A processing method and computer device

Abstract

The processing method computer equipment involved in this application provides a monitoring environment for the artificial intelligence model, from which the monitoring data of the predetermined monitoring indicators of the artificial intelligence model at different times are obtained, and based on these monitoring data, the changes in the model monitoring indicators at different times are determined, and finally whether the artificial intelligence model needs to be updated is determined by whether the changes in the monitoring indicators at the different times meet the change conditions. This application analyzes the predictive ability and rationality of the underlying prediction function of the model by monitoring the changes in the predetermined monitoring indicators of the model at different times, and determines whether the artificial intelligence model needs to be updated (characterizing whether the model's predictive ability meets expectations) by further determining whether the change meets the change conditions, thereby determining the appropriate update time for the artificial intelligence model, thereby providing the model with an intelligent, appropriate-time-based feedback loop, which can effectively guarantee the model's processing performance in the future and ensure its credibility.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more particularly to a processing method and computer equipment. Background Art

[0002] With the development of science and technology, we have entered a new era of artificial intelligence (AI) with technologies such as machine learning (ML) as its core. Artificial intelligence with technologies such as machine learning as its core can be applied to many fields such as science and technology, finance, law, medicine, and military.

[0003] Among them, it is necessary to carry out relevant intelligent applications in the required fields based on the constructed artificial intelligence model. However, as time goes by, such as after a few days, weeks, months or years, the input data of the model may change significantly, resulting in the model not being able to adapt well to these input data, which may make the model a model with poor performance and untrustworthy in the future. Summary of the Invention

[0004] In view of this, the present application provides a processing method and computer device for determining the update timing of the artificial intelligence model, so that the update of the artificial intelligence model can be triggered at the appropriate time to ensure the model's future processing performance and its credibility.

[0005] To this end, this application discloses the following technical solutions:

[0006] A processing method comprising:

[0007] Obtain monitoring data of predetermined monitoring indicators of the artificial intelligence model at different times;

[0008] Determining, based on the monitoring data, changes in the predetermined monitoring indicator at the different times;

[0009] Determining whether changes in the predetermined monitoring indicator at the different times meet a change condition, and obtaining a determination result;

[0010] Based on the determination result, it is determined whether the artificial intelligence model needs to be updated, and a first output result is generated; the first output result can be used to indicate whether the artificial intelligence model needs to be updated.

[0011] In the above method, preferably, obtaining monitoring data of predetermined monitoring indicators of the artificial intelligence model at different times includes at least one of the following:

[0012] Obtaining first monitoring data of important features of the artificial intelligence model at different times; the important features include: features included in the input data of the artificial intelligence model, the impact of which on the model prediction result meets the impact condition;

[0013] Obtaining second monitoring data of the model performance of the artificial intelligence model at different times;

[0014] Obtain third monitoring data of the analog-to-digital input of the artificial intelligence model at different times.

[0015] The above method is preferably wherein:

[0016] The second monitoring data includes the signature rate data of the artificial intelligence model; the signature rate of the model is: the ratio of the number of results of different polarities in the different prediction results of the model for different input data, where the polarity is the polarity of the prediction result pre-set for the input data;

[0017] The third monitoring data includes: feature values ​​of features included in the input data of the artificial intelligence model.

[0018] In the above method, preferably, determining whether the change of the predetermined monitoring indicator generated at the different times satisfies the change condition and obtaining the determination result includes at least one of the following:

[0019] determining, based on the first monitoring data, whether changes in the important feature generated at the different times satisfy a first change condition, and obtaining a first sub-result;

[0020] determining, based on the second monitoring data, whether changes in the model performance at the different times satisfy a second change condition, and obtaining a second sub-result;

[0021] determining, based on the third monitoring data, whether changes in the model input generated at the different times satisfy a third change condition, and obtaining a third sub-result;

[0022] The determining whether the artificial intelligence model needs to be updated and generating a first output result includes:

[0023] Based on at least one of the first sub-result, the second sub-result, and the third sub-result, determine whether the artificial intelligence model needs to be updated, and generate a first output result.

[0024] The above method preferably further comprises:

[0025] Determine whether there is a target feature among the features of the model input data that does not meet the association condition with the prediction result of the artificial intelligence model; if so, generate a second output result including the target feature; the second output result can be used to indicate that a model update is required based on the target feature;

[0026] and / or,

[0027] According to different associations between different features and the prediction results of the artificial intelligence model, model prediction rules are generated so that the artificial intelligence model is updated based on the model prediction rules.

[0028] A computer device comprising:

[0029] a memory for storing at least one set of instruction sets;

[0030] A processor is configured to call and execute the instruction set in the memory, and perform the following operations by executing the instruction set:

[0031] Obtain monitoring data of predetermined monitoring indicators of the artificial intelligence model at different times;

[0032] Determining, based on the monitoring data, changes in the predetermined monitoring indicator at the different times;

[0033] Determining whether changes in the predetermined monitoring indicator at the different times meet a change condition, and obtaining a determination result;

[0034] Based on the determination result, it is determined whether the artificial intelligence model needs to be updated, and a first output result is generated; the first output result can be used to indicate whether the artificial intelligence model needs to be updated.

[0035] In the above-mentioned computer device, preferably, the processor obtains monitoring data of predetermined monitoring indicators of the artificial intelligence model at different times, including at least one of the following:

[0036] Obtaining first monitoring data of important features of the artificial intelligence model at different times; the important features include: features included in the input data of the artificial intelligence model, the impact of which on the model prediction result meets the impact condition;

[0037] Obtaining second monitoring data of the model performance of the artificial intelligence model at different times;

[0038] Obtain third monitoring data of the analog-to-digital input of the artificial intelligence model at different times.

