Classification model optimization method and device, classification method and device, equipment and medium

Abstract

The invention discloses a classification model optimization method and device, a classification method and device, equipment and a medium. The method comprises the steps of performing repair data identification based on a current round of training set and a previous round of training set, and determining a repair sample set; updating the auxiliary model based on the current round training set and the repair sample set, determining a current training loss function value, and determining a newly added clean sample set of the current round; based on the accumulated clean sample set of the previous round and the newly added clean sample set, determining an accumulated clean sample set of the current round, performing model optimization based on the accumulated clean sample set of the current round, and determining a target model; based on the current training loss function value and the classification model, performing data conflict repair on the current round of training set, and determining a repaired training set; and repeatedly executing the repair data identification based on the current round of training set and the previous round of training set until a target classification model is obtained. According to the method, collaborative optimization can be carried out on data restoration and model optimization training, and a classification model with high precision is obtained.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a classification model optimization method, classification method, apparatus, device, and medium. Background Technology

[0002] Relational data is used to optimize and train classification models, typically in classification scenarios such as finance and healthcare. However, relational data often contains numerous errors during acquisition, making it unsuitable for accurate training of classification models. Current techniques usually involve data repair of the relational data before training the classification model. This method separates data repair and model optimization into two distinct stages, leading to a disconnect between the two and a decrease in the accuracy of the optimized model. Therefore, how to collaboratively optimize data repair and model optimization to obtain a highly accurate classification model is a pressing technical problem that needs to be solved. Summary of the Invention

[0003] This invention provides a classification model optimization method, classification method, apparatus, device, and medium to address the technical problem of how to collaboratively optimize data repair and model optimization training to obtain a classification model with high accuracy.

[0004] A classification model optimization method, comprising: Based on the current training set and the previous training set, identify the repair data and determine the repair sample set; The auxiliary model is updated based on the current training set and the repaired sample set. The current training loss function value corresponding to each training sample in the current training set is determined, and the new clean sample set for the current round is determined. Based on the cumulative clean sample set of the previous round and the newly added clean sample set of the current round, determine the cumulative clean sample set of the current round, optimize the target model based on the cumulative clean sample set of the current round, and determine the target model of the current round. Based on the current training loss function value corresponding to each training sample in the current round training set and the classification model with the same structure as the target model, data conflict repair is performed on the current round training set to determine the repaired training set. The repaired training set is updated to the current round training set. The process of identifying repaired data based on the current round training set and the previous round training set is repeated to determine the repaired sample set until the target model is optimized and the target classification model is obtained.

[0005] A classification method, comprising: Obtain the data to be classified; The data to be classified is input into the target classification model to predict the target category corresponding to the data to be classified. The target classification model is determined based on the classification model optimization method described above.

[0006] A classification model optimization device, comprising: The repair sample set determination module identifies the repair data based on the current training set and the previous training set to determine the repair sample set. The auxiliary model update module updates the auxiliary model based on the current training set and the repaired sample set, determines the current training loss function value corresponding to each training sample in the current training set, and determines the new clean sample set for the current round. The target model optimization module determines the cumulative clean sample set for the current round based on the cumulative clean sample set from the previous round and the newly added clean sample set for the current round. Based on the cumulative clean sample set for the current round, it optimizes the target model to determine the target model for the current round. The data conflict repair module performs data conflict repair on the current training set based on the current training loss function value corresponding to each training sample in the current round training set and the classification model with the same structure as the target model, and determines the repaired training set. The target classification model determination module is used to update the repaired training set to the current round training set, repeatedly perform the repair data identification based on the current round training set and the previous round training set, determine the repair sample set, until the target model is optimized and the target classification model is obtained.

[0007] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the classification model optimization method described above, or the processor implements the classification method described above when executing the computer program.

[0008] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described classification model optimization method, or, when executed by a processor, implements the above-described classification method.

[0009] The aforementioned classification model optimization method, classification method, apparatus, equipment, and medium identify repaired data based on the current and previous training sets to determine the repaired sample set. Based on the current and repaired sample sets, the auxiliary model is updated to determine a relatively clean new sample set needed for training the target model. The target model is then trained and updated more accurately using the relatively clean data selected from the current dataset, avoiding interference from inaccurate data in the current dataset. After each round of model optimization, data conflict repair is performed on the current training set. Further model optimization and updates are then performed based on the repaired training set. This process of data conflict repair during model optimization promotes synergy between data conflict repair and target model optimization training, integrating data conflict repair with the actual situation of model optimization and updates, preventing data conflict repair from becoming disconnected from model optimization, and enabling accurate model training. This method effectively coordinates data conflict repair and model optimization and updating through the interaction of the auxiliary model and the target model. It has an adaptive mechanism that integrates data conflict repair and model optimization and updating in a closed loop, dynamically and reasonably performing data conflict repair, thereby promoting accurate model optimization and updating. It achieves the goal of directly and accurately training the target classification model based on the training set with data errors. This classification model optimization method can be applied to a variety of classification scenarios, with wide scenario transferability and high application value. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a classification model optimization method in one embodiment of the present invention; Figure 2 This is another flowchart of the classification model optimization method in one embodiment of the present invention; Figure 3 This is another flowchart of the classification model optimization method in one embodiment of the present invention; Figure 4 This is another flowchart of the classification model optimization method in one embodiment of the present invention; Figure 5 This is another flowchart of the classification model optimization method in one embodiment of the present invention; Figure 6 This is another flowchart of the classification model optimization method in one embodiment of the present invention; Figure 7This is a flowchart of a classification method in one embodiment of the present invention; Figure 8 This is a schematic diagram of a classification model optimization device in one embodiment of the present invention; Figure 9 This is a schematic diagram of a computer device according to an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the working principle of a classification model optimization method in one embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The classification model optimization method provided in this invention includes: identifying repaired data based on the current training set and the previous training set to determine a repaired sample set; updating the auxiliary model based on the current training set and the repaired sample set to determine the current training loss function value for each training sample in the current training set and to determine the newly added clean sample set for the current round; determining the cumulative clean sample set for the current round based on the cumulative clean sample set from the previous round and the newly added clean sample set for the current round; optimizing the target model based on the cumulative clean sample set for the current round to determine the target model for the current round; repairing data conflicts in the current training set based on the current training loss function value for each training sample in the current training set and a classification model with the same structure as the target model to determine the repaired training set; updating the repaired training set to the current training set; and repeating the process of identifying repaired data based on the current training set and the previous training set to determine the repaired sample set until the target model optimization is completed, thus obtaining the target classification model. This method effectively coordinates data conflict repair and model optimization and update. It can reasonably repair data conflicts based on the actual situation of model optimization and update, thereby promoting accurate model optimization and update. It has broad scenario transferability and high application value.

[0014] In one embodiment, such as Figure 1 As shown, a classification model optimization method is provided, which is then applied to... Figure 9 Taking a computer device as an example, the explanation includes the following steps: S101: Based on the current training set and the previous training set, identify the repair data and determine the repair sample set; S102: Update the auxiliary model based on the current training set and the repaired sample set, determine the current training loss function value corresponding to each training sample in the current training set, and determine the new clean sample set for the current round; S103: Based on the cumulative clean sample set of the previous round and the newly added clean sample set of the current round, determine the cumulative clean sample set of the current round, optimize the target model based on the cumulative clean sample set of the current round, and determine the target model of the current round. S104: Based on the current training loss function value corresponding to each training sample in the current round training set and the classification model with the same structure as the target model, perform data conflict repair on the current round training set and determine the repaired training set; S105: Update the repaired training set to the current round training set, repeat the process of identifying repaired data based on the current round training set and the previous round training set, determine the repaired sample set, until the target model is optimized and the target classification model is obtained.

