WO2022252960A1 - Method and apparatus for training prediction model, and computer device and storage medium - Google Patents

Abstract

Provided are a method and apparatus for training a prediction model, and a computer device and a storage medium. The method comprises: acquiring the current streaming data and an online prediction model; if concept drift occurs in the current streaming data, determining a degree value of the concept drift; and according to the degree value of the concept drift, performing online training on the online prediction model in combination with the current streaming data and target streaming data, so as to obtain a target prediction model, wherein the target streaming data is sampled from a plurality of pieces of historical streaming data.

Description

Prediction model training method, device, computer equipment and storage medium

Cross References to Related Applications

This application is based on a Chinese patent application with application number 202110610089.3 and a filing date of June 1, 2021, and claims the priority of this Chinese patent application. The entire content of this Chinese patent application is hereby incorporated by reference into this application.

technical field

The present disclosure relates to the technical field of artificial intelligence, and in particular to a training method, device, computer equipment, non-transitory computer-readable storage medium, computer program product and computer program for a prediction model.

Background technique

In the project of building a smart city, a large number of artificial intelligence models may be used to construct various intelligent scenarios, such as traffic flow forecasting, traffic flow forecasting, air quality forecasting, monitoring and early warning, etc. The input data processed by these models are usually in the form of streams As the time and environment of data generation change, the distribution of data and the information reflected will also change. This phenomenon that the distribution or carried information changes over time is called concept drift. The conceptual drift of the data will cause the performance of the artificial intelligence model to decline, resulting in the inability of the static artificial intelligence model to continuously meet the prediction requirements in the intelligent scene, thereby affecting the prediction effect of the artificial intelligence model.

In related technologies, when the concept drift of reference data is used to assist in the training of the prediction model, both prediction performance and prediction stability cannot be considered, thereby affecting the online prediction effect of the prediction model.

Contents of the invention

The present disclosure aims to solve one of the technical problems in the related art at least to a certain extent.

Therefore, the purpose of this disclosure is to propose a training method, device, computer equipment, non-transitory computer-readable storage medium, computer program product and computer program for a prediction model, which can take into account the learning speed of the prediction model for new knowledge, and Avoid the occurrence of catastrophic forgetting, effectively improve the prediction performance and prediction stability of the prediction model, enable the prediction model to continuously meet the prediction needs in intelligent scenarios, and improve the prediction effect of the prediction model.

In order to achieve the above purpose, the method for training the prediction model proposed by the embodiment of the first aspect of the present disclosure includes: obtaining current flow data and an online prediction model; if concept drift occurs in the current flow data, determine the degree of concept drift; and according to the degree value of the concept drift, combine the current flow data and target flow data to perform online training on the online prediction model to obtain a target prediction model, wherein the target flow data is obtained from multiple historical Sampled from streaming data.

The prediction model training method proposed in the embodiment of the first aspect of the present disclosure determines the training timing of the online prediction model training by combining the value of the degree of concept drift, so as to support the online learning process, if the concept drift of the current streaming data is large , so that the online training process has a faster model learning speed, so that the model can quickly grasp the new knowledge represented by the stream data that produces concept drift, so as to ensure the online real-time performance of the prediction model; and if the current concept drift of the stream data Smaller, it makes the online training process maintain a relatively low learning rate to prevent the occurrence of catastrophic forgetting, so as to fully take into account the learning speed of the prediction model for new knowledge and the stability of the prediction model, so that the prediction model can continue Meet the forecasting needs in intelligent scenarios and improve the forecasting effect of the forecasting model.

In order to achieve the above purpose, the prediction model training device proposed by the embodiment of the second aspect of the present disclosure includes: an acquisition module, used to obtain the current flow data and an online prediction model; a first determination module, used to obtain the current flow data When concept drift occurs, determine the degree value of concept drift; and a training module, configured to perform online training on the online prediction model in combination with the current flow data and target flow data according to the degree value of concept drift, so as to obtain A target prediction model; wherein, the target flow data is obtained by sampling from a plurality of historical flow data.

The prediction model training device proposed in the embodiment of the second aspect of the present disclosure determines the training timing of the online prediction model training by combining the value of the degree of concept drift, so as to support the online learning process, if the concept drift of the current streaming data is large , so that the online training process has a faster model learning speed, so that the model can quickly grasp the new knowledge represented by the stream data that produces concept drift, so as to ensure the online real-time performance of the prediction model; and if the current concept drift of the stream data Smaller, it makes the online training process maintain a relatively low learning rate to prevent the occurrence of catastrophic forgetting, so as to fully take into account the learning speed of the prediction model for new knowledge and the stability of the prediction model, so that the prediction model can continue Meet the forecasting needs in intelligent scenarios and improve the forecasting effect of the forecasting model.

The embodiment of the third aspect of the present disclosure provides a computer device, including a memory, a processor, and a computer program stored in the memory and operable on the processor. On the one hand, the training method of the prediction model proposed by the embodiment.

The embodiment of the fourth aspect of the present disclosure proposes a non-transitory computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the method for training the prediction model as proposed in the embodiment of the first aspect of the present disclosure is implemented. .

The embodiment of the fifth aspect of the present disclosure provides a computer program product. When the instruction processor in the computer program product executes, the method for training the prediction model as proposed in the embodiment of the first aspect of the present disclosure is executed.

The embodiment of the sixth aspect of the present disclosure provides a computer program, the computer program includes computer program code, when the computer program code is run on the computer, the computer executes the prediction model as proposed in the embodiment of the first aspect of the present disclosure training method.

Additional aspects and advantages of the disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the disclosure.

Description of drawings

The above and/or additional aspects and advantages of the present disclosure will become apparent and understandable from the following description of the embodiments in conjunction with the accompanying drawings, wherein:

FIG. 1 is a schematic flowchart of a method for training a prediction model proposed by an embodiment of the present disclosure;

FIG. 2 is a schematic structural diagram of a training device for a prediction model proposed by an embodiment of the present disclosure;

FIG. 3 is a schematic flowchart of a method for training a prediction model proposed by another embodiment of the present disclosure;

FIG. 4 is a schematic flowchart of a method for training a prediction model proposed by another embodiment of the present disclosure;

FIG. 5 is a schematic flowchart of a method for training a prediction model proposed by another embodiment of the present disclosure;

FIG. 6 is a schematic structural diagram of a training device for a prediction model proposed by another embodiment of the present disclosure;

7 is a schematic structural diagram of a training device for a prediction model proposed by an embodiment of the present disclosure;

FIG. 8 is a schematic structural diagram of a training device for a prediction model proposed in another embodiment of the present disclosure;

Figure 9 shows a block diagram of an exemplary computer device suitable for use in implementing embodiments of the present disclosure.