[0039] The above computer device is preferably:

[0040] The second monitoring data includes the signature rate data of the artificial intelligence model; the signature rate of the model is: the ratio of the number of results of different polarities in the different prediction results of the model for different input data, where the polarity is the polarity of the prediction result pre-set for the input data;

[0041] The third monitoring data includes: feature values ​​of features included in the input data of the artificial intelligence model.

[0042] In the above-mentioned computer device, preferably, the processor determines whether the change of the predetermined monitoring indicator generated at the different times satisfies the change condition, and obtains a determination result, including at least one of the following:

[0043] determining, based on the first monitoring data, whether changes in the important feature generated at the different times satisfy a first change condition, and obtaining a first sub-result;

[0044] determining, based on the second monitoring data, whether changes in the model performance at the different times satisfy a second change condition, and obtaining a second sub-result;

[0045] determining, based on the third monitoring data, whether changes in the model input generated at the different times satisfy a third change condition, and obtaining a third sub-result;

[0046] The processor determines whether the artificial intelligence model needs to be updated and generates a first output result, including:

[0047] Based on at least one of the first sub-result, the second sub-result, and the third sub-result, determine whether the artificial intelligence model needs to be updated, and generate a first output result.

[0048] In the above-mentioned computer device, preferably, the processor is further configured to:

[0049] Determine whether there is a target feature among the features of the model input data that does not meet the association condition with the prediction result of the artificial intelligence model; if so, generate a second output result including the target feature; the second output result can be used to indicate that a model update is required based on the target feature;

[0050] and / or,

[0051] According to different associations between different features and the prediction results of the artificial intelligence model, model prediction rules are generated so that the artificial intelligence model is updated based on the model prediction rules.

[0052] It can be seen from the above scheme that the processing method disclosed in this application provides a monitoring environment for the artificial intelligence model, from which the monitoring data of the predetermined monitoring indicators of the artificial intelligence model at different times are obtained, and based on these monitoring data, the changes in the predetermined monitoring indicators of the model at different times are determined, and finally whether the changes in the predetermined monitoring indicators at the different times meet the change conditions is used to determine whether the artificial intelligence model needs to be updated. This application analyzes the predictive ability and rationality of the underlying prediction function of the model by monitoring the changes in the predetermined monitoring indicators of the model at different times, and determines whether the artificial intelligence model needs to be updated (characterizing whether the model's predictive ability meets expectations) by further determining whether the change meets the change conditions, thereby determining the appropriate update time for the artificial intelligence model, thereby providing the model with an intelligent, appropriate-time-based feedback loop, which can effectively guarantee the model's future processing performance and ensure its credibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0054] Figure 1 This is a flow chart of the processing method provided in the embodiment of the present application;

[0055] Figure 2 This is another flow chart of the processing method provided in the embodiment of the present application;

[0056] Figure 3 is a schematic diagram of a curve of feature importance / importance level provided in an embodiment of the present application;

[0057] Figure 4 This is a schematic diagram of a marking rate curve provided in an embodiment of the present application;

[0058] Figure 5 This is a schematic diagram of the processing logic for determining the model update timing based on model monitoring provided by an embodiment of the present application;

[0059] Figure 6 This is another flow chart of the processing method provided in the embodiment of the present application;

[0060] Figure 7 This is another flow chart of the processing method provided in the embodiment of the present application;

[0061] Figure 8This is an example diagram of an application of the processing method of the present application for determining the timing of model update and processing model update provided by an embodiment of the present application;

[0062] Figure 9 This is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0064] The embodiments of the present application disclose a processing method and computer device, which provides a monitoring environment for artificial intelligence models to monitor some important indicators related to the model, such as model characteristics (important features of the model), model performance and model input (feature values ​​of input data), etc. Based on the monitoring data transmitted by the monitoring environment, the underlying decision-making ability and model behavior of the model will be analyzed / interpreted to determine the rationality of the potential function of the model when using it for prediction, and accordingly determine the update timing of the model. The processing method and computer device of the present application will be described in detail below through specific embodiments.

[0065] In an optional embodiment of the present application, a processing method is disclosed, which can be applied and run in computer devices such as portable computers (such as notebooks), desktop computers or large and medium-sized computers, background servers or cloud platform servers in general / special computing or configuration environments.

[0066] like Figure 1 As shown, the processing method disclosed in this embodiment may include the following processing procedures:

[0067] Step 101: Obtain monitoring data of predetermined monitoring indicators of the artificial intelligence model at different times.

[0068] In this embodiment, the artificial intelligence model refers to an ML model built based on the core technology of artificial intelligence - machine learning, such as a linear classification model, a support vector machine model, a deep learning model, etc.

[0069] The predetermined monitoring indicators of the artificial intelligence model are selected from some important indicators related to the artificial intelligence model that can directly or indirectly reflect the underlying behavioral capabilities of the model, which may include but are not limited to indicators in three aspects: model characteristics (important features of the model), model performance and model input (characteristic values ​​of input data).

[0070] In specific implementation, a monitoring environment can be built for the model to be tested, and the model indicator data can be monitored and output in the monitoring environment. Based on the monitoring environment, the following aspects of model data can be monitored, but not limited to:

[0071] 1) Important features of the monitoring model;

[0072] Here, model features specifically refer to the features of the model's input data; model input data usually has multi-dimensional features, and each dimension of features contributes more or less to the decision-making results in the model's prediction decision; for example, assuming that the model's input data is the personal information of website users, then in this embodiment, the model features (that is, the features of the model's input data) can be gender, age, email address, hobbies, etc., and the characteristic value of each feature may affect the decision-making results of the model prediction process to a greater or lesser extent.