[0015] In this context, the current training set refers to the relational data used during the current round of model optimization training. The previous training set refers to the relational data used during the previous round of model optimization training. Relational data refers to data that includes the mapping relationship between attributes and categories. For example, in an income prediction scenario, attributes include a user's education level, educational attainment, working hours, and occupation, while the category is the user's income level. Another example is in a medical resource allocation scenario, where attributes include various examination indicators of patients, and the category is the type of ward the patient needs to be assigned to. In this example, the current training set is obtained by performing data repair on the training samples of at least one tuple and / or the data corresponding to at least one attribute in the previous training set. Here, the training sample refers to a set of relational data.

[0016] As an example, in step S101, after the previous round of model training, the computer device repairs the data of the previous training set to obtain the current training set. It then compares and analyzes the data corresponding to the training samples of the same tuples in the previous and current training sets to determine whether the data has been repaired. If the data corresponding to a training sample of a certain tuple changes, it is determined that the data corresponding to the training sample of that tuple has been repaired. All training samples that have undergone data repair form a repaired sample set. For example, in an income prediction scenario, the computer device compares and analyzes the data corresponding to the attributes affecting income of tuple A in the current and previous training sets. If data changes, it is determined that the data corresponding to the training sample of tuple A has been repaired.

[0017] Here, the auxiliary model refers to the classification model used to assist in training the target model. The newly added clean sample set in the current round refers to the set of training samples that the auxiliary model has determined to have high training accuracy in the current round. The current training loss function value refers to the loss function value generated by the auxiliary model during its own parameter update and optimization process.

[0018] As an example, in step S102, the computer device trains the auxiliary model using each training sample from the current training set and the repair sample set, updates the model parameters corresponding to the auxiliary model, and determines the loss function value corresponding to each training sample, setting this loss function value as the current training loss function value. Based on the current training loss function value corresponding to each training sample, the computer device filters the training samples in the current training set to obtain multiple newly added clean sample sets corresponding to the filtered training samples. For example, the computer device uses multiple training samples with lower current training loss function values ​​as the newly added clean sample set for the current round. In this example, the auxiliary model is updated to determine a relatively clean newly added clean sample set required for training the target model.

[0019] The cumulative clean sample set from the previous round refers to the sample set used in the previous round of optimizing and training the target model. The cumulative clean sample set from the current round refers to the sample set used in the current round of optimizing and training the target model.

[0020] As an example, in step S103, the computer device merges the cumulative clean sample set from the previous round with the newly added clean sample set in the current round to form the cumulative clean sample set for the current round. The cumulative clean sample set for the current round is then input into the target model, and the standard stochastic gradient descent algorithm is used to optimize and update the model parameters corresponding to the target model, resulting in the target model for the current round. In this example, relatively clean data selected from the current round dataset is used to train and update the target model more accurately, avoiding interference from inaccurate data in the current round dataset. In this example, the cumulative clean sample set for the current round... for: The current round is round K. This is the cumulative clean sample set from the previous round. This is the newly added clean sample set for the current round.

[0021] For example, in the income prediction scenario, the computer equipment controls the model parameters corresponding to the target model to optimize and update them, so as to obtain the target model for the current round. The target model in the income prediction scenario is used to determine the category corresponding to the user's income level based on attribute data including the user's education level, work hours and occupation.

[0022] For example, in the scenario of medical resource allocation, the computer equipment controls the model parameters corresponding to the target model to optimize and update them, so as to obtain the target model for the current round. The target model in the medical resource allocation scenario is used to determine the category of the ward type area that the patient needs to be allocated based on attribute data, including various examination indicators of the patient.

[0023] The repaired training set refers to the dataset after data conflict repair of the current round of training set.

[0024] As an example, in step S104, the computer device filters the attributes and training samples of the current training set based on the current training loss function value corresponding to each training sample in the current round training set and the classification model with the same structure as the target model, determines the data corresponding to the attributes that need to be repaired and the data corresponding to the training samples, and uses a preset conflict repair tool to repair the data corresponding to the attributes that need to be repaired and the data corresponding to the training samples, respectively, to obtain the repaired training set.

[0025] For example, in an income prediction scenario, the computer device filters the attributes in the current training set, including education level, educational attainment, and occupation, based on the current training loss function value corresponding to each training sample in the current training set and the classification model with the same structure as the target model. It then determines the data corresponding to the attributes that need to be repaired and filters the training samples for each tuple to determine the data corresponding to the training samples that need to be repaired. The computer device uses a preset conflict repair tool to repair the data corresponding to the attributes that need to be repaired and the data corresponding to the training samples that need to be repaired, respectively, to obtain the repaired current training set.

[0026] In this example, the default conflict resolution tools include, but are not limited to, Rock, Raha, Baran, and HoloClean.

[0027] In this example, during model training, after each round of model optimization and update, data conflict resolution (CR) is performed on the current training set. This allows for further model optimization and update based on the resolved training set, thus enabling data conflict resolution during the model optimization and update process. This ensures that data conflict resolution and model training work together, combining data conflict resolution with the actual situation of model optimization and update, avoiding a disconnect between data conflict resolution and model optimization and update, and enabling accurate model training.

[0028] As an example, in step S105, the computer device updates the repaired training set to the current round training set, and repeatedly performs data identification based on the current round training set and the previous round training set to determine the repaired sample set, until the target model is optimized and the target classification model is obtained. That is, after updating the repaired training set to the current round training set, the computer device repeats steps S101 to S104 until the target model optimization and update are determined to be complete. The optimized and updated target model is determined as the target classification model, which is used to classify the data in the application scenario to which the training set belongs. In this example, when the computer device determines that the current training loss function value and / or the loss function value corresponding to the target model in the current round have converged, it determines that the target model optimization is complete, stops repeating steps S101 to S104, and obtains the target classification model.

[0029] For example, if the training set pertains to an income prediction scenario, a target classification model is used to classify users' income levels based on data corresponding to attributes such as education level, educational attainment, and occupation within the income prediction scenario. As another example, if the training set pertains to a medical resource allocation scenario, a target classification model is used to classify the type of ward a patient needs based on data corresponding to various examination indicators.

[0030] In this embodiment, repair data is identified based on the current training set and the previous training set to determine the repair sample set. The auxiliary model is then updated based on the current training set and the repair sample set to identify a relatively clean new sample set needed to train the target model. The target model is then trained and updated more accurately using the relatively clean data selected from the current dataset, avoiding interference from inaccurate data in the current dataset. After each round of model optimization and update, data conflict repair is performed on the current training set. Further model optimization and update are then performed based on the repaired training set. This process of data conflict repair during model optimization and update ensures that data conflict repair and target model optimization training work synergistically. By integrating data conflict repair with the actual situation of model optimization and update, it avoids a disconnect between data conflict repair and model optimization, enabling accurate model training. This method effectively coordinates data conflict repair and model optimization and updating through the interaction of the auxiliary model and the target model. It has an adaptive mechanism that integrates data conflict repair and model optimization and updating in a closed loop, dynamically and reasonably performing data conflict repair, thereby promoting accurate model optimization and updating. It achieves the goal of directly and accurately training the target classification model based on the training set with data errors. This classification model optimization method can be applied to a variety of classification scenarios, with wide scenario transferability and high application value.

[0031] In one embodiment, such as Figure 2As shown, step S102, which involves updating the auxiliary model based on the current training set and the repaired sample set, determines the current training loss function value for each training sample in the current training set and determines the newly added clean sample set for the current round, includes: S201: Determine the impact estimate for each training sample in the repair sample set. The impact estimate is the evaluation value corresponding to the degree of influence of the training samples in the repair sample set on the auxiliary model update. S202: Based on the impact estimates corresponding to all training samples in the repair sample set and the number of tuples corresponding to the previous training set, perform forgetting update processing on the auxiliary model to determine the forgotten auxiliary model; S203: Input each training sample in the current training set into the forgotten auxiliary model in sequence, perform multiple rounds of parameter fine-tuning and updating of the forgotten auxiliary model, and determine the current training loss function value corresponding to each training sample in the current training set after the update. S204: Based on the current training loss function value and the repair sample set corresponding to each training sample in the current round training set, determine the new clean sample set for the current round.

