Business forecasting method and apparatus
By combining adversarial and non-adversarial prediction models, the problem of the impact of attack samples on business prediction is solved, achieving higher prediction accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ALIPAY (HANGZHOU) INFORMATION TECH CO LTD
- Filing Date
- 2022-06-28
- Publication Date
- 2026-06-16
AI Technical Summary
Existing business forecasting methods are susceptible to attack samples during model training and prediction, resulting in low accuracy of prediction results.
By using an adversarial prediction model that includes attack sample data and a non-adversarial prediction model that does not include attack sample data, business data is predicted, and the two prediction results are fused to improve prediction accuracy.
By combining the prediction results of both offensive and non-offensive approaches, we can fully leverage their respective advantages and improve the accuracy and robustness of business forecasts.
Smart Images
Figure CN115169666B_ABST
Abstract
Description
Technical Field
[0001] This specification relates to one or more embodiments in the field of computer technology, and more particularly to business forecasting methods and apparatus. Background Technology
[0002] Business forecasting is an important technique in the field of AI. It involves inputting sample data into a neural network to perform machine learning, obtaining a corresponding predictive model, and then making business predictions based on this model.
[0003] However, both model training and business prediction can be affected by attack samples, where even small perturbations can cause significant deviations in the model's predictions. If the training samples include attack samples, the resulting model will inevitably impact the accuracy of its predictions for normal samples. Conversely, if the training samples do not include attack samples, the resulting model will be unable to make good predictions for attack samples.
[0004] Therefore, current business forecasting methods are not very accurate. Summary of the Invention
[0005] This specification describes a business forecasting method and apparatus through one or more embodiments, which can improve the accuracy of business forecasting.
[0006] According to the first aspect, a business forecasting method is provided, including:
[0007] Extract business data from business events to predict business trends;
[0008] The business is predicted based on the first sample data and the business data to obtain a first business prediction result; wherein, the first sample data includes at least one attack sample data, the attack sample data being used to characterize data that can reduce the accuracy of the prediction result of the business prediction.
[0009] The business is predicted based on the second sample data and the business data to obtain a second business prediction result; wherein, the second sample data does not include attack sample data;
[0010] The first business forecast result and the second business forecast result are combined to obtain the final business forecast result for the business.
[0011] In one possible implementation, the step of predicting the business based on the first sample data and the business data to obtain a first business prediction result includes:
[0012] An adversarial prediction model is pre-trained using first sample data; wherein the first sample data for training the adversarial prediction model includes at least one adversarial training sample set, and each adversarial training sample set includes an attack sample data and a preset label value.
[0013] The business data is input into the adversarial prediction model, and the first business prediction result is output.
[0014] In one possible implementation, the step of predicting the business based on the second sample data and the business data to obtain a second business prediction result includes:
[0015] A non-adversarial prediction model is trained in advance using second sample data;
[0016] The business data is input into the non-adversarial prediction model, and the second business prediction result is output.
[0017] In one possible implementation, fusing the first business prediction result and the second business prediction result to obtain the final business prediction result includes:
[0018] The probability value for determining that the business data is attack sample data;
[0019] The probability values are used to fuse the first business prediction result and the second business prediction result to obtain the final business prediction result for the business.
[0020] One possible implementation further includes:
[0021] An attack sample identification model is pre-trained; wherein each sample training set used to train the attack sample identification model includes: a sample data and a label value representing whether the sample data is an attack sample data;
[0022] The probability value for determining that the business data is the attack sample data includes:
[0023] The business data is input into the attack sample identification model, and the probability value of the business data being attack sample data is output.
[0024] In one possible implementation, fusing the first business prediction result and the second business prediction result using the probability value to obtain the final business prediction result includes:
[0025] The first corrected prediction result is obtained by multiplying the probability value of the business data being attack sample data with the first business prediction result.
[0026] The second corrected prediction result is obtained by multiplying the difference between the probability value of 1 and the business data being attack sample data by the second business prediction result.
[0027] The sum of the first corrected prediction result and the second corrected prediction result is calculated to obtain the final business prediction result for the service.
