to accurately characterize the conditional dependency between dynamic changes and stable attribute information. As a result, the expressive power of user state representation is limited, and the stability of state assessment results and the accuracy of service adaptation results are also affected.

[0043] Based on this, embodiments of this application provide a method for evaluating the state of a target object, comprising: acquiring a temporal feature sequence and a static feature vector authorized by the target user, wherein the temporal feature sequence is used to characterize the dynamic change information of the target user within a continuous time window, and the static feature vector is used to characterize the stable attribute information of the target user within a continuous time window; performing temporal encoding on the temporal feature sequence to obtain a temporal state representation corresponding to each time step; performing static encoding on the static feature vector to obtain a static state representation; generating a fusion gating vector based on the temporal state representation; performing weighted processing on the static state representation based on the fusion gating vector, and fusing the weighted static state representation with the temporal state representation of the corresponding time step to obtain a fused state representation corresponding to each time step; and arranging the fused state representations corresponding to each time step in the order of time steps to obtain a fused state sequence, and generating a state evaluation result of the target user based on the fused state sequence. According to embodiments of this application, the dynamic change information and stable attribute information of the target user within a continuous time window can be encoded separately, and a fusion gating vector can be generated based on the temporal state representation of each time step to dynamically control the injection intensity of the static state representation at the corresponding time step. Compared to simply concatenating temporal and static features or fusing them with fixed weights, the embodiments of this application can adaptively adjust the degree of influence of stable attribute information according to different temporal states. This allows static state representations to participate in the fusion process with varying intensities at different time steps, thereby more accurately depicting the conditional dependency between dynamic user information and stable attribute information. The resulting fused state sequence retains the evolutionary characteristics of user states over time while introducing stable attribute constraints or supplementary information that match the current temporal context. This improves the accuracy of target user state representation and the stability of state evaluation results, contributing to enhanced reliability of application results such as marketing service adaptation, resource allocation reference, or state level determination.

[0044] It should be noted that the target object status assessment method, apparatus, device, medium, and program product identified in this application can be used in the fields of artificial intelligence technology and fintech, and can also be used in various other fields besides artificial intelligence technology and fintech. The application fields of the target object status assessment method, apparatus, device, medium, and program product provided in the embodiments of this application are not limited.

[0045] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0046] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points for users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process enters the expert decision-making process.

[0047] Figure 1 The illustration shows an application scenario diagram of the target object state assessment method, apparatus, device, medium, and program product according to embodiments of this application.

[0048] like Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a database 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the database 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101 and the second terminal device 102 to interact with the server 105 and / or the database 103 through the network 104 to receive or send information, etc.

[0049] The first terminal device 101 and the second terminal device 102 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal device. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, including but not limited to applications, mini-programs, etc., installed on the terminal device, and can also be in web page form.

[0050] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101 and the second terminal device 102 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data. The server can be a backend server for the aforementioned target application, used to provide backend services to the clients of the target application.

[0051] Database 103 is a professional storage system for storing and managing data. It can store various types of data related to the target application, such as user account information, business transaction records, and content resource data. It supports structured, semi-structured, or unstructured data storage and has management capabilities such as data creation, deletion, modification, querying, backup, and recovery. In this application scenario, database 103 can be connected to server 105 via a communication link. Server 105 can retrieve the required data from database 103 for processing based on requests from the first terminal device 101 and the second terminal device 102. It can also synchronously store new data generated by the operations of the first terminal device 101 and the second terminal device 102 into database 103, thereby achieving data persistence and efficient retrieval.

[0052] It should be noted that the target object status assessment method provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-102. Accordingly, the target object status assessment device provided in this application embodiment can generally be set in server 105 and / or terminal devices 101-102.

[0053] It should be understood that Figure 1 The number of terminal devices, networks, databases, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, databases, and servers can be included.

[0054] The following will be based on Figure 1 The described scene, through Figures 2-4 The target object state evaluation method of the embodiments of this application will be described in detail.

[0055] Figure 2 A flowchart illustrating a target object state evaluation method according to an embodiment of this application is shown schematically.

[0056] like Figure 2 As shown, the target object state evaluation method 200 of this embodiment includes steps S210 to S240.

[0057] In step S210, the time-series feature sequence and static feature vector authorized by the target user are obtained. The time-series feature sequence is used to characterize the dynamic changes of the target user within a continuous time window, and the static feature vector is used to characterize the stable attribute information of the target user within the continuous time window.

[0058] In some embodiments, the continuous time window can be set according to the needs of the user status assessment task, such as several days, several weeks, several months, or other continuous observation intervals. The continuous time window can be divided into multiple time steps, each time step corresponding to a preset time granularity. The system can collect and organize dynamic data within the authorized range of the target user according to the preset time granularity, so that each time step corresponds to a set of features that can reflect the dynamic changes of the user within that time unit. The resulting time-series feature sequence not only contains the status information of each time step, but also retains the chronological relationship between different time steps, thus enabling it to describe the changing trend of the user's status over time.

[0059] In embodiments of this application, dynamic change information may include changes in access behavior, interaction behavior, transaction behavior, service usage, account status, asset status, response records, or other information that changes over time by a target user within a continuous time window. For example, in a marketing service adaptation scenario, dynamic change information may include the number of page visits, service inquiries, function usages, activity participations, transaction frequency, or service response status of a user at different time steps. The above examples are only used to illustrate the types of dynamic change information; in practical applications, dynamic data relevant to the current business scenario can be selected based on the status assessment objectives.

[0060] In some embodiments, the temporal feature sequence can be obtained by processing dynamic data through time alignment, statistical summarization, numerical transformation, missing data imputation, or scaling. Since dynamic data from different sources may have different occurrence times, statistical periods, or recording methods, the dynamic data can first be grouped into corresponding time steps according to a uniform time granularity, and then the temporal features corresponding to each time step can be generated based on the grouping results. This processing method ensures that the features corresponding to multiple time steps have a consistent arrangement structure, facilitating the subsequent extraction of dynamic evolution information by the temporal encoder.

[0061] In the embodiments of this application, the static feature vector can be generated based on attribute information within the authorized range of the target user. Stable attribute information can refer to information that changes with low frequency and long period within a continuous time window, or that has a long-term descriptive effect on the user's state. For example, stable attribute information may include basic user attributes, account type, service preference category, risk tolerance level, authentication status, long-term usage habit category, or other relatively stable information. The static feature vector can be formed through numericalization, normalization, category mapping, or vectorization so that the subsequent static encoder can convert it into a static state representation suitable for fusion with the temporal state representation.

[0062] In some embodiments, stable attribute information is not required to remain completely unchanged within a continuous time window, but rather to have a lower frequency of change or a longer period of change compared to dynamically changing information. For example, some attribute information may be updated only once over a longer period, or may only change when the user actively maintains their profile, adjusts service preferences, or changes their account status. The system can determine whether an attribute is suitable for use as stable attribute information in the construction of a static feature vector based on its frequency of change, period of change, or update pattern within a continuous time window. In this way, the static feature vector can more accurately represent the stable characteristics of the target user within the current observation period.

[0063] It should be noted that the system must obtain authorization from the target user before acquiring the target user's time-series feature sequences and static feature vectors. Authorization may include the target user's confirmation of the relevant data collection scope, data type, purpose of use, processing method, and authorization validity period. The system will only acquire data related to the user status assessment task within the scope of the target user's authorization and will generate time-series feature sequences and static feature vectors based on the data within the authorized scope. Data outside the authorized scope will not be collected, accessed, or processed by the system.

