Model training method and device, electronic equipment and computer storage medium

Abstract

The embodiment of the invention discloses a model training method and device, electronic equipment and a computer storage medium. The method comprises the following steps: inputting multiple pieces of network service data into a preset model, carrying out semantic coding processing on network service feature information through the preset model so as to obtain semantic feature information corresponding to each piece of network service data, and carrying out time sequence feature coding processing on network service time sequence information through the preset model so as to obtain semantic feature information corresponding to each piece of network service data; obtaining time sequence representation information corresponding to each piece of network service data; constructing network service input data corresponding to each piece of network service data according to the semantic feature information and the time sequence representation information so as to obtain a plurality of pieces of network service input data; and training a preset model based on the plurality of network service input data to obtain a target prediction model, the target prediction model being used for generating network service response variable information of the to-be-predicted network service in the future time interval.

Description

Technical Field

[0001] This specification relates to the field of information technology, and in particular to a model training method, apparatus, electronic device, and computer storage medium. Background Technology

[0002] In related technologies, network service operation no longer relies solely on traditional human control and experience-based judgment, but increasingly on intelligent strategy optimization and data-driven decision-making. The exposure, conversion, and adjustment behaviors of each network service unit on the relevant network service platform exhibit strong time-series dependence and intervention response characteristics.

[0003] Taking e-commerce scenarios as an example, the number of online services such as e-commerce products is limited, some products have been online for a short time, and the training samples are sparse. At the same time, traditional methods often rely on a single sales sequence, which limits the prediction accuracy, resulting in low generalization ability and prediction accuracy of the prediction model. Summary of the Invention

[0004] This specification provides a model training method, apparatus, electronic device, and computer storage medium, which can improve the generalization ability and prediction accuracy of prediction models.

[0005] Firstly, the embodiments of this specification provide a model training method, including:

[0006] Acquire multiple network service data and input the above multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information.

[0007] The semantic feature information of the network service is semantically encoded by the above-mentioned preset model to obtain the semantic feature information corresponding to each of the above-mentioned network service data.

[0008] The above-mentioned network service time-series information is processed by time-series feature encoding using the above-mentioned preset model to obtain the time-series representation information corresponding to each of the above-mentioned network service data.

[0009] Based on the semantic feature information and the temporal representation information mentioned above, construct the network service input data corresponding to each of the above network service data to obtain multiple network service input data corresponding to the above multiple network service data.

[0010] The preset model is trained based on the input data of the above-mentioned multiple network services to obtain the target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in the future time interval. The network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted.

[0011] In one possible implementation, the aforementioned network service feature information includes network service text features, and the aforementioned semantic feature information includes a first semantic vector corresponding to the aforementioned network service text features.

[0012] The semantic encoding process performed on the aforementioned network service feature information using the aforementioned preset model yields the semantic feature information corresponding to each of the aforementioned network service data, including:

[0013] The network service text features included in the above network service feature information are input into the first embedding module in the above preset model. The first embedding module performs semantic encoding processing on the network service text features to obtain the first semantic vector corresponding to the network service text features output by the first embedding module.

[0014] In one possible implementation, the aforementioned network service feature information includes network service table features, and the aforementioned semantic feature information includes a second semantic vector corresponding to the aforementioned network service table features.

[0015] The semantic encoding process performed on the aforementioned network service feature information using the aforementioned preset model yields the semantic feature information corresponding to each of the aforementioned network service data, including:

[0016] The network service table features included in the above network service feature information are converted into natural language fragments using the above-mentioned preset model.

[0017] The aforementioned natural language fragments are input into the second embedding module of the aforementioned preset model. The second embedding module performs semantic encoding processing on the aforementioned natural language fragments to obtain the second semantic vector corresponding to the network service table features output by the second embedding module.

[0018] In one possible implementation, the aforementioned network service time series information includes multiple network service time series data, which at least include: main time series data, future time series data, and related time series of the main time series.

[0019] The aforementioned time series representation information includes multiple network service time series representation data corresponding to the aforementioned multiple network service time series data. The aforementioned multiple network service time series representation data includes at least: the main time series representation data corresponding to the aforementioned main time series data, the future time series representation data corresponding to the aforementioned future time series data, and the related time series representation data of the aforementioned related time series data.

[0020] In one possible implementation, the aforementioned network service time-series information is processed by time-series feature encoding using the aforementioned preset model to obtain time-series representation information corresponding to each network service data, including:

[0021] For each network service time series data in the above network service time series information, if the above network service time series data is a multivariate sequence, then extract multiple univariate network service time series from the above network service time series data.

[0022] The above-mentioned univariate network service time series are normalized to obtain multiple normalized univariate time series;

[0023] Each normalized univariate time series is segmented to obtain the network service time series segment set corresponding to the above normalized univariate time series;

[0024] By fusing the multiple network service time series fragments corresponding to the above-mentioned normalized univariate time series, the time series representation information corresponding to the above-mentioned network service data is obtained.

[0025] In one possible implementation, the above-mentioned fusion processing of multiple network service time series fragment sets corresponding to the multiple normalized univariate time series is performed to obtain the time series representation information corresponding to the network service data, including:

[0026] The above-mentioned normalized univariate time series corresponding to multiple network service time series fragment sets are subjected to vertical fusion processing to obtain the interaction feature vector between variables at each time step. The above vertical fusion processing is used to model the dependency relationship between different variables of network services at the same time step.

[0027] The horizontal fusion process is performed on the multiple network service time series fragments corresponding to the above-mentioned normalized univariate time series to obtain the time dimension evolution feature vector of each variable. The above-mentioned horizontal fusion process is used to model the temporal dependency relationship between the same variable at different time steps corresponding to the above-mentioned network services.

[0028] Based on the interaction feature vectors between the variables and the evolution feature vectors of the time dimension, network service time series representation data corresponding to the network service time series data are generated, and time series representation information corresponding to the network service data is obtained.

[0029] In one possible implementation, the semantic feature information includes a first semantic vector corresponding to the text features of the network service and a second semantic vector corresponding to the table features of the network service. The temporal representation information includes multiple network service temporal representation data corresponding to the multiple network service time series data. The multiple network service temporal representation data includes at least: the main temporal representation data corresponding to the main time series, the future temporal representation data corresponding to the future time series data, and the related temporal representation data of the related time series data.

[0030] Based on the semantic feature information and temporal representation information, the above-mentioned network service input data corresponding to each network service data is constructed, resulting in multiple network service input data corresponding to the multiple network service data, including:

[0031] Based on the first semantic vector and the second semantic vector included in the above semantic feature information, corresponding network service text prompts and network service table prompts are constructed respectively.

[0032] Based on the aforementioned relevant time-series representation data included in the time-series representation information, a network service sequence prompt is constructed;

[0033] The network service context information is determined based on the aforementioned main time series representation data and the aforementioned future time series data included in the aforementioned time series representation information;

[0034] By concatenating the above-mentioned network service text prompts, network service table prompts, network service sequence prompts, and network service context information, network service input data corresponding to each of the above-mentioned network service data is generated, resulting in multiple network service input data corresponding to the above-mentioned multiple network service data.

[0035] In one possible implementation, the above-mentioned training of the preset model based on the input data from the multiple network services yields a target prediction model, including:

[0036] Input the above-mentioned network service input data into the above-mentioned preset model;

[0037] The above-mentioned preset model is used to fuse and model the above-mentioned network service text prompts, network service table prompts and network service sequence prompts in the above-mentioned network service input data to obtain network service control vectors. The above-mentioned network service control vectors are used to model covariate information.

[0038] The network service context information in the network service input data is modeled using the above-mentioned preset model to generate a network service prediction representation of the target variable. The network service prediction representation is used to model the temporal context features.

[0039] Based on the above network service prediction representation and the above network service control vector, the above preset model is iteratively optimized using adversarial training to obtain the target prediction model.

[0040] In one possible implementation, based on the aforementioned network service prediction representation and the aforementioned network service control vector, the aforementioned preset model is iteratively optimized using an adversarial training method to obtain the target prediction model, including:

[0041] Based on the aforementioned network service control vector, the generator included in the aforementioned preset model is used to generate network service intervention variables.

[0042] The discriminator included in the above-mentioned preset model is used to determine whether the above-mentioned network service intervention variables conform to the distribution of real network service intervention variables, and the discrimination result is obtained.

[0043] Based on the above discrimination results, calculate the adversarial loss and update the parameters of the discriminator.

[0044] Based on the distribution differences between the aforementioned network service intervention variables and the aforementioned real network service intervention variables, the parameters of the aforementioned generator are updated;

[0045] The generator and discriminator are alternately optimized and trained until the preset training termination condition is met, thereby generating the target prediction model.

[0046] Secondly, embodiments of this specification provide a method for predicting network service response variable information, including:

[0047] Acquire network service data to be predicted and output the network service data to be predicted to the target prediction model. The network service data to be predicted includes feature information of the network service to be processed and time series information of the network service to be processed.

[0048] The target semantic feature information of the network service to be processed is semantically encoded by the target prediction model described above to obtain the target semantic feature information corresponding to the network service data to be predicted.

[0049] The target prediction model described above is used to perform time series feature encoding on the time series information of the network service to be processed, thereby obtaining the target time series representation information corresponding to the network service data to be predicted.

[0050] Based on the above target semantic feature information and the above target temporal representation information, construct the target network service input data corresponding to the above network service data to be predicted;

[0051] The target prediction model is used to infer network service response variable information output by the target prediction model based on the target network service input data. The network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted.

