A method for index selection based on robust fractal autoencoders
By combining the global and local reconstruction paths of the robust fractal autoencoder, and dynamically adjusting the selection window and task-oriented supervision, the problems of redundancy and insufficient adaptability in the autoencoder index selection method are solved, and more efficient and stable index selection and evaluation are achieved.
Patent Information
- Application Number
- CN202611030572.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-12
- Publication Date
- 2026-11-10
AI Technical Summary
Existing autoencoder-based metric selection methods lack diversity constraints during the training phase, leading to redundant selected metric information. They also lack weight protection mechanisms, resulting in unstable selection results after multiple screenings. Furthermore, the lack of label guidance makes it difficult for the selected metrics to be adapted to downstream evaluation tasks.
A robust fractal autoencoder is adopted, which combines global and local reconstruction paths, dynamically adjusts the selection window, exponentially maps weights, introduces task-oriented supervision loss, optimizes the weight update of the feature selection layer, and ensures the redundancy and adaptability of the indicators.
It achieves lower redundancy and more representative index selection, improves the robustness of high-dimensional sparse data processing and the consistency of results from multiple independent runs, and improves the accuracy of downstream tasks.
Smart Images

Figure CN123068906A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer data processing technology, and specifically to an index screening method, apparatus, electronic device, and storage medium based on a robust fractal autoencoder. Background Technology
[0002] In the engineering field, constructing a scientific and comprehensive indicator system is a core prerequisite for conducting performance evaluation, risk prediction, and scheme comparison. With the deep integration of big data and information technology, the data dimensions involved in the evaluation objects are exploding. From complex system evaluation and energy structure security assessment to credit risk systems and operational monitoring, the evaluation indicators often exhibit high dimensionality, strong correlation, and nonlinear mapping characteristics. When dealing with such high-dimensional indicators, directly introducing all information into the evaluation model not only significantly increases the time and space complexity of computation but may also lead to biased evaluation conclusions due to collinearity issues among indicators. Therefore, achieving indicator system reduction while preserving the original physical meaning of the characteristics through indicator screening techniques has become a key problem that needs to be solved.
[0003] Traditional methods for feature reduction mainly fall into three categories: subjective selection, statistical dimensionality reduction, and heuristic search-based reduction. Subjective selection relies on expert surveys for qualitative indicator selection, but its objectivity and scientific validity are easily influenced by individual expert experience. Statistical dimensionality reduction extracts principal components through linear transformations, effectively compressing dimensionality, but the mapped features lose the physical meaning of the original indicators and cannot be used for subsequent evaluation. Furthermore, some schemes attempt to use random search strategies combined with fuzzy rough set theory for indicator reduction; however, random search strategies converge slowly and are prone to getting trapped in local optima when dealing with high-dimensional spaces. Fuzzy rough set theory has limited capacity for handling large-scale data, and its reduction results are highly dependent on the selection of membership functions and the setting of discretization thresholds, resulting in poor generalization performance.
[0004] With the widespread application of deep learning theory in feature extraction, using autoencoder neural networks to handle indicator selection problems has gradually become a major direction of technological evolution. The core logic of the relevant technical system is to construct a multi-layer architecture using a Restricted Boltzmann Machine (RBM). Through layer-by-layer unsupervised pre-training, the latent distribution of the original indicators is captured and the network weights are initialized. Then, the backpropagation algorithm is used to globally fine-tune the expanded network, minimizing the mean square error between the original input indicators and the reconstructed output. After model training, the absolute value distribution of the weight matrix from the input layer to the first hidden layer is calculated to quantitatively evaluate the contribution of each original evaluation indicator to the core feature nodes, thereby achieving the removal of redundant indicators. Although the above-mentioned scheme based on a RBM autoencoder has initially achieved effective extraction of nonlinear features, there is still room for further optimization in terms of the robustness of the selection process, the redundancy of the selection indicators, and the adaptability to the task. In addition, the related technology proposes a fractal autoencoder framework, which adds a sub-neural network to the traditional autoencoder. While training the global neural network to reconstruct the data using all features to evaluate features that can represent global information, the sub-network is trained to reconstruct the original data using only the top-ranked local feature subsets in the global weights. This is to test the diversity of features and filter out highly correlated redundant features.