[0064] In step S220, the temporal feature sequence is temporally encoded to obtain the temporal state representation corresponding to each time step; the static feature vector is statically encoded to obtain the static state representation.

[0065] In the embodiments of this application, temporal encoding of the temporal feature sequence refers to inputting multiple temporal features arranged in time step order into a temporal encoding structure, enabling the model to extract the dynamic change patterns of the target user within a continuous time window and generate a corresponding temporal state representation for each time step. Each temporal state representation may contain not only the dynamic feature information of the current time step, but also contextual information related to the preceding and following time steps.

[0066] In some embodiments, temporal coding can be implemented using model structures capable of processing sequential data, such as recurrent neural networks, gated recurrent structures, long short-term memory structures, temporal convolutional structures, attention structures, or combinations thereof. The temporal coding structure can read the temporal feature sequence sequentially according to time steps and generate temporal state representations corresponding to each time step based on the changing relationships between adjacent time steps, trend relationships over a longer time range, or fluctuation relationships within a local time window. Thus, the dynamic change information of the target user can be transformed from the original feature form into a state representation form more suitable for subsequent gating generation and state evaluation.

[0067] In the embodiments of this application, static encoding of static feature vectors refers to inputting static feature vectors representing stable attribute information of the target user into a static encoding structure to generate a static state representation. The static encoding structure can uniformly express stable attribute information of different categories, dimensions, or value ranges, converting the stable attribute information into a vector form adapted to the temporal state representation. The static state representation can be used as a weighted injection of stable attribute expressions in subsequent fusion processing, participating in the generation of the fusion state representation corresponding to each time step.

[0068] In some embodiments, the static coding structure can employ a fully connected layer, a nonlinear transformation layer, a normalization layer, an embedding mapping layer, or a combination of the above structures. For categorical stable attribute information, categorical mapping or vectorization can be performed first before inputting it into the static coding structure; for numerical stable attribute information, scaling or standardization can be performed first before generating a static state representation. Through the above processing, the inconsistency in expression caused by formal differences between different stable attribute information can be reduced, making the static state representation have a more stable semantic expressive capability.

[0069] In some embodiments, temporal coding and static coding can be performed by different coding structures to maintain the differentiated representation of dynamically changing information and stable attribute information during the coding stage. Temporal coding focuses on learning temporal dependencies, stage changes, and local fluctuations within a continuous time window, while static coding focuses on learning the combination relationships and semantic representations between stable attribute information. By encoding them separately, the two types of heterogeneous features can be extracted in their respective suitable representation spaces. Then, a subsequent fusion gating mechanism dynamically controls the participation strength of the static state representation based on the temporal state representation, thereby improving the feature representation quality in the user state evaluation process.

[0070] In step S230, a fusion gating vector is generated based on the temporal state representation; the static state representation is weighted based on the fusion gating vector, and the weighted static state representation is fused with the temporal state representation of the corresponding time step to obtain the fusion state representation corresponding to each time step.

[0071] In the embodiments of this application, a fusion gating vector is generated based on the temporal state representation. For each time step within a continuous time window, the intensity of the static state representation's participation in the fusion process at that time step is determined according to the temporal state representation corresponding to that time step. The fusion gating vector can be understood as a set of adjustment coefficients corresponding to the static state representation, used to control the degree to which information from different dimensions in the static state representation is introduced at the current time step. In this way, stable attribute information no longer participates in the calculation of all time steps with a fixed intensity, but can be adaptively adjusted according to the dynamic state of the current time step.

[0072] In some embodiments, the fusion gating vector can be generated by a gating generation structure based on the temporal state representation of the current time step. This gating generation structure can receive the temporal state representation of the current time step and can also combine it with contextual information within the vicinity of that time step to obtain a gating result that better reflects the current dynamic state. For example, when the current time step exhibits state fluctuations, trend changes, or local pattern changes, the gating generation structure can adjust the values ​​of each dimension in the fusion gating vector, giving higher participation strength to dimensions in the static state representation that are more relevant to the current state.

[0073] In the embodiments of this application, the static state representation is weighted based on the fusion gating vector. This weighting can be applied to the entire static state representation or to each dimension of the static state representation individually. For the dimension-by-dimensional weighting, each element in the fusion gating vector can correspond to one or a group of dimensions in the static state representation, allowing different stable attribute information to have different fusion strengths at the current time step. Therefore, the system can more precisely control the way stable attribute information functions in different temporal states.

[0074] In some embodiments, the corresponding time step can be determined by a time step mapping relationship. Since the static state representation itself typically does not change sequentially with each time step, the position of the static state representation participating in the fusion at each time step can be clearly defined through the time step mapping relationship. For each time step, the system can take the temporal state representation corresponding to that time step and fuse it with the static state representation weighted by the fusion gating vector to obtain the fused state representation corresponding to that time step. This process can be repeated at each time step within a continuous time window.

[0075] In the embodiments of this application, the fused state representation can reflect the conditional dependency between dynamically changing information and stable attribute information. That is, the degree of influence of the static state representation on the state expression at the current time step is determined by the current temporal state representation, rather than remaining consistent across all time steps. For example, in a marketing service adaptation scenario, when a user's recent interactions change significantly, information related to the user's long-term attributes or preferences can participate in the current state expression with varying degrees of intensity; when the user's dynamic changes tend to stabilize, the way stable attribute information functions can also be adjusted accordingly.

[0076] In some embodiments, the fusion gating vector can also be used to control the participation of feedback modulation terms in the fusion process. The feedback modulation terms can be generated based on the temporal state representation at the current time step and are used to enhance or suppress the current temporal state representation. The system can use the fusion gating vector to weight the feedback modulation terms, so that the fused state representation not only reflects the supplementary role of the static state representation to the temporal state representation, but also reflects the modulating role of the static attribute background on the dynamic state expression. In this way, the expressive power and stability of the fused state representation can be further improved.

[0077] In the embodiments of this application, after completing the fusion processing for each time step within a continuous time window, multiple fused state representations corresponding to different time steps can be obtained. Each fused state representation has a static information injection strength that matches the dynamic state of the current time step, thus avoiding the indiscriminate calculation of static attribute information across all time steps. This time-step dynamic fusion method can more accurately describe the process of the target user's state changing over time and provide a foundation for the subsequent generation of fused state sequences and state evaluation results.

[0078] In step S240, the fusion state representations corresponding to each time step are arranged in the order of time steps to obtain a fusion state sequence, and the state evaluation result of the target user is generated based on the fusion state sequence.

[0079] In the embodiments of this application, the state evaluation result of the target user is generated based on the fused state sequence. This can be achieved by first performing sequence aggregation processing on the fused state sequence to obtain the overall state representation of the target user. Sequence aggregation processing converts the fused state representations corresponding to multiple time steps into a representation that can summarize the overall state within a continuous time window. This overall state representation can reflect both the dynamic evolution trend of the target user within the observation interval and the comprehensive influence of stable attribute information after gating adjustment at different time steps.