[0052] The aforementioned target prediction model is a target prediction model generated by the model training method provided by the first aspect or any possible implementation of the first aspect.

[0053] Thirdly, embodiments of this specification provide a model training apparatus, including:

[0054] The first acquisition module is used to acquire multiple network service data and input the multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information.

[0055] The first processing module is used to perform semantic encoding processing on the network service feature information through the above-mentioned preset model to obtain the semantic feature information corresponding to each of the above-mentioned network service data.

[0056] The second processing module is used to perform time-series feature encoding on the time-series information of the network services through the preset model to obtain the time-series representation information corresponding to each network service data.

[0057] The first construction module is used to construct network service input data corresponding to each of the above network service data based on the above semantic feature information and the above temporal representation information, so as to obtain multiple network service input data corresponding to the above multiple network service data.

[0058] The training module is used to train a preset model based on the input data of the above-mentioned multiple network services to obtain a target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in the future time interval. The network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted.

[0059] Fourthly, embodiments of this specification provide a network service response variable information prediction device, comprising:

[0060] The second acquisition module is used to acquire network service data to be predicted and output the network service data to be predicted to the target prediction model. The network service data to be predicted includes network service feature information and network service time series information to be processed.

[0061] The third processing module is used to perform semantic encoding processing on the network service feature information to be processed through the above target prediction model to obtain the target semantic feature information corresponding to the network service data to be predicted.

[0062] The fourth processing module is used to perform time-series feature encoding processing on the time-series information of the network service to be processed through the above target prediction model to obtain the target time-series representation information corresponding to the network service data to be predicted.

[0063] The second construction module is used to construct the target network service input data corresponding to the network service data to be predicted based on the target semantic feature information and the target temporal representation information.

[0064] The inference module is used to perform inference based on the target network service input data through the target prediction model to obtain the network service response variable information output by the target prediction model. The network service response variable information is used to guide the generation of the intervention strategy for the network service to be predicted.

[0065] The aforementioned target prediction model is generated by the model training method provided by the first aspect or any possible implementation of the first aspect.

[0066] Fifthly, embodiments of this specification provide an electronic device, including: a processor and a memory;

[0067] The processor is connected to the memory.

[0068] The aforementioned memory is used to store executable program code;

[0069] The processor reads the executable program code stored in the memory to run the program corresponding to the executable program code, so as to execute the method provided by the first aspect of the embodiments of this specification or any possible implementation of the first aspect.

[0070] Sixthly, embodiments of this specification provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the method provided by the first aspect of the embodiments of this specification or any possible implementation thereof.

[0071] Seventhly, embodiments of this specification provide a computer program product containing instructions that, when run on a computer or processor, cause the computer or processor to execute the model training method provided by the first aspect or any possible implementation of the first aspect of the embodiments of this specification, or the network service response variable information prediction method provided by the second aspect.

[0072] In this embodiment, multiple network service data are acquired and input into a preset model. Each network service data includes network service feature information and network service temporal information. The preset model performs semantic encoding on the network service feature information to obtain semantic feature information corresponding to each network service data. The preset model also performs temporal feature encoding on the network service temporal information to obtain temporal representation information corresponding to each network service data. Based on the semantic feature information and temporal representation information, network service input data corresponding to each network service data is constructed, resulting in multiple network service input data. The preset model is trained based on these multiple network service input data to obtain a target prediction model. This target prediction model generates network service response variable information for the network service to be predicted within a future time interval. This network service response variable information guides the generation of intervention strategies for the network service to be predicted. Therefore, by jointly encoding and inputting network service feature information and temporal representation information, more expressive input data is constructed, which helps to accurately characterize the behavioral patterns of network services in the time dimension, improves the model's perception and response capabilities to changes in network service states, and thus effectively improves the generalization ability and prediction accuracy of the prediction model. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 An exemplary system architecture diagram of a model training method provided in the embodiments of this specification;

[0075] Figure 2 A schematic flowchart illustrating a model training method provided in an embodiment of this specification;

[0076] Figure 3 A schematic flowchart illustrating a model training method provided in an embodiment of this specification;

[0077] Figure 4 A schematic flowchart illustrating a model training method provided in an embodiment of this specification;

[0078] Figure 5 A schematic flowchart illustrating a model training method provided in an embodiment of this specification;

[0079] Figure 6 A schematic diagram of the architecture of a model training method provided in the embodiments of this specification;

[0080] Figure 7 A flowchart illustrating an encoding processing method provided in an embodiment of this specification;

[0081] Figure 8 A flowchart illustrating a method for processing time-series data of a network service, provided as an embodiment of this specification;

[0082] Figure 9 A flowchart illustrating a training method for a time-series causal effect estimation module provided in an embodiment of this specification;

[0083] Figure 10 A flowchart illustrating a method for predicting network service response variable information provided in an embodiment of this specification;

[0084] Figure 11 This is a schematic diagram of the structure of a model training device provided in the embodiments of this specification;

[0085] Figure 12 This is a schematic diagram of the structure of a network service response variable information prediction device provided in the embodiments of this specification;

[0086] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this specification. Detailed Implementation

[0087] To make the features and advantages of the embodiments of this specification more apparent and understandable, the technical solutions of the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the embodiments of this specification.

[0088] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those in this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the embodiments in this specification as detailed in the appended claims. Furthermore, in the description of the embodiments in this specification, unless otherwise stated, " / " means "or," for example, A / B can mean A or B; the word "and / or" in the text is merely a description of the relationship between related objects, indicating that three relationships can exist, for example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Additionally, in the description of the embodiments in this specification, "multiple" refers to two or more.

[0089] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0090] In related technologies, network service operation no longer relies solely on traditional human control and experience-based judgment, but increasingly on intelligent strategy optimization and data-driven decision-making. The exposure, conversion, and adjustment behaviors of each network service unit on the relevant network service platform exhibit strong time-series dependence and intervention response characteristics.

[0091] Taking online financial service products as an example, current methods mainly focus on individual wealth management products, predicting net subscriptions and redemptions for the next 7, 15, and 30 days within a 30 basis point fluctuation range of the current yield. However, this task faces several challenges: First, the limited number of products and the short online lifespan of some products result in sparse training samples and limited generalization ability of existing models; second, traditional methods often rely on single time-series variables, making it difficult to fully explore the factors behind complex behaviors, leading to insufficient prediction accuracy; third, when predicting volume and price under different yield conditions, it is necessary to possess causal correction and counterfactual inference capabilities to accurately model the impact of yield changes on subscription and redemption behavior.

[0092] Therefore, there is an urgent need to introduce a large model framework with time series modeling and causal reasoning capabilities, which can integrate multi-source heterogeneous data and effectively improve the model's generalization ability and prediction accuracy in sparse sample scenarios, thereby better supporting the dynamic management and strategy optimization of financial products.

[0093] This specification provides a model training method to address the technical problems of low generalization ability and low prediction accuracy of the aforementioned models.

[0094] Please see Figure 1 , Figure 1 This is an exemplary system architecture diagram of a model training method provided in the embodiments of this specification.

[0095] like Figure 1 As shown, the system architecture may include a terminal 101, a network 102, and a server 103. The network 102 serves as the medium for providing a communication link between the terminal 101 and the server 103. The network 102 may include various types of wired or wireless communication links, such as wired communication links including fiber optic cables, twisted-pair cables, or coaxial cables, and wireless communication links including Bluetooth communication links, Wireless-Fidelity (Wi-Fi) communication links, or microwave communication links, etc.

[0096] Terminal 101 can interact with server 103 via network 102 to receive messages from or send messages to server 103. Alternatively, terminal 101 can interact with server 103 via network 102 to receive messages or data sent to server 103 by other users. Terminal 101 can be hardware or software. When terminal 101 is hardware, it can be various electronic devices, including but not limited to tablet computers, laptops, and desktop computers. When terminal 101 is software, it can be installed in the aforementioned electronic devices and can be implemented as multiple software programs or software modules (e.g., to provide distributed services) or as a single software program or software module; no specific limitation is made here.

[0097] In the embodiments of this specification, during the training preparation phase, terminal 101 first obtains multiple network service data and inputs them into a preset model. Each network service data includes network service feature information and network service temporal information. Further, terminal 101 can perform semantic encoding processing on the network service feature information using the preset model to obtain semantic feature information corresponding to each network service data; it can also perform temporal feature encoding processing on the network service temporal information using the preset model to obtain temporal representation information corresponding to each network service data; subsequently, terminal 101 constructs network service input data corresponding to each network service data based on the semantic feature information and temporal representation information, thus obtaining multiple network service input data corresponding to multiple network service data; after the input data is prepared, the preset model can be trained using the multiple network service input data to obtain a target prediction model. The target prediction model is used to generate network service response variable information for the network service to be predicted within a future time interval, and the network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted.

[0098] Server 103 can be a server that provides various services. It should be noted that server 103 can be hardware or software. When server 103 is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When server 103 is software, it can be implemented as multiple software programs or software modules (e.g., used to provide distributed services), or as a single software program or software module; no specific limitations are made here.

[0099] Alternatively, the system architecture may not include server 103. In other words, server 103 may be an optional device in the embodiments of this specification. That is, the method provided in the embodiments of this specification can be applied to a system structure that only includes terminal 101. The embodiments of this specification do not limit this.

[0100] It should be understood that Figure 1The number of terminals, networks, and servers shown is only illustrative; the number can be any number of terminals, networks, and servers depending on the implementation requirements.

