Data prediction method and device

Abstract

The invention discloses a data prediction method and device, and the method comprises the steps: obtaining the historical sequence data of an object in a target business scene and the scene description information of the target business scene, and the historical sequence data comprises the data of the object at at least one historical time point; field matching information of the object is determined based on the scene description information, and the field matching information is used for representing the matching degree of a target business scene corresponding to the object and multiple reference business fields; determining a first prediction sequence feature of the object by using a special prediction model corresponding to the reference service field based on the historical sequence data; and determining prediction sequence data of the object based on the field matching information and a first prediction sequence feature determined by the special prediction model corresponding to each reference service field, the prediction sequence data including prediction data of the object at at least one prediction time point.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a data prediction method and device. Background Art

[0002] Time series forecasting (abbreviated as time series forecasting) refers to predicting the data of an object at several future time points based on the known time series data of the object. Time series forecasting can provide a reliable data basis for business operations and decision-making. For example, in the power industry, the power consumption in the future can be predicted based on the known time series data of power consumption.

[0003] In order to perform time series prediction on objects in a business scenario, a prediction model applicable to the business scenario needs to be built in advance. However, different industries involve different business fields, and even the same business field may involve multiple business scenarios. Therefore, if a prediction model for time series prediction is built separately for each business scenario, it will not only lead to a high complexity of time series prediction, but also lead to more resource consumption. Summary of the invention

[0004] On the one hand, the present application provides a data prediction method, comprising:

[0005] Obtaining historical sequence data of an object in a target business scenario and scenario description information of the target business scenario, wherein the historical sequence data includes data of the object at at least one historical time point;

[0006] Based on the scenario description information, determining domain matching information of the object, wherein the domain matching information is used to characterize a matching degree between a target business scenario corresponding to the object and a plurality of reference business domains;

[0007] Based on the historical sequence data, determining a first prediction sequence feature of the object using a dedicated prediction model corresponding to the reference business field;

[0008] Based on the domain matching information and the first prediction sequence feature determined by the dedicated prediction model corresponding to each reference business domain, the prediction sequence data of the object is determined, and the prediction sequence data includes prediction data of the object at at least one prediction time point.

[0009] In a possible implementation, determining the prediction sequence data of the object based on the domain matching information and the first prediction sequence feature determined by the dedicated prediction model corresponding to each reference business domain includes:

[0010] Processing the fusion features of the first prediction sequence features corresponding to each reference business field using the comprehensive prediction model to obtain the second prediction sequence features of the object;

[0011] Based on the domain matching information and the second predicted sequence feature, predicted sequence data of the object is determined.

[0012] In yet another possible implementation, determining the predicted sequence data of the object based on the domain matching information and the second predicted sequence feature includes:

[0013] Based on the matching degree between each reference business field represented by the field matching information and the target business scenario, weighted summing is performed on the first prediction sequence features determined by the dedicated prediction model corresponding to each reference business field to obtain a comprehensive feature after weighted summing of the features;

[0014] Based on the comprehensive feature and the second predicted sequence feature, predicted sequence data of the object is determined.

[0015] In yet another possible implementation, determining the prediction sequence data of the object based on the domain matching information and the first prediction sequence feature determined by the dedicated prediction model corresponding to each reference business domain includes:

[0016] Based on the matching degree between each reference business field represented by the field matching information and the target business scenario, weighted summing is performed on the first prediction sequence features determined by the dedicated prediction model corresponding to each reference business field to obtain a comprehensive feature after weighted summing of the features;

[0017] Based on the comprehensive features, determining a second predicted sequence feature of the object using a comprehensive prediction model;

[0018] Based at least on the second predicted sequence feature, predicted sequence data for the subject is determined.

[0019] In yet another possible implementation, determining the predicted sequence data of the object based at least on the second predicted sequence feature includes:

[0020] Determining predicted sequence data of the object using an output layer module based on the comprehensive feature and the second predicted sequence feature;

[0021] or,

[0022] Based on the first prediction sequence features and the second prediction sequence features corresponding to each reference business field, the prediction sequence data of the object is determined using an output layer module.

[0023] In yet another possible implementation, determining the domain matching information of the object based on the scene description information includes:

[0024] Based on the scenario description information, determining the domain classification information of the object using a domain classification model, wherein the domain classification information is used to characterize the probability that the target business scenario belongs to each of the multiple reference business domains;

[0025] Based on the domain classification information and the historical sequence data, domain matching information of the object is determined using a domain distribution model, wherein the domain matching information is used to characterize the degree to which the dedicated prediction models corresponding to the multiple reference business domains are suitable for processing the object.

[0026] In another possible implementation, the dedicated prediction model corresponding to the reference business field and the comprehensive prediction model are obtained in the following manner:

[0027] In a first training phase, based on at least one historical sequence data sample of an object sample in the reference business field, a dedicated prediction model corresponding to the reference business field is trained;

[0028] In the second training stage, based on at least one historical sequence data sample of object samples in different sample business scenarios and the scenario description information sample corresponding to the sample business scenarios, combined with the dedicated prediction models of each reference business field obtained through training, a comprehensive prediction model is trained.

[0029] In another possible implementation, the at least one historical sequence data sample based on the object samples in different sample business scenarios and the scenario description information sample corresponding to the sample business scenario are combined with the dedicated prediction models of each reference business field obtained by training to obtain a comprehensive prediction model, including:

[0030] For each historical sequence data sample in each sample business scenario, based on the scenario description information sample of the historical sequence data sample, determine the domain matching information sample of the object sample, wherein the domain matching information sample is used to characterize the matching degree between the sample business scenario and the multiple reference business domains;

[0031] Based on the first sequence data sample in the historical sequence data sample, using the dedicated prediction model corresponding to the reference business field to determine the first prediction sequence feature sample of the object sample;

[0032] Determine the prediction sequence data sample of the object sample based on the field matching information sample and the first prediction sequence feature sample determined by the dedicated prediction model corresponding to each reference business field, combined with the comprehensive prediction model;

[0033] Based on the second sequence data sample in the historical sequence data sample corresponding to the object sample and the predicted sequence data sample, the model parameters of the comprehensive prediction model are adjusted, and the step of determining the first predicted sequence feature sample of the object sample is returned to be executed until the training requirements are met, wherein the second sequence data sample is a sequence data sample located after the first sequence data sample in the historical sequence data sample.

[0034] In yet another possible implementation, the first training stage further includes:

[0035] Training a domain classification model based on at least one scene description information sample and a reference business domain annotated by the scene description information sample;

[0036] The determining of the domain matching information sample of the object sample based on the scene description information sample of the historical sequence data sample includes:

[0037] Based on the scene description information sample of the historical sequence data sample, determining the domain classification information sample of the object sample using the domain classification model;

[0038] Based on the domain classification information sample and the historical sequence data sample, determining the domain matching information sample of the object sample using the domain distribution model to be trained;

[0039] While adjusting the model parameters of the comprehensive prediction model, it also includes: adjusting the model parameters of the field distribution model.