[0080] In some embodiments, sequence aggregation processing can employ average aggregation, weighted aggregation, maximum aggregation, terminal state aggregation, attention aggregation, or a combination of the above. Average aggregation can be used to balance the fused state representations at each time step; weighted aggregation can assign different weights based on the importance of different time steps; maximum aggregation can highlight significant state changes in certain dimensions; terminal state aggregation can focus on the state representation closer to the current time step; attention aggregation can adaptively determine the contribution of each time step according to the evaluation objective. Different aggregation methods can be flexibly selected according to the specific evaluation task.

[0081] In the embodiments of this application, after obtaining the overall state representation, at least one state evaluation value can be generated based on the overall state representation. The state evaluation value can be a score, vector, category probability, level index, ranking index, or other model output that can characterize the state of the target user. One state evaluation value can correspond to one evaluation target, and multiple state evaluation values ​​can each correspond to multiple evaluation targets. For example, state level-related scores, service adaptation-related scores, resource configuration reference-related values, or reach priority-related indicators can be generated separately to support state evaluation needs in different scenarios.

[0082] In some embodiments, a status assessment result for the target user can be further generated based on the status assessment value. The status assessment result may include a status level, service adaptation result, resource allocation reference result, service order reference result, or other results used to assist service decision-making. The status level can be used to indicate that the target user is currently in a certain status range; the service adaptation result can be used to indicate the degree of matching between the target user and a certain type of service, benefit, content, or plan; the resource allocation reference result can be used to provide a reference for service resource deployment, service quota allocation, or operational strategy selection.

[0083] In the embodiments of this application, the relationship between the state evaluation value and the state evaluation result can be achieved through threshold mapping, interval division, sorting transformation, rule mapping, or model output mapping. For example, when the state evaluation value falls within a preset interval, a corresponding state level can be generated; when multiple service adaptation scores are generated, the service adaptation result can be determined based on the score ranking; when a resource configuration reference value is generated, it can be converted into a corresponding resource configuration reference interval. Through this conversion process, the model output can form a state evaluation result that is more convenient for subsequent system use.

[0084] In some embodiments, the fused state sequence can also be used to generate interpretive reference information. For example, the time steps or feature dimensions that have a significant impact on the state evaluation results can be determined based on the fused state representations, fused gating vectors, or aggregation weights corresponding to different time steps. This interpretive reference information can be used to assist in understanding the basis for the formation of the state evaluation results, enabling the user's state evaluation process to not only output results but also provide a certain degree of process reference.

[0085] According to embodiments of this application, dynamic change information within a continuous time window can be constructed into a temporal feature sequence, and stable attribute information can be constructed into a static feature vector. Then, temporal encoding and static encoding are performed separately to obtain the temporal state representation and static state representation for each time step. Subsequently, a fusion gating vector is generated based on the temporal state representation, dynamically controlling the injection intensity of the static state representation at different time steps, and generating the fused state representation corresponding to each time step. Thus, dynamic change information and stable attribute information are no longer simply concatenated in a fixed manner, but are conditionally fused according to the current temporal state, enabling the fused state sequence to more accurately depict the process of target user state changes over time, improving the accuracy, stability, and service adaptation reliability of the state assessment results.

[0086] The target object state evaluation method of this application will be described below by way of specific embodiments.

[0087] In view of the differences between dynamic data and attribute information in terms of time attributes, frequency of change and data form, if the two are directly mixed, it is easy to cause unclear boundaries between time-series features and stable attribute features, which will affect the accuracy of subsequent time-series coding, static coding and gating fusion. In the embodiments of this application, when obtaining feature data authorized by the target user, time-series feature sequences and static feature vectors can be constructed separately.

[0088] Specifically, dynamic data authorized by the target user can be acquired according to a preset time granularity, and time alignment processing can be performed on the dynamic data to ensure that dynamic data from different sources, at different times of occurrence, or with different statistical periods can be grouped into corresponding time steps within a continuous time window. Subsequently, time-series features corresponding to multiple time steps are generated based on the aligned dynamic data, and a time-series feature sequence is formed according to the time step order. At the same time, attribute information authorized by the target user can be acquired, and stable attribute information can be determined from the attribute information based on the frequency or period of change of the attribute information within the continuous time window. For attribute information with low frequency of change, long period of change, or relatively stable descriptive function within the continuous time window, it can be used as stable attribute information to participate in the generation of static feature vectors.

[0089] In this way, dynamically changing information is organized into a temporal feature sequence with a time order, and stable attribute information is organized into a static feature vector, achieving differentiated expression of the two types of heterogeneous features from the input construction stage. Therefore, subsequent processing can extract dynamic evolution patterns from the temporal feature sequence and stable attribute expressions from the static feature vector, respectively. Furthermore, during the fusion stage, a fusion gating vector controls the participation intensity of the static state representation at different time steps, avoiding feature expression confusion caused by the mixed construction of dynamic data and stable attribute information. This improves the structural clarity of the input features during user state assessment and the stability of subsequent state assessment results.

[0090] For example, taking the promotion of personal pension business as an example, for each customer, their past continuous pension information can be obtained. The time-series features statistically analyzed at the monthly granularity within a month constitute a time-series feature sequence. .in, This is the length of the time window; for example, a value of 24. The number of types of time-series features. The Middle line, number Column elements This indicates that the customer was in the first... The first month The values ​​of time-series features.

[0091] Simultaneously, the static characteristics of the customer can be obtained to form a static feature vector. .in This represents the number of types of static features. The Middle element This indicates the customer's first Class static properties.

[0092] To address the issues that temporal feature sequences and static feature vectors belong to different data formats, sharing the same input path can easily lead to feature semantic mixing, and there is a lack of clear correspondence between static feature vectors and multiple time steps, embodiments of this application, after forming the temporal feature sequence and static feature vector, can establish a temporal input channel corresponding to the temporal feature sequence and a static input channel corresponding to the static feature vector, respectively. The temporal input channel is used to carry the temporal feature sequence arranged in time step order, enabling the subsequent temporal encoder to extract dynamic change patterns according to the sequential order within a continuous time window; the static input channel is used to carry the static feature vector, enabling the subsequent static encoder to independently extract the semantic expression of stable attribute information. By setting up dual input channels, the differentiated input structure of dynamic change information and stable attribute information can be maintained before encoding, avoiding the weakening of their respective feature expressions after being directly mixed at the input stage.

[0093] In some embodiments, a time-step mapping relationship can be established between static feature vectors and multiple time steps. This time-step mapping relationship is used to determine the fusion position of the static state representation at each time step, that is, to determine how the static state representation participates in the fusion processing of each time step within a continuous time window. Through this mapping relationship, the static state representation can establish a clear fusion correspondence with the temporal state representations corresponding to multiple time steps without changing its own stable attribute expression.

[0094] In some embodiments, the time-step mapping relationship can be established based on the time-step number, time order, time granularity, or time index within a continuous time window. For example, when a continuous time window is divided into multiple time steps, a corresponding fusion position can be assigned to each time step, enabling the static state representation to participate in computation during the fusion processing stage of each time step. This mapping relationship does not require the static feature vector itself to change with time steps, but rather is used to determine the corresponding position between the static state representation and the temporal state representation of each time step during the fusion stage. Thus, the static state representation can serve as a stable attribute expression, with its participation intensity adjusted by the corresponding fusion gating vector in different time steps.