[0101] Please see Figure 2 , Figure 2 This is a flowchart illustrating a model training method provided in an embodiment of this specification. The execution entity in this embodiment can be a terminal executing the model training method, a processor within the terminal executing the model training method, or a training service for a feature extraction model within the terminal executing the model training method. For ease of description, the following example uses a processor within a terminal as the execution entity to illustrate the specific execution process of the model training method.

[0102] Please see Figure 2 , Figure 2 This is a flowchart illustrating a model training method provided in an embodiment of this specification. Figure 2 As shown, model training methods can include at least:

[0103] S202: Obtain multiple network service data and input the above multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information.

[0104] Optionally, network service data can be data related to network service products. In e-commerce scenarios, network service products can be online goods, stores, etc. The corresponding network service data may include, but is not limited to, product title, product category, product brand, price range, main image information, listing / delisting time, store level, participation in marketing activities, inventory status, and promotional strategies. Network service time-series information may include, but is not limited to, historical clicks, page views, add-to-cart volume, sales volume, conversion rate, collection / return rate, and time-series information of various behaviors at different time periods. In content recommendation scenarios, network service products can be content items such as videos and articles to be recommended. The corresponding network service data may include, but is not limited to, content title, content summary, tag information, creator identity, content type, publication time, channel, and keywords. Network service time-series information may include, but is not limited to, historical play counts, likes, comments, shares, user dwell time, completion rate, and trends in user interaction behavior over time.

[0105] In one embodiment, within a financial scenario, the online service product can be a wealth management product, fund, or insurance product offered by a financial institution. The online service feature information can specifically include textual features and / or tabular features. Taking a financial scenario as an example, the textual features of the online service can include, but are not limited to, product attribute information and time-series feature information. Product attribute information can include product name, product type, term type, risk level, return type, minimum investment amount, product description text, etc.; time-series feature information can include historical return series, unit net asset value change series, return trend, launch time, and other textual descriptions or structured time attributes that evolve over time. The aforementioned tabular features of the online service can include, but are not limited to: statistical features, such as current balance, redeemable amount, subscription and redemption volume in the past N days, subscription-to-redemption ratio, and return statistical indicators (mean, variance, etc.); statistical features of holding users, such as the number of current holding users, the proportion of customers with different risk levels, user retention rate, conversion rate, etc.; and recall similar sequence features, such as statistics on subscription and redemption behavior of similar wealth management products recalled based on historical behavior, and feature alignment information of products with similar return trends, etc.

[0106] In one embodiment, the network service time series information includes multiple network service time series data, which at least include: main time series data, future time series data, and related time series of the main time series.

[0107] The main time series data represents the historical behavior records of the online service product itself, such as core indicators like subscription volume, redemption volume, and net subscription / redemption volume of wealth management products, serving as the most direct basis for prediction targets. Related time series can include other auxiliary series that are behaviorally or semantically related to the main time series, such as subscription / redemption behavior subsequences for different customer subgroups (e.g., risk level, region, age group), historical behavior sequences of similar products (e.g., similar return structures), and migration behavior sequences of the same account between different products. Future time series can guide the model in making predictive inferences for future points in time. These are prior-known covariate information, specifically including work / holiday sequences and system activity schedules. Future time series may not contain complete historical covariates, requiring the model to have the ability to handle partially known and partially missing information, or to simulate intervention effects using conditional sampling, multi-branch prediction, or other methods during the training / inference phase.

[0108] It is worth noting that the main time series constitutes the core of the historical time series and is used to model the evolution of the network service product's own behavior; the relevant time series serves as auxiliary context, enhancing the stability and generalization ability of behavior prediction; and the future time series provides the model with conditional inference capabilities, used to simulate behavioral responses under external intervention conditions. These three types of series together constitute a complete network service time series input system, supporting high-precision prediction and causal decision generation.

[0109] In one embodiment, the preset model can be a deep learning model with semantic and temporal modeling capabilities, such as a large-scale temporal causal model containing embedding layers, Transformer structures, or variant structures.

[0110] S204: The semantic feature information of the above network services is semantically encoded using the above-mentioned preset model to obtain the semantic feature information corresponding to each of the above network service data.

[0111] In one embodiment, the aforementioned network service feature information can first be semantically encoded through the first embedding module in the pre-trained language large model to extract the initial semantic vector representation and obtain the semantic feature information corresponding to each of the aforementioned network service data.

[0112] Optionally, the aforementioned semantic feature information includes a first semantic vector corresponding to the aforementioned web service text features, and / or a second semantic vector corresponding to the aforementioned web service table features. The first semantic vector can represent the numerical representation of the web service text features in the semantic space, capturing the semantic features and implicit structural information of the text. The second semantic vector can be an embedding vector representation obtained by aligning the original non-linguistic modality (web service table features) to the semantic space supported by the language model.

[0113] S206: The time sequence information of the above network services is processed by time sequence feature encoding through the above preset model to obtain the time sequence representation information corresponding to each of the above network service data.

[0114] In some embodiments, the aforementioned preset model can construct time dependencies based on the input network service time-series information, capturing patterns such as periodicity, trends, or abrupt changes, thereby generating expressive time-series representation information. This time-series representation information can be used as input to the model's subsequent prediction module, fused with semantic feature information, to further generate network service response variable information, such as future subscription / redemption volume predictions, conversion rate change predictions, and yield response estimates.

[0115] Optionally, the aforementioned time series representation information includes multiple network service time series representation data corresponding to the aforementioned multiple network service time series data. The aforementioned multiple network service time series representation data includes at least: the master time series representation data corresponding to the aforementioned master time series data, the future time series representation data corresponding to the aforementioned future time series data, and the related time series representation data of the aforementioned related time series data.

[0116] The primary time series representation data is obtained by time-series encoding of the primary time series of the network service product. This primary time series representation data reflects the dynamic trends, periodic characteristics, and abrupt changes in the network service product's behavior over a past period. It provides the model with the primary behavioral drivers for predicting the target and forms the basis for predicting response variables (such as future subscription volume and playback volume). The related time series representation data is encoded from auxiliary time series data related to the primary time series. This related time series representation data can include subgroup characteristics (such as subscription and redemption behavior of high-net-worth users), references to similar products (such as historical trends of similar products), and derived indicator sequences (such as rolling averages and volatility). This related time series representation data can enhance the model's generalization ability, help model complex indirect behavioral drivers, provide richer contextual behavioral information, and improve prediction accuracy. Future time series representation data is generated by encoding currently known future covariate time series (such as holiday markers, preset returns, etc.). Specifically, it can reflect external prior factors that may affect changes in target behavior within future time intervals, including known covariates such as holidays and platform activity schedules. Future time series representation data can be used to enhance the forward-looking predictive ability of the model.

[0117] S208: Based on the above semantic feature information and the above temporal representation information, construct the network service input data corresponding to each of the above network service data, and obtain multiple network service input data corresponding to the above multiple network service data.

[0118] Among them, the aforementioned semantic feature information and temporal representation information can respectively represent the feature representation of network service products at the static semantic level and the dynamic behavioral level.

[0119] For example, the aforementioned semantic features and temporal representations can be fused using methods such as concatenation, attention fusion, multilayer perceptron fusion, or gating mechanisms to generate a unified multimodal input representation. This input representation, serving as the input data for the aforementioned network services, can simultaneously reflect the product's semantic attributes and historical behavioral patterns, improving the model's ability to predict future behavioral changes.

[0120] It is understandable that the input data from the aforementioned multiple network services can constitute the training sample set of the model, serving as the input for the subsequent model training process.

[0121] S210: Based on the input data of the above-mentioned multiple network services, the preset model is trained to obtain the target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in the future time interval. The network service response variable information is used to guide the generation of the intervention strategy for the network service to be predicted.

[0122] The target prediction model can be used to generate prediction results for the network service product to be predicted within a specified future time interval. These prediction results can include network service response variable information, such as: predictions of subscription volume, click-through rate, conversion rate, and behavioral responses to changes in yield for different product parameters (e.g., rate of return) over the next 7, 15, or 30 days.

[0123] Furthermore, the aforementioned network service response variable information can be used to guide the generation of corresponding intervention strategies. For example, in financial scenarios, based on the net subscription and redemption volume forecasts under different yield tiers, it can be determined whether to fine-tune the product yield; in e-commerce scenarios, product exposure strategies, resource allocation, or promotional schemes can be adjusted based on click or conversion forecasts, thereby improving overall platform revenue or user satisfaction.

[0124] In this embodiment, multiple network service data are acquired and input into a preset model. Each network service data includes network service feature information and network service temporal information. The preset model performs semantic encoding on the network service feature information to obtain semantic feature information corresponding to each network service data. The preset model also performs temporal feature encoding on the network service temporal information to obtain temporal representation information corresponding to each network service data. Based on the semantic feature information and temporal representation information, network service input data corresponding to each network service data is constructed, resulting in multiple network service input data. The preset model is trained based on these multiple network service input data to obtain a target prediction model. This target prediction model generates network service response variable information for the network service to be predicted within a future time interval. This network service response variable information guides the generation of intervention strategies for the network service to be predicted. Therefore, by jointly encoding and inputting network service feature information and temporal representation information, more expressive input data is constructed, which helps to accurately characterize the behavioral patterns of network services in the time dimension, improves the model's perception and response capabilities to changes in network service states, and thus effectively improves the generalization ability and prediction accuracy of the prediction model.

[0125] Please see Figure 3 , Figure 3 This is a flowchart illustrating a model training method provided in an embodiment of this specification. Figure 3As shown, model training methods can include at least:

[0126] S302: Acquire multiple network service data and input the above multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information.

[0127] Specifically, S302 is the same as S202, and will not be repeated here.