[0040] In another aspect, the present application further provides a data prediction device, comprising:

[0041] A data acquisition unit, configured to acquire historical sequence data of an object in a target business scenario and scenario description information of the target business scenario, wherein the historical sequence data includes data of the object at at least one historical time point;

[0042] A domain matching unit, configured to determine domain matching information of the object based on the scenario description information, wherein the domain matching information is used to characterize a matching degree between a target business scenario corresponding to the object and a plurality of reference business domains;

[0043] A feature prediction unit, configured to determine a first prediction sequence feature of the object based on the historical sequence data and using a dedicated prediction model corresponding to the reference business field;

[0044] A data prediction unit is used to determine the prediction sequence data of the object based on the field matching information and the first prediction sequence characteristics determined by the dedicated prediction model corresponding to each reference business field, and the prediction sequence data includes the prediction data of the object at at least one prediction time point. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the originals and elements are not necessarily drawn to scale.

[0046] Figure 1 A schematic diagram of a flow chart of the data prediction method provided in this application;

[0047] Figure 2 A schematic diagram of another flow chart of the data prediction method provided in this application;

[0048] Figure 3 A schematic diagram of a framework of an implementation process of the data prediction method in this application;

[0049] Figure 4 A schematic diagram of another flow chart of the data prediction method provided in this application;

[0050] Figure 5 A schematic diagram of a process for training a comprehensive prediction model in this application;

[0051] Figure 6 An example diagram of training different models in stages in this application;

[0052] Figure 7 A flowchart of the model training method provided in this application;

[0053] Figure 8 A schematic diagram of a structure of a data prediction device provided in this application;

[0054] Fig. 9 A schematic diagram of the composition architecture of an electronic device provided in this application. DETAILED DESCRIPTION

[0055] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method section of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0056] The embodiments of the present application are described below in conjunction with the accompanying drawings. Those skilled in the art will appreciate that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0057] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, which is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0058] like Figure 1 , showing a flow chart of the data prediction method provided by the present application, the method of this embodiment can be applied to electronic devices, which may be mobile phones, tablets, servers or device nodes in service systems, etc., without limitation.

[0059] The data prediction method of this embodiment may include:

[0060] S101, obtaining historical sequence data of objects in a target business scenario and scenario description information of the target business scenario.

[0061] In this application, the target business scenario can be any business scenario that requires time series prediction of the objects involved. The objects in the target business scenario are the objects that require time series prediction. There is no restriction on the specific scenario of the target business scenario and the business field to which it belongs. The objects that require time series prediction may be different for different target business scenarios. For example, the target business scenario can be a product sales prediction scenario, then the object that needs to be predicted is product sales; if the target business scenario is a meteorological rainfall prediction scenario, then the object that needs to be predicted is rainfall.

[0062] The scenario description information of the target business scenario is used to describe the scenario characteristics in the target business scenario. The specific data prediction content involved in the target business scenario can be determined through the scenario description information. For example, the scenario description information may include object information such as the name and characteristics of the object that requires data prediction in the target business scenario, as well as part or all of the information such as the purpose of data prediction in the target business scenario. Of course, the scenario description information may also include other information related to the target business scenario, without limitation.

[0063] For example, taking the target business scenario of predicting the sales volume of a certain product A as an example, the scenario description information can be: predicting the future sales of product A to provide data basis for product production or supply.

[0064] In this application, the sequence data of an object is also referred to as a time series or time series data, which includes the data of the object at at least one time point. In this application, the time point at which the data of the object is known (or has been determined) is called a historical time point, and the time point at which the data of the object needs to be predicted is called a predicted time point.

[0065] Correspondingly, the historical sequence data of an object is also referred to as the historical time series of the object, which includes data of the object at at least one historical time point.

[0066] S102: Determine domain matching information of the object based on the scene description information.

[0067] The domain matching information is used to characterize the matching degree between the target business scenario corresponding to the object and multiple reference business domains. For example, the domain matching information may include: the matching degree between the target business scenario corresponding to the object and each reference business domain, or the probability that the target business scenario belongs to each target reference business domain. Of course, the matching degree between the target business scenario and multiple reference business domains can also be characterized in other forms.

[0068] The reference business field may be a pre-selected representative business field involving data prediction.

[0069] It should be noted that the reference business domain can be a specific data prediction business scenario, for example, the reference business domain can be a business scenario for sales forecasting of a certain type of product, a business scenario for traffic flow forecasting, or a business scenario for product supply forecasting.

[0070] The reference business field can also be a data prediction business field that includes business scenarios for multiple time series data prediction. For example, considering that some industry fields may involve time series predictions for multiple objects, it naturally involves business scenarios for multiple time series predictions. However, the specific similarities of time series predictions for different objects in the same industry field can be applied to time series predictions for multiple objects in the same industry field using the same prediction model. In this case, the industry field can be selected as a reference business field involving time series predictions. For example, taking the financial prediction field as an example, financial predictions can involve stock predictions and exchange rate predictions, etc., but stock predictions and financial predictions can use the same prediction model for time series predictions. Therefore, financial predictions are used as a reference business field.

[0071] S103, based on the historical sequence data, using a dedicated prediction model corresponding to a reference business field to determine a first prediction sequence feature of the object.

[0072] In the present application, each reference business domain corresponds to a dedicated prediction model suitable for time series prediction of objects in the reference business domain. For each reference business domain, it is necessary to determine the first prediction sequence feature of the object based on the historical sequence data using the dedicated prediction model corresponding to the reference business domain. The first prediction sequence feature is used to characterize the data feature of the prediction data of the object at at least one prediction time point.

[0073] It is understandable that the model parameters of the dedicated prediction models corresponding to different reference business fields are different, and therefore, the first prediction sequence characteristics determined by the dedicated prediction models corresponding to different reference business fields will also be different.

[0074] The dedicated prediction model corresponding to the reference business field can be any machine learning model or a combination of machine learning models. For example, the dedicated prediction model can be a deep learning model such as a convolutional neural network model, or other types of models, without limitation. Moreover, the model structures of the dedicated prediction models corresponding to different reference business fields can be the same or different, and can be selected based on the prediction requirements of different reference business fields, without limitation.

[0075] In a possible implementation, for any reference business domain, the dedicated prediction model corresponding to the reference business domain may be trained based on at least one historical sequence data sample of an object sample in the reference business domain. The specific implementation process of training the dedicated prediction model based on the at least one historical sequence data sample may not be limited.