[0005] The aforementioned autoencoder-based index selection method still has the following shortcomings in practical applications. First, when quantitatively evaluating the contribution of indices by post-calculating the absolute value distribution of the weight matrix after model training, the model does not introduce constraint logic for the diversity among the original indices during the training phase. When there are two indices with highly correlated physical meanings and similar numerical characteristics in the evaluation system, the model tends to assign similar weights to them, resulting in the selected core index subset still containing redundant information. Second, fractal autoencoders lack a protection mechanism for weight distribution. When using L1 regularization to induce weight sparsity, the weights are easily affected by the regularization term and excessively decay, even being incorrectly eliminated to zero in the early stages of training. This leads to the random elimination of some weakly correlated but valuable indices under different random seed initializations. At the same time, the proportion of its local feature subset to the global features is small. In each round of training, only a small portion of the weights can be trained based on the reconstruction loss of the sub-network. The network is prone to falling into different local optima under different random initialization paths, resulting in low overlap of the index subsets selected in multiple independent training tasks and a lack of process repeatability. Third, the metrics selected by traditional autoencoders do not understand the downstream tasks. The only basis for selecting metrics is the amount of information contained in the metrics. Due to the lack of input labels, it is easy to retain metrics with a large amount of information but mediocre evaluation value, which is not suitable for specific evaluation scenarios. Summary of the Invention
[0006] The present invention aims to provide an index screening method based on a robust fractal autoencoder to address the shortcomings of existing technologies. The technical problem to be solved by the present invention is achieved through the following technical solution.
[0007] In a first aspect, the present invention provides an index selection method based on a robust fractal autoencoder, comprising: Collect the original indicator dataset of the scenario to be evaluated and the evaluation labels corresponding to the samples. Map the values of each original indicator in the original indicator dataset to a preset range to eliminate the influence of units and obtain standardized indicator data. Configure the feature selection layer, encoder, decoder, and task-oriented supervision module of the autoencoder, and initialize the original weights of the feature selection layer, wherein the feature selection layer is configured to map the original weights to exponential weights, so that the exponential weights corresponding to each original index are always greater than zero; In multi-round iterative training, the size of the dynamic selection window is determined according to the current iteration round number. The size of the dynamic selection window shrinks from close to the total number of the original indicators to the preset target number of selections as the iteration round number increases. The original indicators with the highest index weights and a quantity equal to the size of the dynamic selection window are extracted from the feature selection layer to form a local indicator subset. For each iteration, a global reconstruction result is obtained by encoding and decoding all standardized indicator data weighted by the exponential weights through a global reconstruction path, and a local reconstruction result is obtained by encoding and decoding only a subset of the local indicators through a local reconstruction path. The features weighted by the exponential weights or the bottleneck layer features output by the encoder are then input into the task-oriented supervision module to obtain the prediction result of the evaluation task. Based on the global reconstruction error between the global reconstruction result and the standardized index data, the local reconstruction error between the local reconstruction result and the standardized index data, the sparse regularization term applied to the index weights, and the task-oriented supervision loss between the prediction result and the evaluation label, a joint loss function is calculated, and the original weights of the feature selection layer are updated by gradient descent, so that the selected indicators have both reconstruction representativeness and discriminative power for downstream evaluation tasks; Multiple independent rounds of iterative training are performed on the same original index dataset. After the preset reconstruction accuracy threshold is met or the maximum number of iterations is reached, the selection set with the highest overlap among the multiple independent trainings is taken, and the original indexes with the highest index weight ranking in the selection set and whose number is equal to the target selection number are output as the selected evaluation index system.
[0008] Furthermore, the task-oriented supervision loss is either classification cross-entropy loss or regression loss. By introducing the evaluation label, indicators that fluctuate greatly but do not play a decisive role in the evaluation purpose are eliminated, while indicators that play a decisive role in the final decision are given higher weights.
