An unfolding multi-view unsupervised feature selection method for scene interaction

Abstract

The application relates to the field of computer vision, and particularly discloses an unfolding multi-view unsupervised feature selection method for scene interaction, which comprises the following steps: S1, acquiring multi-view data output by a feature extraction module of a VR or AR device; S2, constructing a multi-view unsupervised feature selection model and determining an overall objective function; S3, realizing feature selection and de-outlier value based on low-dimensional representation and reconstruction; S4, constructing a high-order similarity graph and fusing to obtain a consensus high-order similarity graph; S5, updating model parameters by using an alternating iteration algorithm; S6, accelerating model operation by using a deep unfolding algorithm; and S7, screening features based on model output. The technical scheme provided by the application solves the multi-view, semantic mining and time delay problems of the existing method LSAG, effectively improves the VR or AR interaction precision and stability, and reduces the cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly relates to a spread multi-view unsupervised feature selection method for scene interaction. BACKGROUND

[0002] Scene interaction is a technical process in which a user interacts with elements in a scene through gesture control, voice instructions or somatosensory device operation in a virtual reality (VR) or augmented reality (AR) environment, and obtains visual animation, tactile vibration or auditory prompt feedback. As a basic technology for virtual-real fusion human-computer interaction, the response efficiency and feedback correctness of scene interaction directly affect the user operation immersion and the overall interaction performance of the device.

[0003] At present, VR or AR scene interaction is affected by factors such as hardware device performance, environmental interference, algorithm adaptability and user differences, and is prone to problems such as motion recognition deviation, response lag or inaccurate scene feedback, which reduces the user operation efficiency and immersive experience. To solve this problem, a common technical solution is to use a multi-sensor fusion system, which integrates a grayscale camera for simultaneous localization and mapping (SLAM), an RGB camera for environmental perception, an inertial measurement unit for posture measurement and an eye tracking camera for capturing line of sight information. For the image or video data collected by the camera, a pre-trained convolutional neural network is used to extract deep features and convert them into a high-dimensional feature vector. For the time series data generated by the inertial measurement unit or eye tracking device, the mean, variance and other statistical features are calculated within a sliding window and spliced into a feature vector. Then, the feature vectors of different data sources are input into a downstream machine learning model to perform feature matching, loop detection and map construction, etc. tasks to improve the robustness of device positioning and environmental perception and improve the overall performance of scene interaction.

[0004] The high-dimensional feature vector converted from multi-sensor data has problems such as high data dimension, information redundancy and abnormal value interference, and directly inputting the downstream task model will reduce the processing efficiency and accuracy of the model, and further affect the interaction performance of the VR or AR device. To solve this problem, a multi-view unsupervised feature selection method is introduced, which takes the high-dimensional feature vector as input, mines the consistency and complementarity between different views, automatically selects the most discriminative feature subset without manual annotation, realizes data dimension reduction and key information enhancement. The representative method in the field of multi-view unsupervised feature selection is LSAG, which adopts double consensus latent semantic learning, improves the representation quality of consensus latent semantics by fusing data projection and graph projection, and guides feature selection. At the same time, in the double anchor graph learning, the local structure of the data is extracted through the feature weighting subspace learning module and projected into the consensus latent semantic space.

[0005] But for the special application scenario of VR or AR device scene interaction, LSAG has the following limitations:

[0006] LSAG directly learns the consensus semantic structure from multiple view data and the local structure of the projected data to avoid noise interference with the anchor graph structure, but it does not make full use of the diversity of multi-view data, and does not explicitly consider the information loss caused by abnormal values and the projection process, which will lose unique information of each view, and will cause operation interference due to abnormal values, which is easy to cause positioning drift, jitter and even tracking loss, which seriously damages the user's immersive experience;

[0007] LSAG extracts the latent semantic information from two angles, but the extracted latent semantic information is only limited to describing the local structure of the data, ignoring the high-order manifold structure that can more essentially reflect the internal association of the data, resulting in the device being unable to effectively mine deep semantic information, and may produce operation misjudgment under complex environmental changes or atypical user behavior;

[0008] LSAG uses a conventional iterative algorithm to solve each parameter, and although this kind of algorithm can theoretically converge to a better solution, the calculation convergence speed is slow, the iteration number is large, and the time consumption is long, while the VR or AR application requires a delay of milliseconds, and this calculation time consumption cannot be accepted, and the long processing time will cause a delay between user operation and visual feedback, which not only reduces the user experience, but also easily causes motion sickness.