[0076] For ease of distinction, this application refers to the historical sequence data used in the training model as historical sequence data samples, and the objects corresponding to the historical sequence data samples as object samples.

[0077] It should be noted that the order of steps S102 and S103 is not limited to Figure 1 As shown, in practical applications, the order of these two steps can be interchanged or performed simultaneously.

[0078] S104, determining prediction sequence data of the object based on the domain matching information and the first prediction sequence features determined by the dedicated prediction models corresponding to the reference business domains.

[0079] The predicted sequence data includes predicted data of the object at at least one predicted time point.

[0080] In the present application, the prediction sequence data of the object is obtained by combining the first prediction sequence features determined by the dedicated prediction models corresponding to the multiple reference business fields, which is equivalent to using the dedicated prediction models of multiple reference business fields to realize the time series data prediction of the object in the target business scenario.

[0081] The inventor of the present application has found through research that, since time series predictions in different business fields have certain correlations, for time series predictions in any business scenario, the business scenario can be mapped to a combination of time series predictions in one or more business fields. On this basis, after pre-constructing a plurality of representative reference business field dedicated prediction models, for any business scenario, at least some of the reference business field dedicated prediction models can be combined into a prediction model suitable for the business scenario.

[0082] On the basis of the above, since the domain matching information can characterize the degree of matching between the target business scenario corresponding to the object and multiple reference business fields, therefore, combined with the domain matching information, the dedicated prediction models of each reference business field can be integrated to determine the first prediction sequence characteristics and determine the prediction sequence data of the object in the target business scenario.

[0083] From the above content, it can be seen that, unlike the current practice of building a separate prediction model for each business scenario for data prediction, the present application only needs to pre-build a dedicated prediction model for each of the reference business fields. On this basis, for any target business scenario, the present application will determine the field matching information based on the scenario description information of the target business scenario, and based on the historical sequence data of the object in the target business scenario, use the dedicated prediction model of each reference business field to determine the first prediction sequence feature of the object. Since the field matching information can characterize the degree of matching between the target business scenario and each reference business field, therefore, combined with the field matching information, and using the first prediction sequence feature determined by the dedicated prediction model of each reference business field, it is possible to reasonably determine the prediction sequence data of the object, so that the scheme of the present application is applicable to time series prediction in any business scenario, and naturally there is no need to build a prediction model for each time series prediction scenario, thereby reducing the complexity of time series prediction and reducing the resource consumption caused by building a prediction model for each time series prediction scenario.

[0084] In the present application, there are many possibilities for the specific implementation of determining the domain matching information of the object.

[0085] For example, in a possible implementation, the domain classification information of the object can be determined based on the scenario description information using a domain classification model. The domain classification information is used to characterize the probability that the target business scenario belongs to multiple reference business domains. For example, the domain classification information can be a vector including the probability that the target business scenario belongs to multiple reference business domains. In this implementation, the degree of matching between the target business scenario and the reference business domain is represented by the probability that the target business scenario belongs to the reference business domain.

[0086] The domain classification model may be a pre-trained classification model, and there is no restriction on the specific implementation of the trained domain classification model. For example, the domain classification model may be trained based on at least one scene description information sample and a reference business domain annotated by each scene description information sample.

[0087] There are many possible types of specific classification models used by the field classification model, and there is no restriction on this.

[0088] In another possible implementation, considering that the purpose of determining the degree of matching between the target business scenario and each reference business field is to determine which reference business field's dedicated prediction model is more suitable for time series prediction of the object in the target business scenario, based on this, in order to be able to more accurately determine the degree of suitability of the dedicated prediction model corresponding to each reference business field for the object in the target business scenario, the present application can also determine the domain matching information of the object based on the domain classification information and the historical sequence data of the object using the domain classification model after determining the domain classification information of the object.

[0089] The domain matching information is used to characterize the degree to which the dedicated prediction models corresponding to the multiple reference business domains are suitable for processing the object. It can be understood that the more suitable the dedicated prediction model corresponding to the reference business domain is for processing the object, the higher the degree of matching between the target business scenario corresponding to the object and the reference business domain.

[0090] Among them, the domain distribution model, based on the domain classification information corresponding to the object determined by the domain classification model and combined with the historical sequence data of the object, can more accurately determine the dedicated prediction model suitable for processing the object, thereby providing a more reliable feature analysis basis for the subsequent determination of the prediction sequence data based on the first prediction sequence features determined by the dedicated prediction models corresponding to each reference business field.

[0091] In the present application, there may be many specific implementations of determining the prediction sequence data of the object based on the first prediction sequence feature determined by the domain matching information and the dedicated prediction model of each reference business domain, and there is no limitation to this.

[0092] In order to improve the accuracy of the time series data of the object, the present application can also pre-train a comprehensive prediction model, which is a comprehensive prediction model applicable to all business fields. Based on this, the present application can also determine the prediction sequence data based on the field matching information and the first prediction sequence features determined by the dedicated prediction models of multiple reference business fields, combined with the comprehensive prediction model.

[0093] The comprehensive prediction model may be obtained in a variety of ways. In a possible implementation, the comprehensive prediction model may be obtained by training based on at least one historical sequence data sample of object samples in different sample business scenarios and a scene description information sample corresponding to the sample business scenario.

[0094] Several possible implementation methods for determining the prediction sequence data of an object in combination with a comprehensive prediction model are introduced below.

[0095] In a first possible implementation, the present application may first use a comprehensive prediction model to process the fusion features of the first prediction sequence features corresponding to each reference business field to obtain the second prediction sequence features of the object. Then, based on the field matching information and the second prediction sequence features, the prediction sequence data of the object is determined.

[0096] The fusion feature of the first prediction sequence features corresponding to each reference business field may be the sum of the first prediction sequence features determined by the dedicated prediction models of each reference business field, such as adding the first prediction sequence features determined by the dedicated prediction models corresponding to each reference business field to obtain the fusion feature, and inputting the fusion feature into the comprehensive prediction model. The fusion feature may also be a feature obtained by merging or integrating the first prediction sequence features corresponding to each reference business field by the comprehensive prediction model.

[0097] The following takes a specific implementation as an example to introduce the first possible implementation method.

[0098] like Figure 2 , shows another flow chart of the data prediction method provided by the embodiment of the present application. The method of this embodiment may include:

[0099] S201, obtaining historical sequence data of objects in a target business scenario and scenario description information of the target business scenario.

[0100] The historical sequence data includes data of the object at at least one historical time point.

[0101] S202, based on the scene description information, determine the domain classification information of the object using a domain classification model.

[0102] The domain classification information is used to characterize the probability that the target business scenario belongs to multiple reference business domains.