Detailed ways

Embodiments of the present disclosure are described in detail below, examples of which are illustrated in the drawings, in which the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary only for explaining the present disclosure and should not be construed as limiting the present disclosure. On the contrary, the embodiments of the present disclosure include all changes, modifications and equivalents coming within the spirit and scope of the appended claims.

FIG. 1 is a schematic flowchart of a method for training a prediction model proposed by an embodiment of the present disclosure.

Wherein, it should be noted that the execution subject of the prediction model training method in the embodiment of the present disclosure is a prediction model training device, which can be realized by software and/or hardware, and which can be configured in an electronic device. Devices may include, but are not limited to, terminals, servers, and so on.

Among them, the artificial intelligence model used to predict the flow of people, traffic, air quality, monitoring and early warning in the smart city scene can be called a prediction model, and the artificial intelligence model can be, for example, a neural network model, machine learning models, etc., without limitation.

In the embodiment of the present disclosure, it is precisely to solve the technical problem that in the related art, when the concept drift of reference data is used to assist in the training of the prediction model, the prediction performance and the stability of the prediction cannot be taken into account, thereby affecting the online prediction effect of the prediction model. , provides a training method for the prediction model, which can take into account the learning speed of the prediction model for new knowledge and avoid the occurrence of catastrophic forgetting, effectively improve the prediction performance and prediction stability of the prediction model, so that the prediction model can continuously meet the Forecast demand in intelligent scenarios and improve the forecasting effect of the forecasting model.

As shown in FIG. 1 , the method for training the prediction model includes steps S101-S103.

S101: Obtain current streaming data and an online prediction model.

Among them, the prediction model used in the online environment to predict the flow of people, traffic, air quality, monitoring and early warning in the smart city scene can be called an online prediction model, that is to say, the online A predictive model is a predictive model that has been deployed in a scenario and performs corresponding predictive tasks.

Among them, the current flow data can be the flow data collected at the current time point. The flow data refers to the flow data generated in the form of flow. With the change of the time of data generation and the environment, the distribution of data and the information reflected will also change. The data that will change, the flow data can be used as the input of the online prediction model to trigger it to execute the corresponding prediction task, then in the embodiment of the present disclosure, it can support the dynamic collection of the current flow data, and support directly based on the current Online training of the online forecasting model is carried out according to the degree of concept drift of the stream data, and the stream data of the new training sample is formed by using the continuously acquired current stream data and the target stream data obtained by sampling the historical stream data to trigger the forecasting model Online training ensures the prediction performance of the online prediction model and controls the training speed of the online prediction model.

S102: If concept drift occurs in the current stream data, determine a degree value of concept drift.

After obtaining the current stream data and the online prediction model, it is possible to judge in real time whether concept drift has occurred in the current stream data. If concept drift has occurred in the current stream data, determine the degree of concept drift. The value of the degree of concept drift indicates: current The extent to which concept drift occurs with streaming data.

For example, the current flow data can be compared with multiple historical flow data to obtain the change between the current flow data and multiple historical flow data, and then the change can be quantified to obtain the degree of concept drift value, and compare the value of the degree of concept drift with a set threshold of the degree of concept drift, and determine whether to trigger online training of the prediction model according to the result of the comparison.

That is to say, the embodiment of the present disclosure determines the training timing of the online prediction model training by combining the value of the degree of concept drift, so as to support the online learning process. The process has a fast model learning speed, so that the model can quickly grasp the new knowledge represented by the flow data that produces concept drift, so as to ensure the online real-time performance of the prediction model; The training process maintains a relatively low learning rate to prevent the occurrence of catastrophic forgetting, so as to comprehensively take into account the learning speed of the prediction model for new knowledge and the stability of the prediction model.

S103: According to the degree value of concept drift, combine the current flow data and the target flow data to conduct online training on the online prediction model to obtain the target prediction model, wherein the target flow data is obtained by sampling from multiple historical flow data .

After the degree value of concept drift is determined, the online prediction model combined with the current flow data and target flow data can be adaptively triggered to obtain the target prediction model according to the degree value of concept drift.

Wherein, the prediction model obtained through training may be referred to as a target prediction model.

Among them, the historical flow data can be understood as: a current time point corresponding to the current flow data. Multiple flow data accumulated between can be called historical flow data.

In the embodiment of the present disclosure, the training timing of the online prediction model training is determined by combining the value of the degree of concept drift, so as to support the online learning process. The fast model learning speed enables the model to quickly grasp the new knowledge represented by the streaming data that produces concept drift, so as to ensure the online real-time performance of the prediction model; and if the concept drift of the current streaming data is small, the online training process can be maintained. Relatively low learning rate to prevent the occurrence of catastrophic forgetting, so as to comprehensively take into account the learning speed of the prediction model for new knowledge and the stability of the prediction model, so that the prediction model can continuously meet the prediction needs in intelligent scenarios and improve Predictive performance of the predictive model.

An example is described as follows, as shown in FIG. 2 , which is a schematic structural diagram of a training device for a prediction model proposed by an embodiment of the present disclosure. The description of the following embodiments of the present disclosure may be combined with what is shown in FIG. 2 , which is not limited thereto. in,

Data pool: can be used to store all historical flow data.

Real-time data unit: a unit for online real-time streaming data input. This real-time data unit can receive real-time streaming data input, and input real-time streaming data to the reasoning service unit for model reasoning. At the same time, it will also receive real-time streaming data The data is input to the data drift monitoring unit, which is used to calculate whether concept drift occurs in the real-time streaming data.

Playback buffer pool: used to store the latest multiple historical stream data, the multiple historical stream data may specifically be part of the stream data in all the historical stream data in the above data pool, and the online learning module can sample from the playback buffer pool to obtain the target Streaming data for online training of predictive models.

Sample flow data for training: It can be randomly selected from the data pool, and the initial prediction model is trained offline, so that the trained prediction model is pushed to the model warehouse.

Model training: use the number of sample streams for training to conduct offline training on the initial prediction model, so as to push the trained prediction model to the model warehouse, and in the actual prediction scenario, push the latest prediction model in the model warehouse to the inference service unit for online prediction, and thus the prediction model used by the inference service unit to provide inference services may be referred to as an online prediction model.

Data drift detection data set: Receive current stream data and cache it to determine whether concept drift occurs and the degree value of concept drift, and refer to the degree value of concept drift to assist in online learning of online prediction models.

Concept drift evaluation module: use the current flow data received by the data drift monitoring unit to calculate the degree of concept drift of the current flow data relative to the historical flow data, to determine whether the degree of concept drift of the current flow data is greater than the degree of concept drift threshold.