[0073] In the embodiment of the present application, the monitoring of model features mainly focuses on features included in the input data of the model that can have an impact on the model prediction results that meets the influencing conditions, and these features are used as important features of the model.

[0074] The influencing condition may be, by way of example and not limitation, that the importance or level of the feature reaches a set importance threshold or level threshold (for example, mapping the features into multiple levels according to their influence on the model decision result, with different levels representing different degrees of influence on the model decision result); or, the influencing condition may also be that the feature belongs to the top N features after ranking the features in descending order of importance / level, where N is a non-zero natural number.

[0075] 2) Monitor model performance;

[0076] Specifically, one or more indicators of the model that can be used to reflect the model performance can be monitored, such as the model's prediction accuracy, precision, model recall rate, etc.

[0077] 3) Monitor model input.

[0078] The model input here refers to the characteristic values ​​of the model input data, such as the actual values ​​of gender, age, email address, and hobbies in the personal information of website users.

[0079] Step 102: Based on the monitoring data, determine the changes of the predetermined monitoring indicators at different times.

[0080] For some appropriately selected monitoring indicators, the data changes and fluctuations that occur can usually directly or indirectly reflect the changes in the underlying decision-making ability and model behavior of the model, which in turn enables the rationality of the model's prediction function to be explained. Therefore, this embodiment analyzes the changes in the indicator data of the above three selected monitoring indicators at different times to analyze whether the rationality of the model's potential prediction function can meet expectations and whether the model needs to be updated.

[0081] The different times may be different times between a certain historical moment and the current moment determined based on a fixed set duration. For example, assuming that the fixed set duration is one month, three months, or one year, the different times may be different times between the previous month / previous three months / previous year and the current moment. In this embodiment, the different times may be determined based on a sliding window method (with a fixed sliding step, such as the above-mentioned one month, three months, or one year) with the current time as the end of the time.

[0082] Alternatively, in other embodiments, the different times may also be different times covered by the time period from a set fixed starting moment to the current moment. For example, assuming that the fixed starting moment is a moment half a year ago, the different times are the time between the moment half a year ago and the current moment. The time span of the different times will gradually increase with the passage of time, such as increasing to 7 months, 8 months, etc., and then the starting moment can be updated according to the strategy when it increases to the upper limit of the time span (such as one year). The method of determining the different times is not limited here.

[0083] During implementation, the output of the monitoring environment at different times can be used as data to analyze the important characteristics of the model, model performance, and changes in model input at different times.

[0084] Step 103: Determine whether the changes of the predetermined monitoring indicators generated at the different times meet the change conditions, and obtain a determination result.

[0085] Afterwards, we can further determine whether the changes in the model's important features, model performance, and model inputs at different times meet the change conditions. If so, it is considered that the model's underlying decision-making ability and model behavior have undergone major changes, and accordingly, it is considered that the rationality of its underlying prediction function cannot be guaranteed. If the change conditions are not met, it is considered that the rationality of the model's underlying prediction function can be guaranteed.

[0086] The change condition may be configured as follows: the change in the indicator exceeds a set maximum allowable change, where the maximum allowable change may be a change in the model indicator data mapped based on a maximum change tolerance of the model's decision-making capability.

[0087] In implementation, the specific value of the "maximum allowable variation" in this condition can be determined based on the actual application environment of the model project, the accuracy / precision requirements of the model, and combined with experience.

[0088] Step 104: Based on the determination result, determine whether the artificial intelligence model needs to be updated, and generate a first output result; the first output result can be used to indicate whether the artificial intelligence model needs to be updated.

[0089] By making conditional judgments on the changes of monitoring indicators at different times based on the set change conditions, a judgment result (i.e. the above-mentioned determination result) can be obtained, which at least indicates whether the changes of monitoring indicators at different times meet the change conditions or do not meet the change conditions.

[0090] The determination result of whether the change condition is met or not will trigger the generation and display of an output result, namely the first output result. The first output result can at least be used to indicate whether the artificial intelligence model needs to be updated.

[0091] Among them, if the monitoring indicator meets the change condition, it means that the underlying decision-making ability and model behavior of the model have undergone major changes, and the model capability and model credibility cannot be guaranteed. Therefore, a first output result is generated and displayed that can be used to indicate that the model needs to be updated. Subsequently, the computer device that provides the model maintenance function can automatically start the model update process by responding to the first output result, or the model update process can be manually started on the computer device under manual intervention; on the contrary, if the monitoring indicator does not meet the change condition, it means that the underlying decision-making ability and model behavior of the model have not undergone major changes, and the model capability and model credibility can be guaranteed. Therefore, a first output result is generated and displayed that can be used to indicate that the model does not need to be updated. Of course, no information can be output, implying that the model does not need to be updated.

[0092] It should be noted that in actual applications, there may be situations where the machine cannot clearly determine whether a model update is needed. For example, the absolute difference between the change in the monitoring indicator and the set maximum allowable change is less than the set threshold, resulting in increased difficulty in judgment, or the changes in some indicators of different indicators meet the change conditions, while other indicators do not meet them. In this case, it means that it is time to discuss with human intelligence. The machine can trigger the communication process with human intelligence and finally determine whether a model update is needed under the intervention of human intelligence. Whether the model clearly gives the monitoring and analysis results of whether a model update is needed, or the model triggers the discussion process with human intelligence and determines whether the model needs to be updated based on the intervention of human intelligence, it is within the scope of protection of this application.