[0028] According to the second aspect, a business forecasting device is provided, comprising: a business data acquisition module, a first forecasting result determination module, a second forecasting result determination module, and a business forecasting result determination module;
[0029] The business data acquisition module is configured to acquire business data for predicting business from business events;
[0030] The first prediction result determination module is configured to predict the business based on the first sample data and the business data obtained by the business data acquisition module, and obtain a first business prediction result; wherein, the first sample data includes at least one attack sample data, and the attack sample data is used to characterize data that can reduce the accuracy of the prediction result of the business prediction.
[0031] The second prediction result determination module is configured to predict the service based on the second sample data and the service data obtained by the service data acquisition module, and obtain a second service prediction result; wherein, the second sample data does not include attack sample data;
[0032] The business prediction result determination module is configured to merge the first business prediction result obtained by the first prediction result determination module and the second business prediction result obtained by the second prediction result determination module to obtain the final business prediction result of the business.
[0033] In one possible implementation, when the first prediction result determination module obtains a first business prediction result by predicting the business based on the first sample data and the business data, it is configured to perform the following operations:
[0034] An adversarial prediction model is pre-trained using first sample data; wherein the first sample data for training the adversarial prediction model includes at least one adversarial training sample set, and each adversarial training sample set includes an attack sample data and a preset label value.
[0035] The business data is input into the adversarial prediction model, and the first business prediction result is output.
[0036] According to a third aspect, a computing device is provided, comprising: a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method described in any one of the first aspects above.
[0037] According to a fourth aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in any one of the first aspects.
[0038] According to the method and apparatus provided in the embodiments of this specification, when performing business forecasting, business data for forecasting is first obtained from business events. Then, business forecasts are performed based on first sample data, second sample data, and the business data, respectively. Finally, the first business forecast result obtained based on the first sample data and the second business forecast result obtained based on the second sample data are fused to obtain the final business forecast result. Since the first sample data includes attack sample data, while the second sample data does not, the first business forecast result is attack-resistant, while the second business forecast result is not. Thus, by fusing the first and second business forecast results to determine the final business forecast result, the advantages of both attack-resistant and non-attack-resistant results can be fully considered, thereby improving the accuracy of business forecasting. Attached Figure Description
[0039] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0040] Figure 1 This is a flowchart of a business forecasting method provided in one embodiment of this specification;
[0041] Figure 2 This is a flowchart of a method for determining a first business forecast result provided in one embodiment of this specification;
[0042] Figure 3 This is a flowchart of a method for determining a second business forecast result provided in one embodiment of this specification;
[0043] Figure 4 This is a flowchart of another business forecasting method provided in one embodiment of this specification;
[0044] Figure 5This is a flowchart of yet another business forecasting method provided in one embodiment of this specification;
[0045] Figure 6 This is a flowchart of another business forecasting method provided in one embodiment of this specification;
[0046] Figure 7 This is a schematic diagram of a business forecasting device provided in one embodiment of this specification. Detailed Implementation
[0047] In the fields of data mining and artificial intelligence, we typically input feature data into pre-trained models for business prediction; this input feature data is called a sample. Among these sample data, there are some samples where even a small perturbation under certain constraints can cause the model to make incorrect predictions; these are called attack samples. For example, in facial recognition, wearing glasses with a special pattern can bypass the facial recognition model, causing incorrect facial recognition results; such an image is an attack sample.
[0048] The presence of attack samples affects both the robustness and accuracy of the model's business predictions. As shown in Table 1, when the training model does not include attack samples, the resulting normal business model performs well on normal test samples but performs poorly on attack test samples. Conversely, when the training model includes attack samples, the resulting adversarial business model improves its prediction performance on attack test samples but significantly reduces its prediction performance on normal test samples. Therefore, a business prediction scheme that balances robustness and accuracy is needed.
[0049] Table 1
[0050] Normal test sample Attack test samples Normal business model 95.52％ 0％ Adversarial business model 87.65％ 54.67％
[0051] Therefore, this solution considers integrating adversarial and non-adversarial business forecasting results. This approach can fully leverage the advantages of both adversarial and non-adversarial results, thereby improving the accuracy of business forecasting.