[0095] In some embodiments, after establishing the mapping relationship between the temporal input channel, the static input channel, and the time step, the temporal feature sequence can be input into the temporal encoder based on the temporal input channel, and the static feature vector can be input into the static encoder based on the static input channel. Simultaneously, the time step mapping relationship can be passed to subsequent fusion processing steps to determine the fusion position of the static state representation at each time step. In this way, the temporal encoder focuses on extracting the temporal dependencies and state evolution features of dynamically changing information, while the static encoder focuses on extracting the semantic representation of stable attribute information. The time step mapping relationship ensures that the two types of encoding results can accurately correspond during the fusion stage.

[0096] Based on the above approach, the problem of lack of structural distinction when directly concatenating temporal and static features into the input can be solved, as well as the problem of lack of positional basis for static state representations during multi-time-step fusion. By establishing separate input channels, dynamic change information and stable attribute information can be expressed independently during the encoding stage; by establishing time-step mapping relationships, the fusion position of static state representations at each time step can be clearly determined; and by inputting the two types of features into the corresponding temporal encoder and static encoder respectively, the specificity of the encoding results can be improved.

[0097] For example, a temporal encoder can be composed of a bidirectional long short-term memory network (BiLSTM) and a multi-scale temporal self-attention layer, used to extract hidden state sequences from a temporal feature matrix X.

[0098] First, it can Input a BiLSTM layer. Assume the number of hidden units in the BiLSTM is... Forward LSTM processes data sequentially over time. Output the forward hidden state sequence , of which each Reverse LSTM processing along time sequence Output the reverse hidden state sequence , of which each .

[0099] For each time step The forward and backward hidden states are concatenated to obtain the concatenated vector. Since the concatenated dimension is 2h, to maintain consistency with subsequent modules and reduce the number of parameters, a learnable linear transformation is used to compress it to [the desired dimension]. Dimension, obtain the initial hidden state :

[0100]

[0101] in, This is the weight matrix. This is the bias vector. From this, we obtain the initial hidden state sequence. , .

[0102] Then, Input a multi-scale temporal self-attention layer. This attention layer contains... A parallel attention head, The value is 3, corresponding to the short-term, medium-term, and long-term time scales, respectively.

[0103] Each attention point is... As input, before calculating the key matrix and value matrix, first... Apply a one-dimensional convolution operation in the time dimension. The convolution dilation rate used by each attention head is [missing value]. For example, different heads The values ​​are 1, 3, and 6, used to capture time-series patterns of adjacent months, quarterly spans, and semi-annual spans, respectively. The specific value of the inflation rate can be determined based on the length of the time window. Adjustments are typically set as a set of positive integers that increase in a geometric or arithmetic order.

[0104] With the first Taking the size as an example, the calculation process is as follows:

[0105] First of all The applied expansion rate is Temporal convolution yields the convolved sequence. .

[0106] Then with Calculate the query matrix Key matrix Value matrix .in, For a learnable parameter matrix, For the dimension of attention head.

[0107] Next, the attention weight matrix can be calculated. and obtain the head output. , .

[0108] Finally, the outputs of all attention heads are concatenated along the feature dimensions to obtain the concatenation matrix. Then through a linear transformation matrix Mapping back to dimension The enhanced hidden state sequence is obtained. ,remember ,in, For the first A comprehensive time-series representation of the current state over a month.

[0109] In the embodiments of this application, the state evolution stage of the target user within a continuous time window can be determined based on the temporal state representation of the current time step, and a fusion gating vector for controlling the fusion strength of the static state representation can be generated, so that the static state representation can adaptively participate in the fusion process according to the dynamic state of different time steps.

[0110] Figure 3 The flowchart illustrates a method for generating fusion gating vectors based on temporal state representations according to some exemplary embodiments of this application.

[0111] like Figure 3 As shown, the method for generating fusion gated vectors based on temporal state representation includes steps S310 to S340.

[0112] In step S310, multiple state prototype vectors obtained through training are acquired. The state prototype vectors are used to characterize the different state evolution stages of the target user within a continuous time window.

[0113] The state evolution stages here are not limited to specific business labels, but rather refer to typical state patterns automatically learned by the model based on temporal state representations during training. For example, in different application scenarios, state evolution stages can correspond to different change patterns in user state, such as from stable to active, from weak response to strong response, and from low adaptation to high adaptation. Multiple state prototype vectors can be used as learnable parameters, continuously updated according to the state evaluation task during model training, so that each state prototype vector gradually develops the ability to express a typical temporal state pattern.

[0114] In some embodiments, the number of state prototype vectors can be set according to the complexity of the state evaluation task, the length of the continuous time window, or the granularity of state partitioning. A smaller number of state prototype vectors allows for coarser-grained state stage partitioning; a larger number allows for finer-grained expression of state evolution differences. Since the state prototype vectors are obtained through training, there is no need to rely on manual pre-labeling of the state stage corresponding to each time step, reducing the subjectivity of manual stage partitioning and making the state prototype vectors more closely reflect the state evolution patterns in actual data.

[0115] In step S320, the state assignment weight corresponding to the current time step is determined based on the similarity between the temporal state representation of the current time step and the prototype vectors of each state.

[0116] Specifically, the temporal state representation at the current time step can be compared with each state prototype vector to calculate the similarity between the current time step and each state prototype vector. Similarity calculation methods can include inner product similarity, cosine similarity, distance transform similarity, or other methods that can measure vector proximity. Subsequently, multiple similarity results can be normalized to obtain the state assignment weights for the current time step corresponding to multiple state prototype vectors.

[0117] In some embodiments, state assignment weights can be understood as a soft allocation result of the current time step belonging to different state evolution stages. Compared to directly classifying the current time step to a single state stage, state assignment weights can express the situation where the current time step may simultaneously possess characteristics of multiple state stages. For example, the current time step may have some characteristics of a stable stage as well as some characteristics of a transition to an active stage. Expressing this through state assignment weights avoids overly rigid state stage divisions, making the subsequent generation process of the fusion gating vector smoother and more continuous.

[0118] In step S330, multiple state prototype vectors are weighted based on state attribution weights to obtain the state basis vector corresponding to the current time step.

[0119] Specifically, each state prototype vector can be multiplied by its corresponding state assignment weight, and the weighted vectors can be summed to obtain the state basis vector. This state basis vector comprehensively reflects the relationship between the current time step and multiple state evolution stages, and can serve as the basic representation for generating the fusion gating vector. Compared to directly generating gating results based solely on the temporal state representation of the current time step, the state basis vector introduces information about the state evolution stages, enabling the gating generation process to reflect the phased changes in the user's state within a continuous time window.

[0120] In some embodiments, the state basis vector can be used to represent the basic strength distribution of each dimension of the static state representation at the current time step that should participate in the fusion. Since the state basis vector is obtained by weighting multiple state prototype vectors, its expression is related not only to the temporal state representation at the current time step, but also to the typical state evolution pattern obtained through training. Therefore, when the target user is in different state stages or state transition stages, the state basis vector can exhibit different distribution characteristics, providing a stage-aware input for subsequent gating mapping processing.

[0121] In step S340, the state basis vector is subjected to gating mapping to obtain the fused gating vector corresponding to the current time step.

[0122] Gating mapping can be a process of converting state basis vectors into gating vectors that match the dimensions of the static state representation or the static representation to be fused. This process can be achieved through linear transformations, nonlinear transformations, normalization, activation function processing, or a combination of these methods. The generated fusion gating vectors can be used to weight the static state representation in subsequent steps, thereby controlling the injection strength of the static state representation in the current time step.