[0093] This embodiment analyzes the predictive ability and rationality of the underlying prediction function of the model by monitoring the changes in the predetermined monitoring indicators of the model at different times, and determines whether the artificial intelligence model needs to be updated (characterizing whether the model's predictive ability meets expectations) by further determining whether the change meets the change conditions. In this way, the appropriate time to update the artificial intelligence model is determined, thereby providing the model with an intelligent feedback loop based on the right time, which can effectively ensure the model's processing performance in the future and ensure its credibility.

[0094] The following details the specific implementation process of monitoring and analyzing model data for the above three monitoring indicators to determine the rationality of the model prediction function and then determine whether the model needs to be updated (and the corresponding model update timing).

[0095] See Figure 2 In this embodiment, based on the above three monitoring indicators, the processing method can determine whether the model needs to be updated (i.e., the update timing of the model) through the following processing process:

[0096] Step 201: Obtain first monitoring data of important features of the artificial intelligence model at different times, second monitoring data of model performance at different times, and third monitoring data of analog-to-digital input at different times.

[0097] For monitoring important features of the model, optionally, this embodiment uses LIME to detect important features of the model at each prediction. LIME is an explanation tool for machine learning models that can be used to explain the predictions of any machine learning model. It uses the ML-based system input as the model input and observes how the prediction changes by perturbing the input. The output of the ML model is explained based on the mapping between the input perturbation and the model prediction change. In monitoring important features of the model, the LIME tool can be used to obtain important features of the model (i.e., features that can cause greater changes in model predictions when perturbed) by perturbing the input features and observing the changes in predictions brought about by the perturbation, and at the same time obtain feature importance monitoring data. To facilitate observation and analysis, this embodiment uses fusion technology to combine all features of all inputs of the model, and ultimately generates a single ranked list of important features in each prediction (e.g., sorting features in descending order of importance / importance level).

[0098] When monitoring model performance, due to the high randomness of the model's daily accuracy, precision, or recall rate data, it is difficult to objectively and accurately obtain the model performance based on these data. Correspondingly, it is difficult to analyze the changes in model performance based on the changes in these data. Therefore, this application proposes to use the model's flagrate to reflect the model performance, and to reflect the changes in the model performance by the changes in the model flagrate.

[0099] The mark rate of the model refers to the ratio of the number of results of different polarities in the different prediction results of the model for different input data.

[0100] The polarity is the polarity of the prediction result set in advance for the input data. The polarity of the prediction result can be divided in different ways, such as according to different division methods, the prediction result can be divided into positive polarity, negative polarity, good polarity, bad polarity, or 0 polarity, etc.

[0101] Assume that the prediction results of the model correspond to three categories: category 1, category 2, and category 3. One or more of the three categories can be classified as positive polarity (or other classification methods, such as "good", "0", etc.), and the other categories can be classified as negative polarity (or other classification methods, such as "bad", "1", etc.). More specifically, for example, category 1 and category 3 can be classified as positive polarity, while category 2 can be classified as negative polarity, etc. The classification of the polarity of the prediction results can be flexibly set by technical personnel and is not restricted here.

[0102] For a specific model operation scenario, when the model marking rate changes, this embodiment considers that the model performance has changed.

[0103] In addition, the applicant believes that any significant changes in the model input data may result in the model's existing underlying decision-making capabilities no longer being able to adapt well to the input data after the significant changes. Therefore, any significant changes in the model input data should be studied to determine whether the model needs to be updated or whether discussions with human intelligence are needed.

[0104] For monitoring of model input, the feature values ​​of the model input data can be detected in real time. For example, for a website's intelligent recommendation system or intelligent question-answering system, when user personal information is used as system input or as part of the system input to participate in system decision-making, the personal information of users using the system can be monitored in real time, and the focus is on discovering major changes in system user information, such as a change from the majority of system users being "professional one" with a certain technical background to the majority being "professional two" without a technical background, or a change from the majority of system users being female to the majority being male, etc. (these can all be regarded as major changes in input data).

[0105] Step 202: Based on the first monitoring data, determine the changes of the important features at the different times; based on the second monitoring data, determine the changes of the model performance at the different times; based on the third monitoring data, determine the changes of the model input at the different times.

[0106] Among them, the changes in important features of the model can be determined based on the feature importance monitoring data of each prediction of the model obtained by using LIME. As mentioned above, in order to facilitate observation and analysis, this embodiment uses fusion technology to combine all features of all inputs of the model and generate a single-ranked list of important features in each prediction. Therefore, based on the single-ranked list of important features in each prediction, it can be determined whether the important features of the model have changed and the degree of change.

[0107] More specifically, this embodiment selects the top N features (i.e., TOP N important features) from each single-rank list, and generates the importance / importance level curve of the top N features in 2-Dim (two-dimensional) space, such as Figure 3 As shown in the figure, each feature has a curve based on the function fr_i(t), where i represents the different features, t represents time (for example, day / week / month) on the x-axis (horizontal axis), and fr_i(t) represents the importance / level of the feature on the y-axis (vertical axis). As the feature importance / level changes during model operation, the curve will also change in height. By observing and analyzing the changes in the feature importance / level curve, we can understand the changes in important features.

[0108] Similarly, a marking rate curve is drawn for the model marking rate, such as Figure 4 An example of a landmark rate curve is provided. The X-axis (horizontal axis) represents time (e.g., day / week / month), and the Y-axis (vertical axis) represents the landmark rate. By observing the landmark rate curve, changes in the model landmark rate can be analyzed, and changes in the model landmark rate reflect changes in model performance. Furthermore, an eigenvalue curve is plotted for the monitored model input data. By observing and analyzing the changes in the eigenvalue curve, changes in the model input can be understood.