[0103] For example, the scenario description information is input into a domain classification model to obtain the domain classification information predicted by the domain classification model. The domain classification information may include probabilities corresponding to multiple reference business domains, and the probability corresponding to each reference business domain represents the probability that the target business scenario belongs to the reference business domain. The sum of the probabilities of multiple reference business domains may be 1.

[0104] S203, based on the domain classification information and the historical sequence data, using the domain distribution model to determine the domain matching information of the object.

[0105] The domain matching information is used to characterize the suitability of the dedicated prediction models corresponding to multiple reference business domains for the processing objects.

[0106] For example, the domain allocation information and historical sequence data are input into the domain distribution model to obtain the domain matching information output by the domain distribution model. The domain matching information may include the fitness of the dedicated prediction model corresponding to each reference business domain. The fitness corresponding to each dedicated prediction model is the fitness of the dedicated prediction model for processing the object. For example, the fitness may be weighted.

[0107] In the present application, the specific model structure of the domain distribution model and its specific processing of the domain classification model and historical sequence data may not be restricted.

[0108] In one possible implementation, the domain distribution model is represented by It can be expressed as the following formula:

[0109] (Formula 1);

[0110] in, is the historical time series data of the object, The scenario description information of the target business scenario corresponding to the object, Represents model parameters.

[0111] Represents the domain classification information output by the domain classification model, ,in, Represents the unnormalized probability corresponding to each reference business domain determined by the domain classification model.

[0112] is a normalization function, such as, The probabilities corresponding to each reference business domain obtained by the domain classification model are normalized.

[0113] in, is a data function, which is used to determine the previous elements and their corresponding index values. In this application, the index value can be regarded as a weight value or a probability value. is less than , and is a natural number not less than 1.

[0114] Distribute the intermediate processing results of the model to the domain.

[0115] It should be noted that this embodiment is described by taking an implementation method of determining field matching information as an example. The other implementation methods of determining field matching information mentioned above are also applicable to this embodiment and will not be described in detail.

[0116] S204: For each reference business field, based on historical sequence data, a dedicated prediction model corresponding to the reference business field is used to determine a first prediction sequence feature of the object.

[0117] It should be noted that the order of steps S202, S203 and S204 is not limited to Figure 1 As shown, in actual application, step S204 may be performed first and then steps S202 and S203, or steps S204 and S202 may be performed simultaneously, and there is no limitation to this.

[0118] S205, using the comprehensive prediction model to process the fusion features of the first prediction sequence features corresponding to each reference business field to obtain the second prediction sequence features of the object.

[0119] The second prediction sequence feature is determined by the comprehensive prediction model and is used to characterize the feature of the prediction sequence data of the object. For example, the second prediction sequence feature is a feature representation of the prediction sequence data of the object predicted by the comprehensive prediction model, that is, a feature vector corresponding to the prediction sequence data of the object.

[0120] For example, the first prediction sequence features corresponding to the reference business field may be summed up and then input into the comprehensive prediction model to obtain the second prediction sequence features of the object output by the comprehensive prediction model.

[0121] S206, based on the matching degree between each reference business field represented by the field matching information and the target business scenario, perform weighted summation on the first prediction sequence features determined by the dedicated prediction model corresponding to each reference business field to obtain a comprehensive feature after the weighted summation of the features.

[0122] For example, the domain matching information may include the suitability (such as weight) of the dedicated prediction models of each reference business domain for processing the object, and the suitability may be used to perform a weighted summation on the first prediction sequence features determined by each dedicated prediction model.

[0123] For the sake of distinction, the feature obtained by weighted summing of the features of each first sequence prediction feature is called a comprehensive feature.

[0124] For example, the comprehensive feature can be expressed as the following formula 2:

[0125] Comprehensive features = (Formula 2);

[0126] in, Indicates The first prediction sequence characteristics predicted by a dedicated prediction model for a reference business field; Indicates The degree of match between the reference business areas.

[0127] S207, determining the predicted sequence data of the object based on the comprehensive feature and the second predicted sequence feature.

[0128] It can be understood that the first sequence prediction feature determined by the dedicated prediction model is actually the feature of the predicted sequence data of the object predicted by the dedicated prediction model, and the comprehensive feature is the weighted sum of the first sequence prediction features determined by each dedicated prediction model in combination with the domain matching information. Therefore, the comprehensive feature can more accurately characterize the features of the predicted sequence data of the object. On this basis, the predicted sequence data of the object can be determined more accurately by combining the comprehensive feature and the second predicted sequence feature determined by the comprehensive prediction model.

[0129] In a possible implementation, the output layer module may be used to determine the predicted sequence data of the object based on the comprehensive feature and the second predicted sequence feature. For example, the output layer model may be used to convert the comprehensive feature and the second predicted sequence feature into the predicted sequence feature of the object.

[0130] The output layer module may be pre-trained. In an optional manner, in order to improve the accuracy of determining the prediction sequence data, the output layer module and the comprehensive prediction model may be trained synchronously.

[0131] For ease of understanding Figure 2 The specific implementation can be combined with Figure 3 Provide a brief explanation. Figure 3 A schematic diagram of a framework of an implementation process of the data prediction method in this application.

[0132] Depend on Figure 3It can be seen that the present application inputs the scenario description information of the target business scenario corresponding to the object into the domain classification model. The domain classification information determined by the domain classification model will be input into the domain distribution model, and the corresponding historical sequence data will also be input into the domain distribution model to obtain domain matching information, such as the weight corresponding to the dedicated prediction model of each reference business field or the weight of the first prediction sequence feature corresponding to each reference business field.

[0133] The historical sequence data of the object will also be entered into Among them, is the total number of reference business areas. The first prediction sequence features output by each dedicated prediction model are summed up and input into a comprehensive prediction model applicable to any business field.

[0134] On the basis of the above, the degree of suitability (such as weight) of each dedicated prediction model represented by the domain matching information for processing the object is The first prediction sequence features output by the dedicated prediction models are weighted and summed, and then the comprehensive features obtained by the weighted sum of the features are input into the output layer module together with the second prediction sequence features output by the comprehensive prediction model. The output layer module combines the comprehensive features and the second prediction sequence features to determine the prediction sequence data of the object.

[0135] The second possible implementation method of determining the prediction sequence data of the object in combination with the comprehensive prediction model is introduced below.

[0136] like Figure 4 , shows another flow chart of the data prediction method provided by the present application. The method of this embodiment may include:

[0137] S401, obtaining historical sequence data of objects in a target business scenario and scenario description information of the target business scenario.

[0138] The historical sequence data includes data of the object at at least one historical time point.

[0139] S402, based on the scene description information, determine the domain classification information of the object using a domain classification model.

[0140] The domain classification information is used to characterize the probability that the target business scenario belongs to multiple reference business domains.

[0141] S403, based on the domain classification information and the historical sequence data, the domain matching information of the object is determined using the domain distribution model.