[0128] S304: Input the network service text features included in the above network service feature information into the first embedding module in the above preset model, and perform semantic encoding processing on the above network service text features through the first embedding module to obtain the first semantic vector corresponding to the network service text features output by the first embedding module.

[0129] The first embedding module, a component of the pre-defined model, can be used to convert raw text information into a semantic representation in vector form. This first embedding module can be based on word embedding, sentence embedding, or a pre-trained language model structure.

[0130] Understandably, the first semantic vector is the output of the first embedding module, representing the numerical representation of the text features of the network service in the semantic space, and can be used for subsequent feature fusion and model input.

[0131] S306: Convert the network service table features included in the above network service feature information into natural language fragments using the above-mentioned preset model.

[0132] Converting the network service table features included in the above network service feature information into natural language-like fragments can refer to transforming table data or structured information into sentences that are close to natural language through preset rules or lightweight models. For example, if the original network service table feature is "Current balance: 300,000", the corresponding conversion result can be: "The current balance of this product is 300,000 yuan".

[0133] It is understandable that the purpose of converting the network service table features included in the above network service feature information into natural language-like fragments is to achieve modality alignment, which facilitates subsequent unified processing through a language model.

[0134] S308: Input the above-mentioned natural language fragment into the second embedding module in the above-mentioned preset model, and perform semantic encoding processing on the above-mentioned natural language fragment through the second embedding module to obtain the second semantic vector corresponding to the network service table features output by the second embedding module.

[0135] The second embedding module is a component of the pre-defined model and is used to convert raw text information into a semantic representation in vector form. This module can be based on word embeddings, sentence embeddings, or a pre-trained language model structure. Optionally, the second embedding module and the first embedding module mentioned above can be the same module or different modules.

[0136] S310: The time sequence information of the above network services is processed by time sequence feature encoding through the above preset model to obtain the time sequence representation information corresponding to each of the above network service data.

[0137] Specifically, S310 and S206 are the same as above, and will not be repeated here.

[0138] S312: Based on the above semantic feature information and the above temporal representation information, construct the network service input data corresponding to each of the above network service data, and obtain multiple network service input data corresponding to the above multiple network service data.

[0139] Specifically, S312 is the same as S208, and will not be repeated here.

[0140] S314: Based on the input data of the above-mentioned multiple network services, the preset model is trained to obtain the target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in the future time interval. The network service response variable information is used to guide the generation of the intervention strategy for the network service to be predicted.

[0141] Specifically, S314 is the same as S210, and will not be repeated here.

[0142] The embodiments in this specification perform semantic encoding on the text features and table features of the network service during the training phase, and introduce a natural language-like transformation to achieve modality alignment. This allows static features to be processed through a unified language model embedding space, and the final network service input data can be used as high-quality training samples to effectively drive the learning of the target prediction model.

[0143] Please see Figure 4 , Figure 4 This is a flowchart illustrating a model training method provided in an embodiment of this specification. Figure 4 As shown, model training methods can include at least:

[0144] S402: Acquire multiple network service data and input the above multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information.

[0145] Specifically, S402 is the same as S210, and will not be repeated here.

[0146] S404: For each network service time series data in the above network service time series information, if the above network service time series data is a multivariate sequence, then extract multiple univariate network service time series from the above network service time series data.

[0147] Multivariate time series refers to a time series containing multiple variables at each time step. For example, a financial product might simultaneously record daily subscription volume, redemption volume, and yield. By independently extracting each variable, multiple series containing only one variable changing over time can be formed, i.e., univariate network service time series.

[0148] S406: Normalize the above multiple univariate network service time series respectively to obtain multiple normalized univariate time series.

[0149] Optionally, reversible instance normalization, i.e., standardization with a mean of 0 and a variance of 1, can be applied to each univariate network service time series.

[0150] It should be noted that, in order to make similar sequences (such as redemption behavior sequences from different subgroups) comparable in the feature space, normalized univariate time series can be normalized using shared invertible instance normalization parameters, thereby unifying their numerical scale and improving the model's generalization ability and comparative learning effect.

[0151] S408: Perform sharding on each normalized univariate time series to obtain the network service time series shard set corresponding to the above normalized univariate time series.

[0152] Optionally, for each normalized univariate time series, a sliding window approach can be used for sharding to divide each normalized univariate time series into multiple sequence blocks, thereby generating a corresponding network service time series shard set. There is a one-to-one correspondence between the normalized univariate time series and the network service time series shard set, and each network service time series shard set can include multiple sequence blocks (patch).

[0153] It is worth noting that overlap is allowed between any two adjacent sequence blocks to enhance the feature correlation between consecutive sequence blocks and support fine-grained modeling.

[0154] S410: Perform vertical fusion processing on the multiple network service time series fragment sets corresponding to the above multiple normalized univariate time series to obtain the interaction feature vector between variables at each time step. The above vertical fusion processing is used to model the dependency relationship between different variables of network services at the same time step.

[0155] Understandably, after normalization, to model the coupling relationships between multivariate time series, a mixing process can be used to fuse the sequence features using alternating vertical and horizontal methods. Vertical processing refers to inputting the representations of all different variables at each time step into an attention mechanism to calculate the dependencies between different variables, thereby obtaining the interaction feature vector between variables at each time step.

[0156] S412: Perform horizontal fusion processing on the multiple network service time series fragment sets corresponding to the above multiple normalized univariate time series to obtain the time dimension evolution feature vector of each variable. The above horizontal fusion processing is used to model the temporal dependency relationship between the same variable at different time steps corresponding to the above network services.

[0157] Optionally, horizontal fusion processing can be performed on the multiple network service time series fragments corresponding to the above-mentioned normalized univariate time series. This can refer to attention modeling of the representation of the same variable at different time steps, thereby extracting the temporal evolution feature representation of the variable. Through the alternating modeling of the above-mentioned vertical and horizontal attention mechanisms, the initial model can capture the coupling effect of variables in the network service at the same time step and the dynamic change pattern of univariate variables in the time dimension.

[0158] S414: Based on the above-mentioned interaction feature vector between variables and the above-mentioned time dimension evolution feature vector, generate network service time series representation data corresponding to the above-mentioned network service time series data, and obtain the time series representation information corresponding to the above-mentioned network service data.

[0159] In one embodiment, the aforementioned time series representation information includes multiple network service time series representation data corresponding to the aforementioned multiple network service time series data. The aforementioned multiple network service time series representation data includes at least: the master time series representation data corresponding to the aforementioned master time series data, the future time series representation data corresponding to the aforementioned future time series data, and the related time series representation data of the aforementioned related time series data.

[0160] Optionally, the features from the two directions mentioned above (inter-variable interaction feature vectors and time-dimensional evolution feature vectors) can be fused (e.g., through concatenation, weighted fusion, projection mapping, etc.) to form a unified, high-dimensional temporal feature representation vector, i.e., network service temporal representation data. This network service temporal representation data can be used for subsequent fusion with semantic feature information to construct complete network service input data, supporting further predictive modeling or intervention strategy generation.

[0161] S416: Based on the above semantic feature information and the above temporal representation information, construct the network service input data corresponding to each of the above network service data to obtain multiple network service input data corresponding to the above multiple network service data.

[0162] Specifically, S416 and S208 are the same as above, and will not be repeated here.

[0163] S418: Based on the input data of the above-mentioned multiple network services, the preset model is trained to obtain the target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in the future time interval. The network service response variable information is used to guide the generation of the intervention strategy for the network service to be predicted.

[0164] Specifically, S418 is the same as S210, and will not be repeated here.

[0165] This specification's embodiments perform preprocessing operations such as univariate decomposition, normalization, and sliding window slicing on the original multivariate time series. Based on this, a hybrid structure is introduced, combining vertical fusion modeling of inter-variable relationships and horizontal fusion modeling of temporal evolution dependencies, achieving a deep characterization of the dynamic relationships of the series in both the time and variable dimensions. Furthermore, the main time series data, related time series data, and future time series data are independently represented and fused to generate unified temporal representation information, which is then uniformly fused with semantic feature information to construct highly expressive and structurally complete network service input data. Through this refined sequence slicing and cross-modeling mechanism, the model's ability to model complex scenarios can be significantly enhanced.

[0166] Please see Figure 5 , Figure 5 This is a flowchart illustrating a model training method provided in an embodiment of this specification. Figure 5 As shown, model training methods can include at least:

[0167] S502: Acquire multiple network service data and input the above multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information.

[0168] Specifically, S502 is the same as S202, and will not be repeated here.

[0169] S504: The semantic feature information of the above network services is semantically encoded using the above-mentioned preset model to obtain the semantic feature information corresponding to each of the above network service data.

[0170] Specifically, S504 is the same as S204, and will not be repeated here.

[0171] S506: The time sequence information of the above network services is processed by time sequence feature encoding through the above preset model to obtain the time sequence representation information corresponding to each of the above network service data.

[0172] Specifically, S506 is the same as S206, and will not be repeated here.

[0173] S508: Construct corresponding network service text prompts and network service table prompts based on the first semantic vector and the second semantic vector included in the above semantic feature information.

[0174] Specifically, a corresponding text prompt for the online service is constructed based on the first semantic vector, and a corresponding table prompt for the online service is constructed based on the second semantic vector. These text and table prompts can be in natural language format and are used to provide the language model with a foundation for semantic understanding and structured feature hints, respectively, thereby providing more comprehensive contextual information support for subsequent sequence modeling and response variable prediction.

[0175] Optionally, web service text prompts and web service table prompts can be concatenated into a unified semantic sequence, or they can be fused and modeled separately in subsequent modules.

