Performance curve prediction method and device

Abstract

The invention relates to a performance curve prediction method and device, relates to the technical field of artificial intelligence, and aims to improve the accuracy of a prediction curve. The method comprises the steps of obtaining first data used for representing a change trend of a historical performance curve of a target object, wherein the change trend of the historical performance curve continuously changes along with time; obtaining second data used for representing attributes of the target object; wherein the attributes are overlapped with the time range corresponding to the historical performance curve, and the attributes influence the change trend of the historical performance curve; encoding the first data according to the second data through the prediction model to obtain a fusion feature, the fusion feature being a feature representation of the historical performance curve in a hidden space after the attributes corresponding to the plurality of historical time periods are fused; and decoding the fusion features through the prediction model to obtain a prediction curve of the target object, the prediction curve being used for predicting the change trend of the performance of the target object.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and apparatus for predicting performance curves. Background Technology

[0002] With the development of artificial intelligence technology, data prediction using artificial intelligence models has been widely applied in many fields.

[0003] In related technologies, artificial intelligence models are used to predict the performance curves of objects. For example, a curve-based prediction method is used to predict performance curves. For instance, historical performance curves are collected and input into the artificial intelligence model, allowing the model to output a predicted performance curve based on the historical performance curves, or to output key points within the predicted performance curve.

[0004] However, due to the uncertainty of fluctuations in historical performance curves, the accuracy of prediction curves obtained by artificial intelligence models based on historical performance curves is not high. Summary of the Invention

[0005] This disclosure provides a method and apparatus for predicting performance curves, aiming to improve the accuracy of the predicted performance curves. The technical solution of this disclosure is as follows: According to one aspect of the embodiments of this application, a method for predicting performance curves is provided, the method comprising the following steps: Obtain first data to characterize the changing trend of the historical performance curve of the target object; wherein, the changing trend of the historical performance curve changes continuously over time; Obtain second data for the attributes that characterize the target object; wherein the time range of the attribute overlaps with the historical performance curve, and the attribute affects the trend of the historical performance curve. By encoding the first data with the second data through the prediction model, the fusion feature is obtained. The fusion feature is the feature representation of the historical performance curve in the latent space after fusing the attributes corresponding to multiple historical time periods. By decoding and fusing features through a predictive model, a prediction curve for the target object is obtained. This prediction curve is used to predict the performance trend of the target object.

[0006] In one possible implementation, the prediction model includes an encoder that encodes first data based on second data to obtain fused features, including: The encoder performs feature fusion of historical performance curves with respect to attributes based on the first and second data to obtain fused features.

[0007] In one possible implementation, the encoder performs feature fusion of historical performance curves with respect to attributes based on first and second data to obtain fused features, including: The encoder performs attention cross-interaction on the historical performance curves of the first and second data with respect to the attributes to obtain the first intermediate feature; the first intermediate feature is the feature representation of the first data in the latent space after attention cross-interaction with the second data. The encoder performs self-attention processing based on the first intermediate feature to obtain the fused feature.

[0008] In one possible implementation, the fused features are obtained through N rounds of encoding, where N is a positive integer greater than 1; in the i-th round of encoding: By performing attention cross-cutting of the historical performance curve with respect to the attribute using the first feature and the second feature of the i-th round of encoding, the first intermediate feature of the i-th round of encoding is obtained; The second intermediate feature of the i-th round of encoding is obtained by performing attention cross-reference between the attribute and the historical performance curve through the first intermediate feature and the second feature of the i-th round of encoding. The first intermediate feature of the i-th round of encoding is processed by self-attention to obtain the first fusion feature of the i-th round of encoding. In this process, the first feature of the first round of encoding is the first data, the second feature of the first round of encoding is the second data, the first fusion feature of the i-th round of encoding is the first feature of the (i+1)-th round of encoding, the second intermediate feature of the i-th round of encoding is the second feature of the (i+1)-th round of encoding, and the first fusion feature of the N-th round of encoding is the fusion feature.

[0009] In one possible implementation, the encoder includes a curve attribute attention block, a self-attention block, and an attribute curve attention block; The curve attribute attention block is used to perform attention crossover of historical performance curves relative to attributes; The attribute curve attention block is used to perform attention crossover of attributes relative to historical performance curves; Self-attention blocks are used to perform self-attention processing on the first intermediate feature.

[0010] In one possible implementation, the first intermediate feature of the i-th encoding process is obtained by performing attention cross-cutting of the historical performance curve with respect to the attribute using the first feature and the second feature of the i-th encoding process, including: The first Q matrix is ​​calculated using the first feature of the i-th round of encoding, and the second K matrix and second V matrix are calculated using the second feature of the i-th round of encoding. Based on the first Q matrix, the second K matrix, and the second V matrix, the first intermediate feature of the i-th round of encoding is calculated. By performing attention cross-referencing of attributes with respect to historical performance curves using the first intermediate feature and the second feature of the i-th round of encoding, the second intermediate feature of the i-th round of encoding is obtained, including: The second Q matrix is ​​calculated using the second feature of the i-th round of encoding, and the first K matrix and the first V matrix are calculated using the first intermediate feature of the i-th round of encoding. The second intermediate feature of the i-th round of encoding is then calculated based on the second Q matrix, the first K matrix, and the first V matrix. By performing self-attention processing on the first intermediate feature of the i-th round of encoding, the first fusion feature of the i-th round of encoding is obtained, including: The third Q matrix, third K matrix, and third V matrix are calculated using the first intermediate features of the i-th round of encoding. Based on the third Q matrix, third K matrix, and third V matrix, the first fusion feature of the i-th round of encoding is then calculated.

[0011] In one possible implementation, the prediction curve of the target object is obtained by decoding the fused features through a prediction model, including: Decode and fuse features to obtain the prediction performance values ​​of the target object at multiple prediction time points; Based on the prediction performance values ​​at multiple prediction time points, determine the prediction curve of the target object within the prediction time period; Among them, the values ​​of multiple prediction time points in the prediction curve are equal to the prediction performance values ​​of multiple prediction time points; The predicted performance values ​​for multiple prediction time points are absolute predicted values, or the predicted performance values ​​for multiple prediction time points are relative to the historical performance values ​​of the corresponding historical time points in the historical performance curve.

[0012] In one possible implementation, the historical performance curve of the target object is sampled to obtain first data, including: By sampling any historical performance curve from multiple historical time periods, multiple sampling points corresponding to the historical performance curves of multiple historical time periods are obtained. The first data is obtained by taking multiple sampling points corresponding to the historical performance curves of multiple historical time periods. The sampling points corresponding to the historical performance curves of multiple historical time periods are arranged in chronological order in the first data.

[0013] In one possible implementation, obtaining second data of the target object includes: Retrieve the attributes of the target object across multiple historical time periods; Feature extraction is performed on the attributes corresponding to each historical time period to obtain the attribute features of each historical time period; where the attribute features of the historical time period are the representations of the attributes of the corresponding historical time period in the latent space. Based on the attribute characteristics of multiple historical time periods, second data is obtained. The attribute characteristics of different time periods in the second data correspond to different location codes.

[0014] In one possible implementation, the prediction model is trained based on the historical performance curve of the sample object and the attributes corresponding to the historical performance curve. During training, the prediction model generates a prediction curve based on the first sample data corresponding to the historical curve of the sample and the second sample data corresponding to the sample attribute. The weight parameters in the prediction model are adjusted based on the training loss; the training loss is determined based on the application effect of the prediction curve.

[0015] According to two aspects of the present disclosure, a performance curve prediction device is provided, which includes the following modules.

[0016] The first acquisition module is used to acquire first data to characterize the changing trend of the historical performance curve of the target object; wherein, the changing trend of the historical performance curve changes continuously over time. The second acquisition module is used to acquire the second data of the target object; wherein, the second data is used to characterize the attributes corresponding to the historical performance curve of the target object, and the attributes affect the changing trend of the historical performance curve; The encoding module is used to encode the first data based on the second data through the prediction model to obtain the fusion feature. The fusion feature is the feature representation of the historical performance curve in the latent space after fusing the attributes corresponding to multiple historical time periods. The decoding module is used to decode the fused features through the prediction model to obtain the prediction curve of the target object. The prediction curve is used to predict the performance trend of the target object.

[0017] In one possible implementation, the encoding module is used to obtain fused features by performing feature fusion of historical performance curves with respect to attributes based on first data and second data using an encoder.

[0018] In one possible implementation, the encoding module includes: an attention cross unit, used to perform attention cross on the historical performance curves of the first data and the second data relative to the attributes through the encoder to obtain a first intermediate feature; the first intermediate feature is the feature representation of the first data in the latent space after attention cross with the second data. The self-attention processing unit is used to perform self-attention processing on the first intermediate feature by the encoder to obtain the fused feature.

[0019] In one possible implementation, the fused features are obtained through N rounds of encoding, where N is a positive integer greater than 1; in the i-th round of encoding: By performing attention cross-cutting of the historical performance curve with respect to the attribute using the first feature and the second feature of the i-th round of encoding, the first intermediate feature of the i-th round of encoding is obtained; The second intermediate feature of the i-th round of encoding is obtained by performing attention cross-reference between the attribute and the historical performance curve through the first intermediate feature and the second feature of the i-th round of encoding. The first intermediate feature of the i-th round of encoding is processed by self-attention to obtain the first fusion feature of the i-th round of encoding. In this process, the first feature of the first round of encoding is the first data, the second feature of the first round of encoding is the second data, the first fusion feature of the i-th round of encoding is the first feature of the (i+1)-th round of encoding, the second intermediate feature of the i-th round of encoding is the second feature of the (i+1)-th round of encoding, and the first fusion feature of the N-th round of encoding is the fusion feature.

[0020] In one possible implementation, the encoder includes a curve attribute attention block, a self-attention block, and an attribute curve attention block; The curve attribute attention block is used to perform attention crossover of historical performance curves relative to attributes; The attribute curve attention block is used to perform attention crossover of attributes relative to historical performance curves; Self-attention blocks are used to perform self-attention processing on the first intermediate feature.