[0052] like Figure 1 As shown in the embodiments of this specification, a business forecasting method is provided, which may include the following steps:
[0053] Step 101: Obtain business data from business events to predict business trends;
[0054] Step 103: Based on the first sample data and the business data, predict the business to obtain a first business prediction result; wherein, the first sample data includes at least one attack sample data, which is used to characterize data that can reduce the accuracy of the prediction result of the business prediction;
[0055] Step 105: Based on the second sample data and the business data, predict the business to obtain the second business prediction result; wherein, the second sample data does not include attack sample data;
[0056] Step 107: Merge the first business prediction result and the second business prediction result to obtain the final business prediction result for the business.
[0057] In this embodiment, when performing business forecasting, business data to be forecasted is first obtained from business events. Then, business forecasts are performed based on first sample data, second sample data, and the business data, respectively. Finally, the first business forecast result obtained based on the first sample data and the second business forecast result obtained based on the second sample data are merged to obtain the final business forecast result. Since the first sample data includes attack sample data, while the second sample data does not, the first business forecast result is attack-resistant, while the second business forecast result is not. Thus, by fusing the first and second business forecast results to determine the final business forecast result, the advantages of both attack-resistant and non-attack-resistant results can be fully considered, thereby improving the accuracy of business forecasting.
[0058] The following describes the appendix in conjunction with specific embodiments. Figure 1 The steps in the process will be explained.
[0059] First, in step 101, business data for predicting business needs is obtained from business events.
[0060] Business forecasting can include image recognition, risk account prediction, risk transaction prediction, and environmental condition prediction. For image recognition, a business event could be account verification via facial recognition; for risk account and risk transaction prediction, the corresponding business events could be risk prevention and control events; and for environmental condition prediction, the corresponding business events could be environmental and climate observation events. Business data may or may not include offensive data.
[0061] For example, when the business prediction task is facial recognition, the user's facial images obtained through methods such as taking pictures and image acquisition are the business data. If the user in the acquired facial image is wearing glasses with a special pattern, this special pattern can bypass the facial recognition model and adversely affect facial recognition. Therefore, this facial image is considered malicious business data. Conversely, if the user in the facial image is not wearing any objects on their face, and there is no significant difference from the historical data acquisition, then such a facial image is considered normal business data.
[0062] Then, in step 103, the business is predicted based on the first sample data and the business data to obtain the first business prediction result.
[0063] This step considers the robustness of business forecasting, specifically the scenario where the business data consists of attack sample data, thereby improving the accuracy of business forecasting based on attack sample data. Therefore, when making business forecasts based on the first sample data, this first sample data includes at least one attack sample data. This attack sample data is data capable of reducing the accuracy of the forecast results.
[0064] For example, the image of a person wearing special patterned glasses in the description of step 101.
[0065] For example, in the field of risk control, if a risky account executes multiple legitimate transactions over a period of time after committing an illegal transaction in order to cover up the previous illegal transaction, then the transaction data of the risky account during this period is considered attack sample data.
[0066] In one possible implementation, such as Figure 2 As shown, step 103, when predicting the service based on the first sample data and the service data to obtain the first service prediction result, can be achieved through the following steps:
[0067] Step 201: Train an adversarial prediction model in advance using the first sample data; wherein, the first sample data for training the adversarial prediction model includes at least one adversarial training sample set, and each adversarial training sample set includes an attack sample data and a preset label value.
[0068] Step 203: Input the business data into the adversarial prediction model and output the first business prediction result.
[0069] In this embodiment, when making business predictions based on the first sample data, an adversarial prediction model is first trained using the first sample data. Then, this adversarial prediction model is used to predict the business data to obtain the first business prediction result. Since the first sample data used to train the adversarial prediction model includes attack sample data, the adversarial prediction model exhibits excellent robustness under attack sample testing. That is, even when the business data is an attack sample, the adversarial prediction model can still obtain a relatively accurate business prediction result.
[0070] In step 201, at least one adversarial training sample set is used to train the adversarial prediction model. Each adversarial training sample set includes an attack sample data and a preset label value. For example, if the attack sample data is an image of a user's face wearing a special pattern, then the preset label value can identify the user as user A.
[0071] It is easy to understand that the first sample data used to train the adversarial prediction model should not be entirely adversarial training sample sets, as this would severely affect the prediction results for non-attack samples.