[0123] According to embodiments of this application, the trained state prototype vector can be introduced into the generation process of the fusion gating vector, enabling the fusion gating vector to not only depend on the temporal state representation of the current time step, but also reflect the state evolution stages of the target user within a continuous time window. This solves the problem of the static state representation fusion strength lacking stage awareness and being difficult to adjust with changes in user state, allowing the subsequently generated fusion state representation to more accurately reflect the conditional relationship between dynamically changing information and stable attribute information, thereby improving the accuracy and stability of user state assessment results.

[0124] In some embodiments, multiple state prototype vectors can serve as learnable parameters in the state evaluation model, and are updated along with the model parameters during training. Let the state prototype vectors be p. k Where k = 1, 2, ..., K, and K is the number of state prototype vectors; the temporal state of the current time step is represented by h. tTo ensure that the temporal state representation and the state prototype vector are in a suitable representation space for calculating similarity, the temporal state representation at the current time step can be projected using a learnable linear transformation matrix to obtain the projection vector:

[0125]

[0126] in, It is represented as a learnable linear transformation matrix; q t This is the projection vector corresponding to the current time step.

[0127] In some embodiments, it can be based on the projection vector q t With each state prototype vector p k The inner product similarity between the parameters determines the state assignment weight for each state evolution stage at the current time step. Specifically, the inner product similarity can be divided by the temperature parameter. The state assignment weights are obtained by normalizing the weights using a normalized exponential function.

[0128]

[0129] in, This represents the weight assigned to the k-th state prototype vector at the current time step t. The temperature parameter is used. This state assignment weight is used to characterize the degree of proximity between the temporal state representation at the current time step and different state evolution stages.

[0130] In some embodiments, after obtaining the state attribution weights, multiple state prototype vectors can be weighted and summed based on the state attribution weights to generate the state basis vector corresponding to the current time step:

[0131]

[0132] Where, α t This is the state basis vector corresponding to the current time step. This state basis vector integrates the matching relationships between the current time step and multiple state prototype vectors, and can be used as the basic representation for subsequent gating mapping processing to generate the fused gating vector corresponding to the current time step.

[0133] In some embodiments, the state prototype vector is automatically updated during training through backpropagation of the state evaluation task. That is, the learning objective of the state prototype vector does not solely rely on manually labeled state stages, but rather uses the evaluation loss between the state evaluation results and the labeled results to backpropagate the error signal to the state prototype vector, the temporal encoder, the static encoder, and the gating generation structure, enabling the state prototype vector to gradually develop the ability to express typical state evolution patterns. In this way, the model can automatically learn discriminative state patterns within a continuous time window without forcibly pre-setting specific business stage labels.

[0134] In some embodiments, temperature parameters Annealing can be used to adjust this during training. In the early stages of training, a larger temperature parameter can be set to smooth the distribution of state assignment weights, allowing the current time step to reference multiple state prototype vectors simultaneously and preventing them from prematurely concentrating on a few state patterns. As training epochs increase, the temperature parameter can be gradually decreased, making the distribution of state assignment weights clearer and prompting different state prototype vectors to form more explicit expressions of state evolution stages. This process helps improve the distinguishability of state prototype vectors and enhances the stage-awareness of the fused gating vectors.

[0135] In some embodiments, to enhance the discriminative power among multiple state prototype vectors, discriminative constraints can be introduced during training. For example, each state prototype vector can be normalized first, and then a constraint term for the distance between different state prototype vectors can be calculated to maintain the differences between different state prototype vectors in the representation space. Accordingly, discriminative constraints of the following form can be adopted:

[0136]

[0137] in, These represent the normalized state prototype vectors. This constraint reduces the likelihood of multiple state prototype vectors converging into similar regions, allowing different state prototype vectors to express different state evolution patterns.

[0138] In some embodiments, to ensure the continuity of state assignment weights between adjacent time steps, a state assignment continuity constraint can be introduced. For example, the difference between the state assignment weights corresponding to adjacent time steps can be calculated and used as part of the training constraint:

[0139]

[0140] Among them, a tThis represents the state assignment weight vector for multiple state prototype vectors at time step t, where T represents the number of time steps within a continuous time window. This constraint can suppress abnormally drastic changes in state assignment weights between adjacent time steps, making the state evolution process smoother and consistent with the characteristic of continuous changes in user states.

[0141] In some embodiments, the training objective of the state evaluation model can be composed of the state evaluation loss, the state prototype vector distinguishing constraint, and the state ownership continuity constraint, for example:

[0142]

[0143] in, and This is the weight coefficient for the regularization term. For example, all values ​​can be 0.01, which are empirical values ​​and can be adjusted based on the results of the validation set.

[0144] Figure 4 The flowchart illustrates a method for gating mapping of state basis vectors according to some exemplary embodiments of this application.

[0145] like Figure 4 As shown, the method for gating and mapping state basis vectors includes steps S410 to S430.

[0146] In step S410, a conditional context representation corresponding to the current time step is constructed based on the temporal state representation of the current time step, the state change amount between adjacent time steps, and the local temporal context.

[0147] The temporal state representation at the current time step can be used to characterize the dynamic state of the target user at the current time step; the state change between adjacent time steps can be used to characterize the magnitude of change of the current time step relative to the previous or next time step; the local temporal context can be used to characterize the local change pattern within several time steps near the current time step. By combining the above information, a conditional context representation that reflects the current state, short-term changes, and local trends can be obtained.

[0148] In some embodiments, the state change between adjacent time steps can be obtained through the difference, distance, change ratio, or other methods that can characterize the degree of state change between the temporal state representations of two adjacent time steps. For example, the state change of the current time step can be obtained by differentiating the temporal state representation of the current time step from that of the previous time step; alternatively, a more complete local change representation can be constructed by combining adjacent time steps. For the first or last time step in a continuous time window, the corresponding state change can be generated using a preset vector, boundary replication, neighboring time step substitution, or other boundary processing methods.

[0149] In some embodiments, local temporal context can be obtained by aggregating, convolving, sliding windowing, or attention-weighting multiple temporal state representations near the current time step. For example, temporal state representations of several neighboring time steps can be selected centered on the current time step, and a local temporal context can be generated through a local aggregation structure; alternatively, a local temporal transformation can be performed on the complete temporal state sequence so that each time step obtains a corresponding local context representation. Local temporal context can supplement local trend information that is difficult to express by a single time step, enabling subsequent gating mapping processing to more fully perceive the change patterns near the current time step.

[0150] In step S420, the conditional context representation is input into the gating network to obtain the scaling factor corresponding to the current time step.

[0151] In some embodiments, each element of the scaling factor can correspond one-to-one with each dimension of the state basis vector, or it can correspond to different groups of dimensions of the state basis vector. When a certain element of the scaling factor has a high value, it indicates that the corresponding dimension in the state basis vector has a high degree of preservation or enhancement at the current time step; when a certain element of the scaling factor has a low value, it indicates that the corresponding dimension in the state basis vector needs to be weakened at the current time step. Through this method of adjusting dimensions one by one or grouped by dimensions, the fused gating vector can reflect not only the state evolution stage, but also the local changes near the current time step.

[0152] In step S430, the state basis vector is adjusted element by element based on the scaling factor to obtain the fusion gate vector corresponding to the current time step.