[0009] Furthermore, determining the size of the dynamic selection window based on the current iteration round number includes: determining the size of the dynamic selection window corresponding to the current iteration round number according to a shrinkage rule that causes the size of the dynamic selection window to decrease monotonically from the total number of the original indicators to the target screening number with the current iteration round number, and to remain at the target screening number after reaching a preset shrinkage stop period.
[0010] Furthermore, the step of exponentially mapping the original weights to exponential weights includes: taking an exponent with the natural constant as the base of the original weights of the feature selection layer to obtain the exponential weights, such that during the application of the sparse regularization term, the decay rate of the exponential weights when their values are small is slower than the decay rate when a linear penalty is directly applied to the original weights; both the encoder and the decoder are single-layer linear networks without activation functions.
[0011] Furthermore, the sparse regularization term is an L1 norm regularization term applied to the exponential weights; the joint loss function is obtained by weighting and summing the global reconstruction error, the local reconstruction error, the sparse regularization term, and the task-oriented supervision loss according to their respective balance coefficients.
[0012] Furthermore, the step of selecting the screening set with the highest overlap among multiple independent training sessions includes: calculating the overlap of indicators for each pair of screening sets output from the multiple independent training sessions, and selecting the screening set with the highest average overlap with the remaining screening sets as the screening set with the highest overlap.
[0013] A second aspect of the present invention provides an index screening device based on a robust fractal autoencoder, comprising: The standardization module is configured to collect the original indicator dataset of the scenario to be evaluated and the evaluation labels corresponding to the samples, and map the values of each original indicator in the original indicator dataset to a preset range to eliminate the influence of units, thereby obtaining standardized indicator data; The configuration initialization module is configured to configure the feature selection layer, encoder, decoder, and task-oriented supervision module of the autoencoder, initialize the original weights of the feature selection layer, and map the original weights to exponential weights so that the exponential weights corresponding to each original index are always greater than zero; The window subset module is configured to determine the size of a dynamic selection window based on the current iteration number during multi-round iterative training. The size of the dynamic selection window shrinks from close to the total number of the original indicators to a preset target selection number as the number of iterations increases. It also extracts the original indicators with the highest index weights from the feature selection layer, whose number is equal to the size of the dynamic selection window, to form a local indicator subset. The dual-path reconstruction module is configured to, for each iteration, use the global reconstruction path to encode and decode all standardized indicator data weighted by the exponential weights to obtain the global reconstruction result, and use the local reconstruction path to encode and decode only the local indicator subsets to obtain the local reconstruction result. The features weighted by the exponential weights or the bottleneck layer features output by the encoder are then input into the task-oriented supervision module to obtain the prediction result of the evaluation task. The joint optimization module is configured to calculate a joint loss function based on the global reconstruction error, the local reconstruction error, the sparse regularization term applied to the exponential weights, and the task-oriented supervised loss between the prediction result and the evaluation label, and update the original weights of the feature selection layer through gradient descent; The indicator output module is configured to perform multiple independent rounds of iterative training on the same original indicator dataset. After meeting the preset reconstruction accuracy threshold or reaching the maximum number of iterations, it takes the selection set with the highest overlap among the multiple independent training sessions and outputs the original indicators in the selection set whose index weight ranks first and whose quantity is equal to the target selection quantity, as the selected evaluation indicator system.
[0014] Furthermore, the task-oriented supervision loss is either classification cross-entropy loss or regression loss; the encoder and decoder configured by the configuration initialization module are both single-layer linear networks without activation functions.
[0015] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the method described in any of the preceding descriptions.
[0016] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in any of the preceding claims.
