Machine learning model interpretation method, device and storage medium

Abstract

The present disclosure provides a machine learning model interpretation method, device and storage medium, which relate to the field of artificial intelligence technology. The method includes: training a first machine learning model based on network performance data, obtaining feature weights corresponding to each feature contained in the network performance data, and selecting model features based on the feature weights; training a second machine learning model based on the model features, and obtaining a decision tree structure corresponding to the second machine learning model; determining a judgment selection path in the decision tree structure as an explanatory work path; the method, device and storage medium of the present disclosure reduce the complexity of the algorithm by arranging the complexity of the problem to be solved and trimming the features; interpretable work is performed through a decision tree model, the process is simple, the operation is convenient and efficient, and the decision basis of the machine learning model can be accurately obtained, taking into account simplicity and accuracy.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a machine learning model interpretation method, device, and storage medium. Background Art

[0002] In recent years, machine learning technology has flourished and has been widely used in all walks of life. In the application of machine learning models, classification or regression prediction is usually performed through machine learning modeling, and feedback is given for the input information and the conclusion is output, which is a "black box" application. In business scenarios such as operator mobile network optimization, it is necessary not only to build a machine learning model to realize the classification of network quality, but also to understand the weight analysis, judgment strategy and threshold value selection of many input features within the model, master the mathematical relationship between network quality and many features, and then optimize the key features in a targeted manner to completely solve the network quality problem. Therefore, a technical solution that can interpret machine learning models is needed. Summary of the invention

[0003] In view of this, a technical problem to be solved by the present invention is to provide a machine learning model interpretation method, device and storage medium.

[0004] According to a first aspect of the present disclosure, a machine learning model interpretation method is provided, including: obtaining network performance data, and training a first machine learning model based on the network performance data; after the training of the first machine learning model is completed, obtaining feature weights corresponding to each feature contained in the network performance data; selecting model features from all features contained in the network performance data based on the feature weights; processing the network performance data based on the model features to construct a training set; using the training set to train a second machine learning model to obtain a decision tree structure corresponding to the second machine learning model; wherein the decision nodes in the decision tree structure are used to characterize model feature judgments and are provided with feature judgment information; determining a judgment selection path in the decision tree structure according to the feature judgment information as an explanatory working path for the second machine learning model.

[0005] Optionally, the second machine learning model includes: a decision tree model; determining the judgment selection path in the decision tree structure according to the feature judgment information includes: obtaining multiple child decision nodes of the decision node to be judged in the decision tree structure; determining a judgment target node from the multiple child decision nodes based on the feature judgment information of each child decision node; and determining the path between the decision node to be judged and the judgment target node as the judgment selection path.

[0006] Optionally, the feature judgment information includes: a Gini coefficient and the number of sample coverages of this node; and determining a decision target node from the multiple child decision nodes includes: determining a decision target node from the multiple child decision nodes based on the Gini coefficient and the number of sample coverages.

[0007] Optionally, the number of the multiple child decision nodes is two; determining a decision target node from the multiple child decision nodes based on the Gini coefficient and the sample coverage number includes: selecting a child decision node with a smaller Gini coefficient and a larger sample coverage number from the two child decision nodes as the decision target node.

[0008] Optionally, the Pearson correlation coefficient between the network performance data is calculated; and the network performance data is screened based on the Pearson correlation coefficient.

[0009] Optionally, selecting model features from all features included in the network performance data based on the feature weights includes: sorting all the features based on the feature weights; and selecting multiple features with the highest feature weights as the model features based on the sorting results.

[0010] Optionally, the first machine learning model includes: a random forest model; and the network performance data includes: cell performance data.

[0011] According to a second aspect of the present disclosure, a machine learning model interpretation device is provided, comprising: a first training module, used to obtain network performance data, and train a first machine learning model based on the network performance data; a weight acquisition module, used to obtain feature weights corresponding to each feature contained in the network performance data after the first machine learning model is trained; a feature selection module, used to select model features from all features contained in the network performance data based on the feature weights; a data processing module, used to process the network performance data based on the model features to construct a training set; a second training module, used to train a second machine learning model using the training set, and obtain a decision tree structure corresponding to the second machine learning model; wherein the decision nodes in the decision tree structure are used to characterize model feature judgments, and are provided with feature judgment information; a path selection module, used to determine a judgment selection path in the decision tree structure according to the feature judgment information, as an explanatory working path of the second machine learning model.