[0009] Therefore, how to propose an efficient, robust and low-delay feature selection method to improve the perception accuracy and stability of VR or AR device scene interaction in complex environments, while greatly reducing the computational overhead and response delay, has become a key and difficult point in the field of VR or AR scene interaction technology. SUMMARY

[0010] The present application aims at the above problems of the prior art, and provides an unfolding multi-view unsupervised feature selection method for scene interaction, which solves the problems of insufficient utilization of LSAG multi-view diversity, influence of outliers and information loss, limitations of semantic mining, and time delay exceeding the standard, effectively improves the perception accuracy and stability of scene interaction of VR or AR equipment in a complex environment, and reduces the calculation overhead and response delay.

[0011] To achieve the above object, the present application provides an unfolding multi-view unsupervised feature selection method for scene interaction, comprising:

[0012] S1, acquiring multi-view data output by a feature extraction module of a VR or AR equipment: let the data be a feature vector synchronized with a road time stamp, wherein, is the number of views, i.e. the number of sensors, and the view at any time t has feature vectors, and the feature vectors at the time are spliced into a data matrix , , a total data matrix of the views ;

[0013] S2, constructing a multi-view unsupervised feature selection model and determining an overall objective function: the objective function L is:

[0014] ;

[0015] wherein, is the original data input for the current view, is a projection matrix, is a low-dimensional representation of the data, is an outlier term, is an o-order similarity matrix of each view, is a consensus high-order similarity matrix of all views, is a view weight allocated according to the contribution of each view, is a Laplacian matrix of the consensus high-order similarity matrix, , , is a hyperparameter;

[0016] S3, realizing feature selection and outlier removal based on the low-dimensional representation and reconstruction: learning the projection matrix mapping to a low-dimensional space to obtain , and inversely reconstructing and by , identifying the non-structural sparse part in the reconstruction error as an outlier matrix , and imposing on Norm constraints, for Apply Norm sparsity constraints enable feature selection and outlier removal;

[0017] S4. Construct and fuse high-order similarity graphs to obtain a consensus high-order similarity graph: based on Build similarity graphs of each order for each view , ... In the formula, the first-order similarity matrix elements The second-order similarity matrix is ​​calculated using the Gaussian kernel function. of Elements are quantified by verifying the consistency of their neighborhood sets, using a higher-order similarity matrix. The corresponding normalized Laplace matrix Then, based on the data quality of each view, the weight coefficient b is adaptively learned. v The weighted high-order similarity graphs of each view are merged into a consensus high-order similarity graph S*.

[0018] S5. Update model parameters using an alternating iterative algorithm: alternating update , , In the formula Gradient descent is used for updating. Update using the hard threshold method according to The relationship between the size of n and the iteration method is used to update S* by optimizing each row vector. v Solve by transforming it into a quadratic programming problem with linear constraints;

[0019] S6. Accelerate model computation through deep expansion algorithm: Map the alternating iterative optimization process to the forward propagation process of a lightweight neural network. The iteration steps correspond to fixed layers of the network. The algorithm hyperparameters are transformed into learnable weight parameters of the network. Construct a feature selection block that includes a low-dimensional representation module, an outlier removal module, a feature selection module, a high-order similarity module, and a view weight module to complete parameter updates and model acceleration.

[0020] S7. Feature Filtering Based on Model Output: Based on the feature matrix obtained from the model... Scoring each feature and selecting the top-k features yields the input data X for the downstream task. v '∈ℝ nxfv f v The number of features to select.

[0021] Preferably, in S4, the elements s of the first-order similarity matrix S ijThe specific calculation method is:

[0022] ;

[0023] In the formula, The neighbor set, The neighbor set, and σ is a Gaussian kernel function parameter.