[0176] Understandably, constructing prompts refers to transforming vectors into semantic inputs that are acceptable to the language model.

[0177] S510: Construct network service sequence prompts based on the aforementioned relevant time series representation data included in the aforementioned time series representation information.

[0178] It is understandable that the aforementioned relevant time series data can be composed of multiple subsequence data that are related to the main time series data. For example, in financial application scenarios, subsequences related to the main time series data (such as total subscription volume) may include: subscription / redemption behavior sequences by customer group (such as subscription volume curves for institutional clients and retail investors), or subscription and redemption behavior sequences of similar services related to the current network service during historical periods.

[0179] Optionally, based on the aforementioned relevant time-series representation data, the sequence suggestion generation module can construct sequence suggestions for network services. For example, a sequence suggestion could be, "In the past 7 days, the overall subscription volume of similar wealth management products has increased, especially institutional clients' subscriptions surged on T-2 day, with an average daily volatility of 5.3%." This sequence suggestion can summarize the behavioral patterns of relevant sequences in a natural language-like manner, helping the language model to better understand the market or customer context of the current service in subsequent modeling.

[0180] Optionally, sequences highly similar to the current main time series data can be recalled from a pre-defined large-scale sequence library based on similarity calculation (such as cosine similarity). Then, prompt content can be constructed through Retrieval Augmented Generation (RAG) to improve the relevance and interpretability of the prompts.

[0181] S512: Determine network service context information based on the main time series representation data and the future time series data included in the aforementioned time series representation information.

[0182] Optionally, the master time series representation data is used to characterize the state evolution of the network service to be predicted within a historical time interval, while future time series data provides limited known information about future covariates. By combining the master time series representation data and future time series data, network service context information reflecting the contextual semantics of the current time step can be constructed. This context information can be used for time dependency modeling in subsequent target variable prediction, improving the ability to model time series consistency.

[0183] S514: Concatenate the above network service text prompts, the above network service table prompts, the above network service sequence prompts, and the above network service context information to generate network service input data corresponding to each of the above network service data, and obtain multiple network service input data corresponding to the above multiple network service data.

[0184] In one embodiment, the pre-obtained network service text prompts, network service table prompts, network service sequence prompts, and network service context information are concatenated according to a set structure (such as alignment by field paragraphs or concatenation dimensions) to form complete network service input data, so that they can be uniformly input into the large model for further modeling.

[0185] S516: Input the above-mentioned multiple network service input data into the above-mentioned preset model.

[0186] Optionally, the input data from the aforementioned multiple network services can be input into the main module of a preset model for processing. This main module can be a pre-trained large language model main module with frozen parameters.

[0187] S518: The above-mentioned network service text prompts, network service table prompts and network service sequence prompts in the above-mentioned network service input data are fused and modeled using the above-mentioned preset model to obtain the network service control vector, which is used to model covariate information.

[0188] Specifically, a pre-defined model can be used to fuse and model text prompts, table prompts, and sequence prompts for online services. A multi-head attention structure is used to capture the semantic interaction relationships between different covariates, ultimately forming a control vector for online services. This vector represents the information of the covariate (X) and is subsequently used for intervention allocation prediction and control in causal modeling.

[0189] S520: The network service context information in the network service input data is modeled using the above-mentioned preset model to generate a network service prediction representation of the target variable. The network service prediction representation is used to model the temporal context features.

[0190] In one embodiment, a pre-defined model receives contextual information from the aforementioned network service input data, uses temporal location encoding and self-attention mechanisms to jointly model the main time series and future time series, extracts their time-dependent features, and then generates a network service prediction representation for the target variable. This network service prediction representation is the representation vector of the outcome variable Y, which is used in the model for modeling and predicting response variables such as subscription volume and click-through rate.

[0191] S522: Based on the above network service control vector, the generator included in the above preset model is used to generate network service intervention variables.

[0192] Optionally, the network service control vector is processed by the generator G included in the preset model to generate the intervention variable A at the current time step t. t The goal of generator G is to learn to simulate possible intervention strategies from covariates, thereby estimating and generating intervention variables.

[0193] S524: The discriminator included in the above-mentioned preset model is used to determine whether the above-mentioned network service intervention variables conform to the distribution of real network service intervention variables, and the discrimination result is obtained.

[0194] Optionally, the pre-defined model also includes a discriminator D, whose goal is to determine whether the network service intervention variable conforms to the historical distribution of the true intervention variable. The discriminator D can receive the network service prediction representation and the network service intervention variable as input, and through training, predict the probability that the network service intervention variable belongs to the true intervention distribution, outputting the discrimination result.

[0195] S526: Based on the above discrimination results, calculate the adversarial loss and update the parameters of the discriminator.

[0196] Specifically, based on the discrimination results, the authenticity score of the generated intervention variable can be calculated. The discriminator is trained using cross-entropy loss to improve its recognition ability; after each training round, the discriminator parameters are updated according to this loss to enhance the discriminator's ability to distinguish between generated intervention variables and real intervention variables.

[0197] S528: Update the parameters of the generator based on the distribution differences between the above network service intervention variables and the above real network service intervention variables.

[0198] Optionally, the KL divergence (D) between the distribution of intervention variables generated by generator G and the distribution of the true intervention variables is considered. KL The difference is calculated, and the generator parameters are updated to make the generated intervention variables more difficult for the discriminator to distinguish, thereby achieving the adversarial training objective.

[0199] S530: Alternately optimize and train the generator and discriminator until the preset training termination condition is met, and generate the target prediction model.

[0200] Specifically, the parameter optimization process of the discriminator D and generator G is executed alternately, and training continues until the preset termination conditions are met (such as stable validation set performance and convergence of training loss), ultimately yielding the target prediction model. This target prediction model can jointly model future intervention variables and outcome variables, and possesses causal unbiased estimation capabilities.

[0201] This specification's embodiments introduce a multimodal fusion modeling mechanism based on a pre-trained language model, combining covariate control vectors and temporal context feature modeling to achieve joint prediction of network service response variables and intervention variables. During model training, an intervention variable is generated by a generator, and its authenticity is determined by a discriminator. The generator and discriminator are alternately optimized based on adversarial loss and distributional differences, effectively solving the temporal dependency confounding problem in causal modeling. This method achieves full decoupling of covariate information and temporal context, thereby improving the estimation accuracy of intervention variables and the stability of outcome variable predictions, effectively enhancing the model's generalization ability and causal explanatory power.

[0202] Furthermore, the following combination Figure 6 The provided embodiments illustrate the embodiments of this specification.

[0203] In the embodiments of this specification, the preset model may include an input module, a large language model main module, and an output module. First, the input module acquires multiple network service data and inputs these multiple network service data into the preset model. Each of the multiple network service data includes network service feature information and network service time series information. Specifically, the network service feature information may include network service text features and / or network service table features. The network service time series information includes multiple network service time series data, which at least include: main time series data, future time series data, and related time series of the main time series.

[0204] Optionally, for the above-mentioned network service text features, the network service text features can be input into the first embedding module in the preset model, and the network service text features can be semantically encoded by the first embedding module to obtain the first semantic vector corresponding to the network service text features output by the first embedding module.

[0205] Optionally, for the aforementioned web service table features, the web service table features included in the aforementioned web service feature information can be converted into natural language fragments using the aforementioned preset model; the aforementioned natural language fragments are then input into the second embedding module in the aforementioned preset model, and the aforementioned natural language fragments are semantically encoded by the aforementioned second embedding module to obtain the second semantic vector corresponding to the web service table features output by the aforementioned second embedding module. Specifically, the method for semantically encoding the aforementioned natural language fragments can be found in [reference needed]. Figure 7 ,like Figure 7 As shown, the features of the online service table can be divided into the following subcategories: product attributes and statistical features: such as assets under management (AUM), redeemable amount, yield, subscription and redemption volume, etc.; statistical features of holding users: such as asset status, risk preference, investment behavior characteristics, etc.; statistical features of similar products: similarity with other products and subscription and redemption volume of reference products obtained based on similarity retrieval, etc. In addition, the above-mentioned online service table features can be divided into two types: categorical features (such as rating, user type) and numerical features (such as balance, ratio, amount, etc.). For categorical features, the field name and corresponding value can be directly concatenated to form a description in natural language text form, serving as the corresponding text prompt; for numerical features, a semantic reprogramming mechanism can be introduced to achieve modality alignment and feature weight expression: First, for field name C... column Encode the pre-trained language model to obtain the semantic representation of the field as E. column =Encoder(C column = Pooling(Tokenize(C) column In this context, Tokenize(·) represents segmenting the field name into a sequence of tokens that can be processed by a large language model, and Pooling(·) represents pooling the token embedding results (to obtain the overall field semantic vector; then calculate the field query vector Q1 = E). column W Q1 W Q1 The projection weight matrix represents the query vector, used to map the semantic vector to the query space of the attention mechanism. Simultaneously, a prototype semantic space is pre-defined, represented by several representative typical field descriptions, and the prototype key vector K1 = E is obtained through embedding encoding. column W K1 Among them, W K1The projection weight matrix represents the key vector, and the prototype value vector V1 = E obtained through embedding encoding. column W V1 Among them, W V1 This represents the projection weight matrix of the value vector. Further, cross-attention is calculated between the field semantics and the prototype vector using the following formula:

[0206]

[0207] Where d represents the feature dimension of query vector Q1 and key vector K1. It can be used as a scaling factor to improve the numerical stability of attention weight calculation; softmax(·) represents normalization of the similarity score; Z attn This indicates the output of cross-attention.