Online learning module: The online data in the data pool, the playback cache pool, and the degree of concept drift are used for online training of the prediction model, and the trained prediction model is transmitted to the model warehouse, and the current flow data used for online learning is synchronized to the Playback cache pool.

Model warehouse: store prediction models obtained from offline training and prediction models obtained from online training.

Reasoning service unit: Obtain the latest prediction model from the model warehouse, use the latest prediction model to perform model reasoning on online data, and output prediction results.

FIG. 3 is a schematic flowchart of a prediction model training method proposed by another embodiment of the present disclosure. For the description of FIG. 3 , the above-mentioned FIG. 2 may be combined together, including steps S301-S307.

S301: Obtain current streaming data and an online prediction model.

The model training unit in Figure 2 above can be used to conduct offline model training using all the historical flow data in the data pool, and the prediction model is obtained after training, which will be used as the model basis for subsequent online training, and the model training unit will transfer the prediction model to to the model repository.

The inference service unit in FIG. 2 above obtains the latest prediction model from the model warehouse and uses it as the inference model of the system.

S302: If concept drift occurs in the current flow data, determine a degree value of concept drift.

The real-time data unit may input the received real-time stream data into the data drift monitoring unit to calculate whether concept drift occurs in the real-time stream data, and determine the degree of concept drift if concept drift occurs in the current stream data.

S303: If the value of the degree of concept drift is greater than or equal to the threshold of the degree of concept drift, refer to the set rule to sample the target stream data from the historical stream data stored in the playback buffer pool.

For example, current flow data can be compared with historical flow data to detect the degree of data drift, assuming that the judgment threshold for drift detection is set to Th (this Th can be called the concept drift threshold), if Drift detection output (the drift detection output can specifically be that the concept drift degree value is greater than or equal to the concept drift degree threshold) Th, then refer to the set rule to sample the target flow data from the historical flow data stored in the playback buffer pool.

S304: Perform online training on the online prediction model according to the current flow data and the target flow data.

In some embodiments, the online prediction model can be trained online for a first set number of times according to the current stream data and the target stream data, wherein the first set number of times is the maximum value within the set number of times range.

Among them, if the value of the degree of concept drift of the current stream data is greater than or equal to the threshold of the degree of concept drift, it can trigger the online training of the online prediction model directly, and train the set number of times. The set number of times can be called The first set times.

For example, if the degree of concept drift of the current streaming data is greater than or equal to the threshold of the degree of concept drift, it can directly trigger the online training of the online prediction model, and the number of training iterations is Inc_Train_Num (the number of training iterations is Inc_Train_Num, that is, is called the first set number of times), so as to effectively ensure that the prediction model can quickly respond to changes in the distribution of stream data.

The above-mentioned first set number of times is the maximum value within the range of set times.

S305: If the degree of concept drift is less than the threshold of concept drift, store the current stream data in the playback buffer pool, and determine whether the data volume of the stream data stored in the playback buffer pool satisfies the set condition.

For example, if the drift detection output (the drift detection output can be specifically the concept drift degree value) is less than the concept drift degree threshold Th, then the current stream data can be stored in the playback buffer pool, and the data of the stream data stored in the playback buffer pool can be determined quantity.

S306: If the amount of data does not meet the set condition, continuously acquire the current flow data, and dynamically update the amount of data.

The setting condition may be that the amount of data has reached the maximum amount of data that can be stored in the playback buffer pool, and there is no limit to this.

That is to say, if the degree of concept drift of the current stream data is less than the threshold value of the degree of concept drift, the current stream data will be stored in the playback buffer pool, and the data volume of the stream data stored in the playback buffer pool will be determined in real time. If If the amount of data does not reach the maximum amount of data that can be stored in the playback buffer pool, the current streaming data will be continuously obtained and the amount of data will be updated dynamically.

S307: If the amount of data satisfies the set conditions, refer to the set rules to sample the target stream data from the historical stream data stored in the playback buffer pool, and perform online training on the online prediction model according to the target stream data.

When determining the data volume of stream data stored in the playback cache pool in real time, if the data volume reaches the maximum data volume that can be stored in the playback cache pool, then refer to the set rule to sample from the historical stream data stored in the playback cache pool Obtain the target flow data, and perform online training on the online prediction model based on the target flow data.

For example, if it is detected that the degree of concept drift of the current flow data is less than the threshold value of the degree of concept drift, then after the storage of the playback buffer pool is full, the target flow data is sampled from the historical flow data stored in the playback buffer pool, and the The target flow data is input into the online learning module to train the online prediction model online and iteratively update the online prediction model to ensure the effect of the online prediction model.

In some embodiments, the online prediction model can be trained online for a second set number of times according to the target flow data, wherein the second set number of times is less than the first set number of times, and the second set number of times is based on the set number of times range The maximum value, the minimum value, the value of the degree of concept drift, and the threshold value of the degree of concept drift are calculated.

In the embodiment of the present disclosure, when the degree of concept drift of the current stream data is less than the threshold of the degree of concept drift, online training is triggered by the amount of data. Considering that the degree of concept drift of the current stream data is small, the iterations of training can be appropriately reduced The number of iterations Inc_Train_Num, the number of iterations Inc_Train_Num in this case may be referred to as the second set number of times. For the calculation method of the second set number of times, please refer to the following.

For example, in the above-mentioned online learning module shown in Figure 2, the data set can be used to train the online prediction model online and repeat Inc_Train_Num times. In the embodiment of the present disclosure, it is also possible to introduce a playback cache pool and increase The technical characteristics of unknown labels help to increase the training stability of the online prediction model, reduce the number of model retraining, and improve training efficiency.

In the embodiment of the present disclosure, it is also supported to put the latest received current stream data for training into the playback buffer pool in real time, and put the target prediction model obtained from online training into the model warehouse to meet the needs of subsequent scenarios .

In the embodiment of the present disclosure, because the online training of the online prediction model is triggered according to the degree of concept drift, the trigger mechanism of online training is more flexible and applicable, and it is possible to timely avoid the error caused by the concept drift of streaming data. The performance of the model decreases, effectively improving the prediction performance and prediction stability of the prediction model, so that the prediction model can continuously meet the prediction needs in intelligent scenarios, and improve the prediction effect of the prediction model. If the concept drift of the current streaming data is relatively large, online training of the online prediction model can be directly triggered, thereby effectively ensuring that the prediction model can quickly respond to changes in the distribution of streaming data. If the degree of concept drift in detecting the current flow data is small, after the storage of the playback buffer pool is full, the target flow data is sampled from the historical flow data stored in the playback buffer pool, and the target flow data is input into the online learning module, The online prediction model is trained online, and the online prediction model is updated iteratively to ensure the effect of the online prediction model. The degree of concept drift of the current streaming data is small, and online training is triggered by the amount of data. Considering that the degree of concept drift of the current streaming data is small, the number of training iterations can be appropriately reduced, thus ensuring the stability of the prediction model training sex.