[0017] The embodiments of the present invention have the following advantages: The technical problem to be solved by the present invention is that the existing autoencoder-based index selection method lacks diversity constraints during the training phase, resulting in redundant selected index information; lacks weight protection mechanism and global search mechanism, resulting in unstable selection results; and lacks label guidance, making it difficult for the selected index to adapt to downstream evaluation tasks. This invention achieves the following beneficial effects by introducing local subnetwork reconstruction based on dynamic windows during global reconstruction, mapping the original weights to exponential weights that are always greater than zero, shrinking the dynamic selection window from the entire set of indicators to the target subset size with each iteration, and using optional task-oriented supervision loss: First, through dual-path reconstruction, the global and local reconstruction errors can be compared during the training phase to discover and eliminate redundant indicators, resulting in lower redundancy and stronger representativeness of the selected indicators; Second, the exponential mapping changes the decay curve characteristics of the weights, preventing key indicators from being incorrectly eliminated in the early stages of training due to excessive weight decay, thus improving robustness in handling high-dimensional sparse datasets and the consistency of results from multiple independent runs; Third, the dynamic window shrinking guides the model to smoothly transition from global space exploration to core subset refinement, overcoming the defect of prematurely fixing the search field and getting trapped in local optima, further improving the consistency of results from multiple independent runs; Fourth, the optional task-oriented supervision loss enables the selected indicators to not only restore the original data information but also to exhibit better accuracy in downstream classification or regression tasks. Attached Figure Description
[0018] Figure 1 A technical architecture diagram of an index screening method based on a robust fractal autoencoder according to an embodiment of the present invention is shown.
[0019] Figure 2 A flowchart of the index screening method according to an embodiment of the present invention is shown;
[0020] Figure 3 A schematic diagram is shown illustrating a subset of indicators selected from a combat effectiveness assessment scenario using the method described in an embodiment of the present invention;
[0021] Figure 4 A block diagram of a computing device used to implement embodiments of the present invention is shown. Detailed Implementation
[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0023] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0024] like Figure 1 As shown in the embodiment of the present invention, the index selection method based on a robust fractal autoencoder is based on the construction of a deep learning framework with dual reconstruction paths. This framework mainly consists of a weight exponentiation module, a dual-path reconstruction network, and a dynamic window controller, and optionally includes a task-oriented supervision module. Let the total number of original indices to be selected be m, the target selection number be k, and the number of samples be N. The feature selection layer is provided with a learnable original weight vector w of length m. This original weight vector w does not directly participate in feature weighting, but is instead exponentially mapped to exponential weights. The mapping relationship is as follows:
[0025] in, This represents the original weights of the feature selection layer. This represents the exponential weights after the exponential mapping. This represents the natural constant. This mapping ensures that the exponential weights corresponding to all metrics are always greater than zero, effectively preventing key metrics from being incorrectly eliminated in the early stages of training. At the same time, this exponential form changes the gradient decay rate, making the exponential weights decay more slowly when their values are small than when a linear penalty is directly applied to the original weights when applying the sparsity regularization term, thus providing greater robustness when dealing with sparse data.
[0026] See Figure 1 The technical architecture of this invention includes: input standardized index data 110, index selection layer 120, global encoder 130 and global decoder 140 located on the global reconstruction path, sub-network 150 located on the local reconstruction path, and optional task-oriented supervision module 160; the index selection layer 120 applies exponentially mapped exponential weights to the input standardized index data 110, the global encoder 130 and global decoder 140 perform global reconstruction using all weighted indices, the sub-network 150 performs local reconstruction using only a subset of local indices, and the optional task-oriented supervision module 160 performs collaborative optimization using label information when the supervision mode is enabled.
[0027] This invention constructs two parallel reconstruction paths. The first is a global reconstruction path, which uses the full set of indicator data for encoding and decoding via a global neural network, aiming to learn the overall distribution characteristics of the original data and provide regularization constraints for feature selection. The second is a local diversity reconstruction path, which only uses the local indicator subset with the highest index weight within the current dynamic selection window for encoding and decoding via a sub-neural network, forcing the selected local indicator subset to have the ability to reconstruct the overall data, thereby identifying and eliminating indicators with redundant information. In one embodiment, the encoder is a single-layer linear network without activation functions, which reduces the dimensionality of the m-dimensional data after weighting by the feature selection layer to k-dimensionality; the decoder is a single-layer linear network without activation functions, which raises the dimensionality of the k-dimensional hidden features back to m-dimensionality for comparison with the original input data to calculate the reconstruction error.