[0072] Meanwhile, in step 105, the business is predicted based on the second sample data and the business data to obtain a second business prediction result.
[0073] This step considers the accuracy of business forecasting, specifically by fully taking into account the case where the business data is non-attack sample data, thereby improving the accuracy of forecasting based on non-attack sample data. Therefore, when making business forecasts based on the second sample data, this second sample data should not include attack sample data; that is, all second sample data should be non-attack sample data.
[0074] For example, in the field of facial recognition, this non-attack sample data is the facial images recorded or collected in various scenarios;
[0075] For example, in the field of risk control, this non-attack sample data is the user's normal transaction record data over a period of time;
[0076] For example, in the field of environmental and climate prediction, the non-aggressive sample data refers to data such as temperature, humidity, and air quality for the same time period in previous years, as well as for a period of time before the time to be predicted.
[0077] In one possible implementation, such as Figure 3 As shown, step 105, when predicting the business based on the second sample data and the business data to obtain the second business prediction result, can be achieved through the following steps:
[0078] Step 301: Train a non-adversarial prediction model in advance using the second sample data;
[0079] Step 303: Input the business data into the non-adversarial prediction model and output the second business prediction result.
[0080] In this embodiment, when making business predictions based on the second sample data, a non-adversarial prediction model is first trained using the second sample data. Then, the second business prediction result is obtained by using this non-adversarial prediction model to predict the business data. Since the second sample data used to train the non-adversarial prediction model does not include attack sample data, the non-adversarial prediction model can achieve excellent accuracy under non-attack sample testing.
[0081] In step 301, the second sample data used to train the non-adversarial prediction model can consist of several non-adversarial training sample sets. Each non-adversarial training sample set includes one non-adversarial sample data and one predicted label value. Furthermore, the second sample data used to train the non-adversarial prediction model is entirely composed of non-adversarial sample data; this ensures the accuracy of the non-adversarial prediction model in predicting non-attack business data.
[0082] It is easy to understand that both the first sample data and the second sample data are data collected from historical business events, and both the first sample data and the second sample data can be used to describe the business to be predicted.
[0083] Finally, in step 107, the first business prediction result and the second business prediction result are merged to obtain the final business prediction result of the business.
[0084] In this step, after obtaining the first business prediction result based on the attack sample data and the second business prediction result based on the non-attack sample data, the two business prediction results are fused to determine the final business prediction result. For example, different weights can be assigned to the first and second business prediction results to fully utilize the advantages of both business prediction results and obtain a prediction result that balances accuracy and robustness.
[0085] For example, in one possible implementation, such as Figure 4 As shown, step 107 can be achieved in the following way:
[0086] Step 401: Determine the probability value that the business data is attack sample data;
[0087] Step 403: Use the probability value to fuse the first business prediction result and the second business prediction result to obtain the final business prediction result of the business.
[0088] In this embodiment, when determining the final business prediction result of the business data, the probability value of the business data being attack sample data can first be determined. Then, based on the obtained probability value, the first business prediction result and the second business prediction result are fused to obtain the final business prediction result. Since the first business prediction result and the second business prediction result are obtained based on attack sample data and non-attack sample data, respectively, by determining the probability that the business data is attack sample data, the credibility of the two business prediction results can be determined. This provides a reasonable basis for fusing the first and second business prediction results, thereby making the determined final business prediction result more reliable.
[0089] Step 401 will be explained.
[0090] Step 401, in determining the probability that business data is attack sample data, can pre-train an attack sample identification model. Each training set used to train this model includes one sample data point and a label value representing whether the sample data is an attack sample data point. Then, by inputting the business data into the attack sample identification model, the probability that the business data is an attack sample data point can be output.
[0091] Step 403 will be explained.
[0092] After determining the probability value of the business data being attack sample data, step 401 can consider assigning weights to the first business prediction result and the second business prediction result based on the probability value, so as to achieve a balance between accuracy and robustness and further improve the overall accuracy of business prediction.