[0142] It should be noted that this embodiment is described by taking an implementation method of determining field matching information as an example. The other implementation methods of determining field matching information mentioned above are also applicable to this embodiment and will not be described in detail.

[0143] S404: For each reference business field, based on historical sequence data, a dedicated prediction model corresponding to the reference business field is used to determine a first prediction sequence feature of the object.

[0144] For the above steps 401 to S404, reference may be made to the relevant introduction of the previous embodiment and will not be repeated here.

[0145] S405 , based on the matching degree between each reference business field represented by the field matching information and the target business scenario, weighted summation is performed on the first prediction sequence features determined by the dedicated prediction model corresponding to each reference business field to obtain a comprehensive feature after the weighted summation of the features.

[0146] Among them, the specific implementation of the comprehensive feature is the same as the previous Figure 2 The embodiments are similar and will not be described in detail.

[0147] S406, based on the comprehensive features, using the comprehensive prediction model to determine the second prediction sequence features of the object.

[0148] and Figure 2 In the embodiment, the second prediction sequence characteristics are determined directly based on the first prediction sequence characteristics corresponding to each reference business field. In this embodiment, it is necessary to first determine the weighted sum of the characteristics of the first prediction sequence characteristics corresponding to each reference business field, and input the comprehensive characteristics obtained by the weighted sum of the characteristics into the comprehensive prediction model.

[0149] As can be seen from the previous introduction, the comprehensive features can more accurately represent the characteristics of the predicted sequence data of the object. Therefore, the comprehensive prediction model is used to further analyze and process the comprehensive features, so that the second prediction sequence features determined by the comprehensive prediction model can more accurately characterize the characteristics of the predicted sequence data of the object.

[0150] S407, determining predicted sequence data of the object based at least on the second predicted sequence feature.

[0151] In the embodiment of the present application, the predicted sequence data of the object can be determined based only on the second predicted sequence feature, that is, the second predicted sequence feature is converted into predicted sequence data. For example, the second predicted sequence feature is input into the output layer model to obtain the predicted sequence data of the object.

[0152] In one possible implementation, in order to further improve the accuracy of the predicted sequence data of the object, the present application can also determine the predicted sequence data of the object based on the comprehensive feature and the second predicted sequence feature using an output layer module, such as adding the comprehensive feature and the second predicted sequence and inputting them into the output layer model, or inputting the comprehensive feature and the second predicted sequence feature into the output layer module together.

[0153] In another possible implementation, in order to further improve the accuracy of the predicted sequence data of the object, the predicted sequence data of the object may be determined using an output layer module based on the first predicted sequence features and the second predicted sequence features corresponding to each reference business field. For example, the first predicted sequence features and the second predicted sequence features corresponding to each reference business field are added and input into the output layer module to obtain the predicted sequence data.

[0154] In the case of determining the prediction sequence data of an object in combination with a comprehensive prediction model, in order to improve the accuracy of the predicted prediction sequence data of the object, the present application may also train the comprehensive prediction model after training dedicated models for each reference business field.

[0155] Specifically, in the first training stage, for each reference business field, based on at least one historical sequence data sample of an object sample in the reference business field, a dedicated prediction model corresponding to the reference business field is trained.

[0156] In the second training stage, based on at least one historical sequence data sample of object samples in different sample business scenarios and the scenario description information sample corresponding to the sample business scenarios, combined with the dedicated prediction models of each reference business field obtained through training, a comprehensive prediction model is trained.

[0157] Among them, in the second training stage, the model parameters of the dedicated prediction model of each reference business field are fixed and will not be adjusted during the training.

[0158] In the second training stage, each historical sequence data sample of each object sample can be processed according to the processing process of the historical sequence data of the processing object, and the predicted sequence data samples of each object sample predicted by the prediction model are consistent with the actual sequence data samples as the training goal, so as to train the comprehensive prediction model.

[0159] The following is an introduction to a method for implementing training a comprehensive prediction model.

[0160] like Figure 5 , shows a schematic diagram of an implementation flow of training a comprehensive prediction model provided in an embodiment of the present application. This embodiment may include:

[0161] S501, obtaining at least one historical sequence data sample of object samples in multiple different sample business scenarios and a scenario description information sample corresponding to each sample business scenario.

[0162] The historical sequence data sample of the object sample includes: historical data of the object sample at at least one sample time point. Since the historical sequence data sample of the object sample is known sequence data, each sample time point involved in the historical sequence data sample is also a historical time point.

[0163] For ease of distinction, this application refers to the objects involved in model training as object samples, the business scenarios corresponding to the object samples as sample business scenarios, and the historical sequence data of the objects as historical sequence data samples.

[0164] Among them, the scenario description information sample corresponding to the sample business scenario is used to describe the prediction characteristics of the time series data in the sample business scenario, such as the name of the object sample, etc. The specific details are similar to the previous scenario description information and will not be repeated here.

[0165] S502: For each historical sequence data sample in each sample business scenario, determine a domain matching information sample of the object sample based on the scenario description information sample of the historical sequence data sample.

[0166] Among them, the domain matching information sample is used to characterize the matching degree between the sample business scenario and multiple reference business domains.

[0167] It can be understood that, just for the sake of distinction, the domain matching information corresponding to the object sample is called the domain matching information sample. Based on this, the process of determining the domain matching information sample is similar to the previous process of determining the domain matching information, and will not be repeated here.

[0168] S503, for each historical sequence data sample of each object sample, based on the first sequence data sample in the historical sequence data sample, using the dedicated prediction model corresponding to each reference business field, respectively determine the first prediction sequence feature sample of the object sample.

[0169] The first sequence data sample includes: the historical data at at least one consecutive sample time point in the historical sequence data sample. For example, if the historical sequence data sample includes historical data at 20 sample time points, the historical data at the first 10 sample time points can be used to form the first sequence data sample.

[0170] The second sequence data samples in the historical sequence data samples except the first sequence data samples can be used as the real data sequence corresponding to the time series data that needs to be predicted by the object sample.

[0171] S504, based on the domain matching information sample and the first prediction sequence feature sample determined by the dedicated prediction model corresponding to each reference business domain, the prediction sequence data sample of the object sample is determined in combination with the comprehensive prediction model.

[0172] The predicted sequence data sample includes data at at least one sample prediction time point. The at least one sample prediction time point includes each sample time point in the historical sequence data sample that is located after at least one sample time point corresponding to the first sequence data sample. That is, for each historical sequence data sample, at least one sample prediction time point corresponding to the predicted sequence data sample is a sample time point corresponding to the second sequence data sample in the historical sequence data sample.

[0173] Among them, the process of determining the prediction sequence data samples of the object samples in combination with the comprehensive prediction model can be similar to the specific implementation process of determining the prediction sequence data of the object in combination with the comprehensive prediction model.