[0153] Specifically, the scaling factor can be multiplied element-wise, weighted element-wise, or subjected to other vector adjustment operations with the state basis vector to obtain the fused gating vector. This fused gating vector can be used to subsequently control the injection intensity of the static state representation at the current time step. Since the fused gating vector originates from both the state basis vector and the conditional context representation, it possesses both the ability to perceive state evolution stages and the ability to perceive local temporal changes.

[0154] According to embodiments of this application, based on the phased basic gating information provided by the state basis vector, the state change amount and local temporal context of the current time step can be introduced to further conditionally adjust the basic gating information. This processing method can solve the problem of insufficient perception of local dynamic changes when gating is generated solely based on the state evolution stage, enabling the fused gating vector to more accurately adapt to the dynamic state of the current time step, thereby improving the expressive power of the subsequent fused state representation and the stability of the user state evaluation results.

[0155] In some embodiments, the state change between adjacent time steps can be expressed as: For example, we can let:

[0156]

[0157] When the current time step is the first time step in a continuous time window, it can be Set it as a zero vector, or use a preset boundary handling method to determine the state change. By introducing... Gated mapping processing can perceive the degree of state change of the current time step relative to adjacent time steps.

[0158] In some embodiments, a local temporal context can be generated based on the local processing results of the temporal state representation sequence in the time dimension. Let the local temporal context be c. t It can be obtained by sliding extraction along the time dimension using a one-dimensional convolutional layer with a kernel size of 5 and an inflation rate of 1. The kernel size of 5 is chosen to cover two months' windows before and after, striking a trade-off between capturing local trends and maintaining computational efficiency.

[0159] In some embodiments, the temporal state representation of the current time step, the state changes between adjacent time steps, and the local temporal context can be concatenated along the feature dimension to obtain the conditional context representation. The combination of these three components enables the gating network to simultaneously perceive the current state, the trend of state changes, and the pattern characteristics within a local window.

[0160] Then, Input a gated network consisting of two fully connected layers. The weight matrix of the first layer is... Bias is Following this is the ReLU activation function, which outputs an intermediate vector. , The weight matrix of the second layer is: Bias is Followed by a Sigmoid activation function, the output scaling factor is given. :

[0161]

[0162] in, This represents the Sigmoid function, whose output value is between 0 and 1.

[0163] In the embodiments of this application, the state basis vector is adjusted element-wise based on a scaling factor to obtain the fusion gate vector corresponding to the current time step:

[0164]

[0165] in, This represents element-wise multiplication of vectors.

[0166] In the embodiments of this application, after obtaining the fusion gating vector corresponding to the current time step, the fusion position of the static state representation at the current time step can be determined based on the time step mapping relationship. This fusion position indicates the correspondence between the static state representation and the temporal state representation at the current time step, enabling the static state representation to participate in the fusion process at each time step within a continuous time window. Since the static state representation typically characterizes the stable attribute information of the target user within a continuous time window, it itself may not change with time steps. However, through the time step mapping relationship, the position of the static state representation participating in the fusion can be determined at each time step, thus providing a basis for generating the fused state representation by time step.

[0167] In some embodiments, the static state representation can first be projected to obtain the static representation to be fused. This projection process ensures that the static representation to be fused and the temporal state representation at the current time step are in a suitable representation space for fusion. Subsequently, the static representation to be fused can be element-wise weighted based on the fusion gate vector to obtain a weighted static representation to be fused. This weighting process controls the participation strength of each dimension in the static representation to be fused at the current time step, ensuring that stable attribute information participates in the fusion according to the fusion gate vector at the current time step. In some embodiments, the weighted static representation to be fused can be fused with the temporal state representation at the current time step according to the fusion position to obtain the fused state representation corresponding to the current time step.

[0168] Through the above processing, the static state representation does not directly participate in the fusion of all time steps with a uniform intensity. Instead, it is first projected to obtain the static representation to be fused, then weighted element-wise by the fusion gating vector corresponding to the current time step, and finally fused according to the temporal state representation of the fusion position and the current time step. This allows stable attribute information to participate in the generation of the fused state representation with different intensities in different time steps, improving the ability of the fused state representation to express the conditional relationship between the dynamic state and stable attribute information of the current time step.

[0169] In some embodiments, after element-wise weighting of the static representation to be fused based on the fusion gating vector, a feedback modulation term can be further generated based on the temporal state representation of the current time step. The feedback modulation term is used to represent the dynamic components in the temporal state representation of the current time step that can be gated and adjusted, so that the fusion process includes not only the injection of the static state representation into the temporal state representation, but also the enhancement or suppression of the current temporal state representation itself.

[0170] In the embodiments of this application, through a linear transformation matrix For the current hidden state Projection is performed to obtain the feedback modulation term. , .

[0171] In some embodiments, the feedback modulation term can be weighted based on the fusion gating vector corresponding to the current time step to obtain a gated feedback representation. The fusion gating vector has already been generated in the preceding steps according to the state evolution stage and local temporal conditions of the current time step. Therefore, using this fusion gating vector to weight the feedback modulation term ensures that the feedback modulation term participates in the fusion process with appropriate strength only when needed in the current temporal context. In other words, the gated feedback representation is not fixedly superimposed on the temporal state representation, but is controlled by the current fusion gating vector, thus reflecting the characteristics of conditional modulation.

[0172] In some embodiments, gated feedback representation can be used to enhance or suppress the temporal state representation at the current time step. When there is a strong correlation between the target user's current dynamic state and stable attribute information, the fusion gate vector can give higher weights to the corresponding dimensions in the feedback modulation term, thereby enhancing the relevant dynamic state components. When some dynamic state components do not match the current stable attribute background or need to be weakened, the fusion gate vector can reduce the feedback modulation intensity of the corresponding dimensions, thereby suppressing or balancing the temporal state representation. In this way, the static state representation not only injects additional information into the fusion process, but can also influence the expression intensity of the current temporal state representation through the gated feedback representation.

[0173] In some embodiments, after obtaining the gated feedback representation, the temporal state representation of the current time step, the weighted static representation to be fused, and the gated feedback representation can be fused. This fusion process can be understood as a combination of three parts: the first part is the original temporal state representation of the current time step, used to retain dynamic change information; the second part is the static representation to be fused after being weighted by the fusion gate vector, used to introduce stable attribute information matching the current time step; and the third part is the feedback modulation term after being weighted by the fusion gate vector, used to conditionally enhance or suppress the current temporal state representation.

[0174] In some embodiments, a normalization process can be further performed after fusion to obtain the fused state representation corresponding to the current time step. Normalization adjusts the numerical distribution of the fused representation, making the scales between different dimensions more stable and reducing distribution fluctuations that may occur after the superposition of multiple information streams. Through normalization, the fused state representation can be more stably input into subsequent sequence permutations, aggregation layers, or evaluation heads, thereby helping to improve the stability of the state evaluation process.

[0175] Correspondingly, the following form can be adopted:

[0176]

[0177] Among them, f t This represents the fused state representation corresponding to the current time step, and LayerNorm(·) represents the layer normalization process. Through layer normalization, the fused state representation can have a more stable numerical distribution, which facilitates the subsequent formation of a fused state sequence according to the time step order.

[0178] According to embodiments of this application, stable attribute information is not only a one-way injection of the timing state representation of the current time step, but can also enhance or suppress the timing signal itself under the control of the fusion gating vector. Compared with simple residual superposition, the feedback modulation term is controlled by the fusion gating vector, which can determine its participation level according to the current timing context, thus avoiding the feedback modulation term acting in a fixed manner in all time steps.