[0028] To address the problem of the model's search scope being prematurely limited, leading to local optima, this invention implements a strategy of dynamically adjusting the search field of view. This allows the number of indicators participating in subnetwork reconstruction, i.e., the size t of the dynamic selection window, to change dynamically with the number of iterations. In one implementation, the size t shrinks from the entire set of indicators to the target subset as the current iteration number increases:
[0029] in, This represents the total number of original indicators. Indicates the target number of filters. Indicates the current iteration round number. `T` represents the shrinking stopping period, and `round` represents the rounding operation. In the early stages of training, the current iteration number is small, and the size `t` is set relatively large and close to the total number of original metrics `m`, allowing the model to optimize more weights within a larger search space. As the iteration number progresses, the size `t` gradually shrinks. When the current iteration number reaches the shrinking stopping period `T`, the size `t` is fixed at the target number of selections `k`. This mechanism guides the model smoothly from the initial global exploration to the later local refinement, effectively avoiding local minima and significantly improving the consistency and repeatability of selection results under different random seeds.
[0030] This invention determines the optimal index set by jointly optimizing the global reconstruction error, local reconstruction error, and sparse regularization term, and optionally introduces a task-oriented supervision loss. Its overall loss function is defined as:
[0031] Among them, the first item The mean squared error of the global reconstruction requires the model to be able to reconstruct the data using all exponentially weighted indicators, and its expression is:
[0032] Second item The mean squared error of subnetwork reconstruction based on dynamic windows requires the model to reconstruct the data using only the t indicators with the highest current exponential weights. Its expression is:
[0033] Third item The L1 norm sparse regularization term applied to the index weights is used to induce sparsity in the weight distribution, requiring the model to select sufficiently concise indices; the fourth term For the optional task-oriented supervision loss, label information is introduced to guide the indicator weights towards a more discriminative direction for optimization. In the formula, This represents the standardized index vector of the i-th sample. This represents the global reconstruction result obtained by the global reconstruction path for the i-th sample. This represents the local reconstruction result obtained by the local reconstruction path for the i-th sample. Indicates the number of samples. and represents the balance coefficients of the sparse regularization term and the supervised loss term, respectively.
[0034] Furthermore, to address the issue of insufficient adaptability to downstream evaluation tasks, this invention integrates an optional supervision module. When the supervision mode is enabled, the model collects the labels corresponding to the samples while collecting the original indicator dataset, and inputs the features weighted by exponential weights or the bottleneck layer features output by the encoder into the supervision module for prediction. The task-oriented supervision loss is calculated by introducing a classification or regression loss function. This process eliminates indicators that, while highly volatile, do not play a decisive role in the evaluation objective, while assigning higher weights to indicators that are crucial for the final decision, ultimately selecting the most suitable indicators for evaluation.
[0035] like Figure 2 As shown in the figure, the specific processing flow for the automated screening of evaluation indicators in this embodiment of the invention is as follows.
[0036] Step 1, Data Preprocessing and Standardization: Collect the original indicator dataset of the scenario to be evaluated, and map the data to a preset interval to eliminate the influence of units; if the supervised mode is enabled, sample labels are collected simultaneously.
[0037] Step 2, Parameter Initialization and Model Configuration: Configure the hidden layer structure of the autoencoder, set the number of target selection indicators, the shrinkage stopping period, and the balance coefficient of each loss term, and initialize the original weights of the feature selection layer.
[0038] Step 3, Model Training: The dynamic selection window size is calculated based on the current training epoch. The top t indicators with the current exponential weights are extracted from the feature selection layer to construct a local indicator subset. The system simultaneously executes data flow between the global path and the sub-network path. The global path processes all weighted indicators, while the sub-network path processes only the top t core indicators. If supervised mode is enabled, the weighted feature or bottleneck layer output is sent to the supervision module for prediction calculation to obtain the task prediction result. The joint loss function is calculated, and the original weights of the feature selection layer are updated using the gradient descent algorithm. Due to the use of exponential mapping, low-weight parameters can still obtain effective gradient flow during backpropagation, ensuring consistency in the selection process.
[0039] Step 4, Index Output: Determine whether the model reconstruction accuracy has reached the expected threshold or the maximum number of iterations. If the conditions are met, output the original index of the index weight ranking first, which will be used as the final evaluation index system.
[0040] Step 5: Perform multiple independent training sessions, and finally select the set with the highest overlap among the multiple training sessions as the final subset of indicators.