[0109] Therefore, model-based monitoring can ultimately output three curves: feature importance curve, mark rate curve, and feature value curve. The significant changes in each curve need to be explained to human intelligence and the possible reasons analyzed, such as Figure 3 There is a significant change in the curve between time 8 and 9. The percentage of the significant change can be calculated (such as how many characteristic curves have undergone important level changes, the proportion of the changed curves to all curves, etc.) to decide whether to change the model.

[0110] Step 203: Determine whether the changes of the important features generated at the different times meet a first change condition, and obtain a first sub-result.

[0111] The first change condition is a condition that can be used to indicate that an important feature of the model has changed significantly.

[0112] An example of this embodiment is provided below. Figure 3 The first condition can be set as follows: at least P = A% of the characteristic curves of each important characteristic intersect with each other, and the intersecting state can be maintained for a set period of time (such as T = 1 month), where A is a set constant, such as 10 or 20. Each intersection of different characteristic curves indicates a significant change in the importance / importance level of different characteristics, from the original importance / importance level of characteristic one being higher / lower than characteristic two to the importance / importance level of characteristic one being lower / higher than characteristic two.

[0113] Afterwards, the changes in the important features at different times can be judged based on the set first change condition to determine whether the important features have changed significantly at different times, and a first sub-result can be obtained. The first sub-result can be but is not limited to a binary result of "0" or "1", where "1" indicates that the change in the important features meets the first change condition, and "0" indicates that the change in the important features does not meet the first change condition.

[0114] Step 204: Determine whether the changes in the model performance at the different times meet a second change condition, and obtain a second sub-result.

[0115] The second change condition is a condition that can be used to indicate that the model performance has changed significantly. Specifically, the change in the mark rate is used to reflect the change in the model performance.

[0116] Optionally, the second change condition may be set to any one or more of the following conditions:

[0117] 1) The mark rate range rises by at least p = B1% of the average value of the maximum mark rate, or the mark rate range falls by at least p = B2% of the average value of the minimum mark rate; and the state after the rise or fall is maintained for a set period of time;

[0118] B1 and B2 are fixed values, such as 10, 20, etc.

[0119] See Figure 4 , the average value of the maximum value of the marking rate is Figure 4 The average value of each peak value of the sawtooth mark rate curve and the average value of the minimum mark rate are Figure 4 The average value of each trough value of the jagged mark rate curve.

[0120] 2) The minimum / maximum average performance (average of the maximum mark rate / average of the minimum mark rate) continues to decrease (downward trend) and is maintained for a set period of time;

[0121] 3) The minimum / maximum average performance (average of the maximum mark rate / average of the minimum mark rate) continues to rise (upward trend) and is maintained for a set period of time;

[0122] 4) N1-day momentum M1 ≥ ±C1% and maintained for a set period of time;

[0123] For example, the 10-day momentum M1 ≥ ± 70%, etc. Where M1 = ((p-p_10) / p_10) * 100, p represents the mark rate of the model at the current time, and p_10 represents the mark rate of the model N1 days ago.

[0124] If at least one of the above conditions is met, a second sub-result of "1" is obtained; otherwise, a second sub-result of "0" is obtained. Here, "1" indicates that the change in model performance at different times meets the second change condition, and "0" indicates that the change in model performance does not meet the second change condition. Of course, other result forms are possible, and we will not list them one by one here, as they are all within the scope of protection of this application.

[0125] Step 205: Determine whether the changes in the model input generated at the different times meet a third change condition, and obtain a third sub-result.

[0126] The third change condition is a condition that can be used to indicate that a significant change has occurred in the model input. Specifically, the change in the characteristic value of the model input data is used to reflect the change in the model input.

[0127] The feature values ​​of the model input data include two types:

[0128] Numeric type: such as user’s ID number, student’s ID number, etc.

[0129] Category type: such as different occupations of different users - lawyers, nurses.

[0130] For a numerical characteristic value, the third change condition may optionally be set to any one or more of the following:

[0131] 1) The characteristic value increases by at least P = D1% of the average value of the maximum characteristic value, or the characteristic value decreases by at least P = D2% of the average value of the minimum characteristic value, and the increase or decrease is maintained for a set time period;

[0132] D1 and D2 are fixed values, such as 10, 20, etc.

[0133] The average of the maximum eigenvalues ​​is the average of the maximum eigenvalues ​​in each unit time (such as 1 day) in a period of time (such as 6 months); the average of the minimum eigenvalues ​​is the average of the minimum eigenvalues ​​in each unit time (such as 1 day) in a period of time (such as 6 months).

[0134] 2) N2-day momentum M2 ≥ ±C2% and maintained for a set period of time;

[0135] For example, the 10-day momentum M2 ≥ ±70%, etc. Here, M2 = ((ni-ni_10) / ni_10)*100, where ni represents the characteristic value of the model input data at the current time, and p1_10 represents the characteristic value of the model input data N2 days ago.

[0136] For the characteristic value of the category type, optionally, the third change condition can be set to any one or more of the following:

[0137] 1) The category range changes by at least P = 10% and is maintained for a period of time;

[0138] For example, the occupations of more than 10% of the user information samples in the “occupation” of the input user information have changed.

[0139] 2) N3-day momentum M3 ≥ ±C3% and maintained for a set period of time;

[0140] For example, the 10-day momentum M3 is ≥ ±70% and maintained for a set period of time. Here, M3 = ((cl-cl_10) / cl_10)*100, where cl is the number of samples corresponding to the current feature (e.g., occupation) category value (e.g., lawyer), and cl_10 is the number of samples corresponding to the feature (e.g., occupation) category value (e.g., lawyer) N3 days ago.