[0021] In one possible implementation, the first intermediate feature of the i-th encoding process is obtained by performing attention cross-cutting of the historical performance curve with respect to the attribute using the first feature and the second feature of the i-th encoding process, including: The first Q matrix is ​​calculated using the first feature of the i-th round of encoding, and the second K matrix and second V matrix are calculated using the second feature of the i-th round of encoding. Based on the first Q matrix, the second K matrix, and the second V matrix, the first intermediate feature of the i-th round of encoding is calculated. By performing attention cross-referencing of attributes with respect to historical performance curves using the first intermediate feature and the second feature of the i-th round of encoding, the second intermediate feature of the i-th round of encoding is obtained, including: The second Q matrix is ​​calculated using the second feature of the i-th round of encoding, and the first K matrix and the first V matrix are calculated using the first intermediate feature of the i-th round of encoding. The second intermediate feature of the i-th round of encoding is then calculated based on the second Q matrix, the first K matrix, and the first V matrix. By performing self-attention processing on the first intermediate feature of the i-th round of encoding, the first fusion feature of the i-th round of encoding is obtained, including: The third Q matrix, third K matrix, and third V matrix are calculated using the first intermediate features of the i-th round of encoding. Based on the third Q matrix, third K matrix, and third V matrix, the first fusion feature of the i-th round of encoding is then calculated.

[0022] In one possible implementation, the prediction curve of the target object is obtained by decoding the fused features through a prediction model, including: Decode and fuse features to obtain the prediction performance values ​​of the target object at multiple prediction time points; Based on the prediction performance values ​​at multiple prediction time points, determine the prediction curve of the target object within the prediction time period; Among them, the values ​​of multiple prediction time points in the prediction curve are equal to the prediction performance values ​​of multiple prediction time points; The predicted performance values ​​for multiple prediction time points are absolute predicted values, or the predicted performance values ​​for multiple prediction time points are relative to the historical performance values ​​of the corresponding historical time points in the historical performance curve.

[0023] In one possible implementation, the first acquisition module is configured to: By sampling any historical performance curve from multiple historical time periods, multiple sampling points corresponding to the historical performance curves of multiple historical time periods are obtained. The first data is obtained by taking multiple sampling points corresponding to the historical performance curves of multiple historical time periods. The sampling points corresponding to the historical performance curves of multiple historical time periods are arranged in chronological order in the first data.

[0024] In one possible implementation, the second acquisition module is used for: Retrieve the attributes of the target object across multiple historical time periods; Feature extraction is performed on the attributes corresponding to each historical time period to obtain the attribute features of each historical time period; where the attribute features of the historical time period are the representations of the attributes of the corresponding historical time period in the latent space. Based on the attribute characteristics of multiple historical time periods, second data is obtained. The attribute characteristics of different time periods in the second data correspond to different location codes.

[0025] In one possible implementation, the prediction model is trained based on the historical performance curves of the sample objects and the attributes of the samples; During training, the prediction model generates a prediction curve based on the first sample data corresponding to the historical curve of the sample and the second sample data corresponding to the sample attribute. The weight parameters in the prediction model are adjusted based on the training loss; the training loss is determined based on the application effect of the prediction curve.

[0026] According to a third aspect of the present disclosure, an electronic device for predicting performance curves is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the method as described in any one of the first aspects above.

[0027] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided such that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform any of the methods described in the first aspect of the present disclosure.

[0028] According to a fifth aspect of the present disclosure, a computer program product is provided, including computer instructions that, when executed by a processor, cause a computer to perform the method described in any one of the first aspects of the present disclosure.

[0029] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects: In summary, in the process of performance curve prediction, in addition to historical performance curves, the attributes of the target object are also introduced. By fusing the data used to characterize the historical performance curves and the data used to characterize the attributes of the target object, the attributes of the target object can influence the encoding process of the historical performance curves of the target object, thereby achieving multimodal prediction of the performance curves.

[0030] In this way, the fused features obtained from the encoding of the two types of data mentioned above can carry features related to the attributes of the target object, thus providing more reference information for decoding the prediction curve and helping to improve the accuracy of the generated prediction curve. Furthermore, this method allows for obtaining the prediction curve in advance, before the exact performance curve is published, so as to determine the appropriate usage plan for the target object based on the prediction curve. Attached Figure Description

[0031] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0032] Figure 1 This is a schematic diagram illustrating an application environment according to an exemplary embodiment; Figure 2 This is a flowchart illustrating a method for predicting performance curves according to an exemplary embodiment; Figure 3 This is a schematic diagram of the encoder structure according to an exemplary embodiment; Figure 4 This is a flowchart illustrating a method for predicting performance curves according to yet another exemplary embodiment; Figure 5 This is a schematic diagram illustrating the principle of causal masking according to an exemplary embodiment; Figure 6 This is a block diagram of a performance curve prediction device according to an exemplary embodiment; Figure 7 This is a structural block diagram of a predictive electronic device applied to performance curves, according to an exemplary embodiment. Detailed Implementation

[0033] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0034] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0035] First, the terms that may be involved in the embodiments of this application will be explained.

[0036] Function: Let A and B be two non-empty sets. If there exists a correspondence f such that for every element x in set A, there is a unique element y in set B, then f is called a function (or mapping) from A to B. In this embodiment, the performance curve is a continuously changing curve along the time dimension, typically with time on the horizontal axis and performance values ​​on the vertical axis, and can be considered a function; correspondingly, the attributes of the historical performance curve include multiple unrelated attribute values, and therefore can be considered as discrete vectors.

[0037] Root Mean Squared Error (RMSE): A statistical metric used to evaluate the prediction error of a prediction curve. It can be used to calculate the training loss of a prediction model so that the model parameters can be adjusted during training.

[0038] Delta Prediction: Predictions are determined by reflecting the daily average variation in performance. Differential prediction yields differential prediction values. Then, add the historical performance value from the previous day. The predictive performance value at a certain time point in the prediction curve can be obtained. .

[0039] Parameter-Efficient Fine-Tuning (PEFT) models are classic term structure models and their extensions that can characterize the shape of performance curves with a small number of parameters. PEFT models include the Nelson-Siegel-Svensson (NSS) model, among others.

[0040] B-Spline Regression (Spline): is a method that performs nonparametric smoothing / regression in the time domain to fit historical performance curves.

[0041] Functional Neural Network (FNN): Used to directly process functional inputs.

[0042] Multi-Layer Perceptron (MLP): A feedforward neural network consisting of fully connected layers and nonlinear activations.

[0043] Auto-Regressive (AR): An autoregressive model is trained on NSS parameters or spline coefficients to predict the parameter / coefficient settings for the next day.

[0044] Self-attention / cross-attention is the core mechanism in the Transformer model; self-attention models dependencies between inputs of the same type, while cross-attention is used to exchange conditional information between inputs of different types.

[0045] Query / Key / Value (Q / K / V): These are three sets of vectors used in the self-attention / cross-attention calculation process to generate attention weights and weighted outputs.

[0046] The Joint Sequence of Embedded Tokens (JSET) is used to concatenate curve tokens and attribute tokens into a time-ordered input sequence for processing in the same attention layer of the prediction model.

[0047] Causal Mask (CM): Used in the process of feature attention based on attention mechanism to restrict the attention of any time step to the token of the previous time step and the token of the current time step, so as to avoid the leakage of tokens of subsequent time steps and ensure the consistency of autoregressive prediction.

[0048] Encoder: Used to fuse the historical performance curve and the corresponding attributes to obtain the fused feature z. The fused feature is the feature identifier in the latent space after the attributes are fused from the historical performance curve.

[0049] Generator / Decoder: Used to decode the latest fused feature z into a time threshold defined by the time threshold. Prediction curve on For example, the time threshold τ is the time limit threshold, and the corresponding value range is [0.1, 20].

[0050] Functional model: The prediction model provided in the embodiments of this application includes the encoder and decoder described above. For example, the functional model can be a Transformer-style function prediction architecture, which can interact with the function sequence and vector set in the same attention in the encoder to obtain fused features z, and output a prediction curve based on the fused features z. This functional model is also called a function transformer (Func-Transformer, FT) model.

[0051] Function discretization: This is used to sample the historical performance curve for each historical time period. For example, in the time interval [0.1,20], sampling is performed with a time sampling step of 0.1 to obtain a total of 200 sampling points, which are then used as the input data (also known as function input) to characterize the historical performance curve.

[0052] Vector set: This is the input data (also called vector input) used to represent the attributes corresponding to the historical performance curve. Optionally, the attributes corresponding to the historical performance curve can include multiple attributes, and there is no order relationship between the different attributes; the different attributes are discrete. The total number of attributes corresponding to the historical performance curve varies in different historical time periods. Each attribute can be regarded as a vector; therefore, the attributes corresponding to the historical performance curve can form a vector set.

[0053] Temporal Contextual Encoding (TCE): Used to embed time information such as "year, month, day, and day of the week" into function inputs and vector inputs.

[0054] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating an application environment according to an exemplary embodiment, such as... Figure 1 As shown, the application environment may include computer device 01.

[0055] In an optional embodiment, the computer device 01 is used to execute a performance curve prediction method. Specifically, the computer device 01 may be, but is not limited to, electronic devices such as smartphones, desktop computers, tablets, laptops, and digital assistants. Optionally, the operating system running on the computer device 01 may include, but is not limited to, Android, iOS, Linux, and Windows.

[0056] Computer device 01 acquires the historical performance curves of the target object and the corresponding attributes. Based on these historical performance curves and their corresponding attributes, a prediction model is used to obtain a predicted performance curve. This predicted curve is used to predict the performance trend of the target object. This method allows for the early prediction of the target object's performance curve, enabling better analysis, use, arrangement, or configuration of the target object based on the predicted performance curve.

[0057] The prediction model can be deployed on computer device 01 or on other devices with strong computing power.

[0058] Optionally, the application environment may also include server 02. The prediction model can be deployed on server 02. For example, server 02 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server.

[0059] In addition, it should be noted that, Figure 1 The example shown is merely one application scenario of the performance curve prediction method provided in this disclosure.

[0060] In the embodiments described in this specification, the computer device 01 and the server 02 can be directly or indirectly connected through wired or wireless communication, and this application does not impose any restrictions on this.

[0061] In an exemplary embodiment, after obtaining the historical performance curve of the target object and the corresponding attributes, the computer device sends the historical performance curve of the target object and the corresponding attributes to the server 02. The server then runs a prediction model to generate a prediction curve based on the historical performance curve and the corresponding attributes.