[0093] For example, such as Figure 5 As shown, step 403, in which the probability values are used to fuse the first business prediction result and the second business prediction result to obtain the final business prediction result, can be achieved through the following steps:
[0094] Step 501: Calculate the product of the probability value that the business data is the attack sample data and the first business prediction result to obtain the first corrected prediction result;
[0095] Step 503: Calculate the product of the difference between 1 and the probability value of the business data being attack sample data and the second business prediction result to obtain the second corrected prediction result;
[0096] Step 505: Calculate the sum of the first revised forecast result and the second revised forecast result to obtain the final business forecast result.
[0097] In this embodiment, when determining the final business prediction result based on the probability value, the first business prediction result, and the second business prediction result, the first corrected prediction result is obtained by first multiplying the probability value by the first business prediction result. Then, the difference between 1 and the probability value is calculated, and this difference is multiplied by the second business prediction result to obtain the second corrected prediction result. Finally, the final business prediction result is obtained by summing the first and second corrected prediction results. Therefore, this embodiment assigns weights to the two prediction results based on the probability of attack sample data using business data. This fully considers the advantages of both prediction results, thereby balancing the robustness and accuracy of the prediction results to achieve the overall improvement in prediction accuracy.
[0098] For example, the final business forecast result can be calculated using the following formula:
[0099] y = a*y1 + (1-a)*y2
[0100] Where y represents the final business prediction result, y1 represents the first business prediction result, y2 represents the second business prediction result, and a represents the probability that the business data is attack sample data, and 0≤a≤1.
[0101] The following section uses facial recognition as an example to illustrate the business prediction method provided in this manual.
[0102] like Figure 6 As shown, in the field of facial recognition, this business prediction method may include the following steps:
[0103] Step 601: Obtain historically collected facial image data to obtain the first sample data.
[0104] The first sample data includes not only non-attack sample data but also attack sample data. For example, the acquired first sample data includes several normal facial images, such as facial images of users without any accessories or facial images of users without special makeup. Moreover, the acquired first sample data also includes several abnormal facial images that could affect facial recognition of users, such as facial images of users who have been made up and are wearing glasses, headbands, jewelry, wigs, or other accessories.
[0105] Step 603: Train an adversarial prediction model using the first sample data.
[0106] This adversarial prediction model is used for face recognition. Because the first sample data included attack sample data during training, the model exhibits relatively strong robustness and can perform well in predicting attack sample data as well. For example, an adversarial prediction model trained in this way can achieve a relatively high recognition accuracy even for face images of people wearing special patterned glasses.
[0107] Step 605: Obtain the second sample data from the historically collected face image data.
[0108] The second sample data includes only non-attack sample data. Attack sample data can be filtered out from the face image data collected in step 601, and the resulting non-attack sample data can be determined as the second sample data.
[0109] Step 607: Train a non-adversarial prediction model using the second sample data.
[0110] Because the second sample data included only non-attack sample data during training, the model performs well when tested on non-attack samples. That is, it achieves high accuracy when recognizing normal face images.
[0111] Step 609: Train an attack sample identification model.
[0112] This attack sample identification model is used to identify whether a face image is an attack sample, that is, to determine whether the face image to be identified is one that will affect the identification result. For example, it determines that if the user is wearing glasses with a special pattern in the face image to be identified, then the probability of the face image to be identified as an attack sample in the output of the attack sample identification model will be higher.
[0113] Step 611: Obtain the face image of the person to be recognized.
[0114] The facial image in question is a captured image of user A's face.
[0115] Step 613: Input the acquired face image into the adversarial prediction model to obtain the first recognition result z1.
[0116] In this step, when the face image is input into the adversarial prediction model for recognition, the recognition result output by the adversarial prediction model is the rating of user A.
[0117] Step 615: Input the acquired face image into the non-adversarial prediction model to obtain the second recognition result z2.
[0118] In this step, when the face image is input into the non-adversarial prediction model for recognition, the recognition result output by the non-Dukang prediction model is the rating of user A.
[0119] Step 617: Input the acquired face image into the attack sample recognition model to obtain the probability k that the face image is an attack sample.
[0120] Step 619: Calculate the prediction result of the face image to be recognized based on the first recognition result z1, the second recognition result z2 and the probability k.