[0012] Optionally, the second machine learning model includes: a decision tree model; wherein the path selection module includes: a node acquisition unit, used to obtain multiple child decision nodes of the decision node to be judged in the decision tree structure; a node selection unit, used to determine a judgment target node from the multiple child decision nodes based on feature judgment information of each child decision node; a path determination unit, used to determine the path between the decision node to be judged and the judgment target node as the judgment selection path.

[0013] Optionally, the feature judgment information includes: a Gini coefficient and the number of sample coverages of this node; wherein the node selection unit is specifically used to determine a judgment target node from the multiple child decision nodes based on the Gini coefficient and the number of sample coverages.

[0014] Optionally, the number of the multiple child decision nodes is two; wherein the node selection unit is specifically used to select a child decision node with a smaller Gini coefficient and a larger number of sample coverage from the two child decision nodes as the decision target node.

[0015] Optionally, the first training module is further used to calculate the Pearson correlation coefficient between the network performance data; and to screen the network performance data based on the Pearson correlation coefficient.

[0016] Optionally, the feature selection module is specifically used to sort all the features based on the feature weights; and select multiple features with the highest feature weights according to the sorting results as the model features.

[0017] According to a third aspect of the present disclosure, there is provided a machine learning model interpretation device, comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to execute the method as described above based on instructions stored in the memory.

[0018] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the instructions are executed by a processor according to the above machine learning model interpretation method.

[0019] The machine learning model interpretation method, device and storage medium disclosed in the present invention train a first machine learning model based on network performance data, obtain feature weights corresponding to each feature contained in the network performance data, and select model features based on the feature weights; train a second machine learning model based on the model features to obtain a decision tree structure corresponding to the second machine learning model; determine a judgment selection path in the decision tree structure as an explanatory work path; reduce the algorithm complexity by arranging the complexity of the problem to be solved and trimming the features; perform interpretable work through a decision tree model, with a simple process, convenient and efficient operation, and the decision basis of the machine learning model can be accurately obtained, taking into account both simplicity and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0021] Figure 1 A flowchart of an embodiment of a method for explaining a machine learning model according to the present disclosure;

[0022] Figure 2 This is a schematic diagram of the importance evaluation of each feature on the model results;

[0023] Figure 3 A schematic diagram of a flow chart for determining and judging a selection path in one embodiment of a machine learning model interpretation method according to the present disclosure;

[0024] Figure 4 A schematic diagram for visualizing nodes in a decision tree structure;

[0025] Figures 5A-5D A schematic diagram of the distribution of feature values;

[0026] Figure 6 A schematic diagram of modules of an embodiment of a machine learning model interpretation device according to the present disclosure;

[0027] Figure 7 A schematic diagram of a module of a path selection module in one embodiment of a machine learning model interpretation system according to the present disclosure;

[0028] Figure 8 A module diagram of another embodiment of a machine learning model interpretation device according to the present disclosure. DETAILED DESCRIPTION

[0029] The present disclosure is described more fully below with reference to the accompanying drawings, in which exemplary embodiments of the present disclosure are described. The technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure. The technical solutions of the present disclosure are described in many aspects below in conjunction with various figures and embodiments.

[0030] In order to solve the black box problem of machine learning models, the industry has conducted research on the interpretability of machine learning. Interpretability refers to the understanding of the internal mechanism of the model and the understanding of the model results, which represents the degree to which humans can understand the reasons for decision-making. Research on the interpretation of machine learning models is in its early stages, and there are still many problems that have not been solved. At present, a typical problem is that the interpretability of machine learning is restricted by the complexity of the model. In order to obtain clear and intuitive interpretability, users often choose models with simple structures, but such models are not necessarily the optimal models for practical problems. When users use complex models such as deep learning, although they can more accurately reflect and solve practical problems, their interpretability is poor.