[0179] In the embodiments of this application, after obtaining the fused state sequence, a state evaluation result for the target user can be generated based on the fused state sequence. The fused state sequence is formed by arranging the fused state representations corresponding to each time step within a continuous time window in the order of the time steps. Therefore, it simultaneously contains dynamic change information of the target user at multiple time steps, stable attribute information after gating adjustment, and the conditional fusion relationship between the two. In order to convert the fused state representations of multiple time steps into a representation that can characterize the overall state of the target user, the fused state sequence can be input into the aggregation layer in the state evaluation model to obtain the overall state representation of the target user.

[0180] In some embodiments, the aggregation layer can be used to summarize multiple fusion state representations in a fusion state sequence. Aggregation methods may include average aggregation, weighted aggregation, attention aggregation, maximum value aggregation, end-state aggregation, or a combination of the above. After processing by the aggregation layer, the fusion state representations corresponding to multiple time steps can be compressed or converted into a single overall state representation. This overall state representation summarizes the comprehensive state of the target user within a continuous time window, reflecting both the dynamic trends over time and preserving the combined impact of stable attribute information after gating adjustments at different time steps.

[0181] In some embodiments, after obtaining the overall state representation, the overall state representation can be input into at least one evaluation head in the state evaluation model to obtain a state evaluation value. The evaluation head can be used to map the overall state representation to a specific state evaluation output. Depending on the actual evaluation objective, the state evaluation model can set one evaluation head or multiple evaluation heads. Different evaluation heads can be used to generate different types of state evaluation values, such as state category scores, service adaptation scores, resource configuration reference scores, or other numerical results related to user state evaluation.

[0182] In some embodiments, at least one evaluation header may include a categorical evaluation header, a regression evaluation header, or a ranking evaluation header. A categorical evaluation header can be used to generate classification scores corresponding to state levels; a regression evaluation header can be used to generate continuous state evaluation values ​​or resource configuration reference values; and a ranking evaluation header can be used to generate scores related to service adaptation order or reach priority. Different types of evaluation headers can share the same overall state representation, or they can be independently mapped based on the overall state representation, thereby supporting the generation of multiple state evaluation results.

[0183] In some embodiments, a status assessment result for a target user can be generated based on a status assessment value. The status assessment result may include at least one of a status level, service adaptation result, or resource configuration reference result. A status level can be used to indicate that the target user is currently in a certain status range or category; a service adaptation result can be used to indicate the degree of compatibility between the target user and a certain type of service, benefit, content, or solution; a resource configuration reference result can be used to provide a reference for subsequent service resource allocation, service sequence determination, or operational strategy selection. These results can be generated individually or in combination to meet the user status assessment needs in different application scenarios.

[0184] Through the above processing, the multi-time-step fusion information in the fused state sequence can be aggregated into an overall state representation of the target user, and further, a state evaluation value can be generated through at least one evaluation head. Since the fused state sequence already contains the conditional fusion relationship between dynamic change information and stable attribute information, the state evaluation result generated based on this fused state sequence can more comprehensively reflect the state changes and stable attribute impacts of the target user within a continuous time window, thereby improving the reliability of state level determination, service adaptation analysis, and resource allocation reference.

[0185] Corresponding to the above-described target object state assessment method, embodiments of this application also provide a target object state assessment device.

[0186] Figure 5 A schematic block diagram of a target object state evaluation apparatus according to an embodiment of this application is shown.

[0187] like Figure 5 As shown, the target object state assessment device 500 of this embodiment includes a data acquisition module 510, an encoding module 520, a fusion module 530, and a result generation module 540.

[0188] The data acquisition module 510 can be used to acquire a time-series feature sequence and a static feature vector authorized by the target user. The time-series feature sequence is used to characterize the dynamic changes of the target user within a continuous time window, and the static feature vector is used to characterize the stable attribute information of the target user within the continuous time window. In one embodiment, the data acquisition module 510 can be used to execute step S210 described above, which will not be repeated here.

[0189] The encoding module 520 can be used to perform temporal encoding on the temporal feature sequence to obtain the temporal state representation corresponding to each time step; and to perform static encoding on the static feature vector to obtain the static state representation. In one embodiment, the encoding module 520 can be used to execute the step S220 described above, which will not be repeated here.

[0190] The fusion module 530 can be used to generate a fusion gating vector based on the temporal state representation; to perform weighted processing on the static state representation based on the fusion gating vector; and to fuse the weighted static state representation with the temporal state representation of the corresponding time step to obtain the fusion state representation corresponding to each time step. In one embodiment, the fusion module 530 can be used to execute step S230 described above, which will not be repeated here.

[0191] The result generation module 540 can be used to arrange the fusion state representations corresponding to each time step in time step order to obtain a fusion state sequence, and generate the state evaluation result of the target user based on the fusion state sequence. In one embodiment, the result generation module 540 can be used to execute the step S240 described above, which will not be repeated here.

[0192] According to an embodiment of this application, the data acquisition module 510 can also be used to acquire historical business data and scenario rule data corresponding to each sub-business scenario in the target business domain; extract scenario recognition features from the historical business data and scenario rule data to distinguish different sub-business scenarios; and associate and store the scenario recognition features with the corresponding sub-business scenario identifiers to obtain a scenario feature library.

[0193] According to an embodiment of this application, the data acquisition module 510 can also be used to acquire dynamic data authorized by the target user according to a preset time granularity and perform time alignment processing, generate time-series features corresponding to multiple time steps based on the aligned dynamic data, and form a time-series feature sequence; and determine stable attribute information from the attribute information based on the frequency or period of change of the attribute information authorized by the target user within a continuous time window, and generate a static feature vector based on the stable attribute information.

[0194] According to an embodiment of this application, the data acquisition module 510 can also be used to establish a time-series input channel corresponding to the time-series feature sequence and a static input channel corresponding to the static feature vector; establish a time-step mapping relationship between the static feature vector and multiple time steps, the time-step mapping relationship being used to determine the fusion position of the static state representation at each time step; and input the time-series feature sequence and the static feature vector into the corresponding time-series encoder and static encoder based on the time-series input channel, the static input channel and the time-step mapping relationship.

[0195] According to an embodiment of this application, the fusion module 530 can also be used to obtain multiple trained state prototype vectors, which are used to characterize different state evolution stages of the target user within a continuous time window; determine the state attribution weight corresponding to the current time step based on the similarity between the temporal state representation of the current time step and each state prototype vector; perform weighted processing on the multiple state prototype vectors based on the state attribution weights to obtain the state basis vector corresponding to the current time step; and perform gating mapping processing on the state basis vectors to obtain the fusion gating vector corresponding to the current time step.

[0196] According to an embodiment of this application, the fusion module 530 can also be used to construct a conditional context representation corresponding to the current time step based on the temporal state representation of the current time step, the state change amount between adjacent time steps, and the local temporal context; input the conditional context representation into the gating network to obtain the scaling factor corresponding to the current time step; and adjust the state basis vector element by element based on the scaling factor to obtain the fusion gating vector corresponding to the current time step.

[0197] According to an embodiment of this application, the fusion module 530 can also be used to determine the fusion position of the static state representation at the current time step based on the time step mapping relationship; perform projection processing on the static state representation to obtain the static representation to be fused; and perform element-wise weighting processing on the static representation to be fused based on the fusion gating vector corresponding to the current time step, and fuse the weighted static representation to be fused with the temporal state representation of the current time step according to the fusion position to obtain the fused state representation corresponding to the current time step.