[0174] For example, in one possible implementation, a comprehensive prediction model can be used to process the fusion features of the first prediction sequence feature samples corresponding to each reference business field to obtain a second prediction sequence feature sample of the object sample. Based on the field matching information sample and the second prediction sequence feature sample, the prediction sequence data sample of the object is determined. For example, based on the degree of matching between each reference business field represented by the field matching information and the sample business scenario corresponding to the object sample, the first prediction sequence feature samples determined by the dedicated prediction model corresponding to each reference business field are weighted and summed to obtain a comprehensive feature sample after the weighted sum of the features; based on the comprehensive feature sample and the second prediction sequence feature sample, the prediction sequence data sample of the object sample is determined.

[0175] In another possible implementation, based on the degree of matching between each reference business domain represented by the domain matching information sample and the sample business scenario, the first prediction sequence feature samples determined by the dedicated prediction model corresponding to each reference business domain may be weighted and summed to obtain a comprehensive feature sample after the weighted sum of the features. Then, based on the comprehensive feature sample, the second prediction sequence feature sample of the object sample is determined using the comprehensive prediction model, and accordingly, based at least on the second prediction sequence feature sample, the prediction sequence data sample of the object sample is determined.

[0176] It can be understood that, through the above steps S502 to S504, for each historical sequence data sample of each object sample in each sample business scenario, a predicted sequence data sample is determined.

[0177] S505, based on the second sequence data sample in the historical sequence data sample corresponding to the object sample and the predicted sequence data sample, adjust the model parameters of the comprehensive prediction model, and return to execute step S502 until the training requirements are met.

[0178] As described above, for any historical sequence data sample, the second sequence data sample is a sequence data sample located after the first sequence data sample in the historical sequence data sample.

[0179] The training requirement may be that the number of training iterations reaches a set number, or that the prediction accuracy meets the requirements, or that the loss function value converges.

[0180] Among them, the prediction accuracy and the loss function value can be based on the second sequence data sample and the prediction sequence data sample in the historical sequence data sample corresponding to the object sample. For example, if the similarity between the second sequence data sample in each historical sequence data sample and the corresponding prediction sequence data sample is higher, the prediction accuracy is higher and the loss function value is smaller. There is no restriction on the specific implementation of calculating the prediction accuracy and loss function value.

[0181] It is understandable that, in combination with the foregoing introduction to the domain matching information, there are also multiple possible implementation methods for determining the domain matching information sample.

[0182] In particular, if the previous domain matching information needs to be obtained with the help of a domain classification model and a domain distribution model, then the first training stage may also include:

[0183] A domain classification model is trained based on at least one scene description information sample and a reference business domain annotated by the scene description information sample.

[0184] On this basis, in the second training stage, the domain matching information sample of the object sample can be determined by: based on the scene description information sample of the historical sequence data sample, using the domain classification model to determine the domain classification information sample of the object sample. Correspondingly, based on the domain classification information sample and the historical sequence data sample, the domain matching information sample of the object sample can be determined using the domain distribution model to be trained.

[0185] The domain classification information sample is used to characterize the probability that the sample business scenario belongs to multiple reference business domains. The domain matching information sample is used to characterize the suitability of the dedicated prediction models corresponding to the multiple reference business domains for processing the object sample.

[0186] On this basis, in the second training stage, adjusting the model parameters of the comprehensive prediction model also includes: adjusting the model parameters of the field distribution model.

[0187] like Figure 6 An example diagram showing the two stages involved in model training in this application.

[0188] Depend on Figure 6 It can be clearly seen that in the first training stage ( Figure 6 In the first stage of the process, for each reference business domain, at least one historical time series (i.e., historical sequence data) of the object samples in the reference business domain is used as a training sample to train a dedicated prediction model for the reference business domain, and obtain At the same time, it will also combine The domain classification model is trained based on the scenario description information corresponding to the reference business domain.

[0189] After the first training phase, the model parameters of the dedicated prediction model of each reference business field and the classification model of the field are kept unchanged, that is, the dedicated prediction model and the classification model of the field are frozen, and the second training phase is entered ( Figure 6 in the second phase).

[0190] In the second training phase, the historical sequence data samples or scene description information samples can be processed with the help of the dedicated prediction model and the domain classification model, and the training of the comprehensive prediction model and the domain distribution model can be completed.

[0191] For ease of understanding, the following is a comparison of the process of processing the historical sequence data samples in the process of training the comprehensive prediction model. Figure 2 and Figure 3 The process of processing historical sequence data is similar to that in the example. Figure 7 , shows a flow chart of a model training method provided in an embodiment of the present application, and this embodiment may include:

[0192] S701, in the first training stage, a domain classification model is trained based on at least one scene description information sample and a reference business domain annotated with the scene description information sample, and for each reference business domain, a dedicated prediction model corresponding to the reference business domain is trained based on at least one historical sequence data sample of an object sample in the reference business domain.

[0193] S702, in the second training stage, obtaining at least one historical sequence data sample of object samples in a plurality of different sample business scenarios and a scenario description information sample corresponding to each sample business scenario.

[0194] S703 , for each historical sequence data sample of the object sample in each sample business scenario, based on the scenario description information sample of the sample business scenario, determine a domain classification information sample using a domain classification model.

[0195] The field classification information sample is used to characterize the probability that the sample business scenarios belong to the multiple reference business fields respectively.

[0196] S704: Based on the domain classification information sample and the historical sequence data sample of the object sample, determine the domain matching information sample of the object sample using a domain distribution model.

[0197] The domain matching information sample is used to characterize the suitability of the dedicated prediction models corresponding to multiple reference business domains for processing the object sample.

[0198] S705: Based on the first sequence data sample in the historical sequence data sample, use the dedicated prediction model corresponding to each reference business field to respectively determine the first prediction sequence feature sample of the object sample.

[0199] S706, using the comprehensive prediction model to process the fusion features of the first prediction sequence feature samples corresponding to each reference business field to obtain a second prediction sequence feature sample of the object sample.

[0200] S707, based on the matching degree between each reference business field represented by the field matching information and the sample business scenario corresponding to the object sample, perform feature weighted summation on the first prediction sequence feature samples determined by the dedicated prediction model corresponding to each reference business field to obtain a comprehensive feature sample after the feature weighted summation.

[0201] S708, based on the comprehensive feature sample and the second prediction sequence feature sample, use the output layer module to determine the prediction sequence data sample of the object sample.

[0202] S709, based on the second sequence data samples in the historical sequence data samples corresponding to each object sample and the predicted sequence data samples, adjust the model parameters of the comprehensive prediction model, the domain distribution model and the output layer module, and return to execute step S703 until the training requirements are met.