[0031] Figure 1 FIG. 1 is a flow chart of an embodiment of a method for explaining a machine learning model according to the present disclosure, as shown in FIG. Figure 1 As shown:

[0032] Step 101, obtain network performance data, and train a first machine learning model based on the network performance data.

[0033] In one embodiment, the network performance data may be cell performance data, etc.; the first machine learning model may be a variety of models, such as a random forest model, etc. Random forest refers to a classifier that uses multiple trees to train and predict samples. The present disclosure may use an existing random forest model, and may use an existing training method to train a random forest model using network performance data.

[0034] Step 102, after the first machine learning model training is completed, obtain the feature weights corresponding to each feature included in the network performance data.

[0035] In one embodiment, the features included in the network performance data include cell number, cell location, cell bandwidth, etc., and feature weights corresponding to each feature such as cell number, cell location, cell bandwidth, etc. are obtained.

[0036] Step 103: Select model features from all features included in the network performance data based on feature weights.

[0037] Step 104: Process the network performance data based on the model features to construct a training set.

[0038] In one embodiment, the network performance data is pruned based on the model features so that each piece of network performance data only includes the model features, and a training set is constructed based on the pruned network performance data.

[0039] Step 105, use the training set to train the second machine learning model to obtain a decision tree structure corresponding to the second machine learning model; wherein the decision nodes in the decision tree structure are used to characterize model feature judgments and are provided with feature judgment information.

[0040] In one embodiment, the second machine learning model may be a decision tree model, etc. The decision tree model is a non-parametric classifier with fast calculation speed, easy to explain results, and strong robustness. The present disclosure may use an existing decision tree model.

[0041] In one embodiment, the learning capabilities of the first machine learning model and the second machine learning model are evaluated to determine whether the learning capability of the second machine learning model is not lower than the learning capability of the first machine learning model. If yes, proceed to step 106. If not, repeat steps 101-104 until a suitable feature is selected or the process ends. For example, if the difference between the evaluation indicators such as the accuracy and precision of the first machine learning model and the evaluation indicators such as the accuracy and precision of the corresponding second machine learning model is less than a preset threshold, it is determined that the learning capability of the second machine learning model is not lower than the learning capability of the first machine learning model.

[0042] Step 106, determine a judgment selection path in the decision tree structure according to the feature judgment information as an explanatory working path of the second machine learning model.

[0043] In one embodiment, the Pearson correlation coefficient between network performance data is calculated, and the network performance data is screened and processed based on the Pearson correlation coefficient. The network performance data is a wireless network performance indicator of a mobile communication network. The characteristics mainly include KPI, KQI, MR, etc., which respectively record network element key indicators, network element user experience indicators, and base station measurement report information, and describe the operating status of the telecommunications system from different levels. The number of indicators is large and shows a strong spatiotemporal correlation. In terms of time, the coupling degree between indicators at adjacent moments is large, and in terms of space, the indicators of adjacent base stations will affect each other.

[0044] The linear relationship between the features is analyzed using the Pearson coefficient, the data is standardized, and the Pearson correlation coefficient is selected. The Pearson correlation coefficient between the network performance data can be calculated using an existing method, and the network performance data can be screened based on the Pearson correlation coefficient. For example, the Pearson correlation coefficient between the network performance data is calculated, and the network performance data whose absolute value of the Pearson correlation coefficient is lower than or higher than or equal to a preset coefficient threshold is deleted.

[0045] In one embodiment, multiple methods can be used to select model features from all features included in the network performance data based on feature weights. For example, all features are sorted based on feature weights, and multiple features with the highest feature weights are selected as model features based on the sorting results. The number of model features can be set, for example, to 20, 30, etc.

[0046] With the parallel characteristics of the random forest model, the input features are fully analyzed and learned, and finally the features with greater impact on the target are found through the integrated weighting of each decision tree. For example, the RandomForestClassifier method in the sklearn tool library is used to build a random forest model. n_estimators is the tree parameter of the decision tree, which is set between 20 and 100, and other parameters use the default values.