[0208] Ultimately, the attention result Z can be used as a basis for further analysis. attn The value x corresponding to the field column This yields the semantically enhanced numerical feature vector, i.e., the second semantic vector H. column H column =linear(Z) attn )×x column , where linear(Z) attn ) represents the attention output Z attn A linear transformation is performed, which involves extracting semantic features through a learnable fully connected layer (LinearLayer). The numerical value of the resulting second semantic vector reflects the strength of the features.

[0209] For processing methods of multiple network service time series data included in the aforementioned network service time series information, please refer to [link to relevant documentation]. Figure 8 ,like Figure 8As shown, the network service time series information includes multiple network service time series data. These multiple network service time series data include at least: main time series data, future time series data, and related time series of the main time series. The processing methods for each network service time series data are similar. First, the multivariate series can be converted into univariate network service time series. For each univariate network service time series, normalization processing is performed to make different variables easier to model later under a uniform scale. The normalization method adopts reversible instance normalization, which independently standardizes each series to a mean of 0 and a variance of 1. Then, each normalized univariate time series is segmented to obtain the network service time series segment set corresponding to the normalized univariate time series. The network service time series segment set can contain multiple sequence blocks. The length of each sequence block can be represented as Lp, the step size of the sequence block is S, and the total sequence length is T. The number of sequence blocks can then be calculated as follows:

[0210]

[0211] Furthermore, an alternating vertical and horizontal attention fusion approach can be adopted. Vertical attention fusion is used to model the dependencies between different variables at the same time step; horizontal attention fusion is used to model the dynamic trends of the same variable across different time steps; finally, the representations of different variables are concatenated according to time steps and mapped to the input dimension of the language model. Next, a Time Large Language Model (TimeLLM)-style recoding mechanism is used to perform cross-attention matching between the above sequence block representations and the language model's pre-set token embeddings (such as "short-term rise," "continuous fall") through linear projection, forming a time series representation with semantic awareness. This process can specifically include:

[0212] First, multiple time-series blocks are used as input, mapped to query vector Q2, key vector K2, and value vector V2, respectively. Then, a multi-head attention mechanism is used to model the interaction dependencies between different variables, resulting in a fused representation. Specifically, in each attention head, the computation is as follows:

[0213]

[0214] Among them, H l This represents the input features of the l-th layer. Here are the parameter matrices for each head, and `Attention(·)` represents the standard scaled dot product attention calculation. The outputs of all attention heads are concatenated and then linearly transformed to form the multi-head attention output:

[0215] MultiHead(Hl )=Concat(head1,head2,…,head n W o

[0216] Next, residual connectivity and layer normalization are used to process the result:

[0217] Z l =LayerNorm(H l +MultiHead(H l )),H l+1 =LayerNorm(Z) l +(Z l ))

[0218] Among them, FFN(·) is a feedforward neural network module, which can be used to improve the nonlinear modeling ability of the model.

[0219] Furthermore, the output features H1, H2, ... H of different subsequences or variables are then... k Concatenate along the time dimension or the variable dimension, and pass through a linear mapping layer W H Projecting onto a vector space with the same dimension as the input of the pre-trained language model, a unified temporal block representation is constructed:

[0220] H patch =Concat(H1,H2,...H k W H

[0221] Among them, H1, H2, ... H k This represents the representation vector of different variables at a certain time step, reflecting the deep semantics of each variable at that time step. H This represents the linear transformation matrix (weight matrix), used to map the concatenated high-dimensional features to the input vector space of the pre-trained language large model, achieving dimensional alignment.

[0222] Next, the sequence block retrieval vector is constructed using the following formula:

[0223]

[0224] Where H patch The characteristic representation of a time sequence block, This represents the mapping weight matrix of the query vector.

[0225] Furthermore, the sequence block key vector and sequence block value vector are constructed using the following formulas:

[0226]

[0227] Where E1 represents the key vector input of prototype semantics, E1 represents the projection weight matrix of the prototype key vector; E2 represents the value vector input of the prototype semantics. The projection weight matrix represents the prototype value vector.

[0228] Furthermore, cross-attention is calculated using the following formula:

[0229]

[0230] Subsequently, the semantic representations of sequence blocks from different sources (main time series, related time series, and future time series) are integrated, that is, multiple network service time series fragment sets are fused to obtain the time series representation information corresponding to the aforementioned network service data. Next, network service input data corresponding to each network service data is constructed based on the semantic feature information and time series representation information, resulting in multiple network service input data. Finally, these multiple network service input data are input into the main module of the large language model, and the prediction results and causal inference results are output through the output module. Specifically, the output module can include multiple prediction heads, which can be used to perform regression or classification prediction tasks on target variables, as well as estimation tasks on future intervention variables. Among them: the first prediction head can be used to output network service response variable information, which can be the prediction results for network service response variables (such as future redemption volume, user behavior, service load, etc.), and modeling is performed based on the aforementioned time series representation information; the second prediction head can be used to output the estimation results of intervention variables (such as changes in yield, promotion strategies, service control signals, etc.), and causal modeling is performed based on the aforementioned covariate control vectors.

[0231] The embodiments described in this specification enable joint modeling of time dependence and intervention response in network service data. This not only improves prediction accuracy but also provides causal reasoning capabilities, facilitating better strategy optimization and response control. The output can be used for downstream business decisions, such as network service resource allocation, precise user operation, or intervention mechanism optimization.

[0232] In one embodiment, the aforementioned preset model further includes a time-series causal effect estimation module. For the training method of the time-series causal effect estimation module, please refer to... Figure 9 ,

[0233] In one embodiment, the aforementioned preset model further includes a time-series causal effect estimation module, used to achieve unbiased estimation of the intervention effect in network service data, thereby addressing the time-dependent confounding problem. See also... Figure 9First, a time-series causal graph model is constructed, which includes: covariates X0 and X1 representing network service status or contextual features at different times; intervention variables A0 and A1 representing operations or policies applied at time points 0 and 1; outcome variable Y2 representing the response observed at subsequent time points; and potential confounding variables U0 and U1 representing unobserved influencing factors.

[0234] like Figure 9 As shown on the left, in the presence of time-dependent confounding, the covariate X0 influences the intervention variable A1, causing the intervention distribution at each time point to depend on past covariates and interventions (e.g., the arrow between X0 and A1). This can lead to bias in estimating the intervention effect. To address this issue, the time representation Φ1 learned by the model needs to be conditionally independent of the current intervention variable A1, i.e., satisfying:

[0235]

[0236] Among them, A t The intervention variable represents the time point t (e.g., t = 1). Let t-1 represent the historical interventions up to and including time point t-1, and V represent the non-temporally static variable. To achieve the aforementioned goal of causal unbiasedness, an adversarial training approach can be used to construct a generator-discriminator network (GAN).

[0237] Optionally, adversarial training can be divided into two phases: a discriminator optimization phase and a generator optimization phase. In the discriminator optimization phase, the loss function can be expressed as:

[0238] L D (θ Φ ,θ D )=CE(D(Φ t (θ Φ );θ D ),A t )

[0239] Where, Φ t Let A represent the network service representation at time t. t This represents the intervention variable at the current time step, the true value label, and CE, which stands for Cross-Entropy loss. The training objective can be to fix the discriminator parameters θ. Φ Update the parameters θ of the large language model Φ This makes it impossible for discriminator D to determine based on Φ. t Effectively distinguish intervention variables, i.e., Φ t The data no longer carries information related to the distribution of the intervention.

[0240] During the generator optimization phase, the training objective is to minimize the following KL divergence:

[0241]

[0242] Among them, D KL The Kullback-Leibler divergence measures the difference between the generator representation and the distribution generated by historical interventions.

[0243] like Figure 9 As shown on the right, in the absence of time-dependent confounding, covariates X0 and X1 have no direct impact on subsequent intervention variable A1; the intervention distribution is determined solely by the historical intervention A0. In this scenario, the generator does not need to specifically adjust the covariate X... t The bias removal operation is performed on the hidden mixed information in the Φ, even if it is directly based on Φ. t Downstream forecasting can also yield relatively unbiased estimates of the intervention effect.

[0244] In the embodiments described in this specification, through the alternating optimization of the generator and discriminator, an intervention-independent time series representation is finally obtained, and a target prediction model with causal effect identification capability is trained. This training method effectively improves the model's generalization ability and intervention variable modeling ability under time-dependent confounding conditions.

[0245] Please refer to Figure 10 This paper provides a method for predicting network service response variable information, which includes at least the following steps:

[0246] S1002: Obtain the network service data to be predicted and output the network service data to be predicted to the target prediction model. The network service data to be predicted includes the feature information of the network service to be processed and the time series information of the network service to be processed.

[0247] S1004: The target semantic feature information of the network service to be processed is semantically encoded by the target prediction model to obtain the target semantic feature information corresponding to the network service data to be predicted.

[0248] S1006: The time-series feature encoding process of the above-mentioned network service time-series information to be processed is performed by the above-mentioned target prediction model to obtain the target time-series representation information corresponding to the above-mentioned network service data to be predicted.

[0249] S1008: Construct the target network service input data corresponding to the above-mentioned target semantic feature information and target temporal representation information.

[0250] S1010: Based on the target network service input data, the target prediction model is used to infer network service response variable information output by the target prediction model. This network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted. The target prediction model is generated by any of the model training methods described in the embodiments of this specification.

[0251] Optionally, the target prediction model can be a volume and price prediction model, which can be used to predict the net subscription and redemption volume of products under different yields over a period of time. Volume and price prediction can provide guidance for financial institutions to refine their operations, reduce product management friction costs (including investment in high-quality assets and optimization of idle funds), and provide long-term returns for products sold on the platform.