For example, receive current streaming data for reasoning, and obtain real results based on business feedback;

(1) According to the current flow data, calculate the value of the degree of concept drift relative to the historical flow data, and record it as Concept_drift_value, and judge whether the value of the concept drift of the current flow data is greater than or equal to the concept drift threshold value, if the current flow data If the degree of concept drift is large, go to step (4), otherwise go to step (2).

(2) And judge whether the playback buffer pool meets the quantity requirement of online learning, if so, go to step (4), otherwise go to step (1).

(3) Calculate the number of online training times Inc_Train_Num, where Inc_Train_max and Inc_Train_min represent the maximum number of iterations (the maximum value in the set number range) and the minimum number of iterations (the minimum value in the set number range) of online training, respectively.

(31) If the degree of concept drift of the current streaming data is large, Inc_Train_Num=Inc_Train_max, so as to accelerate the learning speed.

(32) If the degree of concept drift of the current stream data is small, then reduce the learning speed of online training, and the number of training iterations can be calculated in the way of Inc_Train_Num=(Inc_Train_max-Inc_Train_min)*Concept_drift_value/Th+Inc_Train_Min.

(4) Online training process: sample from the playback buffer pool to obtain target stream data, mix target stream data with current stream data, and conduct online training for the online prediction model.

(5) After repeating step (5) Inc_Train_Num times, push the trained target prediction model into the model warehouse.

(6) Add the current stream data received in this round to the playback buffer pool, and delete the data that enters the playback buffer pool first according to the first-in-first-out principle to ensure that the total number of samples in the playback buffer pool remains unchanged.

(7) Return to step (1).

That is to say, in the embodiment of the present disclosure, after the target prediction model is trained, the current stream data is stored in the playback buffer pool, and the first stream data in the playback buffer pool is deleted, wherein the storage corresponding to the first stream data order, before the corresponding storage order of other stream data, the first stream data and other stream data together constitute the stream data stored in the playback buffer pool, so that after the online training is completed, the latest stream data can be stored in the playback buffer pool in a timely manner. And according to the first-in-first-out principle, the most "old" historical flow data is deleted to ensure that the total amount of sample flow data in the playback buffer pool remains unchanged, and the feature representation capability of the sample flow data in the playback buffer pool is guaranteed.

FIG. 4 is a schematic flowchart of a prediction model training method proposed by another embodiment of the present disclosure. The description of FIG. 4 may be combined with the above-mentioned FIG. 2 , including steps S401-S402.

S401: Determine the weight according to the data volume of the stream data stored in the playback buffer pool and the corresponding storage time of the historical stream data.

S402: From the historical flow data stored in the playback buffer pool, obtain a set amount of historical flow data based on selection probability sampling as the target flow data, and the ratio between the set amount and the amount of the current flow data is a preset value .

The selection probability may be specifically set according to the weight corresponding to the historical flow data, for details, please refer to the following description.

That is to say, the embodiments of the present disclosure consider that in real prediction scenarios, the distribution deviation between the newer flow data and the historical flow data may be relatively large due to factors such as the data volume and data sampling of the newer flow data. Large, thus providing a weighted playback cache pool strategy to avoid the jitter of the online prediction model, which is more conducive to the security and stability of the online prediction model in real prediction scenarios.

For example, the online training of the online prediction model by mixing the latest received current flow data with the historical flow data can effectively alleviate the deviation caused by the distribution jitter of the latest received current flow data, so that the implementation of the present disclosure In this example, the weighted playback buffer pool can be used to store historical flow data, and during the online training process, a set amount of historical flow data is randomly sampled according to the weight of the stored historical flow data and used as target flow data, combined with current flow data To train the predictive model online.

In the embodiment of the present disclosure, it is also possible to adaptively control the ratio between the number of newer flow data and the number of historical flow data in the online learning training data (this ratio can be expressed by parameter α), the size of α determines Whether the online training of the online prediction model is more inclined to the model effect or the stability of the model, so that it can be adaptively set according to the needs of the actual prediction scenario, and there is no limit to this.

The data volume of online learning samples: the number of newer streaming data = 1: α;

Among them, when α is small, online learning will rely more on historical streaming data, so that better model stability can be obtained; while when α is larger, online learning will rely more on newer streaming data, so that it can This makes the online performance of the target prediction model better.

In the embodiment of the present disclosure, the ratio between the number of newer flow data and the number of historical flow data in the online learning training data train_set can be configured as 1:9, that is, α=0.1, the data volume of online learning training data 1/20 of the data that can be stored in the buffer pool can be retrieved without limitation.

In the embodiment of the present disclosure, the method of setting the selection probability according to the weight corresponding to the historical flow data can be illustrated as follows:

Weighted experience replay means that when sampling samples, the most valuable samples are selected first, but the most valuable samples cannot be selected only, otherwise it will cause overfitting. The higher the value, the greater the probability of being drawn, and the lowest value. , there is also a certain probability of being drawn.

In the embodiments of the present disclosure, when performing an online learning task, it is considered that newer stream data is more valuable for online prediction model learning, so that newer stream data can be configured to have a higher probability of being extracted.

Assuming that the number of samples that can be cached in the playback buffer pool is N, the historical flow data in the playback buffer pool is sorted according to the corresponding storage time, which is recorded as x _i , where i∈[1, N], the smaller i is, the historical flow The closer the data is to the current point in time. Assuming that the decline rate of the weight corresponding to the historical flow data is γ=(1-1/N), then the probability (selection probability) of the historical flow data x _i being sampled is:

ε _i = Aγ ⁱ ;

Among them, A is the normalization factor:

Therefore, for the historical stream data stored in the replay buffer pool in the embodiment of the present disclosure, the selection probability of the earliest historical stream data x _N is about 1/e of the latest historical stream data _x1 , where e is a natural constant, Using this method to set the weight of historical flow data will not reduce the probability of early historical flow data being sampled due to changes in the amount of cached data in the playback buffer pool. With the increase of , the sampled weight gradually decreases, which can be effectively applied to the weighted sampling of the historical flow data in the playback buffer pool in the scenario where there is concept drift.

Therefore, in the embodiment of the present disclosure, the weighted playback buffer pool is used to store historical flow data, and the historical flow data obtained from the playback buffer pool based on selection probability sampling is mixed with a small amount of newer flow data received online in real time. A new online training data set is formed to ensure that the online prediction model continuously learns new knowledge from new samples, while consolidating the old knowledge that has been mastered in the past, improving the performance of the online prediction model and training stability.