[0062] The performance curve prediction method provided in this application, based on known historical performance curves and the attributes of the target object, predicts the changing trend of subsequent performance curves, thereby providing more references for the analysis, research, and use of the target object. Historical performance curves can be viewed as continuous functions, while attributes can be considered as multiple discrete data points. Compared to related technologies that predict curves based on historical performance curves, the performance curve prediction method provided in this application additionally introduces the attributes corresponding to the target object, providing more reference information for curve prediction and helping to improve the accuracy of the predicted curves obtained by the prediction model.

[0063] It should be noted that the following diagram illustrates one possible sequence of steps, and it is not strictly required to follow this order. Some steps can be performed in parallel without interdependence. The user information (including but not limited to user device information, user personal information, user behavior information, etc.) and data (including but not limited to data used for display, training data, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0064] Figure 2 This is a flowchart illustrating a method for predicting a performance curve according to an exemplary embodiment. Exemplarily, the method for predicting a performance curve can be... Figure 1 Computer device 01 in the middle executes, or, has Figure 1 The performance curve prediction method is executed by server 02, or by computer device 01 and server 02 in collaboration. The following example illustrates the method using a computer device as the execution entity.

[0065] like Figure 2 As shown, the method for predicting performance curves may include the following steps: Step 210: Obtain first data to characterize the changing trend of the historical performance curve of the target object; wherein the changing trend of the historical performance curve changes continuously over time.

[0066] The target object is the object whose performance change trend is reflected in the historical performance curve. For example, the target object is a business server, and the historical performance curve is used to reflect the performance change of the business object or the performance change trend of the business server over a historical period.

[0067] For example, the target is a product issuer, and the historical performance curve is used to reflect the revenue of the product issuer. The products issued by the product issuer include, but are not limited to, physical products, financial products, or virtual products.

[0068] Optionally, the x-axis of the historical performance curve represents the time unit, the y-axis represents the historical performance value of the target object, and the trend of the historical performance curve is used to reflect the fluctuation of the performance value corresponding to the target over time.

[0069] In one example, the first data is the historical performance curve itself; for example, the first data is a function formula used to characterize the historical performance curve.

[0070] In another example, the first data is obtained by sampling historical performance curves. For example, the first data is obtained by sampling historical performance values ​​at multiple historical time points in the historical performance curve.

[0071] Here, a historical time point is a point in time within a historical time period, or a specific point in time calculated from the beginning of a historical time period. For example, if the historical time period is one day, the historical time point is a specific moment within that day. Another example is if the historical time period is June 1st, and the historical time points are the first hour, the second hour, the first day, the second day, the first week, the first month, the second month, and so on, starting from June 1st.

[0072] For example, when the historical performance curve is used to characterize the performance of the target object at a point in time that is expected from the beginning of the historical time period, the historical time point is a point in time calculated from the beginning of the historical time period.

[0073] For example, a historical time period corresponds to a historical performance curve, and the historical time point is any value within the range of the horizontal axis of the historical performance curve. In other words, the value of the historical time point is related to the time span of the historical performance curve. The time span of the historical performance curve and the time span of the historical time period may be the same or different.

[0074] Optionally, since the first data is in numerical or functional form, it can be directly used as part of the input to the prediction model, i.e., the function input described above. This helps reduce the preprocessing required before inputting the first data into the prediction model, and also helps reduce the computational overhead in generating the prediction curve. Alternatively, the first data can be embedded, and the result of the embedding process can be input into the prediction model.

[0075] To provide more reference information for the curve prediction process, the computer device acquires first data based on historical performance curves corresponding to multiple historical time periods. This first data characterizes the changing trend of the target object's performance curve over a longer period, helping the prediction model analyze and learn the fluctuation characteristics of the target object's performance curve. This, in turn, helps improve the accuracy of the predicted curve obtained by the prediction model. The inference process for the predicted curve is described in steps 230 and 240. The process of acquiring first data based on historical performance curves from multiple historical time periods is described below.

[0076] Step 220: Obtain second data for the attributes used to characterize the target object; wherein the attributes overlap with the time range corresponding to the historical performance curves, and the attributes affect the changing trend of the historical performance curves.

[0077] The attributes of a target object are used to characterize the characteristics of the target object itself; the attributes of a target object directly or indirectly affect the historical performance curve of the target object.

[0078] Optionally, overlapping time ranges between an attribute and its historical performance curve mean that the attribute and the historical performance curve belong to the same historical time period. In other words, the historical performance curve and the attribute are aligned in the time dimension. For example, if the historical performance curve of the target object corresponds to the historical period of December 12th-13th, then the attribute of the target object also corresponds to December 12th-13th.

[0079] For example, if the historical performance curve corresponds to the historical period from 2020 to 2024, then the attributes of the target object are also the attributes from 2020 to 2024. The attributes of the target object can change dynamically from 2020 to 2024. In order to reflect the changes in the attributes of the target object, the attributes of the target object include attributes at multiple time points in the 2020-2024 period.

[0080] Optionally, the target object's attributes include multiple sub-attributes, which are independent of each other, and each sub-attribute can be represented by a vector. For example, the independence of the multiple sub-attributes means that there is no strict order among the multiple sub-attributes, and the order of the multiple sub-attributes does not affect the prediction result of the performance curve.

[0081] For example, when the target object is a business server, multiple sub-attributes are used to characterize the business load of the target object. For instance, if the business server provides services to multiple applications, then the multiple sub-attributes are used to characterize the business load brought to the business object by these multiple applications. Each sub-attribute is represented using a vector; the vector includes the following dimensions: a point in time and the number of times the application is accessed at that point in time.

[0082] For example, when the target object is a product publisher, multiple sub-attributes are used to characterize the products published by the target object. For instance, the product publisher publishes multiple products, and the multiple sub-attributes characterize the performance of each product; wherein each sub-attribute is represented using a vector; the vector includes at least one of the following dimensions: the price of the product, the revenue of the product, or the current time.

[0083] Optionally, the second data is obtained based on a vector set, wherein the vector set includes vectors corresponding to multiple sub-attributes, and the vectors corresponding to the multiple sub-attributes are arranged in an unordered manner to form the vector set.

[0084] For example, the computer device performs embedding processing on the vector set to obtain second data. In this way, the attributes of the target object can be mapped to the same latent space as the historical performance curve, making it easier for the prediction model to predict the performance curve based on the first and second data.

[0085] Step 230: The prediction model encodes the first data based on the second data to obtain the fusion features. The fusion features are the feature representation of the historical performance curve in the latent space after fusing the attributes of the target object.

[0086] Optionally, the prediction model includes an encoder and a decoder, wherein the fused features are obtained by the encoder processing the first data and the second data; the decoder is used to decode the fused features to obtain the prediction curve.

[0087] After acquiring the first data and the second data, the computer device fuses the first data and the second data at least once using a prediction model to obtain fused features. For example, the number of times the first data and the second data are fused is related to the prediction model.

[0088] For example, in the case of a dual-tower prediction model, the first and second data can be fused only once. As another example, in the case of a transformer-based prediction model, the first and second data can be fused multiple times. Refer to Fusion Method 1 and Fusion Method 2 below for details on fusing the first and second data.

[0089] Step 240: Decode the fused features through the prediction model to obtain the prediction curve of the target object. The prediction curve is used to predict the performance change trend of the target object.

[0090] Optionally, during the encoder generation of fused features Then, the encoder will fuse the features. Transmitted to decoder; decoder decodes and fuses features. , represented as .

[0091] in, Indicates the predicted time period. The decoder is used to obtain the prediction curve. .

[0092] For example, the generator / decoder includes, but is not limited to: a small multilayer perceptron decoder, a basis expansion decoder, a Fourier decoder, or a structured decoder.

[0093] In one possible implementation, step 240, obtaining the prediction curve of the target object by decoding the fusion features through the prediction model, can be implemented as follows: decoding the fusion features to obtain the prediction performance values ​​of the target object at multiple prediction time points; and determining the prediction curve of the target object in the prediction time period based on the prediction performance values ​​at multiple prediction time points.

[0094] In this context, multiple prediction time points refer to specific points within a prediction time period, and the prediction curve is used to simulate the performance curve of that prediction time period. Optionally, these multiple prediction time points correspond to multiple historical time points in the historical performance curve. Specifically, the first data is obtained by sampling the historical performance values ​​of multiple historical time periods from the historical performance curve of the historical time period, and these historical time points correspond to the multiple prediction time points in the prediction time period.

[0095] For example, the predicted time point in the predicted time period corresponds to the same period as the historical time point in the historical time period. If the first data is obtained based on the historical performance curves of multiple historical time periods, then the predicted time point in the predicted time period corresponds to the same period as the historical time point in the first historical time period; wherein, the first historical time period is the historical time period closest to the predicted time period. For example, the predicted time period is the 24th, and the historical time period is the 23rd.

[0096] For example, if the historical time period is [0.1, 20] years starting from the 23rd, and the predicted time period is [0.1, 20] years starting from the 24th; if multiple historical time points in the historical time period are 0.1 years, 0.2 years, 0.3 years, ..., 20 years starting from the 23rd, then multiple predicted time points in the predicted time period are 0.1 years, 0.2 years, 0.3 years, ..., 20 years starting from the 24th.

[0097] Optionally, the decoder decodes and fuses the features to obtain prediction performance values ​​at multiple prediction time points, and the computer device plots prediction curves based on the prediction time points and the prediction performance values ​​at the prediction time points.

[0098] For example, the prediction performance value at the prediction time point includes the following two forms: 1. Absolute value form The values ​​at multiple prediction time points on the prediction curve are equal to the prediction performance values ​​at those multiple time points, and these prediction performance values ​​are absolute prediction values. In this case, the prediction performance values ​​output by the prediction model... This refers to the value of the y-axis of the prediction curve at the prediction time point. .

[0099] For example, when the historical performance curve shows no overall trend, the absolute value form of the predicted performance value can more intuitively confirm the y-axis coordinate of the prediction time point on the prediction curve, thus helping to improve the efficiency of generating the prediction curve. Moreover, the absolute value form of the predicted performance value can be directly used to generate the prediction curve without further processing, which helps to reduce the cumulative error introduced by other processing.

[0100] 2. Relative value form (also known as difference value)

[0101] The predicted performance values ​​at multiple prediction time points are relative to the historical performance values ​​at the corresponding historical time points in the historical performance curve. In this case, the predicted performance values ​​output by the prediction model... Add the historical performance values ​​at the corresponding historical time points It equals the y-axis value at the prediction time point in the prediction curve. In other words, .