[0047] After the random forest model training is completed, the feature_importances feature weights are visualized and output to obtain the feature weights corresponding to each feature contained in the network performance data. The top 20 features are selected in descending order of feature weights as model features, such as Figure 2 shown.

[0048] Figure 3 FIG. 1 is a flow chart of determining and judging a selection path in one embodiment of the machine learning model interpretation method disclosed in the present invention, as shown in FIG. Figure 3 As shown:

[0049] Step 301: In a decision tree structure, obtain multiple child decision nodes of a decision node to be determined. Child decision nodes may also be referred to as leaf decision nodes or leaf nodes.

[0050] Step 302: Based on the feature judgment information of each child decision node, a decision target node is determined from multiple child decision nodes.

[0051] Step 302: determine the path between the decision node to be determined and the judgment target node as the judgment selection path.

[0052] In one embodiment, the feature judgment information includes the Gini coefficient, the number of sample coverage of the node, etc. Based on the Gini coefficient and the number of sample coverage, a decision target node is determined from multiple child decision nodes. For example, if the number of multiple child decision nodes is two, a child decision node with a smaller Gini coefficient and a larger number of sample coverage is selected from the two child decision nodes as the decision target node.

[0053] Alternatively, for child decision nodes with a small Gini coefficient and a small number of sample coverage, and child decision nodes with a large Gini coefficient and a large number of sample coverage, a preset scoring rule can be used to calculate the score of each child decision node. There can be multiple scoring rules, for example, setting a first scoring table corresponding to the Gini coefficient, setting a second scoring table corresponding to the number of sample coverage, determining the first score and the second score corresponding to the Gini coefficient and the number of sample coverage of this node by querying the first scoring table and the second scoring table; and taking the child decision node with a higher sum of the first score and the second score as the decision target node.

[0054] In one embodiment, with the powerful screening ability of the random forest model for input features, 20 model features that play a key role in network optimization objectives can be obtained. The decision tree model performs modeling based on these model features. On the one hand, it solves the problem that the model itself has a weak screening ability for massive features. On the other hand, its good readability facilitates the next step of model interpretability.

[0055] For example, the DecisionTreeClassifier method in the sklearn tool library is used to build a decision tree model. To facilitate subsequent interpretation, the max_depth decision tree depth parameter is set to 3-4 shallow levels, and other parameters use the default values.

[0056] Through the export_graphviz method of the sklearn tool library, the trained decision tree model is visualized and output to obtain the decision tree structure corresponding to the decision tree model, such as Figure 4 shown. Figure 4 Each node (decision node) in represents a model feature judgment. The node can intuitively see the Gini coefficient (gini), the number of samples processed by this node (samples), the number of samples classified by this node and other information (feature judgment information). Select a path with a lower Gini coefficient and a larger number of covered samples as the interpretable working path, that is, the explanatory working path.

[0057] like Figure 4As shown in the figure, in the decision tree structure, two child decision nodes of the decision node to be judged are obtained, the Gini coefficient of the left child node is 0.0, and the number of samples covered by this node is 50; the Gini coefficient of the right child node is 0.5, and the number of samples covered by this node is 100. The scores of the two child decision nodes are calculated using the preset scoring rules, and the left child decision node with a higher score is selected from the two child decision nodes as the judgment target node.

[0058] The present disclosure provides a machine learning interpretation method for reducing the dimensionality of the problem complexity. For complex problems that simple machine learning models are not competent for, firstly, an advanced model is constructed to extract key features (model features) through embedded feature selection. Under the premise of ensuring that the connotation association value between the features and the target remains basically unchanged, the originally large and complex feature set is cut into a key feature subset with controllable quantity and complexity, and the dimensionality of the complex problem is reduced. Finally, a highly interpretable model is constructed on the feature subset for ex ante interpretation.

[0059] For example, in the work of wireless network quality optimization, the depth of the decision tree model is set to 3 layers. After visualization output, an intuitive decision tree structure can be obtained. The explanation branch is selected based on the Gini coefficient and sample coverage quantity of each decision node in the decision tree structure, that is, the judgment selection path is determined.