[0032] Among them, the impact estimate is the evaluation value corresponding to the degree of influence of the training samples in the repair sample set on the update of the auxiliary model.

[0033] As an example, in step S201, the computer device processes the training samples of each tuple in the repair sample set using a preset function to determine the estimated impact of the training samples of each tuple on the auxiliary model. In this example, the preset function is... The estimated impact of the t-th tuple on the auxiliary model is: .

[0034] As an example, in step S202, the computer device uses the influence estimate and the number of tuples corresponding to the previous training set to perform forgetting update processing on the model parameters of the auxiliary model, obtaining the forgotten auxiliary model. In this example, the method for forgetting update processing of the model parameters of the auxiliary model in the current round is as follows: .in, For the previous round (kth round) In round 1, where k > 1), the model parameters of the auxiliary model after the model update are... These are the model parameters of the auxiliary model after the model update in the current round (round k, where k > 1). The number of tuples corresponding to the previous training set. To repair the sample set. Understandably, repairing the sample set... The data corresponding to each training sample in the model was conflict-resolved after the previous round of model optimization and update, indicating that the sample set was repaired. In this example, the data corresponding to each training sample has a negative impact on the auxiliary model in the previous round of model optimization and update. By using the forgetting update method for the model parameters of the auxiliary model in the current round, the auxiliary model can quickly forget and repair the negative impact of the sample set in the previous round of model optimization and update, which significantly improves the efficiency of model optimization and update and the stability of the model.

[0035] As an example, in step S203, the computer device will use the current round of training set. Each training sample is sequentially input into the forgotten auxiliary model. Multiple rounds of parameter fine-tuning are performed on the forgotten auxiliary model to obtain the updated auxiliary model and the current training loss function value for each training sample in the current training set. In this example, the computer device performs fewer than 5 rounds of parameter fine-tuning on the forgotten auxiliary model.

[0036] As an example, in step S204, the computer device analyzes the current training loss function value corresponding to each training sample in the repaired sample set and the current round of training set to determine the new clean sample set for the current round. In this example, the computer device can determine the training samples with smaller current training loss function values ​​and all training samples in the repaired sample set as the new clean sample set for the current round. Understandably, the smaller the current training loss function value, the higher the reliability of the training sample; therefore, training samples with smaller current training loss function values ​​are determined as the new clean sample set for the current round. Furthermore, all training samples in the repaired sample set are data that has been repaired after the previous round of model optimization and updates, and also have high reliability; therefore, all training samples in the repaired sample set are determined as the new clean sample set for the current round. Determining training samples with higher reliability as the new clean sample set for the current round improves the rationality of determining the new clean sample set for the current round.

[0037] In this embodiment, by adjusting the impact estimate value corresponding to each training sample in the repair sample set, the model parameters of the auxiliary model in the current round are updated using a forgetting process. This allows the auxiliary model to quickly forget the negative impact of the repair sample set in the previous round of model optimization and update, significantly improving the efficiency and stability of model optimization and update. Furthermore, in this example, the current training loss function value corresponding to each training sample in the repair sample set and the current round of training set is analyzed to reasonably determine the new clean sample set for the current round.

[0038] In one embodiment, such as Figure 3 As shown, step S204, which is to determine the new clean sample set for the current round based on the current training loss function value and the repaired sample set for each training sample in the current round training set, includes: S301: Obtain multiple historical training loss function values ​​corresponding to each training sample in the current round of training set. The multiple historical training loss function values ​​are training loss function values ​​formed by multiple rounds of training before the current round. S302: Based on the current training loss function value and multiple historical training loss function values ​​corresponding to the training samples of all tuples, perform sample screening on all training samples in the current round training set to determine the newly added clean samples in the current round. S303: Determine the new clean sample set for the current round based on all the new clean samples and repaired sample sets in the current round.

[0039] Understandably, in the multiple training rounds prior to the current round, each training sample corresponds to a historical training loss function value in each training round, thus obtaining multiple historical training loss function values ​​corresponding to each training sample in multiple training rounds.

[0040] As an example, in step S301, the computer device obtains the historical training loss function values ​​for each training sample in the training set during the multiple training rounds prior to the current round. In this example, for the training sample of the t-th tuple, if the current round is the K-th round, the historical training loss function values ​​for the K-th rounds prior to the K-th round are obtained. In round 1, the computer device retrieves the historical training loss function values ​​from round 1 of training. Historical training loss function values ​​of the second round of training Historical training loss function value of the third round of training ... the Kth Historical training loss function value of 1 training round .

[0041] As an example, in step S302, the computer device analyzes the current training loss function value and multiple historical training loss function values ​​corresponding to the training samples of all tuples in order to filter all training samples in the current round training set and determine the newly added clean samples in the current round.

[0042] In this example, by using the current training loss function value corresponding to the training samples of all tuples and multiple historical training loss function values, it is determined whether the training loss function value corresponding to the training samples of each tuple is consistently low. Training samples corresponding to tuples with consistently low training loss function values ​​are identified as new clean samples for the current round. Specifically, the computer device identifies training samples corresponding to tuples whose current training loss function value is below a preset loss function threshold (e.g., 20%) and whose multiple historical training loss function values ​​are consistently low as new clean samples for the current round. The method for determining that the historical training loss function value corresponding to the training sample of any tuple is consistently low includes: comparing the historical training loss function value of the tuple to be selected (e.g., the t-th tuple) with that of the remaining tuples in the same round; if the historical training loss function value of the tuple to be selected in multiple rounds of training is lower than that of multiple tuples in the remaining tuples, then the historical training loss function value corresponding to the training sample of the tuple to be selected is determined to be consistently low; or, if the historical training loss function value corresponding to the tuple to be selected in multiple rounds of training is lower than a preset value, then the historical training loss function value corresponding to the training sample of the tuple to be selected is determined to be consistently low. Here, the preset loss function threshold refers to a preset threshold used to determine the magnitude of the training loss function value.

[0043] For example, for the training sample of the t-th tuple in the current training set, the computer device determines the current training loss function value. If the historical training loss function value corresponding to the training sample of the t-th tuple is consistently low and below a preset loss function threshold, the training sample of the t-th tuple is determined to be a newly added clean sample in the current round. In this example, the computer compares the historical training loss function value of the t-th tuple with that of the remaining tuples in the same round. If the historical training loss function value of the t-th tuple across multiple rounds is lower than that of multiple tuples in the remaining tuples (e.g., the historical training loss function value of the t-th tuple across multiple rounds is lower than that of the remaining 80% of the tuples), the historical training loss function value corresponding to the training sample of the t-th tuple is determined to be consistently low; or, the computer determines that the historical training loss function value corresponding to the training sample of the t-th tuple is consistently low from round 1 to round K. The historical training loss function value corresponding to round 1 (the round before the current round). , , , , If there are M consecutive historical training loss function values ​​close to the current round that are lower than a preset value, then the historical training loss function value corresponding to the training sample of the t-th tuple is determined to be consistently low. The computer device iterates through each training sample in the current round's training set according to the above steps to determine all newly added clean samples in the current round. Understandably, if the current training loss function value corresponding to a training sample is lower than the preset loss function threshold, and the historical training loss function value corresponding to the training sample is consistently low, it indicates that the training sample is relatively reliable and the probability of data error is small. Therefore, this training sample is determined as a newly added clean sample in the current round.

[0044] As an example, in step S303, the computer device merges all the newly added clean samples in the current round and each training sample in the repair sample set into a single set to obtain the newly added clean sample set for the current round. For instance, if the computer device determines that all the newly added clean samples in the current round include newly added clean samples with z (z>1) tuples, then the newly added clean samples with z (z>1) tuples and each training sample in the repair sample set are determined as the newly added clean sample set for the current round.