FIG. 5 is a schematic flowchart of a prediction model training method proposed by another embodiment of the present disclosure. The description of FIG. 5 may be combined with the above-mentioned FIG. 2 , including steps S501-S504.

S501: Obtain an annotation label corresponding to the current flow data.

S502: Obtain at least one prediction label supported by the online prediction model.

S503: If the label label belongs to at least one kind of prediction label, perform online training on the online prediction model in combination with the current flow data and the target flow data according to the degree value of concept drift.

S504: If the label label does not belong to at least one kind of prediction label, perform offline update training on the online prediction model according to the degree value of the concept drift, combined with the current flow data and the target flow data.

That is to say, the embodiments of the present disclosure also support adding several tag cache positions during the original online prediction model training, so as to ensure that when receiving real-time current stream data, if the labeled tag corresponding to the current stream data is the original online prediction When the label is not supported by the model, the online prediction model can still continue the online training process, reduce the number of retraining of the online prediction model, and improve the training efficiency, so that the training method of the prediction model in the embodiment of the present disclosure can be widely applied to streaming data Classification problems where the type of corresponding label changes.

For example, refer to FIG. 6 together. FIG. 6 is a schematic structural diagram of a training device for a prediction model proposed by another embodiment of the present disclosure. The description for the embodiment shown in FIG. 5 can be described as follows:

In the online training process of the classification problem, the label label corresponding to the current flow data may be a prediction label not supported by the current online prediction model, so in the embodiment of the present disclosure, in order to make the online prediction model output conform to the classification number, so that the online prediction model can effectively meet the needs of online model services in real business scenarios, it is possible to add several tag cache positions (size of max_label_buffer_size), when the label label does not belong to the prediction label currently supported by the online prediction model, you can directly associate the label label with the reserved label cache location without modifying the output of the online prediction model, so that the model can be directly trained online. After the tag cache location is processed, the target stream data in the playback cache pool is used to train the online prediction model to obtain the target prediction model.

Use the historical flow data, set the label_buffer_size to max_label_buffer_size, and train the online prediction model; sample the target flow data from the playback buffer pool; merge it with the received current flow data to form an online training data set, and additional processing is required for the classification problem. Data label, the steps are as follows:

If the label label is not an unknown label (that is, the label label belongs to at least one prediction label supported by the online prediction model), it will be directly trained online.

If the label label is an unknown label (that is, the label label does not belong to at least one prediction label supported by the online prediction model), calculate the number of unknown labels as unklabel_len, and compare the number of unknown labels with the number of label cache locations label_buffer_size .

If label_buffer_size<unklabel_len, that is, the label cache location is not enough to put all unknown labels, directly train online.

If label_buffer_size>unklabel_len, encode the unknown label unklabel_len sequentially, and update the label cache position size to label_buffer_size=label_buffer_size-unklabel_len,

Use the online training data set to conduct online training on the basis of the latest online prediction model, update the model parameters, and save it as an online prediction model. If the number of iterations meets the set number, it will end to obtain the target prediction model. If iterative If the number of times does not meet the set number, retrain the online prediction model, set label_buffer_size to max_label_buffer_size, and perform online training.

An application example for the embodiments of the present disclosure may be as follows:

Assuming that the prediction task performed by the online prediction model is the classification of city governance events, the online prediction model can be specifically such as the event classification model. The probabilities of , from high to low, give the three most eligible sectors. Event distribution (content and distribution target departments) will change with time and environment. Therefore, the event classification model will monitor data drift in real time and trigger online training immediately when drift occurs.

In the embodiment of the present disclosure, a text classification algorithm can be used. In order to deal with the unknown labels that appear, when training the event classification model, a number of label buffer positions (label_buffer_size) will be added to the output category of the event classification model in advance. When a new label appears When , the new label can be directly associated with the reserved label cache location, without modifying the output of the event classification model, so that the event classification model can be directly trained online. After receiving a certain amount of new samples, the event classification model mixes the new samples with the data randomly sampled from the playback buffer pool for online training.

The main parameters of event classification online learning are shown in Table 1 below:

Table 1

parameter	value
Playback buffer pool size (N)	20000
Data drift detection pool size	100
Latest: online training data ratio (α)	0.1
Concept Drift Detection Algorithm	MMD
Concept Drift Threshold (Th)	0.01
Maximum number of iterations for online training (Inc_Train_max)	5
Minimum number of iterations for online training (Inc_Train_min)	1
Maximum number of unknown label positions (max_label_buffer_size)	5

Through experiments, it is found that on the drifted data, the top3 accuracy of the original event classification model drops by about 8.7% compared to quantitatively using the full amount of data to retrain the event classification model. By adding unknown labels to the original event classification model in advance Cache, samples are sampled from the cache pool during training, and the top3 accuracy of the online learning event classification model on drift data is 2.1% higher than that of the original event classification model. The specific experimental results on the drift data are shown in Table 2 below:

Table 2

Result analysis:

Using the static event classification model, online data can be continuously reasoned, but the performance of the event classification model will gradually degrade over time.

After every 1000 pieces of data are received, the latest full data is used to retrain the event classification model: due to the continuous introduction of newer stream data, its online performance has been significantly improved compared with the static event classification model, but due to the event The classification model training uses the full amount of data, and the calculation time for a single event classification model update is relatively long.

Using the online event classification model update method and adding unknown label cache can greatly reduce the training time. At the same time, the performance of the event classification model is comparable to retraining with full data.

Based on the update of the online event classification model, the weighted playback cache pool is introduced. Due to the increase in data processing work, the update time of the single event classification model has increased; the introduction of the cache pool data has increased the stability of the event classification model. Compared with the online update method, the model performance has been improved to a certain extent.

Assuming that the prediction task performed by the online prediction model is parking lot flow forecasting, the online forecasting model can be specific, such as the parking lot flow forecasting model. The parking lot flow forecasting model uses the inflow and outflow data of the parking lot in the previous 24 hours to predict the next two Inflow and outflow of hours.

In the embodiment of the present disclosure, a neural network model can be used for time series prediction, and the granularity of traffic data is one data point every half hour, that is, the input of the traffic flow prediction model of this parking lot is 48 consecutive time series points, and the output is 4 consecutive time series points. On the basis of the original parking lot flow prediction model, an online training is performed every half hour, using the latest training sample (that is, the latest 52 data points, of which the first 48 points are used as the input of the parking lot flow prediction model, and the last 4 are used as the parking lot output of the traffic forecasting model) and 127 samples randomly sampled from the replay buffer pool (each sample is composed of 52 consecutive time series data points, the first 48 points are used as the input of the parking lot flow forecasting model, and the last 4 are used as the parking lot flow forecasting model output) to form a training set for online training. During the training process, a large learning rate and a small number of training iterations are used to enable the parking lot flow prediction model to quickly learn the current data distribution.