[0060] In one embodiment, the experimental data set selects the cell performance data (KPI) of a certain city within 5 days, and each record contains 667 features, including cell number, cell location, cell bandwidth, etc. There are 12,337 records in this data set, including 1,980 positive samples and 10,357 negative samples, with an imbalance ratio of 1:5.23. The data from T1 to T3 are used as the training set, the data from T4 is used as the test set, and the data from T5 is used as the validation set.

[0061] The learning results of the random forest model based on all features are: Accuracy = 73.72%, Precesion = 72.94%, Recall = 77%. The top 20 features are selected from 667 features as model features, including the number of successful establishment of additional E-RABs, wireless connection success rate, average number of downlink activated users, RRC connection reconstruction ratio, E-RAB data transmission time, downlink PRB average occupancy rate, etc.

[0062] The learning results of the decision tree model based on the above Top20 model features are: Accuracy = 75.79%, Precesion = 74.69%, Recall = 80.37%. It can be seen that after feature pruning and model changes, the learning ability of the decision tree is not weaker than that of the random forest model, and even exceeds it, indicating that the selection of model features is effective. The learning abilities of the two models for features and targets are basically the same, and the subsequent interpretability analysis based on the decision tree is fully reliable.

[0063] In the decision tree model, the explanation branch is selected according to the algorithm formula based on the Gini coefficient and sample coverage (number) of each node. For example, in this set of experimental data, the nodes with samples sample coverage > 1500 and gini Gini coefficient < 0.32 are finally selected, and the path with the final leaf node classification attribute of 1 is shown in Table 1 below. The key features and thresholds obtained by the interpretability work are shown in Table 1 below:

[0064]

[0065]

[0066] Table 1 - Key features and thresholds obtained for interpretability work

[0067] From the above, we can see that when the RRC connection success rate initiated by the UE in the cell is not greater than 0.998, the average delay of the user plane downlink packet is not less than 117.531 milliseconds, and the cell PDCP layer downlink effective throughput is not greater than 7.065, the network quality of the cell is likely to be degraded. Using this conclusion, samples with degraded network quality were screened out in the training set, with an accuracy rate of 81.144%.

[0068] Verify the results: First, compare the distribution of each feature value in the validation set data on day T5 with the distribution of feature values ​​on days T1-T4. Figures 5A-5D By comparing the results, it can be found that the data distribution characteristics of the two are highly consistent. Therefore, the model interpretability conclusions obtained based on the training set have a certain generalization ability.

[0069] According to the interpretability conclusions (the three key indicators and thresholds in Table 1), the sample data of the validation set was classified and judged with an accuracy of 77.83%, which has a certain degree of usability. It is recommended to focus on the operation quality of the above indicators in daily network optimization work. The accuracy comparison of the interpretability work conclusions in the training set and the validation set is shown in Table 2 below:

[0070]

[0071] Table 2- Comparison of the accuracy of interpretability work conclusions in the training set and the validation set

[0072] Based on the above explanation, we can conclude that when the RRC connection success rate initiated by UE in the cell is not greater than 0.998, the average delay of downlink packets on the user plane is not less than 117.531 milliseconds, and the downlink effective throughput rate of the cell PDCP layer is not greater than 7.065, the network quality of the cell is likely to be degraded. This conclusion was used to screen out samples with degraded network quality in the training set, with an accuracy rate of 81.144%.

[0073] In one embodiment, Figure 6 As shown, the present disclosure provides a machine learning model interpretation device 60, including: a first training module 61, a weight acquisition module 62, a feature selection module 63, a data processing module 64, a second training module 65 and a path selection module 66.

[0074] The first training module 61 obtains network performance data and trains a first machine learning model based on the network performance data. After the first machine learning model is trained, the weight acquisition module 62 obtains feature weights corresponding to each feature included in the network performance data. The feature selection module 63 selects model features from all features included in the network performance data based on the feature weights.