[0198] According to an embodiment of this application, the fusion module 530 can also be used to generate a feedback modulation term based on the temporal state representation of the current time step; to perform weighted processing on the feedback modulation term based on the fusion gating vector corresponding to the current time step to obtain a gated feedback representation; and to perform fusion and normalization processing on the temporal state representation of the current time step, the weighted static representation to be fused, and the gated feedback representation to obtain the fusion state representation corresponding to the current time step.

[0199] According to an embodiment of this application, the result generation module 540 can also be used to input the fused state sequence into the aggregation layer in the state evaluation model to obtain the overall state representation of the target user; input the overall state representation into at least one evaluation head in the state evaluation model to obtain the state evaluation value; and generate the state evaluation result of the target user based on the state evaluation value, wherein the state evaluation result includes at least one of state level, service adaptation result or resource configuration reference result.

[0200] According to embodiments of this application, any multiple modules among the data acquisition module 510, encoding module 520, fusion module 530, and result generation module 540 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the data acquisition module 510, encoding module 520, fusion module 530, and result generation module 540 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or other reasonable means of integrating or packaging circuits, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the data acquisition module 510, encoding module 520, fusion module 530, and result generation module 540 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0201] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a target object state assessment method according to an embodiment of this application.

[0202] like Figure 6 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory 602 or a program loaded from a storage portion 608 into a random access memory 603. The processor 601 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0203] Random access memory 603 stores various programs and data required for the operation of electronic device 600. Processor 601, read-only memory 602, and random access memory 603 are interconnected via bus 604. Processor 601 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 602 and / or random access memory 603. It should be noted that the programs may also be stored in one or more memories other than read-only memory 602 and random access memory 603. Processor 601 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0204] According to embodiments of this application, the electronic device 600 may further include an input / output interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card, such as a local area network card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0205] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0206] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 602 described above, and / or random access memory 603, and / or one or more memories other than read-only memory 602 and random access memory 603.

[0207] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0208] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0209] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by processor 601, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0210] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0211] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0212] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A method for evaluating the state of a target object, characterized in that, The method includes: Obtain a time-series feature sequence and a static feature vector authorized by the target user. The time-series feature sequence is used to characterize the dynamic change information of the target user within a continuous time window, and the static feature vector is used to characterize the stable attribute information of the target user within the continuous time window. The temporal feature sequence is temporally encoded to obtain the temporal state representation corresponding to each time step; the static feature vector is statically encoded to obtain the static state representation. A fusion gate vector is generated based on the temporal state representation; the static state representation is weighted based on the fusion gate vector, and the weighted static state representation is fused with the temporal state representation of the corresponding time step to obtain the fused state representation for each time step; and Arrange the fusion state representations corresponding to each time step in the order of time steps to obtain the fusion state sequence, and generate the state evaluation result of the target user based on the fusion state sequence.

2. The method according to claim 1, characterized in that, The acquisition of the time-series feature sequence and static feature vector authorized by the target user includes: According to a preset time granularity, dynamic data authorized by the target user is acquired and time-aligned. Based on the aligned dynamic data, multiple time-series features corresponding to different time steps are generated, forming the time-series feature sequence; and Based on the frequency or period of change of attribute information authorized by the target user within the continuous time window, the stable attribute information is determined from the attribute information, and the static feature vector is generated based on the stable attribute information.

3. The method according to claim 2, characterized in that, The method further includes: Establish a time-series input channel corresponding to the time-series feature sequence and a static input channel corresponding to the static feature vector, respectively; Establish a time-step mapping relationship between the static feature vector and multiple time steps, wherein the time-step mapping relationship is used to determine the fusion position of the static state representation at each time step; and Based on the timing input channel, the static input channel, and the time step mapping relationship, the timing feature sequence and the static feature vector are input into the corresponding timing encoder and static encoder.

4. The method according to claim 1, characterized in that, The generation of the fusion gating vector based on the temporal state representation includes: Multiple state prototype vectors obtained through training are acquired, and the state prototype vectors are used to characterize the different state evolution stages of the target user within the continuous time window. Based on the similarity between the temporal state representation of the current time step and the prototype vectors of each state, the state affiliation weight corresponding to the current time step is determined. Based on the state attribution weights, multiple state prototype vectors are weighted to obtain the state basis vector corresponding to the current time step; and The state basis vector is subjected to gating mapping to obtain the fused gating vector corresponding to the current time step.

5. The method according to claim 4, characterized in that, The step of performing gated mapping on the state basis vector to obtain the fused gated vector corresponding to the current time step includes: Based on the temporal state representation of the current time step, the state change between adjacent time steps, and the local temporal context, construct the conditional context representation corresponding to the current time step; Inputting the conditional context representation into the gating network yields the scaling factor corresponding to the current time step; and The state basis vector is adjusted element-wise based on the scaling factor to obtain the fusion gating vector corresponding to the current time step.

6. The method according to claim 3, characterized in that, The static state representation is weighted based on the fusion gating vector, and then fused with the temporal state representation of the corresponding time step to obtain the fused state representation for each time step, including: Based on the time step mapping relationship, the fusion position of the static state representation in the current time step is determined; The static state representation is projected to obtain the static representation to be fused; and Based on the fusion gating vector corresponding to the current time step, the static representation to be fused is weighted element by element, and the weighted static representation to be fused is fused with the temporal state representation of the current time step according to the fusion position to obtain the fused state representation corresponding to the current time step.

7. The method according to claim 6, characterized in that, The step of fusing the weighted static representation to be fused with the temporal state representation of the current time step to obtain the fused state representation corresponding to the current time step includes: A feedback modulation term is generated based on the timing state representation of the current time step; Based on the fusion gate vector corresponding to the current time step, the feedback modulation term is weighted to obtain a gated feedback representation; and The temporal state representation of the current time step, the weighted static representation to be fused, and the gated feedback representation are fused and normalized to obtain the fused state representation corresponding to the current time step.

8. The method according to claim 1, characterized in that, The process of generating the state assessment result of the target user based on the fused state sequence includes: The fused state sequence is input into the aggregation layer of the state evaluation model to obtain the overall state representation of the target user. The overall state representation is input into at least one evaluation head in the state evaluation model to obtain a state evaluation value; and A status assessment result for the target user is generated based on the status assessment value. The status assessment result includes at least one of the following: status level, service adaptation result, or resource configuration reference result.

9. A target object state assessment device, characterized in that, The device includes: The data acquisition module is used to: acquire a time-series feature sequence and a static feature vector authorized by the target user, wherein the time-series feature sequence is used to characterize the dynamic change information of the target user within a continuous time window, and the static feature vector is used to characterize the stable attribute information of the target user within the continuous time window; The encoding module is used to: perform temporal encoding on the temporal feature sequence to obtain the temporal state representation corresponding to each time step; and perform static encoding on the static feature vector to obtain the static state representation. The fusion module is configured to: generate a fusion gating vector based on the temporal state representation; weight the static state representation based on the fusion gating vector; and fuse the weighted static state representation with the temporal state representation of the corresponding time step to obtain the fused state representation for each time step; and The result generation module is used to: arrange the fusion state representations corresponding to each time step in the order of time steps to obtain a fusion state sequence, and generate the state evaluation result of the target user based on the fusion state sequence.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

Images