[0024] The elements of the second-order similarity matrix S 2 The specific calculation method is:

[0025]

[0026] In the formula, The neighbor set, The neighbor set, The neighbor set, The neighbor set, to satisfy The second-order similarity matrix elements are calculated for the data point pairs with the same k-neighbor set of

[0027] The calculation method of the high-order similarity matrix S O

[0028]

[0029] In the formula is the first-order similarity matrix, The order similarity matrix, and satisfies

[0030] Preferably, in S5, The update formula of

[0031] ;

[0032] In the formula, , Respectively, the updated and current values, is the gradient value of the objective function to μ is the step size parameter.

[0033] Preferably, in S5, ​​​​​​​​​The update formula of is

[0034] ;

[0035] where, is a hyperparameter.

[0036] Preferably, in S5, The update formula of is

[0037] ;

[0038] where, , , , is decomposed into , three parts, , .

[0039] Preferably, in S5, b v The update of corresponds to a quadratic programming problem:

[0040] ;

[0041] ;

[0042] where, b = [b1,...,b v ], M is a matrix, n is a vector, and 1 is an all-one vector.

[0043] Preferably, in S6, in the low-dimensional representation module, The update formula of is

[0044] ;

[0045] where, μ t is a trainable parameter, and t is a step index.

[0046] Preferably, in S6, in the outlier removal module, The update formula of is

[0047] ;

[0048] where, = - , α t , ρ t are trainable parameters, and t is a step index.

[0049] Preferably, in S6, in the feature selection module, The update formula of is

[0050]

[0051] wherein,

[0052]

[0053]

[0054] ;

[0055]

[0056]

[0057]

[0058]

[0059] ;

[0060] wherein, denotes the step index, and denotes the momentum term, correspond to the step size and momentum parameter of the -th step in the update process, respectively.

[0061] Preferably, in S6, the update formula of S* in the high-order similarity module is:

[0062] ;

[0063] ;

[0064] wherein, , is the high-order similarity matrix of the corresponding view, is the view weight, is the hyperparameter, and t is the trainable parameter, and t is the step index.

[0065] Preferably, in S6, the update formula of S* in the view weight module is:

[0066] ;

[0067] wherein, W t = I - 2ηM t , c t = ηn t , η is the step size parameter, M t , n t ​are the matrix and vector under t steps respectively, and I is the unit matrix.

[0068] Therefore, the present application proposes a kind of for scene interaction unrolled multi-view unsupervised feature selection method, its beneficial effects are as follows:

[0069] (1) by low-dimensional representation and outlier elimination module, the high-dimensional data of multi-source sensor input is projected to low-dimensional space, while retaining the core structural information necessary for scene understanding, reduce the hardware load of VR or AR device and optimize its power consumption, the module can suppress the outliers in the original data, to ensure the stability of virtual object in real space and the visual continuity of virtual scene, bring higher scene recognition accuracy and robustness to system, finally in complex dynamic environment for user provides a stable clear immersive field of view.

[0070] (2) with the aid of high-order similarity module, system can build and fuse the high-order similarity graph of each data view, reveal the deep structure association inside data, in VR or AR interactive scene, through structured information input, can provide accurate data basis for downstream complex interaction tasks, such as accurate grasping or gesture recognition, significantly improve the accuracy of device processing user instruction, so that the operation response is more accurate when user interacts, process is more smooth and natural.

[0071] (3) by depth unrolling algorithm, the iterative optimization process in solving model is unrolled into a depth network architecture with fixed number of layers, this method maps each iteration of optimization algorithm to a forward propagation layer of network, and converts part of key variables into trainable parameters in network, this method shortens the operation time to millisecond level, realizes the instant synchronization between user physical action and system visual feedback under the premise of ensuring high precision, creates seamless and deep immersive scene interaction experience for user.