[0075] The data processing module 64 processes the network performance data based on the model features to construct a training set. The second training module 65 uses the training set to train the second machine learning model to obtain a decision tree structure corresponding to the second machine learning model; wherein the decision nodes in the decision tree structure are used to characterize the model feature judgment and are provided with feature judgment information. The path selection module 66 determines the judgment selection path in the decision tree structure according to the feature judgment information as the explanatory working path of the second machine learning model.

[0076] In one embodiment, the first training module 61 calculates the Pearson correlation coefficient between the network performance data and screens the network performance data based on the Pearson correlation coefficient. The feature selection module 63 sorts all features based on feature weights and selects multiple features with the highest feature weights as model features according to the sorting results.

[0077] In one embodiment, the second machine learning model includes a decision tree model; Figure 7As shown, the path selection module 66 includes a node acquisition unit 661, a node selection unit 662 and a path determination unit 663. The node acquisition unit 661 acquires multiple child decision nodes of the decision node to be determined in the decision tree structure. The node selection unit 662 determines a decision target node from multiple child decision nodes based on the feature judgment information of each child decision node. The path determination unit 663 determines the path between the decision node to be determined and the decision target node as the judgment selection path.

[0078] The feature judgment information includes the Gini coefficient, the number of sample coverage of the node, etc. The node selection unit 662 determines a decision target node from multiple child decision nodes based on the Gini coefficient and the number of sample coverage. For example, if the number of multiple child decision nodes is two, the node selection unit 662 selects a child decision node with a smaller Gini coefficient and a larger number of sample coverage from the two child decision nodes as the decision target node.

[0079] In one embodiment, Figure 8 FIG. 1 is a module diagram of another embodiment of a machine learning model interpretation device according to the present disclosure. Figure 8 As shown, the device may include a memory 801, a processor 802, a communication interface 803 and a bus 804. The memory 801 is used to store instructions, the processor 802 is coupled to the memory 801, and the processor 802 is configured to execute the above-mentioned machine learning model interpretation method based on the instructions stored in the memory 801.

[0080] The memory 801 may be a high-speed RAM memory, a non-volatile memory, etc. The memory 801 may also be a memory array. The memory 801 may also be divided into blocks, and the blocks may be combined into virtual volumes according to certain rules. The processor 802 may be a central processing unit CPU, or an application-specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the machine learning model interpretation method disclosed in the present invention.

[0081] In one embodiment, the present disclosure provides a computer-readable storage medium, which stores computer instructions. When the instructions are executed by a processor, the machine learning model interpretation method in any of the above embodiments is implemented.

[0082] The machine learning model interpretation method, device and storage medium provided in the above embodiments train a first machine learning model based on network performance data, obtain feature weights corresponding to each feature contained in the network performance data, and select model features based on the feature weights; train a second machine learning model based on the model features to obtain a decision tree structure corresponding to the second machine learning model; determine a judgment selection path in the decision tree structure as an explanatory work path; reduce the algorithm complexity by arranging the complexity of the problem to be solved and trimming the features; perform interpretable work through a decision tree model, with a simple process, convenient and efficient operation, and the decision basis of the machine learning model can be accurately obtained, taking into account both simplicity and accuracy.

[0083] The method and system of the present disclosure may be implemented in many ways. For example, the method and system of the present disclosure may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present disclosure are not limited to the order specifically described above, unless otherwise specifically stated. In addition, in some embodiments, the present disclosure may also be implemented as a program recorded in a recording medium, which includes machine-readable instructions for implementing the method according to the present disclosure. Therefore, the present disclosure also covers a recording medium storing a program for executing the method according to the present disclosure.

[0084] The description of the present disclosure is given for the purpose of illustration and description, and is not intended to be exhaustive or to limit the present disclosure to the disclosed form. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present disclosure, and to enable those of ordinary skill in the art to understand the present disclosure and thereby design various embodiments with various modifications suitable for specific uses.