[0252] The inventive concept based on the above model training method, such as Figure 11 As shown in the embodiments of this specification, a model training apparatus 1100 for implementing the model training method described above is also provided. The model training apparatus 1100 includes:

[0253] The first acquisition module 1110 is used to acquire multiple network service data and input the multiple network service data into a preset model. Each network service data in the multiple network service data includes network service feature information and network service time sequence information.

[0254] The first processing module 1120 is used to perform semantic encoding processing on the network service feature information through the preset model to obtain the semantic feature information corresponding to each network service data.

[0255] The second processing module 1130 is used to perform time-series feature encoding processing on the time-series information of the network service through the preset model to obtain the time-series representation information corresponding to each network service data.

[0256] The first construction module 1140 is used to construct network service input data corresponding to each of the above network service data based on the above semantic feature information and the above temporal representation information, so as to obtain multiple network service input data corresponding to the above multiple network service data.

[0257] The training module 1150 is used to train a preset model based on the input data of the above-mentioned multiple network services to obtain a target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in the future time interval. The network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted.

[0258] In one possible implementation, the aforementioned network service feature information includes network service text features, and the aforementioned semantic feature information includes a first semantic vector corresponding to the aforementioned network service text features; the aforementioned first processing module 1120 includes:

[0259] The first processing unit is used to input the network service text features included in the network service feature information into the first embedding module in the preset model, and to perform semantic encoding processing on the network service text features through the first embedding module to obtain the first semantic vector corresponding to the network service text features output by the first embedding module.

[0260] In one possible implementation, the aforementioned network service feature information includes network service table features, and the aforementioned semantic feature information includes a second semantic vector corresponding to the aforementioned network service table features; the aforementioned first processing module 1120 includes:

[0261] The conversion unit is used to convert the network service table features included in the network service feature information into natural language fragments through the above-mentioned preset model.

[0262] The second processing unit is used to input the aforementioned natural language fragment into the second embedding module in the aforementioned preset model, and to perform semantic encoding processing on the aforementioned natural language fragment through the aforementioned second embedding module to obtain the second semantic vector corresponding to the network service table features output by the aforementioned second embedding module.

[0263] In one possible implementation, the aforementioned network service time series information includes multiple network service time series data, which at least include: a main time series data, future time series data, and related time series of the main time series; the aforementioned time series representation information includes multiple network service time series representation data corresponding to the aforementioned multiple network service time series data, which at least include: the main time series representation data corresponding to the main time series data, the future time series representation data corresponding to the future time series data, and the related time series representation data of the related time series data.

[0264] In one possible implementation, the second processing module 1130 includes:

[0265] The extraction unit is used to extract multiple univariate network service time series from the network service time series data in the above network service time series information if the above network service time series data is a multivariate sequence.

[0266] The normalization unit is used to normalize the above multiple univariate network service time series respectively, so as to obtain multiple normalized univariate time series;

[0267] The sharding unit is used to shard each normalized univariate time series to obtain the network service time series shard set corresponding to the above normalized univariate time series.

[0268] The fusion unit is used to fuse the multiple network service time series fragment sets corresponding to the above-mentioned multiple normalized univariate time series to obtain the time series representation information corresponding to the above-mentioned network service data.

[0269] In one possible implementation, the aforementioned fusion unit includes:

[0270] The first fusion subunit is used to perform vertical fusion processing on the multiple network service time series fragment sets corresponding to the above multiple normalized univariate time series to obtain the interaction feature vector between variables at each time step. The above vertical fusion processing is used to model the dependency relationship between different variables of network services at the same time step.

[0271] The second fusion subunit is used to perform horizontal fusion processing on the multiple network service time series fragment sets corresponding to the above multiple normalized univariate time series to obtain the time dimension evolution feature vector of each variable. The above horizontal fusion processing is used to model the temporal dependency relationship between the same variable at different time steps corresponding to the above network services.

[0272] The generation subunit is used to generate network service time series representation data corresponding to the network service time series data based on the above-mentioned interaction feature vector between variables and the above-mentioned time dimension evolution feature vector, and to obtain the time series representation information corresponding to the above-mentioned network service data.

[0273] In one possible implementation, the semantic feature information includes a first semantic vector corresponding to the text features of the network service and a second semantic vector corresponding to the table features of the network service. The temporal representation information includes multiple network service temporal representation data corresponding to the multiple network service time series data. The multiple network service temporal representation data includes at least: the main temporal representation data corresponding to the main time series, the future temporal representation data corresponding to the future time series data, and the related temporal representation data of the related time series data.

[0274] The aforementioned first building module 1140 includes:

[0275] The first construction unit is used to construct corresponding network service text prompts and network service table prompts based on the first semantic vector and the second semantic vector included in the above semantic feature information, respectively.

[0276] The second construction unit is used to construct network service sequence prompts based on the relevant time-series representation data included in the aforementioned time-series representation information.

[0277] The determining unit is used to determine network service context information based on the main time series representation data and the future time series data included in the aforementioned time series representation information;

[0278] The splicing unit is used to splice the above-mentioned network service text prompts, network service table prompts, network service sequence prompts, and network service context information to generate network service input data corresponding to each of the above-mentioned network service data, thereby obtaining multiple network service input data corresponding to the above-mentioned multiple network service data.

[0279] In one possible implementation, the training module 1150 described above includes:

[0280] The input unit is used to input the above-mentioned multiple network service input data into the above-mentioned preset model;

[0281] The first modeling unit is used to perform fusion modeling on the network service text prompts, network service table prompts and network service sequence prompts in the network service input data through the above-mentioned preset model to obtain a network service control vector, which is used to model covariate information.

[0282] The second modeling unit is used to perform feature modeling on the network service context information in the network service input data through the preset model, and generate a network service prediction representation of the target variable. The network service prediction representation is used to model the temporal context features.

[0283] The optimization unit is used to iteratively optimize the preset model based on the above network service prediction representation and the above network service control vector using an adversarial training method to obtain the target prediction model.

[0284] In one possible implementation, the optimization unit includes:

[0285] The generation subunit is used to generate network service intervention variables based on the network service control vector and using the generator included in the preset model.

[0286] The sub-unit is determined to determine whether the network service intervention variables conform to the distribution of real network service intervention variables through the discriminator included in the above-mentioned preset model, and to obtain the discrimination result;

[0287] The computational subunit is used to calculate the adversarial loss and update the parameters of the discriminator based on the above discrimination results;

[0288] The update subunit is used to update the parameters of the generator based on the distribution difference between the above-mentioned network service intervention variables and the above-mentioned real network service intervention variables;

[0289] The training subunit is used to alternately optimize and train the generator and discriminator until the preset training termination condition is met, thereby generating the target prediction model.

[0290] The division of modules in the above-described model training device is for illustrative purposes only. In other embodiments, the data processing device can be divided into different modules as needed to complete all or part of the functions of the above-described model training device. The implementation of each module in the model training device provided in the embodiments of this specification can be in the form of a computer program. This computer program can run on a terminal or server. The program modules constituted by this computer program can be stored in the memory of the terminal or server. When the computer program is executed by a processor, it implements all or part of the steps of the model training method described in the embodiments of this specification.

[0291] Based on the inventive concept of the above-mentioned method for predicting network service response variable information, such as Figure 12 As shown in the embodiments of this specification, a network service response variable information prediction device 1200 is also provided for implementing the network service response variable information prediction method described above. The network service response variable information prediction device 1200 includes:

[0292] The second acquisition module 1210 is used to acquire network service data to be predicted and output the network service data to be predicted to the target prediction model. The network service data to be predicted includes network service feature information and network service time series information to be processed.

[0293] The third processing module 1220 is used to perform semantic encoding processing on the network service feature information to be processed through the target prediction model to obtain the target semantic feature information corresponding to the network service data to be predicted.

[0294] The fourth processing module 1230 is used to perform time series feature encoding processing on the time series information of the network service to be processed through the above target prediction model to obtain the target time series representation information corresponding to the network service data to be predicted.

[0295] The second construction module 1240 is used to construct the target network service input data corresponding to the network service data to be predicted based on the target semantic feature information and the target temporal representation information.

[0296] The inference module 1250 is used to perform inference based on the target network service input data through the target prediction model to obtain the network service response variable information output by the target prediction model. The network service response variable information is used to guide the generation of the intervention strategy for the network service to be predicted. The target prediction model is a target prediction model generated by any of the model training methods provided in the embodiments of this specification.

[0297] The division of modules in the above-described network service response variable information prediction device is for illustrative purposes only. In other embodiments, the data processing device can be divided into different modules as needed to complete all or part of the functions of the above-described network service response variable information prediction device. The implementation of each module in the network service response variable information prediction device provided in the embodiments of this specification can be in the form of a computer program. This computer program can run on a terminal or server. The program modules constituted by this computer program can be stored in the memory of the terminal or server. When the computer program is executed by a processor, it implements all or part of the steps of the network service response variable information prediction method described in the embodiments of this specification.

[0298] This specification also provides an electronic device, which may be a server, and its internal structure diagram may be as follows: Figure 13 As shown, this electronic device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. The processor executes computer programs to implement a model training method or a network service response variable information prediction method.

[0299] Those skilled in the art will understand that Figure 13 The structures shown are merely block diagrams of some structures related to the embodiments of this specification, and do not constitute a limitation on the electronic devices to which the embodiments of this specification are applied. Specific electronic devices may include more or fewer components than those shown in the figures, or combine certain components, or have different component arrangements.