[0045] In this embodiment, by comparing the current training loss function value and the historical training loss function value of the training samples of the same tuple in the current round, it is determined whether the loss function value corresponding to each training sample is consistently low. The training samples with consistently low loss function values ​​are identified as newly added clean samples in the current round. The newly added clean samples in the current round and each training sample in the repair sample set are identified as the newly added clean sample set in the current round. This achieves the purpose of reasonably determining the newly added clean sample set used to optimize and update the target model, and can effectively ensure the data diversity of the newly added clean sample set, preventing overfitting of the target model.

[0046] In one embodiment, such as Figure 4 As shown, step S104, which involves performing data conflict repair on the current training set based on the current training loss function value corresponding to each training sample in the current round of training and a classification model with the same structure as the target model, and determining the repaired training set, includes: S401: Based on the current training loss function value corresponding to each training sample in the current round of training set, the training samples in the current round of training set are filtered to determine the pre-training set and the validation set; S402: Based on a classification model with the same structure as the target model, a pre-training set, and a validation set, perform attribute filtering on the candidate attributes corresponding to each training sample in the current round of training set to determine the key attributes corresponding to the current round of training set. S403: Based on a classification model with the same structure as the target model, a pre-training set, and a validation set, perform tuple filtering on the training samples in the current round of training set to determine the key tuples corresponding to the current round of training set; S404: Perform data conflict repair on the key attributes and key tuples corresponding to the current round of training set, and determine the repaired training set.

[0047] The validation set refers to the set of training samples with high confidence in the current training set. The pre-training set refers to the set of training samples in the current training set excluding the validation set.

[0048] As an example, in step S401, the computer device performs confidence processing on the current training loss function value corresponding to each training sample in the current training set to determine the confidence level for each training sample. In this example, the computer device processes the current training set... In the diagram, the current training loss function value corresponding to the training sample of the t-th tuple. of Performing softmax normalization by 1x yields the normalized result as softmax( 1 The normalized result is softmax( 1 The confidence level corresponding to the training sample of the t-th tuple is determined. Understandably, the confidence level corresponding to the training sample of the t-th tuple. The higher the confidence level, the greater the contribution and reliability of the training sample of the t-th tuple in the model optimization and update process, which is relatively reliable and beneficial to the model optimization and update. The confidence level corresponding to the training sample of the t-th tuple... The lower the confidence level, the smaller the contribution and the lower the confidence level of the training sample of the t-th tuple in the model optimization and update process, which is more detrimental to the model optimization and update. Therefore, it is necessary to determine the confidence level corresponding to each training sample so as to reasonably divide the current round of training set according to the confidence level of each training sample.

[0049] In this example, the computer device uses the confidence level of each training sample as the probability of adding it to the validation set. The p training samples with higher confidence levels are added to the validation set, resulting in a validation set suitable for model validation. The computer device then defines the pre-training set as all training samples in the current training set, excluding those in the validation set. In this example, the current training set is partitioned reasonably based on the confidence level of each training sample.

[0050] Candidate attributes refer to all attributes contained in the training samples. Key attributes refer to the attributes in the current training set that require data conflict repair.

[0051] As an example, in step S402, the computer device trains and validates a classification model with the same structure as the target model based on the pre-training set and validation set. This is used to filter attributes among the candidate attributes corresponding to each training sample in the current round of training set, selecting multiple candidate attributes with strong training influence on the classification model. These multiple candidate attributes with strong training influence on the classification model are then identified as the key attributes corresponding to the current round of training set. Understandably, the stronger the training influence of a candidate attribute, the more it indicates the need for data conflict correction to improve model training accuracy.

[0052] Among them, key tuples refer to the tuples in the current training set that require data conflict repair.

[0053] As an example, in step S403, the computer device trains and validates a classification model with the same structure as the target model based on the pre-training set and validation set. This is used to filter tuples for each training sample in the current round of training, identifying multiple tuples with strong training influence on the classification model. These multiple tuples with strong training influence on the classification model are then identified as the key tuples corresponding to the current round of training. Understandably, the stronger the training influence of a tuple, the more it indicates the need for data conflict correction to improve model training accuracy.

[0054] As an example, in step S404, the computer device uses a preset conflict resolution tool to resolve data conflicts in the data corresponding to the key attributes of each tuple training sample in the current training set. Furthermore, the computer device uses the preset conflict resolution tool to resolve data conflicts in the training samples of key tuples in the current training set line by line. After resolving the data conflicts in the training samples of key attributes and key tuples in the current training set, a revised training set corresponding to the current training set is obtained. In this example, the preset conflict resolution tools include, but are not limited to, Rock, Raha, Baran, and HoloClean. Understandably, resolving conflicts in the data corresponding to key attributes and key tuples that have a strong influence on model optimization and training yields a revised training set with higher data reliability, facilitating the next round of more accurate classification model training based on the revised training set.

[0055] In this embodiment, after the current round of training ends, the current training set is further partitioned based on the current training loss function value corresponding to each training sample. This allows for the accurate identification of key attributes and key tuples. Based on these accurately identified attributes and tuples, data conflict repair is performed on the current training set, resulting in a repaired training set. This repaired training set is then used for the next round of more accurate classification model training. This method achieves data repair during model training, synergizing data conflict repair with model training. It effectively avoids the disconnect between data conflict repair and model training, improving the efficiency and accuracy of model training.

[0056] In one embodiment, such as Figure 5 As shown, step S402, which involves filtering candidate attributes for each training sample in the current training set based on a classification model with the same structure as the target model, a pre-training set, and a validation set, to determine the key attributes corresponding to the current training set, includes: S501: Pre-clean the S candidate attributes corresponding to all training samples in the pre-training set, and determine the S attribute-cleaned training sets corresponding to the S candidate attributes respectively. S502: Train a classification model with the same structure as the target model based on the pre-training set to determine the first classification model; verify the accuracy of the first classification model based on the validation set to determine the model accuracy corresponding to the first classification model; S503: Based on the S attribute-cleaned training set corresponding to the S candidate attributes, train a classification model with the same structure as the target model to determine the S secondary classification models corresponding to the S candidate attributes; based on the validation set, verify the accuracy of the S secondary classification models to determine the model accuracy corresponding to the S secondary classification models. S504: Based on the model accuracy corresponding to the first classification model and the model accuracy corresponding to the S second classification models, determine the S accuracy improvement rates. The candidate attributes corresponding to the top M second classification models with the largest accuracy improvement rates are determined as the key attributes corresponding to the current round of training set, where S≥M>1.

[0057] Pre-cleaning refers to the process of cleaning the data in the pre-training set to identify key attributes among candidate attributes. The attribute-cleaned training set refers to the training set obtained by cleaning the data corresponding to a candidate attribute in the pre-training set separately.

[0058] As an example, in step S501, the computer device pre-cleans the attribute data of the S candidate attributes corresponding to all training samples in the training set, determines the attribute-cleaned training set corresponding to each pre-cleaned candidate attribute, and obtains S second classification models corresponding to the S candidate attributes. In this example, for the S candidate attributes contained in the training samples of each tuple in the pre-training set, the computer device performs S pre-cleansing on the pre-training set, pre-cleansing only the data corresponding to one candidate attribute in the pre-training set each time, and obtains attribute-cleaned training sets after cleaning the S candidate attributes individually. For example, in an income prediction scenario, the training samples for each tuple in the pre-training set include four candidate attributes: user's education level, educational attainment, working hours, and occupation. The computer device pre-cleans the data corresponding to the education level for each tuple in the pre-training set to determine the cleaned training set for the corresponding attributes. Similarly, the computer device pre-cleans the data corresponding to the educational attainment for each tuple in the pre-training set to determine the cleaned training set for the corresponding attributes. Likewise, the computer device pre-cleans the data corresponding to the working hours for each tuple in the pre-training set to determine the cleaned training set for the corresponding attributes. Finally, the computer device pre-cleans the data corresponding to the occupation for each tuple in the pre-training set to determine the cleaned training set for the corresponding attributes.