[0041] The implementation process of this invention is illustrated below with a specific embodiment. In this embodiment, game log data generated by an open-source benchmarking platform is used as the original indicator dataset for the scenario to be evaluated. This platform can simulate and generate game logs for micro-level adversarial battles between red and blue teams using multiple agents. 2000 games involving three combat units from each of the red and blue teams are selected. Each game contains 31 evaluation indicators, and each game has a clear score. Since these 31 indicators only consider computability and may not be suitable for game evaluation, indicator screening is performed, with a preset target screening quantity of 12.
[0042] In the data preprocessing stage, after obtaining 31 indicators from 2000 games, all input indicators are mapped to the [0,1] interval to eliminate gradient flooding and regularization bias caused by differences in units and numerical scales between different indicators; and the 2000 simulation data are divided into independent and mutually exclusive training set, validation set and test set according to the stratified sampling principle in a ratio of 72:8:20.
[0043] In the stage of building a robust fractal autoencoder network, a learnable primitive weight vector w of length 31 is initialized at the front end of the network, and its exponential form is used. As exponential weights; the encoder is a single-layer linear network without activation functions, which reduces the 31-dimensional data after weighting by the feature selection layer to k dimensions; the decoder is a single-layer linear network without activation functions, which raises the k-dimensional hidden layer features back to 31 dimensions, which is used to compare with the original input data to calculate the reconstruction error.
[0044] During the loss function configuration phase, the overall objective function is dominated by four collaborative mechanisms: global reconstruction mean squared error, which forces the network to use all 31 weighted indicators to reconstruct the original 31-dimensional data distribution as losslessly as possible; local subnet reconstruction mean squared error, in each forward propagation, the network finds the t indicators with the largest values in the current index weights in real time, and then relies only on these t indicators to complete the data reconstruction, thereby forcing the network to eliminate local redundancy between indicators; L1 sparse regularization, which causes the weights of marginal indicators that do not contribute much to reconstruction and prediction to continuously decay; and supervised classification cross-entropy, which calculates the error between the game result score and the true label, forcing the network to retain the core indicators related to winning from a task-oriented perspective.
[0045] In the initial training cycle, the training set data is fed into the network for 1000 rounds of iterative training. Throughout this process, a dynamic window strategy is implemented synchronously and consistently: when the network calculates the mean square error of local subnetwork reconstruction, the number of indicators t retained and used in the calculation by the subnetwork is not a fixed constant, but a dynamically changing shrinkage value. In the early stages of training, the value of t is set relatively large and close to the original total number of indicators, 31, giving the network room to explore in a broad feature space and preventing the model from mistakenly eliminating potentially valuable indicators at the beginning, thus avoiding the risk of getting trapped in local optima. As the number of iterations progresses, the value of t decreases linearly in the middle and later stages until it is finally fixed at the preset target number of 12, locking in 12 of the most representative core indicators.
[0046] In the final index extraction and sorting stage, after hundreds of rounds of gradient backpropagation, the various losses reach convergence and equilibrium, and the network parameters are fixed. At this point, the exponential weight vector of the trained feature selection layer is directly extracted, and these 31 values are sorted in descending order. The original indices corresponding to the first 12 exponential weights are the final selected indices. Weights ranked 13th and below have values that are extremely close to 0 due to L1 regularization and are considered redundant indices that are discarded. The final subset of selected indices is as follows: Figure 3 As shown.
[0047] In the experimental verification phase, to verify the robustness of the index selection method of the present invention, 20 independent training cycles were conducted using the present invention method, the method based on the restricted Boltzmann machine autoencoder, and the fractal autoencoder method, respectively. The overlap of the indexes selected in each experiment was compared pairwise, and the average index overlap is shown in Table 1. To verify that the indexes selected by the present invention are more suitable for the evaluation task, subsets of indexes selected by different methods were also used for game result score prediction, and the prediction accuracy is shown in Table 2.
[0048] Table 1. Average overlap of screening indicators after pairwise comparisons of different methods after 20 independent training iterations.
[0049] Table 2. Accuracy of predicting game outcome scores using different subsets of indicators selected by different methods.