[0300] In one possible implementation, a computer storage medium is provided that stores instructions, which, when executed on a computer or processor, cause the computer or processor to perform one or more steps in the above embodiments. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium.

[0301] In one possible implementation, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0302] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer storage medium or transmitted through the computer storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0303] It should be noted that the information (including but not limited to network service data), data (including but not limited to data used for analysis, stored data, and displayed data), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the network service data involved in this specification were all obtained under full authorization.

[0304] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and its implementation can be combined arbitrarily.

[0305] The above-described embodiments are merely preferred embodiments of the embodiments in this specification and are not intended to limit the scope of the embodiments in this specification. Various modifications and improvements made by those skilled in the art to the technical solutions of the embodiments in this specification without departing from the design spirit of the embodiments in this specification should fall within the protection scope defined by the claims.

[0306] The foregoing has described specific embodiments of the embodiments described in this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A model training method, comprising: Acquire multiple network service data and input the multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information. The semantic feature information of the network service is semantically encoded using the preset model to obtain the semantic feature information corresponding to each network service data. The network service time-series information is processed by time-series feature encoding using the preset model to obtain the time-series representation information corresponding to each network service data. Based on the semantic feature information and the temporal representation information, construct the network service input data corresponding to each network service data to obtain multiple network service input data corresponding to the multiple network service data; The preset model is trained based on the input data of the multiple network services to obtain the target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in the future time interval. The network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted.

2. The method as described in claim 1, wherein the network service feature information includes network service text features, and the semantic feature information includes a first semantic vector corresponding to the network service text features; The step of semantically encoding the network service feature information using the preset model to obtain the semantic feature information corresponding to each network service data includes: The network service text features included in the network service feature information are input into the first embedding module in the preset model. The first embedding module performs semantic encoding processing on the network service text features to obtain the first semantic vector corresponding to the network service text features output by the first embedding module.

3. The method as described in claim 1, wherein the network service feature information includes network service table features, and the semantic feature information includes a second semantic vector corresponding to the network service table features; The step of semantically encoding the network service feature information using the preset model to obtain the semantic feature information corresponding to each network service data includes: The preset model is used to convert the network service table features included in the network service feature information into natural language fragments. The natural language-like fragment is input into the second embedding module in the preset model. The second embedding module performs semantic encoding processing on the natural language-like fragment to obtain the second semantic vector corresponding to the network service table features output by the second embedding module.

4. The method as described in claim 1, wherein the network service time series information includes multiple network service time series data, and the multiple network service time series data includes at least: Main time series data, future time series data, and related time series of the main time series; The time series representation information includes multiple network service time series representation data corresponding to the multiple network service time series data. The multiple network service time series representation data includes at least: the main time series representation data corresponding to the main time series data, the future time series representation data corresponding to the future time series data, and the related time series representation data of the related time series data.

5. The method as described in claim 4, wherein the step of performing time-series feature encoding processing on the network service time-series information through the preset model to obtain the time-series representation information corresponding to each network service data includes: For each network service time series data in the network service time series information, if the network service time series data is a multivariate sequence, then extract multiple univariate network service time series from the network service time series data; The multiple univariate network service time series are normalized respectively to obtain multiple normalized univariate time series; Each normalized univariate time series is segmented to obtain a set of network service time series segments corresponding to the normalized univariate time series; The multiple network service time series fragment sets corresponding to the multiple normalized univariate time series are fused to obtain the time series representation information corresponding to the network service data.

6. The method as described in claim 5, wherein fusing the multiple network service time series fragment sets corresponding to the multiple normalized univariate time series to obtain the time series representation information corresponding to the network service data includes: The multiple network service time series fragment sets corresponding to the multiple normalized univariate time series are subjected to vertical fusion processing to obtain the interaction feature vector between variables at each time step. The vertical fusion processing is used to model the dependency relationship between different variables of network services at the same time step. The horizontal fusion process is performed on the multiple network service time series fragment sets corresponding to the multiple normalized univariate time series to obtain the time dimension evolution feature vector of each variable. The horizontal fusion process is used to model the temporal dependency relationship of the same variable between different time steps corresponding to the network service. Based on the interaction feature vector between variables and the evolution feature vector of the time dimension, network service time series representation data corresponding to the network service time series data is generated, and the time series representation information corresponding to the network service data is obtained.

7. The method as described in claim 1, wherein the semantic feature information includes a first semantic vector corresponding to the text features of the network service and a second semantic vector corresponding to the table features of the network service, and the temporal representation information includes multiple network service temporal representation data corresponding to the multiple network service time series data, wherein the multiple network service temporal representation data includes at least: The main time series representation data corresponding to the main time series, the future time series representation data corresponding to the future time series data, and the related time series representation data of the related time series data; The step of constructing network service input data corresponding to each network service data based on the semantic feature information and the temporal representation information, to obtain multiple network service input data corresponding to the multiple network service data, includes: Based on the semantic feature information, including the first semantic vector and the second semantic vector, corresponding network service text prompts and network service table prompts are constructed respectively. A network service sequence prompt is constructed based on the relevant time-series representation data included in the time-series representation information; Network service context information is determined based on the main time series representation data and the future time series data included in the time series representation information; The network service text prompt, the network service table prompt, the network service sequence prompt, and the network service context information are concatenated to generate network service input data corresponding to each network service data, thus obtaining multiple network service input data corresponding to the multiple network service data.

8. The method as described in claim 7, wherein training the preset model based on the input data from the plurality of network services to obtain the target prediction model includes: Input the multiple network service input data into the preset model; The preset model is used to fuse and model the network service text prompts, network service table prompts and network service sequence prompts in the network service input data to obtain a network service control vector, which is used to model covariate information. The network service context information in the network service input data is modeled using the preset model to generate a network service prediction representation of the target variable. The network service prediction representation is used to model the temporal context features. Based on the network service prediction representation and the network service control vector, the preset model is iteratively optimized using adversarial training to obtain the target prediction model.

9. The method of claim 8, wherein the step of iteratively optimizing the preset model using adversarial training based on the network service prediction representation and the network service control vector to obtain the target prediction model includes: Based on the network service control vector, the network service intervention variables are generated using the generator included in the preset model; The discriminator included in the preset model determines whether the network service intervention variable conforms to the distribution of the real network service intervention variable, and obtains the discrimination result; Based on the discrimination results, the adversarial loss is calculated and the parameters of the discriminator are updated; The generator parameters are updated based on the distribution differences between the network service intervention variables and the real network service intervention variables. The generator and the discriminator are alternately optimized and trained until a preset training termination condition is met, thereby generating a target prediction model.

10. A method for predicting network service response variable information, comprising: Acquire network service data to be predicted and output the network service data to be predicted to the target prediction model. The network service data to be predicted includes network service feature information and network service time series information to be processed. The target prediction model is used to perform semantic encoding on the feature information of the network service to be processed to obtain the target semantic feature information corresponding to the network service data to be predicted. The target prediction model is used to perform time-series feature encoding on the time-series information of the network service to be processed, so as to obtain the target time-series representation information corresponding to the network service data to be predicted. Based on the target semantic feature information and the target temporal representation information, construct the target network service input data corresponding to the network service data to be predicted; The target prediction model infers based on the target network service input data to obtain network service response variable information output by the target prediction model. The network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted. The target prediction model is generated by the model training method as described in any one of claims 1 to 9.

11. A model training device, comprising: The first acquisition module is used to acquire multiple network service data and input the multiple network service data into a preset model. Each network service data includes network service feature information and network service time sequence information. The first processing module is used to perform semantic encoding processing on the network service feature information through the preset model to obtain the semantic feature information corresponding to each network service data. The second processing module is used to perform time-series feature encoding on the network service time-series information through the preset model to obtain the time-series representation information corresponding to each network service data. The first construction module is used to construct network service input data corresponding to each network service data according to the semantic feature information and the temporal representation information, so as to obtain multiple network service input data corresponding to the multiple network service data. The training module is used to train a preset model based on the input data of the multiple network services to obtain a target prediction model. The target prediction model is used to generate network service response variable information of the network service to be predicted in a future time interval. The network service response variable information is used to guide the generation of intervention strategies for the network service to be predicted.

12. A network service response variable information prediction device, comprising: The second acquisition module is used to acquire network service data to be predicted and output the network service data to be predicted to the target prediction model. The network service data to be predicted includes network service feature information and network service time series information to be processed. The third processing module is used to perform semantic encoding processing on the network service feature information to be processed through the target prediction model to obtain the target semantic feature information corresponding to the network service data to be predicted. The fourth processing module is used to perform time-series feature encoding processing on the time-series information of the network service to be processed through the target prediction model to obtain the target time-series representation information corresponding to the network service data to be predicted. The second construction module is used to construct the target network service input data corresponding to the network service data to be predicted based on the target semantic feature information and the target temporal representation information. The inference module is used to perform inference based on the target network service input data through the target prediction model to obtain the network service response variable information output by the target prediction model. The network service response variable information is used to guide the generation of the intervention strategy for the network service to be predicted. The target prediction model is generated by the model training method as described in any one of claims 1 to 9.

13. An electronic device, comprising: Processor and memory; The processor is connected to the memory; The memory is used to store executable program code; The processor runs a program corresponding to the executable program code stored in the memory to perform the method as described in any one of claims 1-10.

14. A computer storage medium storing a plurality of instructions adapted for loading by a processor and performing the method steps of any one of claims 1-10.

15. A computer program product comprising instructions that, when run on a computer or processor, cause the computer or processor to perform the method as described in any one of claims 1-10.

Images