[0059] The first classification model is a classification model trained on a pre-training set. The accuracy of the first classification model refers to the accuracy with which it performs classification.

[0060] As an example, in step S502, the computer device inputs each training sample from the pre-training set into a classification model with the same structure as the target model, trains the classification model with the same structure as the target model, obtains a first classification model, and uses a validation set to verify the accuracy of the first classification model, thus determining the model accuracy corresponding to the first classification model. For example, in an income prediction scenario, the computer device uses training samples from each tuple in the pre-training set corresponding to the income prediction scenario to train a classification model with the same structure as the target model, obtains a first classification model, and uses a validation set from the pre-training set corresponding to the income prediction scenario to verify the accuracy of the first classification model, thus determining the model accuracy corresponding to the first classification model.

[0061] The second classification model refers to the classification model trained using the cleaned training set corresponding to each candidate attribute. The number of second classification models is the same as the number of candidate attributes. The model accuracy of the second classification model refers to the accuracy with which the second classification model performs the classification.

[0062] As an example, in step S503, the computer device uses the attribute cleaning training set corresponding to the S candidate attributes to train the classification model with the same structure as the target model, determines a second classification model corresponding to each candidate attribute, and then determines the second classification model corresponding to the S pre-cleaned candidate attributes. Based on the validation set, the accuracy of the second classification model corresponding to each candidate attribute is verified to determine the model accuracy of the S second classification models.

[0063] For example, in the income prediction scenario, S=4. The computer equipment uses the attribute cleaning training set corresponding to the education level to train a classification model with the same structure as the target model, determines a second classification model corresponding to the education level, and verifies the accuracy of the second classification model corresponding to the education level based on the validation set, thus determining the model accuracy of the second classification model corresponding to the education level.

[0064] The computer equipment uses the attribute cleaning training set corresponding to the level of education to train a classification model with the same structure as the target model, determines a second classification model corresponding to the level of education, and verifies the accuracy of the second classification model corresponding to the level of education based on the validation set, thus determining the model accuracy of the second classification model corresponding to the level of education.

[0065] The computer equipment uses the attribute cleaning training set corresponding to the working time to train a classification model with the same structure as the target model, determines a second classification model corresponding to the working time, and verifies the accuracy of the second classification model corresponding to the working time based on the validation set, thus determining the model accuracy of the second classification model corresponding to the working time.

[0066] The computer equipment uses the attribute cleaning training set corresponding to occupations to train a classification model with the same structure as the target model, and determines a second classification model corresponding to occupations. Based on the validation set, the accuracy of the second classification model corresponding to occupations is validated, and the model accuracy of the second classification model corresponding to occupations is determined.

[0067] As an example, in step S504, the computer device compares the model accuracy corresponding to the S candidate attributes with the model accuracy corresponding to the first classification model to determine the accuracy improvement of the model accuracy corresponding to the S candidate attributes, and identifies the top M second classification models with the largest accuracy improvement. The candidate attributes corresponding to the top M second classification models are then identified as the key attributes of the current training set. In this example, the accuracy improvement is the difference between the model accuracy and the model accuracy corresponding to the first classification model. For example, in an income prediction scenario, the computer device compares the model accuracy of the second classification model corresponding to education level with the model accuracy corresponding to the first classification model to determine the accuracy improvement after pre-cleaning the data corresponding to the candidate attribute of education level. The computer device compares the model accuracy of the second classification model corresponding to education level with the model accuracy corresponding to the first classification model to determine the accuracy improvement after pre-cleaning the data corresponding to the candidate attribute of education level. The computer device compares the model accuracy of the second classification model corresponding to working hours with the model accuracy corresponding to the first classification model to determine the accuracy improvement after pre-cleaning the data corresponding to the candidate attribute of working hours. The computer equipment compares the accuracy of the second classification model corresponding to occupation with the accuracy of the first classification model to determine the accuracy improvement after pre-cleaning the data corresponding to the candidate attribute of occupation. The computer equipment then further analyzes the accuracy improvement after pre-cleaning the data corresponding to the candidate attributes of education level, work duration, and occupation, and identifies the top M candidate attributes with the largest improvements in accuracy as the key attributes for the current round of training. For example, when M=2, the key attributes are education level and occupation.

[0068] Understandably, the greater the improvement in model accuracy after pre-cleaning the data corresponding to the candidate attribute, the more significant the importance and impact of the candidate attribute on model optimization. Therefore, candidate attributes that have been pre-cleaned with a large improvement in accuracy are identified as key attributes, so as to facilitate subsequent data conflict repair of key attributes and accurate optimization of the target model.

[0069] In this embodiment, the attribute data corresponding to the S candidate attributes in the training samples of the pre-training set are pre-cleaned to determine the attribute-cleaned training sets corresponding to the S candidate attributes. The model is then trained using the attribute-cleaned training sets corresponding to the S candidate attributes. The accuracy improvement of the trained model before and after data pre-cleaning for each candidate attribute is determined. The candidate attributes corresponding to the top M models with the largest accuracy improvement are identified as the key attributes for the current round. This method is reasonable. Before the actual data conflict repair, this method evaluates the potential benefits through pre-cleaning and model training to determine the key attributes, so that the data conflict repair can be effectively carried out based on the key attributes. This effectively reduces secondary noise and does not require manual intervention. The key attributes can be determined by the attribute evaluation mechanism in this method, which can dynamically and efficiently determine the key attributes that need to be repaired.

[0070] In one embodiment, such as Figure 6 As shown, step S403, which involves filtering tuples from training samples in the current training set based on a classification model, pre-training set, and validation set with the same structure as the target model, to determine the key tuples corresponding to the current training set, includes: S601: Determine the sample candidate pool based on the current training loss function values ​​corresponding to Y tuples of training samples in the pre-training set; the sample candidate pool includes L tuples of training samples; where Y≥L>1; S602: Pre-clean the training samples of L tuples in the sample candidate pool, and determine the L sample clean training sets corresponding to the L tuples respectively. S603: Based on the L samples corresponding to L tuples, clean the training set and train a classification model with the same structure as the target model to determine the L third classification models corresponding to the L tuples; based on the validation set, verify the accuracy of the L third classification models to determine the model accuracy corresponding to the L third classification models. S604: Based on the model accuracy corresponding to the first classification model and the model accuracy corresponding to L third classification models, determine the L accuracy improvement rates. The tuples corresponding to the top N third classification models with the largest accuracy improvement rates are determined as the key tuples corresponding to the current round of training set, where N>1.

[0071] The sample candidate pool refers to the candidate pool corresponding to the training samples that need to be pre-cleaned, including multiple training samples that need to be pre-cleaned.

[0072] As an example, in step S601, the computer device adds the a training samples with larger current training loss function values ​​from the Y tuple training samples in the pre-training set to the sample candidate pool. The computer device uses a confidence-weighted sampling method to sample (Y-a) training samples in the pre-training set, excluding the a training samples that have already been added to the sample candidate pool, to obtain b sampled training samples. The b sampled training samples are added to the sample candidate pool to ensure the representativeness and diversity of the training samples in the sample candidate pool, resulting in a sample candidate pool containing L training samples, where a+b=L, L>a>1, and L>b>1.

[0073] Among them, the sample cleaning training set refers to the training set obtained after pre-cleaning the data in the training samples of a single tuple.

[0074] As an example, in step S602, the computer device pre-cleans the data corresponding to the training samples of one tuple in the sample candidate pool each time. Each pre-cleaning results in one tuple being pre-cleaned, while the remaining tuples are not pre-cleaned, thus completing the pre-cleaning of the training samples of L tuples in the sample candidate pool and determining the L cleaned training sets corresponding to the L tuples. For example, in the income prediction scenario, the computer device determines that there are L training samples of L tuples in the sample candidate pool. The computer device first pre-cleans the training sample of the first tuple, and then pre-cleans the remaining L tuples. Training samples with 1 tuple are not pre-cleaned, resulting in a cleaned training set where only the first tuple is pre-cleaned. This process is repeated, with the computer pre-cleaning the training samples with the Lth tuple and not pre-cleaning the training samples with the first L-1 tuples, resulting in a cleaned training set where only the Lth tuple is pre-cleaned, and so on, to obtain L cleaned training sets.