[0121] In this step, the following formula can be used to calculate the score for user A in the face image to be recognized:
[0122] Q = k * z1 + (1 - k) * z2
[0123] In this context, Q represents the rating of user A's face image.
[0124] Of course, it should be noted that the first identification result may include not only user A's rating but also ratings from other users. Similarly, the second identification result may also include ratings from other users. Therefore, in step 619 above, when calculating the rating, a rating can be assigned to each user in each identification result, thus obtaining the ratings for each identified user. Further, based on each user's rating, the highest-rated user is determined as the final identification result.
[0125] like Figure 7 As shown in the embodiments of this specification, a business forecasting device is also provided. The device may include: a business data acquisition module 701, a first forecast result determination module 702, a second forecast result determination module 703, and a business forecast result determination module 704.
[0126] The business data acquisition module 701 is configured to acquire business data for predicting business from business events;
[0127] The first prediction result determination module 702 is configured to predict the business based on the first sample data and the business data obtained by the business data acquisition module 701, and obtain the first business prediction result; wherein, the first sample data includes at least one attack sample data, which is used to characterize data that can reduce the accuracy of the prediction result of the business prediction.
[0128] The second prediction result determination module 703 is configured to predict the business based on the second sample data and the business data obtained by the business data acquisition module 701, and obtain the second business prediction result; wherein, the second sample data does not include attack sample data.
[0129] The business forecast result determination module 704 is configured to merge the first business forecast result obtained by the first forecast result determination module 702 and the second business forecast result obtained by the second forecast result determination module 703 to obtain the final business forecast result.
[0130] In one possible implementation, when the first prediction result determination module 702 obtains a first service prediction result by predicting the service based on the first sample data and the service data, it is configured to perform the following operations:
[0131] An adversarial prediction model is pre-trained using the first sample data; wherein the first sample data for training the adversarial prediction model includes at least one adversarial training sample set, and each adversarial training sample set includes an attack sample data and a preset label value.
[0132] The business data is input into the adversarial prediction model, and the first business prediction result is output.
[0133] In one possible implementation, when the second prediction result determination module 703 obtains a second service prediction result by predicting the service based on the second sample data and the service data, it is configured to perform the following operations:
[0134] A non-adversarial prediction model is trained in advance using second sample data;
[0135] The business data is input into the non-adversarial prediction model, and the second business prediction result is output.
[0136] In one possible implementation, when the business prediction result determination module 704 fuses the first business prediction result and the second business prediction result to obtain the final business prediction result of the business, it is configured to perform the following operation:
[0137] Determine the probability value that the business data is attack sample data;
[0138] The probability values are used to fuse the first business prediction result and the second business prediction result to obtain the final business prediction result for the business.
[0139] In one possible implementation, the business forecasting device further includes:
[0140] A model training module is configured to pre-train an attack sample identification model; wherein each sample training set used to train the attack sample identification model includes: a sample data and a label value representing whether the sample data is attack sample data.
[0141] When determining the probability value that business data is attack sample data, the business prediction result determination module 704 is configured to perform the following operation:
[0142] The business data is input into the attack sample identification model, and the output is the probability value that the business data is attack sample data.
[0143] In one possible implementation, when the business prediction result determination module 704 uses the probability value to fuse the first business prediction result and the second business prediction result to obtain the final business prediction result of the business, it is configured to perform the following operation:
[0144] The first corrected prediction result is obtained by multiplying the probability value of the business data being the attack sample data with the first business prediction result.
[0145] The second corrected prediction result is obtained by multiplying the difference between the probability value of 1 and the probability value of the business data being the attack sample data by the second business prediction result.
[0146] The sum of the first and second revised forecast results is calculated to obtain the final business forecast result for the business data.
[0147] This specification also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the methods in any of the embodiments of the specification.
[0148] This specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method in any of the embodiments of the specification.
[0149] It is understood that the structures illustrated in the embodiments of this specification do not constitute a specific limitation on the business forecasting apparatus. In other embodiments of this specification, the business forecasting apparatus may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.
[0150] The information interaction and execution process between the various units in the above-mentioned device are based on the same concept as the method embodiments in this specification, and the specific details can be found in the descriptions in the method embodiments in this specification, so they will not be repeated here.