[0141] For a numerical or categorical feature value, if it is determined to meet at least one of the corresponding conditions, a third sub-result of "1" is obtained; otherwise, a third sub-result of "0" is obtained. Here, "1" indicates that the changes in the model input at different times meet the third change condition, and "0" indicates that the changes in the model input do not meet the third change condition. Similarly, the results are not limited to "1" or "0" and can also be other forms, all of which are within the scope of protection of this application.

[0142] Step 206: Based on at least one of the first sub-result, the second sub-result, and the third sub-result, determine whether the artificial intelligence model needs to be updated, and generate a first output result.

[0143] Ultimately, the first result oriented towards the important features of the model, the second sub-result oriented towards the model performance and the third sub-result oriented towards the model input can be comprehensively considered to determine whether the artificial intelligence model needs to be updated based on the set strategy.

[0144] For example, in an optional embodiment, if the first sub-result, the second sub-result, and the third sub-result are all "1", the machine can directly give a first output result that can be used to indicate that a model update is required; otherwise, if the first sub-result, the second sub-result, and the third sub-result are all "0", the machine can give a first output result that can be used to indicate that a model update is not required. Of course, in this case, the machine can also not provide any output, and the behavior of the machine not providing any output is implied as no need for a model update; and for the situation where some of the first sub-result, the second sub-result, and the third sub-result are "1" and some are "0", it means that a discussion with human intelligence is required, and a decision on whether the model needs to be updated is made under the intervention of human intelligence. Thus, the process of discussing with human intelligence can be triggered by the machine, for example, the machine outputs a visual interface with prompt information and for receiving information input, prompting human intelligence to discuss, and input relevant instruction information into the interface to finally decide whether the model needs to be updated. The above is only an example of the present application, and flexible strategy settings can be made during implementation.

[0145] See Figure 5 , which shows a processing logic diagram for monitoring the model's important features, model performance, and model input during model operation to determine whether the model needs to be updated and to provide the appropriate time for model update. Among them, xAI stands for Explainable Artificial Intelligence, which refers to explainable artificial intelligence.

[0146] This embodiment achieves model interpretation by monitoring and analyzing data on the three aspects of the above-mentioned important features, model performance and model input. Through model interpretation, it is determined whether the underlying predictive ability of the model is reliable, and then the appropriate update time is determined.

[0147] In another optional implementation of the present application, Figure 6 As shown, the above processing method may further include the following processing:

[0148] Step 207: Determine whether there is a target feature in the features of the model input data that does not meet the association condition with the prediction result of the artificial intelligence model.

[0149] Step 208: If so, generate a second output result including the target feature; the second output result can be used to indicate that a model update is required based on the target feature;

[0150] In this embodiment, for a feature, if its importance / importance level to the model prediction reaches at least the set importance / importance level threshold, then the feature is considered to be associated with the model prediction result (essentially having a high correlation); otherwise, if the importance / importance level of the feature to the model prediction does not reach the set threshold, then the feature is considered to be irrelevant to the model prediction result (essentially having a low correlation).

[0151] As an important feature of the model, that is, a feature that will have a greater impact on the model prediction, it should at least be correlated with the model prediction results.

[0152] In view of this, the above association conditions can be the importance / importance level conditions of the features to the model prediction determined based on the actual logical association between the model features and the model prediction results.

[0153] For example, assuming that the model is used to predict the user's occupation as "lawyer" or "non-lawyer" based on the input user information, and assuming that the features in the user information include "gender", "skills (such as professional skills), and "age", then, based on the actual logical association between different features and the model prediction results, the association condition can be set, but is not limited to, as a condition representing the following information:

[0154] The importance / rank of the “skill” feature to the model’s prediction results should reach an importance / rank threshold Z;

[0155] The importance / rank of the "Gender" feature to the model's prediction results should be lower than the importance / rank threshold Z.

[0156] The above conditions are all formulated based on the actual logical relationship between features and model prediction results, and are consistent with the actual correlation between features and model prediction results.

[0157] Based on this assumption, if the model's important feature monitoring reveals that the model uses "gender" as an important feature to predict whether the user's occupation is "lawyer" during the prediction process, then this feature must not meet the above-mentioned association conditions. Therefore, this feature is identified as a target feature that does not meet the association conditions. This essentially achieves the identification of model deviation features or erroneous features.

[0158] Afterwards, a second output result can be further generated that includes at least the target feature. This second output result not only indicates that a model update is required, but also further points out the deviation features / error features of the model, thereby correspondingly indicating that the model should be updated based on the deviation features / error features of the model to achieve a balance in the importance of each model feature and improve the accuracy or precision of the model prediction. It is easy to understand that if there is no target feature that does not meet the association conditions with the prediction result of the artificial intelligence model, it is not necessary to generate and output the second output result.

[0159] In another optional implementation of the present application, Figure 7 As shown, the above processing method may further include the following processing:

[0160] Step 209: Generate model prediction rules based on different associations between different features and the prediction results of the artificial intelligence model, so that the artificial intelligence model is updated based on the model prediction rules.

[0161] During implementation, different associations between different features and the prediction results of the artificial intelligence model can be pre-established based on the actual logical association between the features of the model input data and the model prediction, and model prediction rules can be generated. The model prediction rules can be rules generated by the logical association between the features obtained by the machine based on big data learning and the model prediction, or they can be rules generated by the machine based on the logical association between manually provided features and model predictions, or they can be rules directly configured by humans, and of course they can also be a combination of any two or three of the above.