[0072] The technical solutions of the present application will be described in further detail below with the aid of drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0073] Fig. 1 is the high-order similarity matrix construction flow chart of the unrolled multi-view unsupervised feature selection method for scene interaction of the present application;

[0074] Fig. 2 is the calculation schematic diagram of each-order similarity matrix construction of the unrolled multi-view unsupervised feature selection method for scene interaction of the present application;

[0075] Fig. 3 is the overall method flow chart of the unrolled multi-view unsupervised feature selection method for scene interaction of the present application. DETAILED DESCRIPTION

[0076] In order to make the technical solutions, advantages and objectives of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below. The described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without any inventive effort fall within the protection scope of the present application.

[0077] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those of ordinary skill in the art to which the present application belongs.

[0078] As shown in Figs. 1-3 , the present application provides an unfolding multi-view unsupervised feature selection method for scene interaction, comprising:

[0079] S1, acquiring multi-view data output by a feature extraction module of a VR or AR device: let the data be a feature vector synchronized with a time stamp, wherein, the number of views is the number of sensors, and the view has feature vectors at an arbitrary time t, and the feature vectors of times are spliced into a data matrix , ;

[0080] S2, constructing a multi-view unsupervised feature selection model and determining an overall objective function: the objective function L is:

[0081] ;

[0082] In the formula, is the original data input for the current view, is a projection matrix, is a low-dimensional representation of the data, is an outlier term, is an o-order similarity matrix of each view, is a consensus high-order similarity matrix of all views, is a view weight allocated according to the contribution of each view, is a Laplacian matrix of the consensus high-order similarity matrix, , , is a hyperparameter;

[0083] S3, realizing feature selection and outlier removal based on the low-dimensional representation and reconstruction: learning the projection matrix to map to a low-dimensional space to obtain​ ,pass and Reverse Restructuring The non-structural sparse portion of the reconstruction error is identified as an outlier matrix. ,right Apply Norm constraints, for Apply Norm sparsity constraints enable feature selection and outlier removal;

[0084] S4. Construct and fuse high-order similarity graphs to obtain a consensus high-order similarity graph: based on Build similarity graphs of each order for each view , ... In the formula, the first-order similarity matrix elements The second-order similarity matrix is ​​calculated using the Gaussian kernel function. of Elements are quantified by verifying the consistency of their neighborhood sets, using a higher-order similarity matrix. , corresponding to the Laplace matrix Then, based on the data quality of each view, the weight coefficient b is adaptively learned. v The weighted high-order similarity graphs of each view are merged into a consensus high-order similarity graph S*.

[0085] The elements s of the first-order similarity matrix S ij The specific calculation method is as follows:

[0086] ;

[0087] In the formula, for of Nearest neighbor set for of The nearest neighbor set, where σ is the parameter of the Gaussian kernel function.

[0088] Second-order similarity matrix S 2 elements The specific calculation method is as follows:

[0089]

[0090] In the formula, for of Nearest neighbor set for of Nearest neighbor set for of a set of neighbors, for a set of neighbors, i.e. to compute the second order similarity matrix elements for the same pair of data points as the set of k-neighbors that satisfy a set of neighbors, i.e. to compute the second order similarity matrix elements for the same pair of data points as the set of k-neighbors that satisfy a set of neighbors, i.e. to compute the second order similarity matrix elements for the same pair of data points as the set of k-neighbors that satisfy

[0091] The high order similarity matrix S O is computed as:

[0092]

[0093] where is the first order similarity matrix, is the order similarity matrix, and satisfies .

[0094] S5, update the model parameters using an alternating iterative algorithm: update , , where is updated using gradient descent, is updated using a hard thresholding method, is updated according to the size relationship between and n to select different iteration methods for updating, S* is updated by optimizing each row vector, and b v is solved by transforming it into a quadratic programming problem with linear constraints;

[0095] The update formula for

[0096] ;

[0097] where, , are the updated and current values of respectively, is the gradient value of the objective function with respect to , and μ is the step size parameter.

[0098] The update formula for

[0099] ;

[0100] where, , and α is a hyperparameter.

[0101] The update formula for

[0102] ;

[0103] where, , , , Decomposed by SVD , Three parts , .

[0104] b v The quadratic programming problem corresponding to the update is:

[0105] ;

[0106] ;

[0107] In the formula, b = [b1,...,b v ], M is a matrix, n is a vector, and 1 is a vector consisting entirely of 1s.