[0075] The third classification model refers to the classification model trained using the cleaned training set corresponding to each tuple. The number of third classification models is the same as the number of cleaned training sets. The model accuracy of the third classification model refers to the accuracy of its classification.

[0076] As an example, in step S603, the computer device uses the L samples corresponding to the L tuples to clean the training set, and trains the classification model with the same structure as the target model to determine the L third classification models corresponding to the L tuples; the validation set is used to verify the accuracy of the L third classification models to determine the model accuracy corresponding to the L third classification models.

[0077] For example, in a revenue prediction scenario, the computer equipment uses a training set with samples pre-cleaned only for the first tuple to train a classification model with the same structure as the target model. This determines the third-class classification model trained with the first tuple pre-cleaned. Based on the validation set, the accuracy of this third-class classification model is validated, determining the model accuracy with the first tuple pre-cleaned. Similarly, the computer equipment uses a training set with samples pre-cleaned only for the Lth tuple to train a classification model with the same structure as the target model. This determines the third-class classification model trained with the Lth tuple pre-cleaned. Based on the validation set, the accuracy of this third-class classification model is validated, determining the model accuracy with the Lth tuple pre-cleaned. This yields the model accuracy corresponding to each training sample pre-cleaned for each of the first to Lth tuples in a single iteration.

[0078] As an example, in step S604, the computer device compares the model accuracy corresponding to the L third-class classification models with the model accuracy corresponding to the first-class classification model, determines the accuracy improvement of each tuple in the L third-class classification models, and identifies the top N tuples with the largest accuracy improvement as the key tuples for the current round of training. In this example, the accuracy improvement is the difference between the model accuracy corresponding to the third-class classification model and the model accuracy corresponding to the first-class classification model.

[0079] For example, in an income prediction scenario, the computer device cleans the training samples of the first to Lth tuples and then trains the model accuracy of the third classification model. The difference between the model accuracy of the third classification model and the model accuracy of the first classification model is compared to determine the accuracy improvement of the third classification model for each tuple. The tuples corresponding to the top N third classification models with the largest accuracy improvement are identified as the key tuples for the current round of training.

[0080] Understandably, the greater the improvement in model accuracy after the training samples of a tuple are pre-cleaned, the more significant the importance and influence of the training samples of that tuple on model optimization. Therefore, the tuples corresponding to the third classification model with a large improvement in accuracy are identified as key tuples, so as to facilitate data conflict repair for key tuples with a large influence on model training and optimization, and to accurately optimize the target model.

[0081] In this embodiment, by pre-cleaning the training samples of L tuples in the pre-training set, the cleaned training set corresponding to the L tuples is determined. The model is then trained using the cleaned training set corresponding to the L tuples, and the accuracy improvement of the model before and after pre-cleaning is determined. The top N tuples with the largest accuracy improvement are identified as the key tuples for the current round. This method is reasonable because it evaluates potential benefits through pre-cleaning and model training before actual data conflict repair, identifies key tuples, and facilitates effective data conflict repair based on these key tuples, effectively reducing secondary noise. It requires no manual intervention; the key tuples can be determined dynamically and efficiently using the tuple selection mechanism in this method.

[0082] In one embodiment, the method for determining whether the target model optimization is complete includes: Repair the sample set if it is an empty set, and / or if the accuracy improvement of the auxiliary model corresponding to multiple rounds of updates is less than a preset threshold. The improvement in the accuracy of the auxiliary model is determined based on the difference in accuracy between two adjacent rounds. The accuracy of the auxiliary model is determined by validating the updated auxiliary model using the validation set.

[0083] The preset magnitude threshold refers to a pre-defined threshold used to determine the magnitude of the improvement in the accuracy of the auxiliary model. Auxiliary model accuracy refers to the level of accuracy achieved by the auxiliary model.

[0084] As an example, when a computer device determines that the optimization update of the target model is complete and identifies the most recently updated target model as the target classification model during the execution of a classification model optimization method, the following two conditions must be met: Scenario 1: During the execution of steps S101 to S105, in step S101, when the repair sample set obtained by identifying the repair data based on the current training set and the previous training set is an empty set, it is determined that there are no training samples that need to be identified and repaired in the previous training set after the previous model training. The previous training set is directly determined as the current training set. The model has converged to a stable state. At this time, it is determined that the update of the target model can be stopped, and the most recently updated target model is determined as the target classification model.

[0085] Scenario 2: During each round of model optimization, after executing step S401 to determine the validation set, the computer equipment uses the validation set determined in each round to verify the accuracy of the auxiliary model after each round of updates, determines the accuracy of the auxiliary model corresponding to the auxiliary model, and obtains the accuracy of the auxiliary model in the continuous model optimization and update process. The difference between the accuracy of the auxiliary model in the next round and the accuracy of the auxiliary model in the adjacent previous round is determined as the improvement of model accuracy. If it is determined that the improvement of model accuracy in multiple consecutive rounds is lower than the preset threshold, it is determined that the marginal benefit region of multiple rounds of repair on model accuracy is stable. At this time, it is determined that the update of the target model can be stopped, and the target model of the most recently updated one is determined as the target classification model.

[0086] In this embodiment, by determining whether the repair sample set is empty and / or whether the accuracy improvement of the auxiliary model corresponding to the multi-round updated auxiliary model is lower than a preset threshold, a reasonable judgment is made on whether the target model has been optimized, so as to ensure that the target model is optimized and updated accurately and stably.

[0087] In another embodiment, such as Figure 7 As shown, a classification method is provided, which is then applied to... Figure 9 Taking a computer device as an example, the explanation includes the following steps: S701: Obtain the data to be classified; S702: Input the data to be classified into the target classification model to predict the target category corresponding to the data to be classified; The target classification model is determined based on the classification model optimization method in the above embodiments.

[0088] Here, the data to be classified refers to the data used for classification.

[0089] As an example, in step S701, the computer device acquires data to be classified corresponding to multiple attributes in the application scenario, so as to classify the data. For example, in an income prediction scenario, the computer device acquires data to be classified corresponding to a user's education level, educational attainment, working hours, and occupation, so that it can subsequently identify these data according to the target classification model corresponding to the income prediction scenario, and determine the categories corresponding to the multiple data to be classified.

[0090] As an example, in step S702, the computer device inputs the data to be classified into the target classification model and outputs the target category corresponding to the data to be classified. For example, in an income prediction scenario, the data to be classified corresponding to a user's education level, educational attainment, working hours, and occupation are input into the target classification model corresponding to the income prediction scenario to predict the category corresponding to the user's income level.

[0091] In this embodiment, a target classification model is used to directly identify and classify the data to be classified, thereby achieving the goal of accurate and efficient classification of the data.

[0092] In this embodiment, as Figure 10 As shown, the classification model optimization method includes a creator model optimization stage and a critic data conflict repair stage. In the creator model optimization stage, the computer device executes steps S101 to S103 and step S105 to perform dual-model assistance based on the auxiliary model and the target model, optimizing and updating the target model to obtain a target classification model with higher classification accuracy. In the critic data conflict repair stage, the computer device executes step S104 to repair data conflicts in the current training set based on the current training loss function value output after the auxiliary model update and a classification model with the same structure as the target model. This repaired training set is then used to coordinate with the creator model optimization stage, repairing the training set for the creator model optimization stage and assisting in accurate model optimization and updates.

[0093] In this embodiment, the creator model optimization stage is also used to execute steps S201 to S204 and steps S301 to S303 to reasonably determine the new clean sample set for the current round.