[0151] Those skilled in the art will recognize that, in one or more of the examples above, the functions described herein can be implemented using hardware, software, widgets, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.
[0152] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects described in this specification. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.
Claims
1. Business forecasting methods, including: Extract business data from business events to predict business trends; The business is predicted based on the first sample data and the business data to obtain a first business prediction result; wherein, the first sample data includes at least one attack sample data, the attack sample data being used to characterize data that can reduce the accuracy of the prediction result of the business prediction. The business is predicted based on the second sample data and the business data to obtain a second business prediction result; wherein, the second sample data does not include attack sample data; The first business prediction result and the second business prediction result are merged to obtain the final business prediction result of the business; The step of predicting the business based on the first sample data and the business data to obtain the first business prediction result includes: An adversarial prediction model is pre-trained using first sample data; wherein the first sample data for training the adversarial prediction model includes at least one adversarial training sample set, and each adversarial training sample set includes an attack sample data and a preset label value. The business data is input into the adversarial prediction model, and the first business prediction result is output. The step of predicting the business based on the second sample data and the business data to obtain a second business prediction result includes: A non-adversarial prediction model is trained in advance using second sample data; The business data is input into the non-adversarial prediction model, and the second business prediction result is output. Wherein, the step of fusing the first business prediction result and the second business prediction result to obtain the final business prediction result includes: The probability value for determining that the business data is attack sample data; The probability values are used to fuse the first business prediction result and the second business prediction result to obtain the final business prediction result for the business.
2. The method according to claim 1, wherein, Further includes: An attack sample identification model is pre-trained; wherein each sample training set used to train the attack sample identification model includes: a sample data and a label value representing whether the sample data is an attack sample data; The probability value for determining that the business data is the attack sample data includes: The business data is input into the attack sample identification model, and the probability value of the business data being attack sample data is output.
3. The method according to claim 1, wherein, The step of fusing the first business prediction result and the second business prediction result using the probability value to obtain the final business prediction result includes: The first corrected prediction result is obtained by multiplying the probability value of the business data being attack sample data with the first business prediction result. The second corrected prediction result is obtained by multiplying the difference between the probability value of 1 and the business data being attack sample data by the second business prediction result. The sum of the first corrected prediction result and the second corrected prediction result is calculated to obtain the final business prediction result for the service.
4. A business forecasting device, including: The module includes a business data acquisition module, a first prediction result determination module, a second prediction result determination module, and a business prediction result determination module. The business data acquisition module is configured to acquire business data for predicting business from business events; The first prediction result determination module is configured to predict the business based on the first sample data and the business data obtained by the business data acquisition module, and obtain a first business prediction result; wherein, the first sample data includes at least one attack sample data, and the attack sample data is used to characterize data that can reduce the accuracy of the prediction result of the business prediction. The second prediction result determination module is configured to predict the service based on the second sample data and the service data obtained by the service data acquisition module, and obtain a second service prediction result; wherein, the second sample data does not include attack sample data; The business prediction result determination module is configured to fuse the first business prediction result obtained by the first prediction result determination module and the second business prediction result obtained by the second prediction result determination module to obtain the final business prediction result of the business. Wherein, when the first prediction result determination module predicts the service based on the first sample data and the service data to obtain the first service prediction result, it is configured to perform the following operations: An adversarial prediction model is pre-trained using first sample data; wherein the first sample data for training the adversarial prediction model includes at least one adversarial training sample set, and each adversarial training sample set includes an attack sample data and a preset label value. The business data is input into the adversarial prediction model, and the first business prediction result is output. Wherein, when the second prediction result determination module predicts the service based on the second sample data and the service data to obtain the second service prediction result, it is configured to perform the following operations: A non-adversarial prediction model is trained in advance using second sample data; The business data is input into the non-adversarial prediction model, and the second business prediction result is output. Specifically, when the business prediction result determination module merges the first business prediction result and the second business prediction result to obtain the final business prediction result, it is configured to perform the following operations: Determine the probability value that the business data is attack sample data; The probability values are used to fuse the first business prediction result and the second business prediction result to obtain the final business prediction result for the business.
5. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-3.
6. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-3.