[0108] S6. Accelerate model computation through deep expansion algorithm: Map the alternating iterative optimization process to the forward propagation process of a lightweight neural network. The iteration steps correspond to fixed layers of the network. The algorithm hyperparameters are transformed into learnable weight parameters of the network. Construct a feature selection block that includes a low-dimensional representation module, an outlier removal module, a feature selection module, a high-order similarity module, and a view weight module to complete parameter updates and model acceleration.

[0109] In the low-dimensional representation module The update formula is:

[0110] ;

[0111] In the formula, μ t t represents the trainable parameter and t represents the step index.

[0112] In the outlier removal module The update formula is:

[0113] ;

[0114] In the formula, = - α t ρ t t represents the trainable parameter and t represents the step index.

[0115] In the feature selection module The update formula is:

[0116]

[0117] in,

[0118]

[0119]

[0120] ;

[0121]

[0122]

[0123]

[0124]

[0125] ;

[0126] In the formula, Indicates step index, and Represents the momentum term. These correspond to the first step in the update process. The step size and momentum parameters.

[0127] The update formula for S* in the higher-order similarity module is:

[0128] ;

[0129] ;

[0130] In the formula, , This is the higher-order similarity matrix of the corresponding views. For view weight, For hyperparameters, θ t t represents the trainable parameter and t represents the step index.

[0131] In the view weight module The update formula is:

[0132] ;

[0133] In the formula, W t =I-2ηM t c t =ηn t η is the step size parameter, M t n t Let be the matrix and vector at step size t, respectively, and I be the identity matrix.

[0134] S7. Feature Filtering Based on Model Output: Based on the feature matrix obtained from the model... Scoring each feature and selecting the top-k features yields the input data X for the downstream task.v ∈ ℝ nxfv , f v is the selected number of features.

[0135] Example One,

[0136] As Figs. 1-3 shown, the present application provides a kind of for scene interaction to verify the performance of unfolding multi-view unsupervised feature selection method, focus on the performance of the method in multi-view data processing, feature selection effect and operation efficiency, combined with the actual application demand of VR or AR device scene interaction, design the following verification experiment, specific steps are as follows:

[0137] I. Experimental data preparation:

[0138] Data source:

[0139] Real scene interaction data collected by VR or AR device multi-sensor synchronously, covering typical interaction scenes, such as gesture control, voice command response, somatosensory operation, etc., integrating multi-view data, the number of views is consistent with the number of sensors, including gray camera data for synchronous positioning and map construction SLAM, RGB camera data for environment perception, inertial measurement unit IMU data for posture measurement, eye tracking camera data for capturing visual information;

[0140] Camera data: extract deep features by pre-trained convolutional neural network, and convert them into high-dimensional feature vectors;

[0141] Inertial measurement unit, eye tracking device data: calculate mean, variance and other statistical features in sliding window, and splice them into feature vectors;

[0142] Data preprocessing: synchronize the time stamps of each view data to ensure that the feature vectors of different view data at the same time are aligned; collect n feature vectors at different times, and construct view data matrix Total data matrix ;

[0143] Comparison method: select LSAG, a representative method in the field of multi-view unsupervised feature selection, as the comparison group, and the present application method as the experimental group, to ensure that the hardware environment (same VR or AR device and supporting computing hardware) and software environment (consistent programming language and machine learning framework) of the two groups are consistent.

[0144] II. Definition of experimental indicators:

[0145] Feature selection effect index: the accuracy of the downstream scene interaction task (such as action recognition, scene matching) as the core index, the accuracy calculation formula is: accuracy = correct response / total interaction times x 100%, evaluate the promotion effect of the selected features on the task performance;

[0146] Operation efficiency index: record the total processing time (unit: millisecond, ms) of the two groups of methods from inputting multi-view data to outputting selected features, verify whether it meets the millisecond level delay requirement of VR or AR device; At the same time, the number of model iteration convergence is counted, and the convergence speed is compared;

[0147] Robustness index: add different intensity of salt and pepper noise outliers (simulate interference caused by environment or sensor abnormality) in the original data, calculate the task accuracy decay rate of the two groups of methods under the interference of abnormal values, decay rate = (no abnormal value accuracy-no abnormal value accuracy) / no abnormal value accuracy x 100%, the lower the decay rate, the stronger the anti-interference ability of the method.