[0102] For example, when the historical performance curve shows an overall upward or downward trend, predicting the performance difference between the prediction time point and the historical time point through the encoder helps to eliminate the influence of historical performance values ​​from historical time points that are far from the prediction time point on the prediction process, and helps to make the distribution of predicted performance values ​​more stable. On the one hand, this prediction is less difficult, which helps to reduce the difficulty of model training; on the other hand, it helps to improve the accuracy of predicted performance values, thereby improving the reference value of the prediction curve.

[0103] Optionally, the predicted performance values ​​at multiple prediction time points are kept in the same form, either all in absolute form or all in relative form, in order to reduce the error caused by the inconsistent form of the predicted performance values ​​at different prediction time points during the process of the decoder generating predicted performance values ​​at multiple prediction time points based on fused features, and help to improve the accuracy of the prediction curve.

[0104] For example, the form of the predicted performance value at the prediction time point can be selected by the user. For example, before step 230, the prediction method of the performance curve further includes: the computer device acquiring indication information for indicating the form of the predicted performance value at the prediction time point; and the computer device adjusting the output mode of the prediction model based on the indication information so that the decoded predicted performance value meets the requirements of the indication information.

[0105] The indication information is sent from other devices to the computer device, or the computer device detects the user's format setting operation and obtains the indication information based on the format setting operation; the format setting operation is used to indicate the format of the prediction performance value at the prediction time point.

[0106] In summary, in the process of performance curve prediction, in addition to historical performance curves, the attributes of the target object are also introduced. By fusing the data used to characterize the historical performance curves and the data used to characterize the attributes of the target object, the attributes of the target object can influence the encoding process of the historical performance curves of the target object, thereby achieving multimodal prediction of the performance curves.

[0107] In this way, the fused features obtained from the encoding of the two types of data mentioned above can carry features related to the attributes of the target object, thus providing more reference information for decoding the prediction curve and helping to improve the accuracy of the generated prediction curve. Furthermore, this method allows for obtaining the prediction curve in advance, before the exact performance curve is published, so as to determine the appropriate usage plan for the target object based on the prediction curve.

[0108] The method for obtaining the first data is described below.

[0109] In one possible implementation, step 210, obtaining first data to characterize the changing trend of the historical performance curve of the target object, can be implemented as follows: the computer device samples the historical performance curves of multiple historical time periods to obtain multiple sampling points corresponding to the historical performance curves of the multiple historical time periods; the computer device obtains the first data based on the multiple sampling points corresponding to the historical performance curves of the multiple historical time periods, and the multiple sampling points corresponding to the historical performance curves of the multiple historical time periods are arranged in chronological order in the first data.

[0110] Here, multiple sampling points refer to multiple historical performance values ​​obtained by sampling from historical performance curves over multiple historical time periods. For example, the sampling point obtained by sampling historical time point 1 is the historical performance value corresponding to historical time point 1 in the historical performance curve.

[0111] Optionally, multiple historical time periods are consecutive historical time periods with corresponding historical performance curves; here, "consecutive" means that there is no missing time period between any two historical time periods, the missing time period corresponds to an additional historical performance curve, and the multiple historical time periods do not include the missing time period.

[0112] In one example, multiple historical time periods are L consecutive historical time periods preceding the prediction time period t. The performance curve prediction method is used to obtain the performance curve of the target object in the prediction time period t. Here, the prediction time period is either the current time period or any unreached time period after the current time. In other words, this embodiment of the application can predict the performance curve of the target object in the prediction time period in advance, even when the performance curve of the target object has not been generated or published based on the actual performance of the target object.

[0113] For example, each historical time period represents a calendar day, a month, or a quarter, etc. For instance, the historical performance curve corresponding to a historical time period is the performance curve of the target object during a day.

[0114] Optionally, the sampling point is the historical performance value of the target object at a certain historical time point on the x-axis of the historical performance curve. The computer device uniformly samples the historical performance curves of multiple historical time periods, and the total number of sampling points obtained from each historical performance curve is equal.

[0115] For example, a computer device samples M points from the historical performance curve at each historical point in time, where M is a positive integer greater than 1. If M equals 200, the computer device samples 200 points from each historical performance curve.

[0116] For example, the M sampling points are obtained by uniformly sampling the historical performance curve, and the interval between any two adjacent image sampling points in the M sampling points is equal on the x-axis; or, the M sampling points are obtained by non-uniformly sampling the historical performance curve, and the time interval between two adjacent sampling points in the M sampling points on the x-axis is not fixed.

[0117] For example, if the historical time period is [0.1, 30] years starting from January 1st, the historical time points corresponding to the M sampling points can be each month in the historical time period (relatively dense sampling), or the historical time points corresponding to the M sampling points can be the 1st month, 3rd month, 6th month, 1st year, 2nd year, 3rd year, 5th year, 7th year, 10th year, 30th year, etc. starting from January 1st (relatively sparse sampling).

[0118] For example, a computer device determines the x-axis coordinates corresponding to sampling points in a historical performance curve based on a learnable grid; wherein the learnable grid is obtained through training and can indicate the x-axis coordinates corresponding to M samples from the historical performance curve.

[0119] For example, the learnable mesh determines the x-axis coordinates of M sampling points based on the fluctuation frequency of historical performance curves. When fluctuations are frequent, the sampling density is increased, and the time interval between sampling points on the x-axis is decreased; when the historical performance curve changes gradually, the time interval between sampling points on the x-axis is increased. In this case, the sampling methods for multiple historical performance curves can be different. The sparsity of the sampling points is dynamically controlled based on the fluctuation trend of the historical performance curves, so that the M sampling points can more accurately represent the fluctuations of the historical performance curves.

[0120] Since historical performance curves are essentially continuous numerical signals with important microscopic morphological features (such as slope and convexity), multiple sampling points obtained from the historical performance curves yield the first data, which preserves the microscopic morphological features of the historical performance curves. The historical performance curve is equivalent to a continuous function with time on the x-axis and historical performance values ​​on the y-axis. The first data obtained in this way can be considered a sequence that retains the complete unfolded form of the function on the time axis.

[0121] Thus, obtaining the first data by sampling historical performance curves helps reduce the loss of numerical precision and smoothness information from historical performance curves, while preserving their topological features. In this way, the predictive model directly processes the raw historical performance values ​​obtained from sampling the historical performance curves, ensuring that the encoding of the first data preserves the topological structure (TS) and differentiability of the historical performance curves in the latent space, thereby helping to more accurately capture fluctuations in the historical performance curves.

[0122] In one example, historical performance curves are recorded on a daily basis, with the first data point representing the trend of the historical performance curve over L days. For instance, a computer device acquires historical performance curves over L days. t is used to represent the date.

[0123] in, This represents the historical performance curve of the target object on day L relative to the current date, and so on. This represents the historical performance curve of the target object relative to the day before the current date. For example, L is a positive integer greater than 1, such as L = 5, 20, 60, etc.

[0124] In one example, the target object has 5 working days per week, and each working day corresponds to a historical return curve. If the target object's performance is highly sensitive to environmental fluctuations, then L equals 5, meaning that the first data is generated using historical performance curves within a shorter period. If the target object's sensitivity to environmental fluctuations is moderate, then L equals 20. By providing historical performance curves within one month, the prediction of the performance curve can balance short-term noise while capturing medium-term trends. Thus, by using a larger number of historical performance curves, the accuracy of the predicted performance curve is improved.

[0125] For example, L can also be equal to 60, which means that the first data is generated based on the historical performance curve of the target object within a quarter. By providing sufficient historical performance data, the predicted performance curve can reflect the long-term performance change trend of the target object.

[0126] Subsequently, the computer device samples M sampling points from each historical performance curve and arranges them according to the corresponding time L. The first data is obtained by sampling M points.

[0127] For example, the first data includes multiple sampling points sampled from multiple historical performance curves and a separator, wherein the separator is used to isolate the multiple sampling points obtained from different historical performance curves. For example, the separator is used to provide the prediction model with prior time information about the first data.

[0128] For example, the delimiter includes at least one of the following: a first delimiter and a second delimiter. The first delimiter is used to characterize the historical time period corresponding to the historical performance curve to which the sampling point belongs. For instance, if the first data includes multiple sampling point groups, and each sampling point group includes multiple sampling points sampled from the same historical performance curve, and one sampling point group corresponds to a historical performance curve for one historical time period, then the same first delimiter corresponds to each group of sampling points.

[0129] For example, the first separator is also called the time anchor. For instance, the first separator of multiple sampling points obtained from sampling the historical performance curve on June 1st is used to characterize June 1st.

[0130] The second separator is used to characterize the historical time point corresponding to the sampling point in the historical performance curve. The value of the sampling point is obtained by sampling the historical time point from the historical performance curve.

[0131] For example, the second delimiter is also called the Time To Mask (TTM) coordinate, or Positional Encoding (PE). Exemplarily, the second delimiter can be denoted as... .

[0132] In one example, each sample point in the first data is represented as [sample point data, TTM coordinates and time anchor], and multiple sample points are arranged serially in chronological order.

[0133] In this way, embedding time information into the first data can provide prior knowledge about time for the prediction model in the process of processing the first data. This helps to better determine the impact of attributes of different historical time periods on historical performance curves during the process of generating prediction curves, thereby helping to improve the accuracy of the generated prediction curves.

[0134] This embodiment is based on Functional Data Analysis (FDA) theory. It obtains the first data by sampling historical performance curves. Compared with related technologies that use historical curve fitting to obtain the input data for the prediction model, it completely removes the explicit historical performance curve fitting / interpolation step and realizes the end-to-end mapping from the original historical performance value to the predicted performance value.

[0135] This simplifies the processing of historical performance curves. Compared to methods that fit historical performance curves using NSS, B-Spline interpolation, etc., this application does not destroy the historical performance curves, or linearly compresses them, thereby helping to avoid smoothing errors and data loss or distortion in historical performance curves caused by traditional fitting processes, and enhancing the ability of the first data to capture complex shapes in historical performance curves.

[0136] Furthermore, the sampling process has relatively low computational overhead, which helps improve the efficiency of obtaining the first data from historical performance curves, thereby helping to shorten the total time spent obtaining the predicted curve during the performance curve prediction process.