[0162] Afterwards, the prediction rule can be further applied to the update of the model. For example, after the model determines a suitable update time based on the processing method of this application, the processing flow of updating the model using the prediction rule is triggered. Alternatively, the model can be updated in real time after the prediction rule is generated or configured.

[0163] A simple example is provided below.

[0164] Combined with actual applications, it can be seen that in user occupation prediction, the "gender" feature has a very low contribution to the prediction of the "lawyer" occupation, while the "gender" feature has a higher contribution to the prediction of the "nurse" occupation. That is, based on a certain gender, such as "female" or "male", it is easier to distinguish users as "nurses" or "non-nurses", but not easy to distinguish users as "lawyers" or "non-lawyers". The "age" feature has a higher contribution to the prediction of the "student" occupation. For example, a certain age of "17 years old" can predict the user as a "student" with a high probability, rather than other occupations, such as "lawyer". Therefore, the following prediction rules can be generated:

[0165] “Gender” is an important feature for the prediction of the “Nurse” category;

[0166] “Gender” is a non-significant feature for the prediction of the “Lawyer” category;

[0167] “Age” is an important feature for the prediction of “Student” class.

[0168] Afterwards, the prediction rules can be used to update the model. For example, assuming that "gender" is used as an important feature for predicting the "lawyer" profession in the artificial intelligence model, the performance of the model is difficult to guarantee. After updating the model based on the above prediction rules, "gender" can be changed from an important feature of the model prediction to a non-important feature, thereby further improving the model's prediction accuracy and precision.

[0169] See Figure 8 , provides a logical example diagram for determining the model update timing and model update processing using the processing method of the present application, wherein 801 represents a person (such as a technician or administrator in a model debugging environment, or a front-end user in an actual operation environment) controlling the operation of the ML model 802 at the front end of the device. After the model 802 is started, data (such as the collected and stored front-end user personal information) is imported from the database 803, and the data in the database 803 is used as the model input to perform prediction processing and the prediction results can be fed back to the front-end device. At the same time, the background will start monitoring the model 802, obtain monitoring data of the model's important features, model performance, and model input, and use the obtained monitoring data as the input of xAI 804, wherein xAI 804 interprets the model based on the monitoring data and outputs analysis information such as a single-rank list of important features, importance / importance level curves, mark rate curves, and characteristic value curves, which are recorded as 805 in the figure. The background finally determines whether the model needs to be updated based on this analysis information, that is, whether the current time is the right time to update the model, and outputs the decision result to human intelligence 806. Of course, when the background cannot clearly decide whether the model needs to be updated, it can trigger a discussion process for human intelligence 806, and combine the intervention of human intelligence to finally determine whether the model needs to be updated.

[0170] In addition, the deviation feature / error feature selection project can be performed in combination with the above corresponding embodiments of the present application, and the deviation features / error features that do not meet the association conditions in the model can be determined through feature analysis 807, and the model can be updated, such as Figure 8 In addition, the generation of prediction rules and the model update processing based on the prediction rules can also be performed in combination with the above corresponding embodiments of the present application, such as Figure 8 Feedback from Rule generator 809 to model 802.

[0171] By determining the update timing of the model and processing the model update, it is possible to determine an appropriate update timing for the model to trigger the model update. When necessary, it can also be combined with discussions with human intelligence to ensure the model's future processing performance and credibility without causing unnecessary model update workload due to inappropriate model update timing.

[0172] Corresponding to the above-mentioned processing method, an embodiment of the present application also provides a computer device, which may be, but is not limited to, a portable computer (such as a notebook), a desktop computer or a large or medium-sized computer, a background server or a cloud platform server in a general / special computing or configuration environment.

[0173] like Figure 9 As shown, the computer device includes:

[0174] Memory 901, used to store at least one set of instruction sets;

[0175] The processor 902 is configured to call and execute the instruction set in the memory, and perform the following operations by executing the instruction set:

[0176] Obtain monitoring data of predetermined monitoring indicators of the artificial intelligence model at different times;

[0177] Determining, based on the monitoring data, changes in the predetermined monitoring indicator at the different times;

[0178] Determining whether changes in the predetermined monitoring indicator at the different times meet a change condition, and obtaining a determination result;

[0179] Based on the determination result, it is determined whether the artificial intelligence model needs to be updated, and a first output result is generated; the first output result can be used to indicate whether the artificial intelligence model needs to be updated.

[0180] In an optional implementation of the embodiment of the present application, the processor 902 obtains monitoring data of a predetermined monitoring indicator of the artificial intelligence model at different times, including at least one of the following:

[0181] Obtaining first monitoring data of important features of the artificial intelligence model at different times; the important features include: features included in the input data of the artificial intelligence model, the impact of which on the model prediction result meets the impact condition;

[0182] Obtaining second monitoring data of the model performance of the artificial intelligence model at different times;

[0183] Obtain third monitoring data of the analog-to-digital input of the artificial intelligence model at different times.

[0184] In an optional implementation of the embodiment of the present application, the second monitoring data includes the signature rate data of the artificial intelligence model; the signature rate of the model is: the ratio of the number of results of different polarities in the different prediction results of the model for different input data, where the polarity is the polarity of the prediction result pre-set for the input data;

[0185] The third monitoring data includes: feature values ​​of features included in the input data of the artificial intelligence model.

[0186] In an optional implementation of the embodiment of the present application, the processor 902 determines whether the change of the predetermined monitoring indicator generated at the different times satisfies the change condition, and obtains a determination result including at least one of the following:

[0187] determining, based on the first monitoring data, whether changes in the important feature generated at the different times satisfy a first change condition, and obtaining a first sub-result;

[0188] determining, based on the second monitoring data, whether changes in the model performance at the different times satisfy a second change condition, and obtaining a second sub-result;

[0189] Based on the third monitoring data, it is determined whether the changes of the model input generated at the different times meet a third change condition to obtain a third sub-result.