[0094] In this embodiment, the critic data conflict repair stage is also used to execute steps S401 to S404, S501 to S504, and S601 to S604. These steps are used to construct an attribute evaluation mechanism and a tuple screening mechanism without human intervention. This is used to accurately determine the key attributes and key tuples that need data conflict repair, improve the accuracy of the data in the current training set, avoid the negative impact of data repair, and achieve efficient and stable key attribute evaluation and key tuple screening. Only high-value data is repaired, which reduces the risk of erroneous modification while maximizing the improvement of model accuracy.

[0095] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0096] In one embodiment, a classification model optimization apparatus is provided, which corresponds one-to-one with the classification model optimization methods in the above embodiments. For example... Figure 8 As shown, the classification model optimization device includes a sample set repair determination module 801, an auxiliary model update module 802, a target model optimization module 803, a data conflict repair module 804, and a target classification model determination module 805. Detailed descriptions of each functional module are as follows: The repair sample set determination module 801 identifies the repair data based on the current training set and the previous training set to determine the repair sample set. The auxiliary model update module 802 updates the auxiliary model based on the current training set and the repaired sample set, determines the current training loss function value corresponding to each training sample in the current training set, and determines the new clean sample set for the current round. The target model optimization module 803 determines the cumulative clean sample set for the current round based on the cumulative clean sample set of the previous round and the newly added clean sample set for the current round, and optimizes the target model based on the cumulative clean sample set for the current round to determine the target model for the current round. The data conflict repair module 804 performs data conflict repair on the current training set based on the current training loss function value corresponding to each training sample in the current round training set and the classification model with the same structure as the target model, and determines the repaired training set. The target classification model determination module 805 is used to update the repaired training set to the current round training set, repeatedly perform the repair data identification based on the current round training set and the previous round training set, determine the repair sample set, until the target model is optimized and the target classification model is obtained.

[0097] In one embodiment, the auxiliary model update module 802 includes: The impact estimate determination submodule is used to determine the impact estimate corresponding to each training sample in the repair sample set. The impact estimate is the evaluation value corresponding to the degree of influence of the training samples in the repair sample set on the auxiliary model update. The forgetting update processing submodule performs forgetting update processing on the auxiliary model based on the impact estimates corresponding to all training samples in the repair sample set and the number of tuples corresponding to the previous training set, and determines the forgotten auxiliary model. The fine-tuning and updating submodule is used to sequentially input each training sample in the current training set into the forgotten auxiliary model, perform multiple rounds of parameter fine-tuning and updating of the forgotten auxiliary model, and determine the updated auxiliary model and the current training loss function value corresponding to each training sample in the current training set. The newly added clean sample set determination submodule determines the new clean sample set for the current round based on the current training loss function value and the repaired sample set corresponding to each training sample in the current round training set.

[0098] In one embodiment, the newly added clean sample set determination submodule includes: The historical training loss function value acquisition unit is used to acquire multiple historical training loss function values ​​corresponding to each training sample in the current round of training set. The multiple historical training loss function values ​​are training loss function values ​​formed by multiple rounds of training before the current round. A new clean sample determination unit is added, which is used to filter all training samples in the current round training set based on the current training loss function value and multiple historical training loss function values ​​corresponding to the training samples of all tuples, and determine the new clean samples for the current round. A new clean sample set determination unit is added to determine the new clean sample set for the current round based on all the new clean samples and repaired sample sets in the current round.

[0099] In one embodiment, the data conflict repair module 804 includes: The set is divided into sub-modules. Based on the current training loss function value corresponding to each training sample in the current round of training set, the training samples in the current round of training set are filtered to determine the pre-training set and the validation set. The attribute filtering submodule, based on a classification model with the same structure as the target model, pre-training set and validation set, performs attribute filtering on the candidate attributes corresponding to each training sample in the current round of training set, and determines the key attributes corresponding to the current round of training set. The tuple filtering submodule, based on a classification model with the same structure as the target model, pre-training set, and validation set, filters tuples from the training samples in the current round of training set to determine the key tuples corresponding to the current round of training set. The data conflict repair submodule is used to repair data conflicts in the key attributes and key tuples of the current round of training set and determine the repaired training set.

[0100] In one embodiment, the attribute filtering submodule includes: The candidate attribute pre-cleaning unit is used to pre-clean the S candidate attributes corresponding to all training samples in the pre-training set, and to determine the S attribute-cleaned training set corresponding to the S candidate attributes respectively. The first model training and validation unit trains a classification model with the same structure as the target model based on the pre-training set to determine the first classification model; and validates the accuracy of the first classification model based on the validation set to determine the model accuracy corresponding to the first classification model. The second model training and validation unit trains a classification model with the same structure as the target model based on a cleaned training set corresponding to S candidate attributes, thereby determining S second classification models corresponding to the S candidate attributes; and validates the accuracy of the S second classification models based on the validation set, thereby determining the model accuracy corresponding to the S second classification models. The key attribute determination unit determines S accuracy improvement rates based on the model accuracy of the first classification model and the model accuracy of the S second classification models. The candidate attributes corresponding to the top M second classification models with the largest accuracy improvement rates are determined as the key attributes of the current training set, where S≥M>1.

[0101] In one embodiment, the tuple filtering submodule includes: The sample candidate pool determination unit determines the sample candidate pool based on the current training loss function values ​​corresponding to Y tuples of training samples in the pre-training set; the sample candidate pool includes L tuples of training samples; where Y≥L>1; The tuple pre-cleaning unit is used to pre-clean the training samples of L tuples in the sample candidate pool, and to determine the L sample clean training sets corresponding to the L tuples respectively. The third model training and validation unit cleans the training set of L samples corresponding to L tuples, trains the classification model with the same structure as the target model, and determines L third classification models corresponding to L tuples; and validates the accuracy of L third classification models based on the validation set to determine the model accuracy corresponding to L third classification models. The key tuple determination unit determines L precision improvement rates based on the model precision of the first classification model and the model precision of L third classification models. The tuples corresponding to the top N third classification models with the largest precision improvement rates are determined as the key tuples corresponding to the current round of training set, where N > 1.

[0102] In one embodiment, a classification device is provided, which corresponds one-to-one with the classification methods in the above embodiments. The classification device includes the following functional modules, detailed below: The module for acquiring data to be classified is used to acquire the data to be classified. The classification module is used to input the data to be classified into the target classification model and predict the target category corresponding to the data to be classified.

[0103] Specific limitations regarding the classification model optimization device can be found in the limitations regarding the classification model optimization method above, and specific limitations regarding the classification device can be found in the limitations regarding the classification method above; they will not be repeated here. The aforementioned classification model optimization device and its modules can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0104] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data used or generated during the execution of classification model optimization methods and classification methods. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a classification model optimization method, or, when executed by the processor, implements a classification method.

[0105] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the classification model optimization method described in the above embodiment, for example... Figure 1 As shown in S101-S105, or Figures 2 to 6 As shown, to avoid repetition, it will not be described again here. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in this embodiment of the classification model optimization device, for example... Figure 8 The functions of the repair sample set determination module 801, auxiliary model update module 802, target model optimization module 803, data conflict repair module 804, and target classification model determination module 805 shown are not described in detail here to avoid repetition.

[0106] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the classification method described in the above embodiments, for example... Figure 7S701-S702, as shown, will not be described again here to avoid repetition. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in this embodiment of the classification device, such as the functions of the data acquisition module and the classification module. To avoid repetition, these will not be described again here.

[0107] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the classification model optimization method described in the above embodiment, for example... Figure 1 As shown in S101-S105, or Figures 2 to 6 As shown, to avoid repetition, it will not be described again here. Alternatively, when the computer program is executed by the processor, it implements the functions of each module / unit in this embodiment of the classification model optimization device, for example... Figure 8 The functions of the repair sample set determination module 801, auxiliary model update module 802, target model optimization module 803, data conflict repair module 804, and target classification model determination module 805 shown are not described again here to avoid repetition. The computer-readable storage medium can be non-volatile or volatile.