[0137] The method for obtaining the second data is described below.

[0138] In one possible implementation, step 220, obtaining second data for characterizing the attributes of the target object, can be implemented as follows: the computer device obtains the attributes of the target object corresponding to multiple historical time periods; the computer device performs feature extraction on the attributes corresponding to each historical time period to obtain the attribute features of each historical time period; wherein, the attribute features of the historical time period are the representation of the attributes of the corresponding historical time period in the latent space; the computer device obtains the second data based on the attribute features of multiple historical time periods, wherein the attribute features of different time periods in the second data correspond to different position codes.

[0139] In this context, the target object's attributes for each time period include at least one sub-attribute. The total number of sub-attributes included in the attributes for different historical time periods may be the same or different.

[0140] Optionally, the attributes corresponding to multiple historical time periods and the historical performance curves corresponding to multiple historical time periods are in one-to-one correspondence. That is, the first data is used to characterize the changing trend of the historical performance curves of multiple historical time periods, and the second data is used to characterize the attributes of the target object in multiple historical time periods; the time range of the attributes of the target object and the historical performance curves of the target object overlap.

[0141] For example, attributes corresponding to multiple historical time periods form an attribute set, and the second data is obtained based on this attribute set. Among these, attributes corresponding to the same historical time period... Each sub-attribute can be considered as a sub-attribute group.

[0142] For example, historical time period Corresponding attributes ,in, Indicates historical time period The i-th sub-attribute in the corresponding attribute. It can be in vector form.

[0143] Optionally, after obtaining the attributes corresponding to multiple historical time periods, the computer device extracts features from the attributes corresponding to each historical time period through a feature extraction layer to obtain the attribute features of each historical time period. The feature dimensions of the attribute features of each historical time period are the same.

[0144] The feature extraction layer can be placed outside the prediction model or embedded in the prediction model. For example, the feature extraction layer is also called the embedding layer, and extracting features from attributes is also called embedding attributes.

[0145] For example, a feature extraction layer is embedded in a prediction model. The computer device inputs attributes from multiple historical time periods into the prediction model, and the feature extraction layer in the prediction model extracts features from the attributes from the multiple historical time periods.

[0146] For example, the feature extraction layer is structured as a shared multilayer perceptron (MLP). For instance, the feature extraction layer is a multilayer perceptron comprising two rectified linear units (ReLUs). Each hidden layer (HL) in the multilayer perceptron may contain 16, 32, 48, or 64 neurons. These neurons are used to map the original high-dimensional features into a compact (e.g., 16-dimensional) latent space, and after ReLU activation, the final projection is the input dimension that the model can process.

[0147] After obtaining the attribute features of each time period, the computer device concatenates and combines the attribute features of multiple historical time periods in chronological order to obtain the second data. For example, if multiple historical time periods are included in chronological order as: historical time period 1, historical time period 2, ..., historical time period L, then the second data is [attribute features of historical time period 1, attribute features of historical time period 2, ..., attribute features of historical time period L].

[0148] In one possible implementation, when an attribute includes multiple sub-attributes, the computer device performs feature extraction on the attribute corresponding to each historical time period to obtain the attribute features of each historical time period, including multiple: for a certain historical time period, feature extraction is performed on multiple sub-attributes in the attribute to obtain multiple sub-attribute features, and the attribute features of the historical time period include multiple sub-attribute features.

[0149] Exemplarily, the method further includes pooling multiple sub-attribute features to obtain the attribute features for the given time period. Exemplarily, this pooling can employ a deep set / set-transformer style pooling. This helps to significantly reduce the length of the second data, adapting to performance curve prediction scenarios with a large number of attributes and improving the efficiency of predictive curve generation. Exemplarily, in this case, the subsequent encoding process can only perform attention crossover of the curve relative to the attributes and self-attention of the curve, as detailed in the embodiments below.

[0150] Optionally, the second data also includes delimiters to represent the time information of each attribute feature, so that the time information is embedded in the second data to help the prediction model distinguish the attribute features of different time periods in the second data.

[0151] For example, the delimiter includes at least one of the following: a first delimiter or a second delimiter. For an explanation of the first delimiter and the second delimiter, refer to the description of the first data in step 230, which will not be repeated here.

[0152] In one example, each attribute feature in the second data is represented as [attribute feature, TTM coordinate, time anchor].

[0153] Both the first and second data sets contain time plots and TTM coordinates. These time anchors and TTM coordinates are provided as key context coordinates to the prediction model. This allows the prediction model to establish alignment anchors during the processing of the first and second data sets based on the time anchors of the sampling points in the first data set and the time anchors of the attribute features in the second data set. This enables the attention crossover process to align the sampling points in the first data set with the attribute features in the second data set.

[0154] Furthermore, in the first data encoding process, the contextual relationship of each sampling point in the first data can be determined based on the TTM coordinates; in the second data encoding process, the contextual relationship of each attribute feature in the second data can be determined based on the TTM coordinates. By providing prior time information, it helps to more accurately complete the fusion encoding of the first data relative to the second data.

[0155] For example, the structure of the feature extraction layer can also be Fourier features, sinusoidal position encoding, or feature extraction can be achieved using a shared embedding table + interpolation. This approach enables the various attribute features in the second data to be aligned with the sampling points in the first data based on time, so as to adjust the implementation of the alignment anchor points without changing the backbone of the prediction model.

[0156] For example, during the acquisition of the second data, sparse keypoint sampling and learning interpolation can be performed on attributes corresponding to multiple historical time periods to obtain the second data.

[0157] The process of encoding the first data based on the second data is described below.

[0158] In one possible implementation, the prediction model includes an encoder. Step 230, encoding first data according to second data to obtain fused features, includes: using the encoder to perform feature fusion of historical performance curves with respect to attributes based on the first and second data, to obtain fused features.

[0159] Optionally, the historical performance curve is used to fuse the features of the attribute into the historical performance curve, so that the features of the attribute can affect the features of the historical performance curve, thereby allowing the attribute to influence the value of the predicted performance value during the decoding and fusion of features.

[0160] For example, the historical performance curve is fused with respect to the attribute feature at least once. The historical performance curve is fused with respect to the attribute feature, which can be implemented through at least one of the following fusion methods 1 and fusion method 2.

[0161] Fusion Method 1: Late Fusion (LF)

[0162] Optionally, in fusion mode 1, the encoder has a dual-tower structure, and the feature fusion of historical performance curves relative to attributes is achieved through late fusion.

[0163] Specifically, the encoder includes a curve encoder, an attribute encoder, and a final fusion layer; wherein, the curve encoder is used to encode the first data to obtain curve features; the attribute encoder is used to encode the second data to obtain attribute features; and the final fusion layer is used to fuse the curve features and attribute features to obtain fused features.

[0164] For example, the model types of curve encoders and attribute encoders include at least one of the following: neural network models, deep neural networks, or models built based on transformers. The model types of curve encoders and attribute encoders can be the same or different.

[0165] For example, the final fusion layer is a gating layer, which obtains fused features by splicing curve features and attribute features. Another example is that the final fusion layer of the encoder is an attention layer. For a detailed introduction to attention layers, please refer to fusion method 2 below.

[0166] Controlling the number of times historical performance curves are fused with attributes in this way helps reduce the computational load in the process of obtaining fused features, thereby improving the efficiency of generating prediction curves.

[0167] Fusion Method 2: Attention Fusion

[0168] Optionally, in fusion method 2, the encoder is a transformer structure. Exemplarily, the encoder includes at least one attention layer, and the historical performance curve relative to the attribute feature fusion is performed at least once in each attention layer.

[0169] Optionally, in fusion method 2, the sampling points in the first data also include TTM coordinates and time plots, and the attribute features in the second data also include TTM coordinates and time anchors; both the first and second data are sequences; wherein the first data includes multiple curve tokens, and the second data includes multiple attribute tokens. The curve tokens can be values ​​of sampling points obtained from historical performance curves, and the attribute tokens can be obtained by feature extraction from the vectors of sub-attributes.

[0170] For example, feature fusion of historical performance curves relative to attributes is achieved at least through: attention crossover of historical performance curves relative to attributes and self-attention.

[0171] The attention crossover between historical performance curves and attributes is also known as Vector-to-Function Attention (VFA). The historical performance curve can be viewed as a function, and the attributes of the target object can be viewed as vectors. Optionally, the attention crossover between historical performance curves and attributes is used to find content related to the historical performance curve from the attributes, thereby determining which attribute features in the second data will affect the changing trend of the historical performance curve, so that the macro-level changing trend of the target object's performance is constrained by micro-level attributes.

[0172] Self-attention is used to achieve intertemporal dependencies and morphological smooth transitions between the embedding sequences of historical performance curves. The embedding sequence of historical performance curves refers to the feature sequence related to the first data during the encoding process.

[0173] For example, feature fusion of historical performance curves relative to attributes may also include attention crossover of attributes relative to historical performance curves.

[0174] The attention cross of an attribute relative to the historical performance curve is also called the function-to-vector attention cross (FVA). The attention cross of an attribute relative to the historical performance curve is used to find content related to the attribute from the historical performance curve, and then determine whether the attribute features are reasonable relative to the historical performance curve.

[0175] In this way, by constraining the micro-level attribute characteristics through the macro-level change trend of the target object curve, it is possible to smooth and denoise abnormal or incorrect attribute characteristics in the second data.

[0176] By employing "vector-function attention crossover" and "vector-function attention crossover", a two-way closed-loop feedback mechanism is formed in the encoding process, which mutually supports the macroscopic historical performance curve and the microscopic attribute features. This helps to achieve high-precision performance curve prediction and has a certain ability to resist attribute noise interference, thus exhibiting good robustness.

[0177] In one possible implementation, the encoder performs feature fusion of historical performance curves relative to attributes based on first data and second data to obtain fused features, including: performing attention cross-processing on the first data and second data with respect to attributes to obtain first intermediate features; the first intermediate features are the feature representation of the first data in the latent space after attention cross-processing with the second data; and performing self-attention processing on the first intermediate features to obtain fused features.

[0178] The specific process of feature fusion is illustrated in the following example.

[0179] In this way, the second data, which represents the irregular and unordered attributes of the target object, can be processed in the same attention layer as the first data, which represents the continuous historical performance curve. This enables the joint processing of multimodal data, allowing the prediction model to predict performance curves based on multimodal data. Attention cross-referencing allows the encoding process to establish the temporal dependency between the first and second data, which helps improve the expressive power of the fused features for the changing trends of historical performance curves and the attributes of the target object, thereby improving the accuracy of the predicted curve obtained by decoding the fused features.