[0148] Three, model parameter setting:

[0149] Method parameters of the application: set the initial value of the hyperparameter, and the specific value is adjusted combined with experimental data; In the depth expansion module, the number of feature selection blocks corresponds to the number of iteration steps of the original iteration algorithm, and the trainable parameters are updated by using a gradient descent type optimizer, and the learning rate is adjusted according to the convergence condition of the experiment;

[0150] LSAG method parameters: refer to the original algorithm configuration parameters, and the iteration convergence threshold is set to a conventional value (such as 1e-6), so that it is in the optimal performance state.

[0151] Four, experimental step execution:

[0152] Model training and feature selection:

[0153] Experimental group: strictly follow the S2-S7 process, first construct a multi-view unsupervised feature selection model and determine the overall objective function, realize feature selection and outlier removal through low-dimensional representation and reconstruction, construct high-order similarity graph of each view based on low-dimensional representation and fuse into consensus high-order similarity graph, update 、 、 Parameters, and then map the iteration process to neural network forward propagation through the depth expansion algorithm, and finally score the features according to the feature matrix Select top-k features;

[0154] Comparison group: process the same multi-view data according to the original algorithm process of LSAG, complete feature selection and output top-k features;

[0155] Downstream task test: input the top-k features screened by the two groups of methods into the downstream scene interaction task model, calculate the task accuracy on the test data set, repeat the test multiple times to take the average value, and reduce the random error;

[0156] Operation efficiency statistics: use a timing tool to record the total time of each data processing of the two groups of methods, and calculate the average processing time of multiple experiments; record the number of LSAG iteration convergences and the number of layers of neural network forward propagation of the method of the application;

[0157] Robustness test: add different intensity salt and pepper noise outliers to the original data to generate a data set containing outliers, repeat the above model training, feature selection and downstream task test steps, and calculate the accuracy decay rate of the two groups of methods.

[0158] Five, experimental result record and analysis:

[0159] The method of the application has higher downstream task accuracy than LSAG under different k values because it mines deep structure of data through high-order similarity graph, reconstructs to retain effective information without outliers, and the selected features are more discriminative; the processing time of the method of the application is much lower than that of LSAG by converting the iteration process into neural network forward propagation through the deep expansion algorithm, which meets the millisecond level delay requirement of VR or AR; LSAG cannot adapt to the timeliness requirement of scene interaction because of too many iteration times and too long processing time; the accuracy decay rate of the application under different outlier intensities is lower than that of LSAG by reconstructing and splitting outliers, which shows that it has stronger anti-environmental interference ability and is more suitable for VR or AR complex interactive environment.

[0160] Therefore, the application provides an expansion type multi-view unsupervised feature selection method for scene interaction, which retains the core structure of data through a low-dimensional representation and outlier removal module, mines deep correlations through a high-order similarity module, and accelerates operation through a deep expansion module, effectively solving the defects of LSAG in multi-view utilization, semantic mining and time delay, and improving the precision, robustness and real-time performance of VR or AR device scene interaction.

[0161] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the application and not to limit them, although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the application.