The main parameters of the online learning of the traffic flow forecasting model are shown in Table 3 below:

table 3

parameter	value
Playback buffer pool size (N)	20000
Data drift detection pool size	1000
Latest: online training data ratio (α)	1/64
Concept Drift Detection Algorithm	ADWin
Concept Drift Threshold (Th)	0.002
Maximum number of iterations for online training (Inc_Train_max)	5
Minimum number of iterations for online training (Inc_Train_min)	1

Experiments on the drifting data show that the online learning depot flow forecasting model has a decline in the RMSE and MAE three depot flow forecasting model evaluation indicators compared with the original depot flow forecasting model. The specific experimental results on the drift data are shown in Table 4 below:

Table 4

Result analysis:

The static parking lot flow forecasting model can continuously reason about online data, but the performance of the parking lot flow forecasting model will gradually degrade over time.

Full update in T+3 days: every 3 days, use the full data to retrain the parking lot flow forecasting model. The effect of the parking lot flow forecasting model is better than the static parking lot flow forecasting model, but due to the complexity of the network and the large amount of historical flow data, The update time of a single parking lot flow forecasting model is long, and it is difficult to meet the online service requirements of the urban parking lot flow forecasting model.

T + 3 days of data online training method + weighted playback cache pool: use the data received every 3 days, combined with some historical flow data in the weighted playback cache pool, to adjust the parking lot traffic prediction model. This method does not require manual participation, and can automatically update the parking lot flow prediction model online; since only a small amount of newer flow data is used to update the parking lot flow prediction model, the calculation time is greatly shortened.

Since the online update method takes more into account the impact of time-adjacent flow data, the effect of the parking lot flow forecasting model has a certain improvement compared to using the full amount of data to update the parking lot flow forecasting model.

The learning speed self-control method can adaptively adjust the learning speed according to the conceptual drift degree of the streaming data. In the case of drifting online streaming data, the performance of the parking lot flow prediction model can be improved to a certain extent.

FIG. 7 is a schematic structural diagram of a training device for a prediction model proposed by an embodiment of the present disclosure.

As shown in Figure 7, the training device 70 of this predictive model comprises:

An acquisition module 701, configured to acquire current flow data and an online prediction model;

The first determination module 702 is configured to determine the degree of concept drift when the concept drift occurs in the current stream data; and

The training module 703 is used to perform online training on the online prediction model according to the degree of concept drift, combining the current flow data and the target flow data, so as to obtain the target prediction model; wherein, the target flow data is obtained from a plurality of historical flow data obtained by sampling.

In some embodiments of the present disclosure, the training module 703 is specifically used for:

If the value of the degree of concept drift is greater than or equal to the threshold of the degree of concept drift, the reference setting rule samples the target flow data from the historical flow data stored in the playback buffer pool;

The online prediction model is trained online according to the current flow data and the target flow data.

In some embodiments of the present disclosure, the training module 703 is specifically used for:

If the value of the degree of concept drift is less than the threshold value of the degree of concept drift, then store the current flow data in the playback buffer pool, and determine the data volume of the stored flow data in the playback buffer pool;

If the amount of data does not meet the set conditions, the current flow data will be continuously obtained and the amount of data will be updated dynamically;

If the amount of data satisfies the set conditions, then refer to the set rules to sample the target stream data from the historical stream data stored in the playback cache pool, and perform online training on the online prediction model based on the target stream data.

In some embodiments of the present disclosure, the training module 703 is specifically used for:

The online prediction model is trained online for a first set number of times according to the current stream data and the target stream data, wherein the first set number of times is a maximum value within the range of the set number of times.

In some embodiments of the present disclosure, the training module 703 is specifically used for:

According to the target stream data, the online prediction model is trained for a second set number of times online, wherein the second set number of times is less than the first set number of times, and the second set number of times is based on the maximum value in the set number of times range, The minimum value, the degree of concept drift value, and the degree of concept drift threshold are calculated.

In some embodiments of the present disclosure, as shown in FIG. 8 , the device 70 further includes:

The processing module 704 is configured to store the current stream data in the playback buffer pool after the target prediction model is obtained through training, and delete the first stream data in the playback buffer pool, wherein the storage order corresponding to the first stream data is different from other stream data Prior to the corresponding storage order, the first stream data and other stream data together constitute the stream data stored in the playback buffer pool.

In some embodiments of the present disclosure, the training module 703 is specifically used for:

From the historical stream data stored in the playback buffer pool, a set amount of historical stream data is obtained based on selection probability sampling as the target stream data, and the ratio between the set amount and the current stream data is a preset value.

In some embodiments of the present disclosure, the selection probability is determined by the weight of historical flow data. As shown in FIG. 8, the device 70 also includes:

The second determination module 705 is configured to determine the weight according to the data volume of the stream data stored in the playback cache pool and the storage time corresponding to the historical stream data.

In some embodiments of the present disclosure, wherein,

The acquiring module 701 is further configured to acquire, after acquiring the current stream data, an annotation label corresponding to the current stream data, and acquire at least one prediction label supported by the online prediction model;

The training module 703 is specifically used for:

If the label label belongs to at least one kind of prediction label, the online prediction model is trained online according to the degree value of the concept drift, combined with the current flow data and the target flow data;

If the label label does not belong to at least one prediction label, the online prediction model is updated and trained offline according to the degree of concept drift, combined with the current flow data and the target flow data.

Corresponding to the training method of the prediction model provided by the embodiments of FIGS. 1 to 6 above, the embodiment of the present disclosure also provides a training device for the prediction model. The training method of the prediction model provided in the embodiment of FIG. 6 corresponds, so the implementation of the training method of the prediction model is also applicable to the training device of the prediction model provided in the embodiment of the present disclosure, and will not be described in detail in the embodiment of the present disclosure.

In the embodiment of the present disclosure, the training timing of the online prediction model training is determined by combining the value of the degree of concept drift, so as to support the online learning process. The fast model learning speed enables the model to quickly grasp the new knowledge represented by the streaming data that produces concept drift, so as to ensure the online real-time performance of the prediction model; and if the concept drift of the current streaming data is small, the online training process can be maintained. Relatively low learning rate to prevent the occurrence of catastrophic forgetting, so as to comprehensively take into account the learning speed of the prediction model for new knowledge and the stability of the prediction model, so that the prediction model can continuously meet the prediction needs in intelligent scenarios and improve Predictive performance of the predictive model.

In order to realize the above-mentioned embodiments, an embodiment of the present disclosure also proposes a computer device, including: a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the program, it realizes the aforementioned The training method of the prediction model proposed in the embodiment.