[0180] In one possible implementation, the fused features are obtained through N rounds of encoding, where N is a positive integer greater than 1.

[0181] In one possible implementation, the i-th round of the fusion encoding process includes the following encoding steps 1-3: Optionally, the encoder includes N attention layers, which are stacked on top of each other. Each attention layer performs the following encoding steps 1, 2, and 3 respectively. The i-th attention layer is used to perform the i-th round of encoding process, the output of the i-th attention layer is the input of the (i+1)-th encoding layer, and the output of the N-th attention layer is the fused feature.

[0182] In one possible implementation, such as Figure 3 As shown, the encoder structure includes a curve attribute attention block 310, an attribute curve attention block 320, and a self-attention block 330; the curve attribute attention block 310 is used to perform attention crossover of historical performance curves with respect to attributes; the attribute curve attention block 320 is used to perform attention crossover of attributes with respect to historical performance curves; and the self-attention block 330 is used to perform self-attention processing of the first intermediate feature.

[0183] This is equivalent to extending the traditional attention layer into a composite computation unit that performs three attention calculations, thereby enabling the fusion of multimodal data within a single attention layer. In this way, deep fusion of the first and second data is achieved through N rounds of encoding.

[0184] The specific functions of the curve attribute attention block 310, attribute curve attention block 320, and self-attention block 330 are illustrated with examples from the following coding steps 1-3.

[0185] Encoding step 1: By performing attention cross-cutting of the historical performance curve with respect to the attribute using the first feature of the i-th round of encoding and the second feature of the i-th round of encoding, the first intermediate feature of the i-th round of encoding is obtained.

[0186] In this process, the first feature of the first round of encoding is the first data, the second feature of the first round of encoding is the second data, the first fusion feature of the i-th round of encoding is the first feature of the (i+1)-th round of encoding, the second intermediate feature of the i-th round of encoding is the second feature of the (i+1)-th round of encoding, and the first fusion feature of the N-th round of encoding is the fusion feature.

[0187] In one possible implementation, in encoding step 1, the encoder calculates the first Q matrix using the first feature of the i-th round of encoding process, and calculates the second K matrix and the second V matrix using the second feature of the i-th round of encoding process. Based on the first Q matrix, the second K matrix, and the second V matrix, the encoder calculates the first intermediate feature of the i-th round of encoding process.

[0188] Optionally, the i-th attention layer includes a curve attribute attention block, and the curve attribute attention block includes a first operator. Second operator and the third operator .

[0189] For example, the first operator Second operator and the third operator These are all mapping operators, or simply operators, and are functions of independent projection. For example, the first operator... Second operator and the third operator Weight matrices Weight matrix The parameters are determined by training the prediction model. The training process of the prediction model is described in the following example.

[0190] For example, the first operator Perform a matrix multiplication operation with the first feature of the i-th round of encoding to obtain the first Q matrix; the second operator Perform matrix multiplication with the second feature of the i-th round of encoding to obtain the second K matrix; the third operator Perform a matrix multiplication operation with the second feature of the i-th round of encoding to obtain the second V matrix.

[0191] The curve attribute attention block calculates the first intermediate feature of the i-th round of encoding based on the first Q matrix, the second K matrix, and the second V matrix. This includes: calculating the attention score using the first Q matrix and the second K matrix; where the attention score is used to characterize the correlation between the token in the first feature of the i-th round of encoding and the token in the second feature of the i-th round of encoding; and performing a softmax normalization operation on the attention score to obtain the processed attention score. The processed attention score is then weighted and summed with the second V matrix to obtain the first intermediate feature of the i-th round of encoding.

[0192] Encoding step 2: The second intermediate feature of the i-th round of encoding is obtained by performing attention cross-reference between the attribute and the historical performance curve through the first intermediate feature of the i-th round of encoding and the second feature of the i-th round of encoding.

[0193] In one possible implementation, the encoder calculates the second Q matrix using the second feature of the i-th round of encoding, and calculates the first K matrix and the first V matrix using the first intermediate feature of the i-th round of encoding. Based on the second Q matrix, the first K matrix, and the first V matrix, the encoder calculates the second intermediate feature of the i-th round of encoding.

[0194] Optionally, the i-th attention layer includes an attribute curve attention block, and the attribute curve attention block includes a fourth operator. Fifth operator and the sixth operator Regarding the fourth operator Fifth operator and the sixth operator The first operator in the explanation and encoding step 1 Second operator and the third operator Similarly, please refer to the above examples for details.

[0195] Similarly, the fourth operator Perform a matrix multiplication operation with the second feature of the i-th round of encoding to obtain the second Q matrix; the fifth operator Perform a matrix multiplication operation with the first intermediate feature of the i-th round of encoding to obtain the first K matrix; the sixth operator Perform a matrix multiplication operation with the first intermediate feature of the i-th round of encoding to obtain the first V matrix.

[0196] In one possible implementation, during the encoding of attributes relative to functions, each attribute token only undergoes attention cross-interaction with the most recently acquired curve token. That is, only the current day's vector context is allowed to conditionally encode the latest function once, thus prohibiting cross-period forward propagation of earlier vectors. This helps reduce the difficulty of setting up causal masks. For information on the role of causal masks, please refer to [link / reference needed]. Figure 5 Corresponding implementation examples.

[0197] Subsequently, the attribute curve attention block calculates the second intermediate feature of the i-th round of encoding based on the second Q matrix, the first K matrix, and the first V matrix. This process is similar to that of the attribute curve attention block calculating the first intermediate feature of the i-th round of encoding based on the first Q matrix, the second K matrix, and the second V matrix. For details, please refer to Encoding Process 1, which will not be elaborated here.

[0198] Encoding step 3: Self-attention processing is performed on the first intermediate feature of the i-th round of encoding to obtain the first fusion feature of the i-th round of encoding.

[0199] In one possible implementation, the third Q matrix, the third K matrix, and the third V matrix are calculated using the first intermediate features of the i-th round of encoding, and the first fusion feature of the i-th round of encoding is calculated based on the third Q matrix, the third K matrix, and the third V matrix.

[0200] Optionally, the i-th attention layer includes a self-attention block, and the self-attention block includes a seventh operator. The eighth operator and the ninth operator Regarding the Seven Calculators The eighth operator and the ninth operator The first operator in the explanation and encoding step 1 Second operator and the third operator Similarly, please refer to the above examples for details.

[0201] Similarly, the seventh operator Perform a matrix multiplication operation with the first intermediate feature of the i-th round of encoding to obtain the third Q matrix; the eighth operator Perform a matrix multiplication operation with the first intermediate feature of the i-th round of encoding to obtain the third K matrix; the ninth operator Perform a matrix multiplication operation with the first intermediate feature of the i-th round of encoding to obtain the third V matrix.

[0202] Subsequently, the self-attention block calculates the first fusion feature of the i-th round of encoding based on the third Q matrix, the third K matrix, and the third V matrix. This process is similar to that of the curve attribute attention block, which calculates the first intermediate feature of the i-th round of encoding based on the first Q matrix, the second K matrix, and the second V matrix. For details, please refer to Encoding Process 1, which will not be elaborated here.

[0203] For example, the self-attention blocks in encoding step 3 can be replaced with a State Space Model (SSM) or a One-Dimensional Temporal Convolution (DTC). For instance, the self-attention blocks in some of the N attention layers can be replaced with SSM, while the attention blocks in other attention layers are not replaced.

[0204] This helps reduce the computational overhead in coding step 3 and improves coding efficiency. Furthermore, it also helps reduce memory usage during the execution of coding step 3.

[0205] In one possible implementation, the above embodiment illustrates the attention mechanism implementation process of encoding steps 1-3. Encoding steps 1-3 can also calculate the K matrix and V matrix based on lightweight convolution and basis unrolling.

[0206] Thus, through the above three coding steps, multimodal interaction between historical performance curves and attributes is achieved.

[0207] In one possible implementation, encoding steps 1 and 2 only apply to tokens within the same historical time period in the first and second features, while encoding step 3 applies to tokens within different historical time periods in the first intermediate feature. For example, if each historical time period represents one day, in the i-th round of encoding, the "function" is first applied to tokens within the same day. The attention crossover of the "set" yields an intraday enhanced function representation; then, function self-attention is performed across the day dimension. In this way, focusing the attention crossover between historical performance curves and attributes on the "same day" helps reduce the risk of information leakage in later historical time periods and also helps reduce the complexity of implementing causal masks.

[0208] Attention fusion enables dynamic computation of Q, K, and V matrices based on first and second data of arbitrary lengths, eliminating the need to limit the length of the first and second data and providing good adaptability for curve prediction in various scenarios. This also eliminates the need for forced alignment of the first and second data, helping to avoid distortion of the input prediction model's data signal.

[0209] Furthermore, since the attention mechanism is insensitive to the length of the second data, even if the total number of sub-attributes included in the attribute is different in different historical time periods or at different historical time points in the same historical time period, as long as the dimensions included in each sub-attribute are the same, the dimensionality requirements of the prediction model for the input data can be met. This makes the curve prediction process unaffected by the fluctuation in the number of sub-attributes included in the attribute, and realizes multimodal prediction of the performance curve based on the historical performance curve and the attributes of the target object.

[0210] The following example illustrates the process of generating a prediction curve. In this example, the prediction model is based on a transformer architecture and includes an encoder and a decoder.

[0211] For example, in this example, the prediction period t is 1 day, specifically June 10, 20xx. The predicted performance value is in relative form, and the form of the predicted performance value is preset through a form setting operation.

[0212] like Figure 4As shown, the method for predicting performance curves includes the following steps: Step 410: Obtain first data to characterize the changing trend of the historical performance curve of the target object; wherein the changing trend of the historical performance curve changes continuously over time.

[0213] The historical performance curves of the computer equipment over the past L days are represented as follows: For example, for each historical performance curve, the computer device samples 100 points and displays the 100... The L sampling points are arranged in chronological order to obtain the first data.

[0214] Step 420: Obtain second data for the attributes used to characterize the target object; wherein the attributes overlap with the time range corresponding to the historical performance curve, and the attributes affect the changing trend of the historical performance curve.