[0050] As shown in Tables 1 and 2, the method of this invention outperforms the comparative methods in terms of both the stability of the selected indicators and the accuracy of the prediction results. The high robustness of the method of this invention stems from the fact that the exponential mapping prevents drastic fluctuations in weight updates, while the dynamic window prevents the model from getting trapped in local optima too early. The combination of the two makes the results more consistent under different runs. The higher accuracy indicates that the indicators selected by the method of this invention are more representative and diverse. As a phenomenon, in the experimental results of the method of this invention, the 12 indicators selected in each round of training will definitely include one of the two indicators describing the firepower strike capability, namely the output intensity per unit time and the attack command density, and the presence of one of them will not result in the other appearing simultaneously. In contrast, in the multi-round results of the restricted Boltzmann machine autoencoder method, the simultaneous selection of these two indicators often occurs. Since these two indicators describe the same capability from different perspectives and have high redundancy, this phenomenon indirectly proves that the indicator redundancy design of the method of this invention is effective. In addition, as shown in Table 2, the prediction accuracy decreases after removing the dynamic window or the downstream task supervision module, indicating that the dynamic window mechanism and the task-oriented supervision module both make substantial contributions to improving the selection quality.
[0051] Figure 4 A block diagram of a computing device that can be used to implement embodiments of the present invention is shown. Figure 4 As shown, the computing device 800 includes a processor 801, a read-only memory 802, and a random access memory 803. The processor 801, read-only memory 802, and random access memory 803 are interconnected via a bus 804. An input / output interface 805 is also connected to the bus 804. An input unit 806, an output unit 807, a storage unit 808, and a communication unit 809 are connected to the computing device 800 via the input / output interface 805. The processor 801 can execute a computer program stored in the storage unit 808 for implementing the aforementioned index screening method. The communication unit 809 is used to exchange information with other devices via a computer network.
[0052] In another embodiment, the indicator screening method of the present invention can be run on an electronic device, the electronic device including one or more processors and a storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned indicator screening method. The method can also be stored as a computer program in a computer-readable storage medium, and when the computer program is executed by a processor, it implements the aforementioned indicator screening method. The computer-readable storage medium includes, but is not limited to, random access memory, read-only memory, erasable programmable read-only memory, optical disk, or magnetic disk, and other media capable of storing program code.
[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for index selection based on a robust fractal autoencoder, characterized in that, include: Collect the original indicator dataset of the scenario to be evaluated and the evaluation labels corresponding to the samples. Map the values of each original indicator in the original indicator dataset to a preset range to eliminate the influence of units and obtain standardized indicator data. Configure the feature selection layer, encoder, decoder, and task-oriented supervision module of the autoencoder, and initialize the original weights of the feature selection layer, wherein the feature selection layer is configured to map the original weights to exponential weights, so that the exponential weights corresponding to each original index are always greater than zero; In multi-round iterative training, the size of the dynamic selection window is determined according to the current iteration round number. The size of the dynamic selection window shrinks from close to the total number of the original indicators to the preset target number of selections as the iteration round number increases. The original indicators with the highest index weights and a quantity equal to the size of the dynamic selection window are extracted from the feature selection layer to form a local indicator subset. For each iteration, a global reconstruction result is obtained by encoding and decoding all standardized indicator data weighted by the exponential weights through a global reconstruction path, and a local reconstruction result is obtained by encoding and decoding only a subset of the local indicators through a local reconstruction path. The features weighted by the exponential weights or the bottleneck layer features output by the encoder are then input into the task-oriented supervision module to obtain the prediction result of the evaluation task. Based on the global reconstruction error between the global reconstruction result and the standardized index data, the local reconstruction error between the local reconstruction result and the standardized index data, the sparse regularization term applied to the index weights, and the task-oriented supervision loss between the prediction result and the evaluation label, a joint loss function is calculated, and the original weights of the feature selection layer are updated by gradient descent, so that the selected indicators have both reconstruction representativeness and discriminative power for downstream evaluation tasks; Multiple independent rounds of iterative training are performed on the same original index dataset. After the preset reconstruction accuracy threshold is met or the maximum number of iterations is reached, the selection set with the highest overlap among the multiple independent trainings is taken, and the original indexes with the highest index weight ranking in the selection set and whose number is equal to the target selection number are output as the selected evaluation index system.