In order to realize the above-mentioned embodiments, the embodiments of the present disclosure also propose a non-transitory computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the training of the prediction model as proposed in the foregoing embodiments of the present disclosure is implemented. method.

In order to realize the above-mentioned embodiments, the embodiments of the present disclosure further propose a computer program product. When the instruction processor in the computer program product executes, the method for training the prediction model as proposed in the foregoing embodiments of the present disclosure is executed.

In order to realize the above-mentioned embodiments, the embodiments of the present disclosure also propose a computer program, the computer program includes computer program code, when the computer program code is run on the computer, the computer executes the predictive model as proposed in the foregoing embodiments of the present disclosure. training method.

Figure 9 shows a block diagram of an exemplary computer device suitable for use in implementing embodiments of the present disclosure. The computer device 12 shown in FIG. 9 is only an example, and should not limit the functions and scope of use of the embodiments of the present disclosure.

As shown in FIG. 9, computer device 12 takes the form of a general-purpose computing device. Components of computer device 12 may include, but are not limited to: one or more processors or processing units 16 , system memory 28 , bus 18 connecting various system components including system memory 28 and processing unit 16 .

Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus structures. For example, these architectures include but are not limited to Industry Standard Architecture (Industry Standard Architecture; hereinafter referred to as: ISA) bus, Micro Channel Architecture (Micro Channel Architecture; hereinafter referred to as: MAC) bus, enhanced ISA bus, video electronics Standards Association (Video Electronics Standards Association; hereinafter referred to as: VESA) local bus and Peripheral Component Interconnection (hereinafter referred to as: PCI) bus.

Computer device 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by computer device 12 and include both volatile and nonvolatile media, removable and non-removable media.

The memory 28 may include a computer system readable medium in the form of a volatile memory, such as a random access memory (Random Access Memory; hereinafter referred to as: RAM) 30 and/or a cache memory 32 . Computer device 12 may further include other removable/non-removable, volatile/nonvolatile computer system storage media. By way of example only, storage system 34 may be used to read and write to non-removable, non-volatile magnetic media (not shown in FIG. 9, commonly referred to as a "hard drive").

Although not shown in FIG. 9, a disk drive for reading and writing to a removable nonvolatile disk (such as a "floppy disk") may be provided, as well as a removable nonvolatile disk (such as a Compact Disk ROM (Compact Disk). Disc Read Only Memory; hereinafter referred to as: CD-ROM), Digital Video Disc Read Only Memory (hereinafter referred to as: DVD-ROM) or other optical media). In these cases, each drive may be connected to bus 18 via one or more data media interfaces. Memory 28 may include at least one program product having a set (eg, at least one) of program modules configured to perform the functions of various embodiments of the present disclosure.

A program/utility 40 having a set (at least one) of program modules 42 may be stored, for example, in memory 28, such program modules 42 including but not limited to an operating system, one or more application programs, other program modules, and program data , each or some combination of these examples may include implementations of network environments. The program modules 42 generally perform the functions and/or methods of the embodiments described in the present disclosure.

The computer device 12 may also communicate with one or more external devices 14 (e.g., a keyboard, pointing device, display 24, etc.), and with one or more devices that enable a user to interact with the computer device 12, and/or with Any device (eg, network card, modem, etc.) that enables the computing device 12 to communicate with one or more other computing devices. Such communication may occur through input/output (I/O) interface 22 . Moreover, the computer device 12 can also communicate with one or more networks (such as a local area network (Local Area Network; hereinafter referred to as: LAN), a wide area network (Wide Area Network; hereinafter referred to as: WAN) and/or public networks, such as the Internet, through the network adapter 20. ) communication. As shown, network adapter 20 communicates with other modules of computer device 12 via bus 18 . It should be appreciated that although not shown, other hardware and/or software modules may be used in conjunction with computer device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives And data backup storage system, etc.

The processing unit 16 executes various functional applications and data processing by running the programs stored in the system memory 28 , such as realizing the training method of the prediction model mentioned in the foregoing embodiments.

It should be noted that the foregoing explanations of the embodiments of the prediction model training method are also applicable to the computer equipment, non-transitory computer-readable storage media, computer program products, and computer programs in the above embodiments, and will not be repeated here.

Other embodiments of the present disclosure will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. The present disclosure is intended to cover any modification, use or adaptation of the present disclosure. These modifications, uses or adaptations follow the general principles of the present disclosure and include common knowledge or conventional technical means in the technical field not disclosed in the present disclosure. . The specification and examples are to be considered exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It should be understood that the present disclosure is not limited to the precise constructions which have been described above and shown in the drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

It should be noted that, in the description of the present disclosure, terms such as "first" and "second" are used for description purposes only, and should not be understood as indicating or implying relative importance. In addition, in the description of the present disclosure, unless otherwise specified, "plurality" means two or more.

Any process or method descriptions in flowcharts or otherwise described herein may be understood to represent modules, segments or portions of code comprising one or more executable instructions for implementing specific logical functions or steps of the process , and the scope of preferred embodiments of the present disclosure includes additional implementations in which functions may be performed out of the order shown or discussed, including substantially concurrently or in reverse order depending on the functions involved, which shall It is understood by those skilled in the art to which the embodiments of the present disclosure pertain.

It should be understood that various parts of the present disclosure may be implemented in hardware, software, firmware or a combination thereof. In the embodiments described above, various steps or methods may be implemented by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one or combination of the following techniques known in the art: Discrete logic circuits, ASICs with suitable combinational logic gates, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

Those of ordinary skill in the art can understand that all or part of the steps carried by the methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. During execution, one or a combination of the steps of the method embodiments is included.

In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing module, each unit may exist separately physically, or two or more units may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules. If the integrated modules are realized in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium.

The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, and the like.

In the description of this specification, descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific examples", or "some examples" mean that specific features described in connection with the embodiment or example , structure, material or characteristic is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Although the embodiments of the present disclosure have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limitations on the present disclosure, and those skilled in the art can understand the above-mentioned embodiments within the scope of the present disclosure. The embodiments are subject to changes, modifications, substitutions and variations.