Claims

1. A machine learning model interpretation method, comprising: Acquire network performance data, and train a first machine learning model based on the network performance data, wherein the network performance data includes: cell performance data; the cell performance data includes a wireless connection success rate and an RRC connection reconstruction ratio; After the first machine learning model is trained, obtaining feature weights corresponding to each feature included in the network performance data; Selecting model features from all features included in the network performance data based on the feature weights; Processing the network performance data based on the model features to construct a training set; The second machine learning model is trained using the training set to obtain a decision tree structure corresponding to the second machine learning model; wherein the decision nodes in the decision tree structure are used to characterize model feature judgment and are provided with feature judgment information; the feature judgment information includes: a Gini coefficient and the number of samples covered by this node; Determining a judgment selection path in the decision tree structure according to the feature judgment information as an explanatory working path of the second machine learning model; Wherein, the second machine learning model includes: a decision tree model; and determining the judgment selection path in the decision tree structure according to the feature judgment information includes: In the decision tree structure, multiple child decision nodes of the decision node to be judged are obtained; based on the Gini coefficient of each child decision node and the sample coverage quantity, a judgment target node is determined from the multiple child decision nodes; and the path between the decision node to be judged and the judgment target node is determined as the judgment selection path.

2. The method according to claim 1, wherein the number of the plurality of child decision nodes is two; and the step of determining a decision target node from the plurality of child decision nodes based on the Gini coefficient and the sample coverage number comprises: A child decision node with a smaller Gini coefficient and a larger number of sample coverage is selected from the two child decision nodes as the decision target node.

3. The method of claim 1, further comprising: Calculating the Pearson correlation coefficient between the network performance data; The network performance data is screened based on the Pearson correlation coefficient.

4. The method according to claim 1, wherein the step of selecting a model feature from all features included in the network performance data based on the feature weight comprises: sorting all the features based on the feature weights; According to the sorting results, multiple features with the highest feature weights are selected as the model features.

5. The method according to any one of claims 1 to 4, wherein: The first machine learning model includes: a random forest model.

6. A machine learning model interpretation device, comprising: A first training module, used to obtain network performance data, and train a first machine learning model based on the network performance data; wherein the network performance data includes: cell performance data; the cell performance data includes a wireless connection success rate and an RRC connection reconstruction ratio; A weight acquisition module, used to acquire feature weights corresponding to each feature included in the network performance data after the first machine learning model training is completed; A feature selection module, configured to select model features from all features included in the network performance data based on the feature weights; A data processing module, used for processing the network performance data based on the model features to construct a training set; A second training module is used to train a second machine learning model using the training set to obtain a decision tree structure corresponding to the second machine learning model; wherein the decision nodes in the decision tree structure are used to characterize model feature judgments and are provided with feature judgment information; the feature judgment information includes: a Gini coefficient and the number of samples covered by this node; A path selection module, used for determining a judgment selection path in the decision tree structure according to the feature judgment information as an explanatory working path of the second machine learning model; Wherein, the second machine learning model includes: a decision tree model; the path selection module includes: A node acquisition unit, used to acquire multiple child decision nodes of the decision node to be determined in the decision tree structure; A node selection unit, configured to determine a decision target node from the plurality of child decision nodes based on the Gini coefficient and the sample coverage quantity of each child decision node; The path determination unit is used to determine the path between the decision node to be determined and the judgment target node as the judgment selection path.

7. The device according to claim 6, wherein the number of the plurality of child decision nodes is two; wherein, The node selection unit is specifically used to select a child decision node with a smaller Gini coefficient and a larger number of sample coverage from the two child decision nodes as the decision target node.

8. The device according to claim 6, wherein: The first training module is also used to calculate the Pearson correlation coefficient between the network performance data; and to screen the network performance data based on the Pearson correlation coefficient.

9. The device according to claim 6, wherein: The feature selection module is specifically used to sort all the features based on the feature weights; and select multiple features with the highest feature weights as the model features according to the sorting results.

10. A machine learning model interpretation device, comprising: Memory; and a processor coupled to the memory, wherein the processor is configured to execute the method according to any one of claims 1 to 5 based on instructions stored in the memory. 11 . A computer-readable storage medium, wherein the computer-readable storage medium non-transitorily stores computer instructions, wherein the instructions are executed by a processor to execute the method according to claim 1 .

Images