2. The method according to claim 1, characterized in that, The task-oriented supervision loss is either classification cross-entropy loss or regression loss. By introducing the evaluation labels, indicators that fluctuate greatly but do not play a decisive role in the evaluation purpose are eliminated, while indicators that play a decisive role in the final decision are given higher weights.
3. The method according to claim 1, characterized in that, The step of determining the size of the dynamic selection window based on the current iteration round number includes: determining the size of the dynamic selection window corresponding to the current iteration round number according to a shrinkage rule that causes the size of the dynamic selection window to decrease monotonically from the total number of the original indicators to the target screening number as the current iteration round number increases, and to remain at the target screening number after reaching a preset shrinkage stop period.
4. The method according to claim 1, characterized in that, The step of mapping the original weights to exponential weights includes: taking an exponent with the natural constant as the base of the original weights of the feature selection layer to obtain the exponential weights, such that during the application of the sparse regularization term, the decay rate of the exponential weights when their values are small is slower than the decay rate when a linear penalty is directly applied to the original weights; both the encoder and the decoder are single-layer linear networks without activation functions.
5. The method according to claim 1, characterized in that, The sparse regularization term is an L1 norm regularization term applied to the exponential weights; the joint loss function is obtained by weighting and summing the global reconstruction error, the local reconstruction error, the sparse regularization term, and the task-oriented supervision loss according to their respective balance coefficients.
6. The method according to claim 1, characterized in that, The step of selecting the screening set with the highest overlap among multiple independent training sessions includes: calculating the overlap of indicators for each pair of screening sets output from the multiple independent training sessions, and selecting the screening set with the highest average overlap with the other screening sets as the screening set with the highest overlap.
7. An index screening device based on a robust fractal autoencoder, characterized in that, include: The standardization module is configured to collect the original indicator dataset of the scenario to be evaluated and the evaluation labels corresponding to the samples, and map the values of each original indicator in the original indicator dataset to a preset range to eliminate the influence of units, thereby obtaining standardized indicator data; The configuration initialization module is configured to configure the feature selection layer, encoder, decoder, and task-oriented supervision module of the autoencoder, initialize the original weights of the feature selection layer, and map the original weights to exponential weights so that the exponential weights corresponding to each original index are always greater than zero; The window subset module is configured to determine the size of a dynamic selection window based on the current iteration number during multi-round iterative training. The size of the dynamic selection window shrinks from close to the total number of the original indicators to a preset target selection number as the number of iterations increases. It also extracts the original indicators with the highest index weights from the feature selection layer, whose number is equal to the size of the dynamic selection window, to form a local indicator subset. The dual-path reconstruction module is configured to, for each iteration, use the global reconstruction path to encode and decode all standardized indicator data weighted by the exponential weights to obtain the global reconstruction result, and use the local reconstruction path to encode and decode only the local indicator subsets to obtain the local reconstruction result. The features weighted by the exponential weights or the bottleneck layer features output by the encoder are then input into the task-oriented supervision module to obtain the prediction result of the evaluation task. The joint optimization module is configured to calculate a joint loss function based on the global reconstruction error, the local reconstruction error, the sparse regularization term applied to the exponential weights, and the task-oriented supervised loss between the prediction result and the evaluation label, and update the original weights of the feature selection layer through gradient descent; The indicator output module is configured to perform multiple independent rounds of iterative training on the same original indicator dataset. After meeting the preset reconstruction accuracy threshold or reaching the maximum number of iterations, it takes the selection set with the highest overlap among the multiple independent training sessions and outputs the original indicators in the selection set whose index weight ranks first and whose quantity is equal to the target selection quantity, as the selected evaluation indicator system.
8. The apparatus according to claim 7, characterized in that, The task-oriented supervision loss is either classification cross-entropy loss or regression loss; the encoder and decoder configured by the configuration initialization module are both single-layer linear networks without activation functions.
9. An electronic device, characterized in that, The electronic device includes: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.