[0190] The processor 902 determines whether the artificial intelligence model needs to be updated and generates a first output result, including: determining whether the artificial intelligence model needs to be updated based on at least one of the first sub-result, the second sub-result and the third sub-result, and generating a first output result.

[0191] In an optional implementation of the embodiment of the present application, the processor 902 may also be configured to:

[0192] Determine whether there is a target feature among the features of the model input data that does not meet the association condition with the prediction result of the artificial intelligence model; if so, generate a second output result including the target feature; the second output result can be used to indicate that a model update is required based on the target feature;

[0193] and / or,

[0194] According to different associations between different features and the prediction results of the artificial intelligence model, model prediction rules are generated so that the artificial intelligence model is updated based on the model prediction rules.

[0195] For the computer device disclosed in the embodiment of the present application, since it corresponds to the processing method applied to the computer device disclosed in the corresponding embodiment above, the description is relatively simple. For relevant similarities, please refer to the description of the processing method part in the corresponding embodiment above, and no further details will be given here.

[0196] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referenced to each other.

[0197] For the convenience of description, the above systems or devices are described as being divided into various modules or units according to their functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0198] Through the description of the above embodiments, it can be seen that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present application or certain parts of the embodiments.

[0199] Finally, it should be noted that, in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0200] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A processing method comprising: Obtaining first monitoring data of important features of the artificial intelligence model at different times, second monitoring data of model performance at different times, and third monitoring data of model input at different times; Determining, based on the first monitoring data, changes in the important features at different times; determining, based on the second monitoring data, changes in the model performance at different times; and determining, based on the third monitoring data, changes in the model input at different times. Determining whether changes in the important features at the different times satisfy a first change condition, to obtain a first sub-result, where the first change condition is a condition that can be used to characterize a significant change in the important features of the model; determining whether the change in the model performance at the different times satisfies a second change condition, to obtain a second sub-result, where the second change condition is a condition that can be used to indicate a significant change in the model performance; determining whether changes in the model input at the different times satisfy a third change condition, to obtain a third sub-result, wherein the third change condition is a condition that can be used to indicate a significant change in the model input; Based on the first sub-result, the second sub-result and the third sub-result, determine whether the artificial intelligence model needs to be updated, and generate a first output result; the first output result can be used to indicate whether the artificial intelligence model needs to be updated.

2. The method according to claim 1, wherein the important features include: The input data of the artificial intelligence model includes features that have an impact on the model's prediction results and meet the influencing conditions.

3. The method according to claim 2, wherein: The second monitoring data includes the signature rate data of the artificial intelligence model; the signature rate of the model is: the ratio of the number of results of different polarities in the different prediction results of the model for different input data, where the polarity is the polarity of the prediction result pre-set for the input data; The third monitoring data includes: feature values ​​of features included in the input data of the artificial intelligence model.

4. The method according to claim 2, further comprising: Determining whether there are target features among the features of the model input data that do not meet the association conditions with the prediction results of the artificial intelligence model; If so, generating a second output result including the target feature; the second output result can be used to indicate that a model update is required according to the target feature; and / or, According to different associations between different features and the prediction results of the artificial intelligence model, model prediction rules are generated so that the artificial intelligence model is updated based on the model prediction rules.

5. A computer device comprising: a memory for storing at least one set of instruction sets; A processor is configured to call and execute the instruction set in the memory, and perform the following operations by executing the instruction set: Obtaining first monitoring data of important features of the artificial intelligence model at different times, second monitoring data of model performance at different times, and third monitoring data of model input at different times; Determining, based on the first monitoring data, changes in the important features at different times; determining, based on the second monitoring data, changes in the model performance at different times; and determining, based on the third monitoring data, changes in the model input at different times. Determining whether changes in the important features at the different times satisfy a first change condition, to obtain a first sub-result, where the first change condition is a condition that can be used to characterize a significant change in the important features of the model; determining whether the change in the model performance at the different times satisfies a second change condition, to obtain a second sub-result, where the second change condition is a condition that can be used to indicate a significant change in the model performance; determining whether changes in the model input at the different times satisfy a third change condition, to obtain a third sub-result, wherein the third change condition is a condition that can be used to indicate a significant change in the model input; Based on the first sub-result, the second sub-result and the third sub-result, determine whether the artificial intelligence model needs to be updated, and generate a first output result; the first output result can be used to indicate whether the artificial intelligence model needs to be updated.

6. The computer device according to claim 5, wherein the important features include: The input data of the artificial intelligence model includes features that have an impact on the model's prediction results and meet the influencing conditions.

7. The computer device of claim 6, wherein: The second monitoring data includes the signature rate data of the artificial intelligence model; the signature rate of the model is: the ratio of the number of results of different polarities in the different prediction results of the model for different input data, where the polarity is the polarity of the prediction result pre-set for the input data; The third monitoring data includes: feature values ​​of features included in the input data of the artificial intelligence model.

8. The computer device according to claim 6, wherein the processor is further configured to: Determine whether there is a target feature among the features of the model input data that does not meet the association condition with the prediction result of the artificial intelligence model; if so, generate a second output result including the target feature; the second output result can be used to indicate that a model update is required based on the target feature; and / or, According to different associations between different features and the prediction results of the artificial intelligence model, model prediction rules are generated so that the artificial intelligence model is updated based on the model prediction rules.

Images