Claims (22)

1. A method for training a predictive model, characterized in that the method comprises:

   Obtain current streaming data and online prediction models;

   If concept drift occurs to the current stream data, determining a degree value of concept drift; and

   performing online training on the online prediction model in combination with the current flow data and the target flow data according to the degree value of the concept drift, so as to obtain the target prediction model,

   Wherein, the target flow data is obtained by sampling from a plurality of historical flow data.
2. The method according to claim 1, wherein the online training of the online prediction model in combination with the current flow data and the target flow data according to the degree value of the concept drift comprises:

   If the value of the degree of concept drift is greater than or equal to the threshold value of the degree of concept drift, then refer to the set rule to sample the target flow data from the historical flow data stored in the playback buffer pool;

   The online prediction model is trained online according to the current flow data and the target flow data.
3. The method according to claim 1, wherein the online training of the online prediction model in combination with the current flow data and the target flow data according to the degree value of the concept drift comprises:

   If the value of the degree of concept drift is less than the threshold value of the degree of concept drift, storing the current stream data in the playback buffer pool, and determining the data volume of the stored stream data in the playback buffer pool;

   If the data volume does not meet the set condition, continuously acquire the current stream data, and dynamically update the data volume;

   If the amount of data satisfies the set condition, refer to the set rule to sample the target stream data from the historical stream data stored in the playback buffer pool, and Predictive models are trained online.
4. The method according to claim 2, wherein the online training of the online prediction model according to the current flow data and the target flow data comprises:

   The online prediction model is trained online for a first set number of times according to the current stream data and the target stream data, wherein the first set number of times is a maximum value within a range of set times.
5. The method according to claim 3, wherein the online training of the online prediction model according to the target flow data comprises:

   According to the target stream data, the online prediction model is trained online for a second set number of times, wherein the second set number of times is smaller than the first set number of times, and the second set number of times is based on The maximum value, the minimum value, the value of the degree of concept drift, and the threshold value of the degree of concept drift in the range of the set number of times are calculated.
6. The method according to any one of claims 1 to 5, characterized in that, after training to obtain the target prediction model, comprising:

   storing the current stream data in a playback buffer pool;

   delete the first-stream data in the playback buffer pool;

   Wherein, the storage order corresponding to the first stream data is before the storage order corresponding to other stream data, and the first stream data and the other stream data together constitute the stream data stored in the playback buffer pool.
7. The method according to claim 2 or 3, wherein the reference to the set rule is obtained by sampling the target flow data from the historical flow data stored in the playback buffer pool, including:

   From the historical flow data stored in the playback buffer pool, a set amount of historical flow data is obtained based on selection probability sampling as the target flow data, the ratio between the set amount and the amount of the current flow data is the default value.
8. The method according to claim 7, wherein the selection probability is determined by the weight of the historical flow data, and the method further comprises:

   The weight is determined according to the data volume of the stream data stored in the playback cache pool and the corresponding storage time of the historical stream data.
9. The method according to any one of claims 1 to 8, characterized in that, after said acquiring the current stream data, further comprising:

   Obtain an annotation tag corresponding to the current stream data;

   Obtain at least one prediction label supported by the online prediction model;

   If the label label belongs to the at least one prediction label, performing online training on the online prediction model in combination with the current flow data and target flow data according to the degree value of the concept drift;

   If the label label does not belong to the at least one prediction label, performing offline update training on the online prediction model in combination with the current flow data and the target flow data according to the degree value of the concept drift.
10. A training device for a predictive model, characterized in that the device comprises:

    Acquisition module, used to obtain current flow data and online prediction model;

    A first determination module, configured to determine the degree of concept drift when the current stream data has concept drift; and

    A training module, configured to perform online training on the online prediction model in combination with the current flow data and target flow data according to the degree value of the concept drift, so as to obtain a target prediction model; wherein, the target flow data is Sampled from multiple historical flow data.
11. The device according to claim 10, wherein the training module is specifically used for:

    If the value of the degree of concept drift is greater than or equal to the threshold value of the degree of concept drift, then refer to the set rule to sample the target flow data from the historical flow data stored in the playback buffer pool;

    The online prediction model is trained online according to the current flow data and the target flow data.
12. The device according to claim 10, wherein the training module is specifically used for:

    If the value of the degree of concept drift is less than the threshold value of the degree of concept drift, storing the current stream data in a playback buffer pool, and determining the data volume of stream data stored in the playback buffer pool;

    If the data volume does not meet the set condition, continuously acquire the current stream data, and dynamically update the data volume;

    If the amount of data satisfies the set condition, refer to the set rule to sample the target stream data from the historical stream data stored in the playback buffer pool, and Predictive models are trained online.
13. The device according to claim 11, wherein the training module is specifically used for:

    The online prediction model is trained online for a first set number of times according to the current stream data and the target stream data, wherein the first set number of times is a maximum value within a range of set times.
14. The device according to claim 12, wherein the training module is specifically used for:

    According to the target stream data, the online prediction model is trained online for a second set number of times, wherein the second set number of times is smaller than the first set number of times, and the second set number of times is based on The maximum value, the minimum value, the value of the degree of concept drift, and the threshold value of the degree of concept drift in the range of the set number of times are calculated.
15. The device according to any one of claims 10 to 14, wherein the device further comprises:

    A processing module, configured to store the current stream data in the playback buffer pool after training the target prediction model, and delete the first stream data in the playback buffer pool, wherein the first stream data corresponds to A storage order, prior to the storage order corresponding to other stream data, the first stream data and the other stream data together constitute the stream data stored in the playback buffer pool.
16. The device according to claim 11 or 12, wherein the training module is specifically used for:

    From the historical flow data stored in the playback buffer pool, a set amount of historical flow data is obtained based on selection probability sampling as the target flow data, the ratio between the set amount and the amount of the current flow data is the default value.
17. The device according to claim 16, wherein the selection probability is determined by the weight of the historical flow data, and the device further comprises:

    The second determination module is configured to determine the weight according to the data volume of the stream data stored in the playback buffer pool and the corresponding storage time of the historical stream data.
18. The device according to any one of claims 10 to 17, wherein,

    The acquiring module is further configured to acquire, after acquiring the current stream data, an annotation label corresponding to the current stream data, and acquire at least one prediction label supported by the online prediction model;

    The training module is specifically used for:

    If the label label belongs to the at least one prediction label, performing online training on the online prediction model in combination with the current flow data and target flow data according to the degree value of the concept drift;

    If the label label does not belong to the at least one prediction label, performing offline update training on the online prediction model in combination with the current flow data and the target flow data according to the degree value of the concept drift.
19. A computer device, characterized in that it comprises a memory, a processor, and a computer program stored on the memory and operable on the processor, when the processor executes the program, it realizes any one of claims 1 to 9. the method described.
20. A non-transitory computer-readable storage medium, when the instructions in the non-transitory computer-readable storage medium are executed by the processor of the electronic device, the electronic device can perform the operation described in any one of claims 1 to 9. described method.
21. A computer program product, characterized in that the computer program product includes computer program code, and when the computer program code is run on a computer, the method according to any one of claims 1 to 9 is realized.
22. A computer program, characterized in that the computer program includes computer program code, and when the computer program code is run on a computer, the computer is made to execute the method according to any one of claims 1 to 9.

Images