[0108] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the classification method described in the above embodiments, for example... Figure 7 S701-S702, as shown, will not be described again here to avoid repetition. Alternatively, when the computer program is executed by the processor, it implements the functions of each module / unit in this embodiment of the classification device, such as the functions of the data acquisition module and the classification module. To avoid repetition, these will not be described again here. The computer-readable storage medium can be non-volatile or volatile.

[0109] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0111] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A classification model optimization method, characterized in that, include: Based on the current training set and the previous training set, identify the repair data and determine the repair sample set; The auxiliary model is updated based on the current training set and the repaired sample set. The current training loss function value corresponding to each training sample in the current training set is determined, and the new clean sample set for the current round is determined. Based on the cumulative clean sample set of the previous round and the newly added clean sample set of the current round, determine the cumulative clean sample set of the current round, optimize the target model based on the cumulative clean sample set of the current round, and determine the target model of the current round. Based on the current training loss function value corresponding to each training sample in the current round training set and the classification model with the same structure as the target model, data conflict repair is performed on the current round training set to determine the repaired training set. The repaired training set is updated to the current round training set. The process of identifying repaired data based on the current round training set and the previous round training set is repeated to determine the repaired sample set until the target model is optimized and the target classification model is obtained.

2. The classification model optimization method according to claim 1, characterized in that, The process of updating the auxiliary model based on the current training set and the repaired sample set, determining the current training loss function value for each training sample in the current training set, and determining the newly added clean sample set for the current round, includes: Determine the impact estimate for each training sample in the repair sample set, wherein the impact estimate is an evaluation value corresponding to the degree of influence of the training samples in the repair sample set on the update of the auxiliary model; Based on the impact estimates corresponding to all training samples in the repair sample set and the number of tuples corresponding to the previous training set, the auxiliary model is subjected to forgetting update processing to determine the forgotten auxiliary model. Each training sample in the current training set is sequentially input into the forgotten auxiliary model. The parameters of the forgotten auxiliary model are fine-tuned and updated multiple times to determine the updated auxiliary model and the current training loss function value corresponding to each training sample in the current training set. Based on the current training loss function value corresponding to each training sample in the current round training set and the repaired sample set, the new clean sample set for the current round is determined.

3. The classification model optimization method according to claim 2, characterized in that, The step of determining the new clean sample set for the current round based on the current training loss function value corresponding to each training sample in the current round training set and the repaired sample set includes: Obtain multiple historical training loss function values ​​corresponding to each training sample in the current training set. These multiple historical training loss function values ​​are training loss function values ​​formed from multiple training rounds prior to the current round. Based on the current training loss function value and multiple historical training loss function values ​​corresponding to the training samples of all tuples, sample screening is performed on all training samples in the current round training set to determine the newly added clean samples for the current round. Based on all the newly added clean samples in the current round and the repaired sample set, determine the newly added clean sample set for the current round.

4. The classification model optimization method according to claim 1, characterized in that, The process of performing data conflict repair on the current training set based on the current training loss function value corresponding to each training sample in the current round of training and a classification model with the same structure as the target model, and determining the repaired training set, includes: Based on the current training loss function value corresponding to each training sample in the current round of training set, the training samples in the current round of training set are screened to determine the pre-training set and the validation set; Based on the classification model with the same structure as the target model, the pre-training set and the validation set, attribute filtering is performed on the candidate attributes corresponding to each training sample in the current round of training set to determine the key attributes corresponding to the current round of training set. Based on the classification model with the same structure as the target model, the pre-training set, and the validation set, tuples are filtered from the training samples in the current round of training to determine the key tuples corresponding to the current round of training. Data conflict repair is performed on the key attributes and key tuples corresponding to the current round of training set to determine the repaired training set.

5. The classification model optimization method according to claim 4, characterized in that, The method, based on a classification model with the same structure as the target model, the pre-training set, and the validation set, performs attribute filtering on candidate attributes corresponding to each training sample in the current round of training to determine the key attributes corresponding to the current round of training, including: Pre-clean the S candidate attributes corresponding to all training samples in the pre-training set, and determine the S attribute-cleaned training sets corresponding to the S candidate attributes respectively. Based on the pre-training set, a classification model with the same structure as the target model is trained to determine a first classification model; based on the validation set, the accuracy of the first classification model is validated to determine the model accuracy corresponding to the first classification model. Based on the S attribute-cleaned training set corresponding to the S candidate attributes, a classification model with the same structure as the target model is trained to determine S second classification models corresponding to the S candidate attributes; based on the validation set, the accuracy of the S second classification models is verified to determine the model accuracy corresponding to the S second classification models. Based on the model accuracy corresponding to the first classification model and the model accuracy corresponding to S second classification models, determine S accuracy improvement rates. Then, determine the candidate attributes corresponding to the top M second classification models with the largest accuracy improvement rates as the key attributes corresponding to the current round of training set, where S≥M>1.

6. The classification model optimization method according to claim 5, characterized in that, The step of filtering training samples in the current training set based on a classification model with the same structure as the target model, the pre-training set, and the validation set to determine the key tuples corresponding to the current training set includes: Based on the current training loss function values ​​corresponding to the Y tuples of training samples in the pre-training set, a sample candidate pool is determined; the sample candidate pool includes L tuples of training samples; where Y≥L>1; The training samples of L tuples in the sample candidate pool are pre-cleaned, and the L sample cleaned training sets corresponding to the L tuples are determined respectively. Based on the L samples cleaned training set corresponding to L tuples, a classification model with the same structure as the target model is trained to determine L third classification models corresponding to L tuples; based on the validation set, the accuracy of the L third classification models is verified to determine the model accuracy corresponding to the L third classification models. Based on the model accuracy corresponding to the first classification model and the model accuracy corresponding to L third classification models, determine L accuracy improvement rates. Then, determine the tuples corresponding to the top N third classification models with the largest accuracy improvement rates as the key tuples corresponding to the current round of training set, where N > 1.

7. The classification model optimization method according to claim 4, characterized in that, The methods for determining whether the target model optimization is complete include: The repair sample set is an empty set, and / or the accuracy improvement of the auxiliary model corresponding to the multiple rounds of updates is less than a preset threshold. The improvement in the accuracy of the auxiliary model is determined based on the difference in accuracy between two adjacent rounds of the auxiliary model; The accuracy of the auxiliary model is determined by validating the updated auxiliary model using the validation set.

8. A classification method, characterized in that, include: Obtain the data to be classified; The data to be classified is input into the target classification model to predict the target category corresponding to the data to be classified. The target classification model is determined based on the classification model optimization method according to any one of claims 1 to 7.

9. A classification model optimization device, characterized in that, include: The repair sample set determination module identifies the repair data based on the current training set and the previous training set to determine the repair sample set. The auxiliary model update module updates the auxiliary model based on the current training set and the repair sample set, determines the current training loss function value corresponding to each training sample in the current training set, and determines the new clean sample set for the current round. The target model optimization module determines the cumulative clean sample set for the current round based on the cumulative clean sample set from the previous round and the newly added clean sample set for the current round. Based on the cumulative clean sample set for the current round, it optimizes the target model to determine the target model for the current round. The data conflict repair module performs data conflict repair on the current training set based on the current training loss function value corresponding to each training sample in the current round training set and the classification model with the same structure as the target model, and determines the repaired training set. The target classification model determination module is used to update the repaired training set to the current round training set, repeatedly perform the repair data identification based on the current round training set and the previous round training set, determine the repair sample set, until the target model is optimized and the target classification model is obtained.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the classification model optimization method as described in any one of claims 1 to 7, or when the processor executes the computer program, it implements the classification method as described in claim 8.

11. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the classification model optimization method as described in any one of claims 1 to 7, or, when the computer program is executed by the processor, it implements the classification method as described in claim 8.

Images