Claims

1. An unsupervised feature selection method for unfolding multi-view applications in scene interaction, characterized in that, include: S1. Obtain multi-view data output by the feature extraction module of the VR or AR device: Let the data be... Feature vectors of time-stamp synchronization. The number of views is the same as the number of sensors. At any time t, the number of views is... Each feature vector will The feature vectors at each time step are concatenated into a data matrix. , Total data matrix of each view ; S2. Construct a multi-view unsupervised feature selection model and determine the overall objective function: The objective function L is: ； In the formula, This is to input raw data for the current view. For the projection matrix, For low-dimensional representation of data, For outlier items, Let O be the similarity matrix of each view. For the consensus high-order similarity matrix of all views, View weights are assigned based on the contributions of each view. Let Laplace's matrix be the consensus high-order similarity matrix. , , For hyperparameters; S3. Feature selection and outlier removal based on low-dimensional representation and reconstruction: learning the projection matrix. Will Mapping to a lower-dimensional space yields ,pass and Reverse Restructuring The non-structural sparse portion of the reconstruction error is identified as an outlier matrix. ,right Apply Norm constraints, for Apply Norm sparsity constraints enable feature selection and outlier removal; S4. Construct and fuse high-order similarity graphs to obtain a consensus high-order similarity graph: based on Build similarity graphs of each order for each view , ... In the formula, the first-order similarity matrix elements The second-order similarity matrix is ​​calculated using the Gaussian kernel function. of Elements are quantified by verifying the consistency of their neighborhood sets, using a higher-order similarity matrix. The corresponding normalized Laplace matrix ; Then, the weight coefficient b is adaptively learned based on the data quality of each view. v The weighted high-order similarity graphs of each view are merged into a consensus high-order similarity graph S. ; S5. Update model parameters using an alternating iterative algorithm: alternating update , , S b v In the formula Gradient descent is used for updating. Update using the hard threshold method according to The relationship between the size of n and the iteration method for updating S is chosen. By optimizing the updates of each row vector, b v Solve by transforming it into a quadratic programming problem with linear constraints; S6. Accelerate model computation through deep expansion algorithm: Map the alternating iterative optimization process to the forward propagation process of a lightweight neural network. The iteration steps correspond to fixed layers of the network. The algorithm hyperparameters are transformed into learnable weight parameters of the network. Construct a feature selection block that includes a low-dimensional representation module, an outlier removal module, a feature selection module, a high-order similarity module, and a view weight module to complete parameter updates and model acceleration. S7. Feature Filtering Based on Model Output: Based on the feature matrix obtained from the model... Scoring each feature and selecting the top-k features yields the input data X for the downstream task. v '∈ℝ nxfv f v The number of features to select; In S4, the elements s of the first-order similarity matrix S ij The specific calculation method is as follows: ； In the formula, for of Nearest neighbor set for of The nearest neighbor set, where σ is the parameter of the Gaussian kernel function; Second-order similarity matrix S 2 elements The specific calculation method is as follows: In the formula, for of Nearest neighbor set for of Nearest neighbor set for of Nearest neighbor set for of Nearest neighbor set, that is, satisfying and Calculate the elements of the second-order similarity matrix for pairs of data points with the same k-nearest neighbor set; High-order similarity matrix S O The calculation method is as follows: In the formula It is a first-order similarity matrix. for A similarity matrix of order 1, satisfying .

2. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S5, The update formula is: ； In the formula, , These are the updated and current versions, respectively. value, For the objective function pair The gradient value, where μ is the step size parameter.

3. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S5, The update formula is: ； In the formula, α is a hyperparameter.

4. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S5, The update formula is: ； In the formula, , , , Decomposed by SVD , Three parts , .

5. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S5, b v The quadratic programming problem corresponding to the update is: ； ； In the formula, b = [b1,...,b v ], M is a matrix, n is a vector, and 1 is a vector consisting entirely of 1s.

6. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S6, in the low-dimensional representation module The update formula is: ； In the formula, μ t t represents the trainable parameter and t represents the step index.

7. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S6, the outlier removal module The update formula is: ； In the formula, = - α t ρ t Here are the trainable parameters, and t is the step index; In S6, in the feature selection module The update formula is: in, ； ； In the formula, Indicates step index, and Represents the momentum term. These correspond to the first step in the update process. The step size and momentum parameters.

8. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S6, the higher-order similarity module S The update formula is: ； ； In the formula, , This is the higher-order similarity matrix of the corresponding views. For view weight, For hyperparameters, θ t t represents the trainable parameter and t represents the step index.

9. The unfolded multi-view unsupervised feature selection method for scene interaction according to claim 1, characterized in that, In S6, in the view weight module The update formula is: ； In the formula, W t =I-2ηM t c t =ηn t η is the step size parameter, M t n t Let be the matrix and vector at step size t, respectively, and I be the identity matrix.

Images