[0203] exist Figure 7 In the second training phase, for each historical sequence data sample corresponding to an object sample, the process of determining the predicted sequence data sample of the object sample can be adopted Figure 3 The framework process shown in the figure is as follows: Figure 7 The process will not be described in detail.

[0204] For ease of understanding, the models that need to adjust model parameters and freeze parameters (i.e., do not need to adjust model parameters) in the second training phase are Figure 3 In the figure, the upper right corner of the box containing each model with frozen parameters is marked with a freeze symbol (such as Figure 3 The word "Frozen" in the middle circle), such as Figure 3 It can be clearly seen A freeze symbol is marked on the boxes corresponding to the dedicated prediction model of the reference business field and the classification model of the field. It can be clearly seen that in the second training stage, there is no need to adjust the model parameters of the dedicated prediction model and the field classification model.

[0205] On the other hand, corresponding to a data prediction method of the present application, the present application also provides a data prediction device.

[0206] like Figure 8 A schematic diagram of the structure of a data prediction device provided by the present application is shown. The device of this embodiment may include:

[0207] A data acquisition unit 801 is used to obtain historical sequence data of an object in a target business scenario and scenario description information of the target business scenario, wherein the historical sequence data includes data of the object at at least one historical time point;

[0208] A domain matching unit 802 is used to determine domain matching information of the object based on the scenario description information, where the domain matching information is used to represent the degree of matching between a target business scenario corresponding to the object and a plurality of reference business domains;

[0209] A feature prediction unit 803 is used to determine a first prediction sequence feature of the object based on the historical sequence data and using a dedicated prediction model corresponding to the reference business field;

[0210] The data prediction unit 804 is used to determine the prediction sequence data of the object based on the field matching information and the first prediction sequence feature determined by the dedicated prediction model corresponding to each reference business field, and the prediction sequence data includes the prediction data of the object at at least one prediction time point.

[0211] In a possible implementation, the data prediction unit may include:

[0212] A first comprehensive prediction unit, used for processing the fusion features of the first prediction sequence features corresponding to each reference business field by using the comprehensive prediction model to obtain the second prediction sequence features of the object;

[0213] The first sequence prediction unit is used to determine the predicted sequence data of the object based on the domain matching information and the second predicted sequence feature.

[0214] Further, the first sequence prediction unit may include:

[0215] A feature summing subunit is used to perform weighted summation of the first prediction sequence features determined by the dedicated prediction model corresponding to each reference business field based on the matching degree between each reference business field represented by the field matching information and the target business scenario, so as to obtain a comprehensive feature after the weighted summation of the features;

[0216] The first sequence prediction subunit is used to determine the predicted sequence data of the object based on the comprehensive feature and the second predicted sequence feature.

[0217] In yet another possible implementation, the data prediction unit may include:

[0218] A feature synthesis unit, configured to perform weighted summation of first prediction sequence features determined by a dedicated prediction model corresponding to each reference business field based on the matching degree between each reference business field represented by the field matching information and the target business scenario, to obtain a comprehensive feature after the weighted summation of features;

[0219] A second comprehensive prediction unit, configured to determine a second prediction sequence feature of the object using a comprehensive prediction model based on the comprehensive feature;

[0220] The second sequence prediction unit is used to determine the predicted sequence data of the object based at least on the second predicted sequence feature.

[0221] Furthermore, the second sequence prediction unit can be specifically used to determine the prediction sequence data of the object using an output layer module based on the comprehensive features and the second prediction sequence features; or, based on the first prediction sequence features and the second prediction sequence features corresponding to each reference business field, determine the prediction sequence data of the object using an output layer module.

[0222] In yet another possible implementation, the domain matching unit may include:

[0223] A domain classification unit, configured to determine domain classification information of the object using a domain classification model based on the scenario description information, wherein the domain classification information is used to characterize the probability that the target business scenario belongs to each of the multiple reference business domains;

[0224] A model matching unit is used to determine the domain matching information of the object based on the domain classification information and the historical sequence data using a domain distribution model, wherein the domain matching information is used to characterize the suitability of the dedicated prediction models corresponding to the multiple reference business domains for processing the object.

[0225] In another possible implementation, the model training unit is used to obtain the dedicated prediction model corresponding to the reference business field and the comprehensive prediction model in the following manner:

[0226] In a first training phase, based on at least one historical sequence data sample of an object sample in the reference business field, a dedicated prediction model corresponding to the reference business field is trained;

[0227] In the second training stage, based on at least one historical sequence data sample of object samples in different sample business scenarios and the scenario description information sample corresponding to the sample business scenarios, combined with the dedicated prediction models of each reference business field obtained through training, a comprehensive prediction model is trained.

[0228] In another possible implementation, the model training unit includes:

[0229] a domain information determination unit, configured to determine, in a second training phase, for each historical sequence data sample in each sample business scenario, a domain matching information sample of the object sample based on the scenario description information sample of the historical sequence data sample, wherein the domain matching information sample is used to characterize the degree of matching between the sample business scenario and the plurality of reference business domains;

[0230] A feature sample determination unit, configured to determine a first prediction sequence feature sample of the object sample based on a first sequence data sample in the historical sequence data sample and using a dedicated prediction model corresponding to the reference business field;

[0231] A sequence sample determination unit, configured to determine a prediction sequence data sample of the object sample based on the field matching information sample and a first prediction sequence feature sample determined by a dedicated prediction model corresponding to each reference business field, in combination with a comprehensive prediction model;

[0232] A training control unit is used to adjust the model parameters of the comprehensive prediction model based on a second sequence data sample in the historical sequence data sample corresponding to the object sample and the predicted sequence data sample, and return to execute the operation of the feature sample determination unit until the training requirements are met, wherein the second sequence data sample is a sequence data sample located after the first sequence data sample in the historical sequence data sample.

[0233] In another possible implementation, the model training unit is further used in the first training stage to train a domain classification model based on at least one scene description information sample and a reference business domain annotated by the scene description information sample;

[0234] The field information determination unit includes:

[0235] A classification determination subunit, configured to determine the domain classification information sample of the object sample using the domain classification model based on the scene description information sample of the historical sequence data sample;

[0236] a matching information determination subunit, configured to determine a domain matching information sample of the object sample using a domain distribution model to be trained based on the domain classification information sample and the historical sequence data sample;

[0237] While the model training unit adjusts the model parameters of the comprehensive prediction model, it is also used to: adjust the model parameters of the domain distribution model.

[0238] The present application also provides an electronic device. Fig. 9 , which shows a schematic diagram of a composition structure of the electronic device, and the electronic device at least includes a processor 901 and a memory 902;

[0239] The processor 901 is configured to execute the data prediction method described in any one of the above embodiments;

[0240] The memory 902 is used to store programs required by the processor to perform operations.