[0215] The computer device retrieves the attributes of the target object from the previous L days, represented as... For example, Let L be the attributes of the target object for the j days prior to June 10, 20xx, where j is a positive integer less than or equal to L. The computer device extracts the attributes of the target object through feature extraction, obtaining the attribute features for the previous L days, and then arranges these L days' attribute features into a second set of data.

[0216] For a detailed description of steps 410 and 420, please refer to the above embodiment.

[0217] Step 430: Combine the first data and the second data to obtain the combined sequence.

[0218] In one possible implementation, the computer device concatenates the first and second data into a joint sequence and inserts a causal mask. This controls the attention processing process to consider only the historical performance curves of earlier historical time periods and / or the attributes corresponding to earlier historical time periods, avoiding the leakage of information from other later historical time periods. This ensures that early information can be forward-propagated through the attention layer to the fused features.

[0219] like Figure 5As shown, the effect of causal masking is illustrated using two attribute tokens (501) from day tL in the second data and three attribute tokens (502) from day t-L+1; and curve tokens (503) from day tL in the first data and curve token (504) from day t-L+1. Under the effect of causal masking, during the attention crossover between curves and attributes, curve token 503 can only notice attribute token 501 from the second data, and curve token 504 can only notice attribute tokens 501 and 502 from the second data. This prevents the leakage of attribute tokens from historical time periods later than tL in the second data. Similarly, under the effect of causal masking, during the attention crossover between attributes and curves, attribute token 502 can only notice curve token 503 from the first data. During the self-attention process, curve token501 can only notice curve token503 and itself, thus preventing other curve tokens in the first data that are later than t-L+1 from being prematurely leaked during the self-attention process.

[0220] In another possible implementation, the first data also includes a separator, which allows for the sorting of the sampling points. Similarly, the second data also includes a separator, which aligns sampling points and attributes within the same historical time period. Please refer to the above embodiment for an explanation of separators.

[0221] After obtaining the joint sequence, the computer device inputs the joint sequence into the prediction model. Subsequently, the encoder of the prediction model encodes the first and second data through N rounds of encoding to obtain the fused features. Specifically, the first feature of the first round of encoding is the first data, the second feature of the first round of encoding is the second data, the first fused feature of the i-th round of encoding is used as the first feature of the (i+1)-th round of encoding, the second intermediate feature of the i-th round of encoding is used as the second feature of the (i+1)-th round of encoding, and the first fused feature of the N-th round of encoding is the fused feature.

[0222] Steps 440-460 are the encoding steps that need to be performed in each round of encoding during the N-round encoding process. The computer device repeats steps 440-460 N times to obtain the fused features. For details of this part, please refer to the above embodiment, which will not be repeated here.

[0223] Step 440: By performing attention cross-interaction of the historical performance curve with respect to the attribute through the first feature of the i-th round of encoding and the second feature of the i-th round of encoding, the first intermediate feature of the i-th round of encoding is obtained.

[0224] Step 450: The second intermediate feature of the i-th round of encoding is obtained by performing attention cross-reference between the attribute and the historical performance curve through the first intermediate feature of the i-th round of encoding and the second feature of the i-th round of encoding.

[0225] Step 460: Perform self-attention processing on the first intermediate feature of the i-th round of encoding process to obtain the first fusion feature of the i-th round of encoding process.

[0226] After step 460, if i is less than N, then i is incremented by 1, and then steps 440-460 are executed again.

[0227] Step 470: Decode the fused features to obtain the prediction performance values ​​of the target object at multiple prediction time points.

[0228] Among them, the predicted performance values ​​at multiple prediction time points are absolute predicted values, or the predicted performance values ​​at multiple prediction time points are relative to the historical performance values ​​at the corresponding historical time points in the historical performance curve.

[0229] Step 480: Plot the prediction curve based on the prediction performance values ​​at multiple prediction time points.

[0230] For example, each time the decoder generates a prediction performance value, it locates the prediction point corresponding to that prediction performance value in the prediction curve and connects that prediction point with the prediction point corresponding to the prediction performance value output by the decoder in the previous step, thus guiding the generation of multiple prediction performance values ​​to obtain the prediction curve.

[0231] For any parts not described in detail in this embodiment, please refer to the embodiments above.

[0232] This embodiment achieves feature fusion between the first and second data through attention encoding, making multimodal performance curve prediction possible. By providing the attributes of the target object during the performance curve prediction process, more reference information is provided for the performance curve prediction process, thereby helping to improve the accuracy of the predicted performance curve.

[0233] In one possible implementation, the prediction model is trained based on the historical performance curve of the sample object and the attributes corresponding to the historical performance curve; during the training process, the prediction model generates a prediction curve based on the first sample data corresponding to the historical performance curve and the second sample data corresponding to the sample attributes; the weight parameters in the prediction model are adjusted based on the training loss; the training loss is determined based on the application effect of the prediction curve.

[0234] Before the performance curve prediction method, the prediction model needs to be trained. During the training process, the prediction model generates a prediction curve based on the first sample data and the second sample data. For an introduction to the first sample data and the second sample data, please refer to the first data and the second data respectively. The process of generating the prediction curve is similar to the execution process of the performance curve prediction method. Please refer to the above embodiment for details.

[0235] For example, the training loss is calculated based on RMSE in an offline training environment. For instance, historical performance curves and sample attributes of samples from the previous 100 days are collected, and a predicted curve for day 61 is predicted based on the historical performance curves and sample attributes of samples from the previous 60 days. The training loss is determined based on RMSE based on the difference between the historical performance curve and the predicted performance curve for day 61.

[0236] For example, the training loss is calculated based on the application effect in an online training environment. For instance, the application prediction curve is simulated to obtain the application effect, which refers to the performance gains obtained after adjusting the target object based on the prediction curve (e.g., reduced server conflicts, increased product publisher capacity, increased revenue, etc.). The performance loss is then determined based on the application effect, and the model parameters of the prediction model are adjusted in reverse. (Model parameters may include the curve attribute attention block in the encoder, the weight factors in the attribute curve attention block and self-attention block, and the parameters in the decoder).

[0237] Empirical results show that, when the prediction performance is in the form of a relative value, this scheme can achieve positive application results (PnL=6.99%) while maintaining a competitive RMSE.

[0238] For example, the training loss may also include at least one of the following: shape regularization loss, quantile loss, Huber loss, etc. Among them, shape regularization loss includes, but is not limited to, at least one of the following: smoothing loss, monotonicity loss, and concavity loss, which are used to characterize the differences between the predicted curve and the label curve in terms of smoothness, monotonicity, and concavity, respectively.

[0239] For example, the training process of the predictive model can be performed periodically, such as fine-tuning the predictive model every K days, or periodically migrating the heavy model to the light model to meet latency requirements, thus helping to reduce the online cost of the predictive model. For example, K is an integer, such as K equals 7, 10, 16, 30, etc.

[0240] Optionally, during training, the decoder can output both absolute and relative prediction performance values ​​at the same prediction time point, enabling joint training and allowing the prediction model to predict both forms of prediction performance values ​​simultaneously. After the prediction model is deployed, the form of the prediction performance values ​​output by the decoder can be controlled through gating mechanisms or weighting.

[0241] Figure 6 A block diagram of a performance curve prediction device is shown according to an exemplary embodiment. (Refer to...) Figure 6 The device 600 may include: The first acquisition module 610 is used to acquire first data to characterize the changing trend of the historical performance curve of the target object; wherein the changing trend of the historical performance curve changes continuously over time.

[0242] The second acquisition module 620 is used to acquire second data for characterizing the attributes of the target object; wherein the attribute overlaps with the time range corresponding to the historical performance curve, and the attribute affects the trend of the historical performance curve.

[0243] The encoding module 630 is used to encode the first data based on the second data through the prediction model to obtain the fusion feature. The fusion feature is the feature representation of the historical performance curve in the latent space after fusing the attributes corresponding to multiple historical time periods.

[0244] The decoding module is used to decode the fused features through the prediction model to obtain the prediction curve of the target object. The prediction curve is used to predict the performance trend of the target object.

[0245] In one possible implementation, the encoding module 630 is used to obtain fused features by performing feature fusion of historical performance curves with respect to attributes based on first data and second data using an encoder.

[0246] In one possible implementation, the encoding module 630 includes: an attention crossover unit, used to perform attention crossover of the historical performance curves of the first data and the second data with respect to the attributes through the encoder to obtain a first intermediate feature; the first intermediate feature is the feature representation of the first data in the latent space after attention crossover with the second data; and a self-attention processing unit, used to perform self-attention processing based on the first intermediate feature through the encoder to obtain a fused feature.

[0247] In one possible implementation, the fused feature is obtained through N rounds of encoding, where N is a positive integer greater than 1. In the i-th round of encoding: attention cross-processing is performed between the first feature and the second feature of the i-th round of encoding on the historical performance curve relative to the attribute to obtain the first intermediate feature of the i-th round of encoding; attention cross-processing is performed between the first intermediate feature and the second feature of the i-th round of encoding on the attribute relative to the historical performance curve to obtain the second intermediate feature of the i-th round of encoding; self-attention processing is performed on the first intermediate feature of the i-th round of encoding to obtain the first fused feature of the i-th round of encoding; wherein, the first feature of the first round of encoding is the first data, the second feature of the first round of encoding is the second data, the first fused feature of the i-th round of encoding is used as the first feature of the (i+1)-th round of encoding, the second intermediate feature of the i-th round of encoding is used as the second feature of the (i+1)-th round of encoding, and the first fused feature of the N-th round of encoding is the fused feature.

[0248] In one possible implementation, the encoder includes a curve attribute attention block, a self-attention block, and an attribute curve attention block; the curve attribute attention block is used to perform attention crossover of historical performance curves with respect to attributes; the attribute curve attention block is used to perform attention crossover of attributes with respect to historical performance curves; and the self-attention block is used to perform self-attention processing of a first intermediate feature.