[0241] It can be understood that the electronic device may further include a display unit 903 and an input unit 904 .

[0242] Of course, the electronic device may also have Fig. 9 There may be more or fewer components, without limitation.

[0243] A computer program product is also provided in an embodiment of the present application, including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements any data prediction method provided in the embodiment of the present application.

[0244] A computer-readable storage medium is also provided in an embodiment of the present application. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any data prediction method provided in the embodiment of the present application.

[0245] It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

[0246] Through the description of the above implementation mode, the technicians in the field can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present application, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0247] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0248] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

Claims

1. A data prediction method, comprising: Obtaining historical sequence data of an object in a target business scenario and scenario description information of the target business scenario, wherein the historical sequence data includes data of the object at at least one historical time point; Based on the scenario description information, determining domain matching information of the object, wherein the domain matching information is used to characterize a matching degree between a target business scenario corresponding to the object and a plurality of reference business domains; Based on the historical sequence data, determining a first prediction sequence feature of the object using a dedicated prediction model corresponding to the reference business field; Based on the domain matching information and the first prediction sequence feature determined by the dedicated prediction model corresponding to each reference business domain, the prediction sequence data of the object is determined, and the prediction sequence data includes prediction data of the object at at least one prediction time point.

2. The data prediction method according to claim 1, wherein the first prediction sequence feature determined based on the domain matching information and the dedicated prediction model corresponding to each reference business domain, determines the prediction sequence data of the object, comprising: Processing the fusion features of the first prediction sequence features corresponding to each reference business field using the comprehensive prediction model to obtain the second prediction sequence features of the object; Based on the domain matching information and the second predicted sequence feature, predicted sequence data of the object is determined.

3. The data prediction method according to claim 2, wherein determining the predicted sequence data of the object based on the domain matching information and the second predicted sequence feature comprises: Based on the matching degree between each reference business field represented by the field matching information and the target business scenario, weighted summing is performed on the first prediction sequence features determined by the dedicated prediction model corresponding to each reference business field to obtain a comprehensive feature after weighted summing of the features; Based on the comprehensive feature and the second predicted sequence feature, predicted sequence data of the object is determined.

4. The data prediction method according to claim 1, wherein the first prediction sequence feature determined based on the domain matching information and the dedicated prediction model corresponding to each reference business domain, determines the prediction sequence data of the object, comprising: Based on the matching degree between each reference business field represented by the field matching information and the target business scenario, weighted summing is performed on the first prediction sequence features determined by the dedicated prediction model corresponding to each reference business field to obtain a comprehensive feature after weighted summing of the features; Based on the comprehensive features, determining a second predicted sequence feature of the object using a comprehensive prediction model; Based at least on the second predicted sequence feature, predicted sequence data for the subject is determined.

5. The data prediction method according to claim 4, wherein determining the predicted sequence data of the object based at least on the second predicted sequence feature comprises: Determining predicted sequence data of the object using an output layer module based on the comprehensive feature and the second predicted sequence feature; or, Based on the first prediction sequence features and the second prediction sequence features corresponding to each reference business field, the prediction sequence data of the object is determined using an output layer module.

6. The data prediction method according to claim 1, wherein determining the domain matching information of the object based on the scene description information comprises: Based on the scenario description information, determining the domain classification information of the object using a domain classification model, wherein the domain classification information is used to characterize the probability that the target business scenario belongs to each of the multiple reference business domains; Based on the domain classification information and the historical sequence data, domain matching information of the object is determined using a domain distribution model, wherein the domain matching information is used to characterize the degree to which the dedicated prediction models corresponding to the multiple reference business domains are suitable for processing the object.

7. According to the data prediction method of claim 2 or 5, the dedicated prediction model corresponding to the reference business field and the comprehensive prediction model are obtained by: In a first training phase, based on at least one historical sequence data sample of an object sample in the reference business field, a dedicated prediction model corresponding to the reference business field is trained; In the second training stage, based on at least one historical sequence data sample of object samples in different sample business scenarios and the scenario description information sample corresponding to the sample business scenarios, combined with the dedicated prediction models of each reference business field obtained through training, a comprehensive prediction model is trained.

8. The data prediction method according to claim 7, wherein the at least one historical sequence data sample based on the object samples in different sample business scenarios and the scenario description information sample corresponding to the sample business scenario are combined with the dedicated prediction models of each reference business field obtained by training to obtain a comprehensive prediction model, comprising: For each historical sequence data sample in each sample business scenario, based on the scenario description information sample of the historical sequence data sample, determine the domain matching information sample of the object sample, wherein the domain matching information sample is used to characterize the matching degree between the sample business scenario and the multiple reference business domains; Based on the first sequence data sample in the historical sequence data sample, using the dedicated prediction model corresponding to the reference business field to determine the first prediction sequence feature sample of the object sample; Determine the prediction sequence data sample of the object sample based on the field matching information sample and the first prediction sequence feature sample determined by the dedicated prediction model corresponding to each reference business field, combined with the comprehensive prediction model; Based on the second sequence data sample in the historical sequence data sample corresponding to the object sample and the predicted sequence data sample, the model parameters of the comprehensive prediction model are adjusted, and the step of determining the first predicted sequence feature sample of the object sample is returned to be executed until the training requirements are met, wherein the second sequence data sample is a sequence data sample located after the first sequence data sample in the historical sequence data sample.

9. The data prediction method according to claim 8, further comprising in the first training stage: Training a domain classification model based on at least one scene description information sample and a reference business domain annotated by the scene description information sample; The determining of the domain matching information sample of the object sample based on the scene description information sample of the historical sequence data sample includes: Based on the scene description information sample of the historical sequence data sample, determining the domain classification information sample of the object sample using the domain classification model; Based on the domain classification information sample and the historical sequence data sample, determining the domain matching information sample of the object sample using the domain distribution model to be trained; While adjusting the model parameters of the comprehensive prediction model, it also includes: adjusting the model parameters of the field distribution model.

10. A data prediction device, comprising: A data acquisition unit, configured to acquire historical sequence data of an object in a target business scenario and scenario description information of the target business scenario, wherein the historical sequence data includes data of the object at at least one historical time point; A domain matching unit, configured to determine domain matching information of the object based on the scenario description information, wherein the domain matching information is used to characterize a matching degree between a target business scenario corresponding to the object and a plurality of reference business domains; A feature prediction unit, configured to determine a first prediction sequence feature of the object based on the historical sequence data and using a dedicated prediction model corresponding to the reference business field; A data prediction unit is used to determine the prediction sequence data of the object based on the field matching information and the first prediction sequence characteristics determined by the dedicated prediction model corresponding to each reference business field, and the prediction sequence data includes the prediction data of the object at at least one prediction time point.

Images