[0249] In one possible implementation, the first intermediate feature of the i-th round of encoding is obtained by performing attention cross-cutting of the historical performance curve with respect to the attribute using the first feature of the i-th round of encoding and the second feature of the i-th round of encoding. This includes: calculating the first Q matrix using the first feature of the i-th round of encoding, and calculating the second K matrix and the second V matrix using the second feature of the i-th round of encoding. The first intermediate feature of the i-th round of encoding is then calculated based on the first Q matrix, the second K matrix, and the second V matrix. The second intermediate feature of the i-th round of encoding is obtained by performing attention cross-reference between the attribute and the historical performance curve using the first intermediate feature and the second feature of the i-th round of encoding. This includes: calculating the second Q matrix using the second feature of the i-th round of encoding, and calculating the first K matrix and the first V matrix using the first intermediate feature of the i-th round of encoding. Based on the second Q matrix, the first K matrix, and the first V matrix, the second intermediate feature of the i-th round of encoding is then calculated. The first fusion feature of the i-th round of encoding is obtained by performing self-attention processing on the first intermediate feature of the i-th round of encoding, including: calculating the third Q matrix, the third K matrix and the third V matrix through the first intermediate feature of the i-th round of encoding, and calculating the first fusion feature of the i-th round of encoding based on the third Q matrix, the third K matrix and the third V matrix.

[0250] In one possible implementation, the decoding module 640 is used to decode the fused features to obtain the predicted performance values ​​of the target object at multiple prediction time points; and to determine the prediction curve of the target object in the prediction time period based on the predicted performance values ​​at the multiple prediction time points; wherein the values ​​of the multiple prediction time points in the prediction curve are equal to the predicted performance values ​​at the multiple prediction time points; and the predicted performance values ​​at the multiple prediction time points are absolute predicted values, or relative predicted values ​​relative to the historical performance values ​​of the corresponding historical time points in the historical performance curve.

[0251] In one possible implementation, the first acquisition module 610 is configured to: sample any one of the historical performance curves in multiple historical time periods to obtain multiple sampling points corresponding to the historical performance curves in multiple historical time periods; and obtain first data based on the multiple sampling points corresponding to the historical performance curves in multiple historical time periods, wherein the multiple sampling points corresponding to the historical performance curves in multiple historical time periods are arranged in chronological order in the first data.

[0252] In one possible implementation, the second acquisition module 620 is configured to: acquire the attributes of the target object corresponding to multiple historical time periods; extract features from the attributes corresponding to each historical time period to obtain the attribute features of each historical time period; wherein the attribute features of the historical time period are the representation of the attributes of the corresponding historical time period in the latent space; and obtain second data based on the attribute features of multiple historical time periods, wherein the attribute features of different time periods in the second data correspond to different position codes.

[0253] In one possible implementation, the prediction model is trained based on the historical performance curves of the sample objects and the sample attributes of the sample objects; during the training process, the prediction model generates a prediction curve based on the first sample data corresponding to the historical performance curves and the second sample data corresponding to the sample attributes; the weight parameters in the prediction model are adjusted based on the training loss; the training loss is determined based on the application effect of the prediction curves.

[0254] Figure 7 This is a block diagram illustrating a predictive electronic device applied to performance curves, based on an exemplary embodiment.

[0255] Based on the same inventive concept, embodiments of the present invention provide a computing device, such as... Figure 7 As shown, it includes at least one processor 701 and a memory 702 connected to at least one processor. In this embodiment of the invention, the specific connection medium between the processor 701 and the memory 702 is not limited. Figure 7 Taking the connection between the processor 701 and the memory 702 via a bus as an example, the bus can be divided into address bus, data bus, control bus, etc.

[0256] In this embodiment of the invention, the memory 702 stores instructions that can be executed by at least one processor 701. By executing the instructions stored in the memory 702, at least one processor 701 can perform the steps included in the aforementioned data processing method based on a blockchain system.

[0257] The processor 701 is the control center of the computing device, connecting various parts of the device via various interfaces and lines. It performs data processing by running or executing instructions stored in the memory 702 and accessing data stored in the memory 702. Optionally, the processor 701 may include one or more processing units. The processor 701 may integrate an application processor and a modem processor. The application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles issuing instructions. It is understood that the modem processor may not be integrated into the processor 701. In some embodiments, the processor 701 and the memory 702 may be implemented on the same chip; in other embodiments, they may be implemented on separate chips.

[0258] Processor 701 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the data processing embodiments based on the blockchain system can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0259] Memory 702, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 702 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 702 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In embodiments of the present invention, memory 702 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0260] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program executable by a computing device, wherein when the program is run on the computing device, the computing device performs the aforementioned... Figures 2 to 5 The method for predicting the performance curve described in any one of the above.

[0261] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0262] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0263] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0264] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0265] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0266] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting performance curves, characterized in that, The method includes: First data is obtained to characterize the changing trend of the historical performance curve of the target object; wherein the changing trend of the historical performance curve changes continuously over time. Obtain second data for characterizing the attributes of the target object; wherein the attributes overlap with the time range corresponding to the historical performance curve, and the attributes affect the trend of change of the historical performance curve; The prediction model encodes the first data according to the second data to obtain the fusion feature, which is the feature representation of the historical performance curve in the latent space after fusing the attributes; The prediction model decodes the fused features to obtain the prediction curve of the target object, which is used to predict the performance change trend of the target object.

2. The method according to claim 1, characterized in that, The prediction model includes an encoder, wherein encoding the first data according to the second data to obtain fused features includes: The encoder performs feature fusion of historical performance curves with respect to attributes based on the first and second data to obtain the fused features.

3. The method according to claim 2, characterized in that, The step of obtaining the fused features by fusing historical performance curves with respect to attributes based on the first and second data using the encoder includes: The encoder performs attention cross-interaction on the historical performance curves of the first data and the second data relative to the attributes to obtain a first intermediate feature; the first intermediate feature is the feature representation of the first data in the latent space after attention cross-interaction with the second data. The encoder performs self-attention processing based on the first intermediate feature to obtain the fused feature.

4. The method according to claim 3, characterized in that, The fusion feature is obtained through N rounds of encoding, where N is a positive integer greater than 1; in the i-th round of encoding: By performing attention cross-cutting of the historical performance curve with respect to the attribute using the first feature and the second feature of the i-th round of encoding, the first intermediate feature of the i-th round of encoding is obtained; The second intermediate feature of the i-th round of encoding is obtained by performing attention cross-reference between the attribute and the historical performance curve using the first intermediate feature and the second feature of the i-th round of encoding. The first fusion feature of the i-th round of encoding is obtained by performing self-attention processing on the first intermediate feature of the i-th round of encoding. Wherein, the first feature of the first round of encoding process is the first data, the second feature of the first round of encoding process is the second data, the first fusion feature of the i-th round of encoding process is the first feature of the (i+1)-th round of encoding process, the second intermediate feature of the i-th round of encoding process is the second feature of the (i+1)-th round of encoding process, and the first fusion feature of the N-th round of encoding process is the fusion feature.

5. The method according to claim 4, characterized in that, The encoder includes a curve attribute attention block, a self-attention block, and an attribute curve attention block. The curve attribute attention block is used to perform attention crossover of historical performance curves relative to attributes; The attribute curve attention block is used to perform attention crossover of attributes relative to historical performance curves; The self-attention block is used to perform self-attention processing on the first intermediate feature.

6. The method according to claim 4, characterized in that, The step of performing attention cross-cutting of historical performance curves relative to attributes using the first feature and the second feature of the i-th round of encoding to obtain the first intermediate feature of the i-th round of encoding includes: The first Q matrix is ​​calculated using the first feature of the i-th round of encoding, and the second K matrix and the second V matrix are calculated using the second feature of the i-th round of encoding. Based on the first Q matrix, the second K matrix, and the second V matrix, the first intermediate feature of the i-th round of encoding is calculated. The step of performing attention cross-referencing of attributes with respect to historical performance curves using the first intermediate feature and the second feature of the i-th round of encoding to obtain the second intermediate feature of the i-th round of encoding includes: The second Q matrix is ​​calculated based on the second feature of the i-th round of encoding process, and the first K matrix and the first V matrix are calculated based on the first intermediate feature of the i-th round of encoding process. The second intermediate feature of the i-th round of encoding process is then calculated based on the second Q matrix, the first K matrix, and the first V matrix. The step of performing self-attention processing on the first intermediate features of the i-th round of encoding to obtain the first fusion feature of the i-th round of encoding includes: The third Q matrix, the third K matrix, and the third V matrix are calculated based on the first intermediate features of the i-th round of encoding. The first fusion feature of the i-th round of encoding is then calculated based on the third Q matrix, the third K matrix, and the third V matrix.

7. The method according to any one of claims 1-6, characterized in that, The step of decoding the fused features through the prediction model to obtain the prediction curve of the target object includes: Decode the fused features to obtain the prediction performance values ​​of the target object at multiple prediction time points; Based on the prediction performance values ​​at the multiple prediction time points, the prediction curve of the target object during the prediction time period is determined; The values ​​of multiple prediction time points on the prediction curve are equal to the prediction performance values ​​of those multiple prediction time points. The predicted performance values ​​at the multiple prediction time points are absolute predicted values, or the predicted performance values ​​at the multiple prediction time points are relative predicted values ​​to the historical performance values ​​at the corresponding historical time points in the historical performance curve.

8. The method according to any one of claims 1-6, characterized in that, The acquisition of the first data used to characterize the changing trend of the historical performance curve of the target object includes: By sampling historical performance curves for multiple historical time periods, multiple sampling points corresponding to the historical performance curves for each of the multiple historical time periods are obtained. The first data is obtained based on the multiple sampling points corresponding to the historical performance curves of the multiple historical time periods, and the multiple sampling points corresponding to the historical performance curves of the multiple historical time periods are arranged in chronological order in the first data.

9. The method according to any one of claims 1-6, characterized in that, The acquisition of second data for characterizing the attributes of the target object includes: Obtain the attributes of the target object in multiple historical time periods; Feature extraction is performed on the attributes corresponding to each historical time period to obtain the attribute features of each historical time period; wherein, the attribute features of the historical time period are the representation of the attributes of the corresponding historical time period in the latent space. The second data is obtained based on the attribute characteristics of the multiple historical time periods, and the attribute characteristics of different time periods in the second data correspond to different location codes.

10. The method according to any one of claims 1-6, characterized in that, The prediction model is trained based on the historical performance curves of the sample objects and the sample attributes of the sample objects. During training, the prediction model generates a prediction curve based on the first sample data corresponding to the historical performance curve of the sample and the second sample data corresponding to the sample attribute. The weight parameters in the prediction model are adjusted based on the training loss; the training loss is determined based on the application effect of the prediction curve.

Images