Intelligent tax assistant interaction method

Abstract

The invention provides an intelligent tax assistant interaction method, which comprises the following steps: if an intelligent tax assistant judges that the use property of an area is office or residence, training existing tax classification sample data by adopting a supervised machine learning algorithm, and establishing a tax classification model for judging tax rate categories corresponding to the office area and the residence area; if the intelligent tax assistant judges that the use property of the region is a mixed property, the mixed property region is subdivided into a plurality of sub-regions through a clustering analysis algorithm, and the sub-regions are determined to be office property or residential property according to the area proportion and use characteristics of the sub-regions; analyzing the influence weight of the area and the tax rate of each region on the comprehensive tax rate through a logistic regression algorithm, calculating the comprehensive tax rate of the office region and the residence region by adopting a weighted average method, and determining the proportion coefficient of each region in tax calculation; and generating an intelligent tax suggestion report in the intelligent tax assistant according to the comprehensive tax rate of the building, the tax amount to be paid and the tax optimization scheme.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an intelligent tax assistant interaction method. Background Art

[0002] After the internal partition renovation of civil properties, the same building may be mixed with office and residential use, which brings significant data processing challenges to tax classification. When dividing the area for different uses and calculating the tax amount, the intelligent tax assistant faces the technical difficulty of how to accurately define the office use and residential use. When the same building is partitioned to form a mixed use of office and residential use, how to accurately divide the area for different uses and apply the corresponding tax rates becomes a thorny issue. The mixed use of office and residential may make it difficult to separate and measure water and electricity costs, further increasing the complexity of tax calculation. Traditional manual measurement and calculation methods are not only time-consuming and labor-intensive, but also difficult to ensure the accuracy and consistency of data. If a unified tax rate is simply determined according to the main use of the building, it may lead to unfair taxation and it is difficult to truly reflect the actual use of the building. Therefore, there is an urgent need for an intelligent technical means that can quickly and accurately identify and divide the space area for different uses inside the building, calculate the tax payable according to the corresponding tax rate standards, and generate a detailed tax analysis report to provide a reliable basis for the tax collection and management work of the tax department. Summary of the invention

[0003] The present invention provides an intelligent tax assistant interaction method, which mainly includes: Through laser scanning and panoramic cameras, the spatial layout diagram and area data of each area after the renovation of the internal partitions of the building are obtained, and image segmentation and target detection are performed on the spatial layout diagram to extract the geometric features and spatial relationships of each area. Combined with the building design specifications and tax classification standards, the use nature of each area is determined, including office, residential and mixed use; If the intelligent tax assistant determines that the area is used for office or residential purposes, it uses a supervised machine learning algorithm to train the existing tax classification sample data and establish a tax classification model to determine the tax rate category corresponding to the office area and the residential area; If the intelligent tax assistant determines that the use nature of the area is mixed, the mixed nature area will be subdivided into several sub-areas through the cluster analysis algorithm, and the sub-areas will be determined to be office or residential based on the area proportion and use characteristics of the sub-areas; Through case-based reasoning algorithms, similar cases to the sub-region are retrieved from the existing tax case database, and relevant attributes, including area, usage function, and occupancy, are compared and analyzed to predict the tax rate tendency of the region and form preliminary tax recommendations; The logistic regression algorithm is used to analyze the influence of the area and tax rate of each region on the comprehensive tax rate, and the weighted average method is used to calculate the comprehensive tax rate of office and residential areas, and the proportion coefficient of each region in the tax calculation is determined; Based on the area, tax rate and proportion coefficient of each area, a dynamic programming algorithm is used to establish a tax calculation model to obtain the tax payable for the building. At the same time, the sensitivity analysis method is used to evaluate the impact of changes in regional area and tax rate on the tax amount, and a tax optimization plan is formed through the intelligent tax assistant; Generate an intelligent tax recommendation report in the Intelligent Tax Assistant based on the building’s comprehensive tax rate, tax payable, and tax optimization plan.

[0004] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses an intelligent tax assistant interaction method, which is used to classify taxes and calculate tax amounts for renovated spaces inside buildings. The spatial layout diagram is obtained by laser scanning and panoramic cameras, and the characteristics of each area are extracted by image segmentation and target detection, and the nature of use is judged in combination with building specifications. For office and residential areas, supervised learning is used to establish a tax classification model; for mixed-nature areas, cluster analysis is used to subdivide sub-areas. Case reasoning is also used to retrieve similar cases from the case library to predict tax rate tendencies; logistic regression is used to analyze the influence weights of various factors on the tax rate, and dynamic programming is used to establish a tax amount calculation model. Finally, an intelligent tax recommendation report is generated to provide decision support for building tax classification and tax amount calculation, thereby improving the intelligence and precision of tax management. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] Figure 1 The present invention is a flowchart of an intelligent tax assistant interaction method.

[0006] Figure 2 A schematic diagram of an intelligent tax assistant interaction method of the present invention. DETAILED DESCRIPTION

[0007] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0008] like Figure 1-2 In this embodiment, a smart tax assistant interaction method may specifically include: Step S101, through laser scanning and panoramic cameras, obtain the spatial layout diagram and area data of each area after the internal partition renovation of the building, perform image segmentation and target detection on the spatial layout diagram, extract the geometric features and spatial relationships of each area, and combine the building design specifications and tax classification standards to determine the use nature of each area, which includes office, residential and mixed use.

[0009] A three-dimensional laser scanner is used to obtain first point cloud data inside the building, and a first texture point cloud model is obtained by aligning the first point cloud data with the panoramic image data; a spatial clustering process is performed on the first texture point cloud model to obtain a second texture point cloud model, and a region growing algorithm is used to extract wall surfaces based on the second texture point cloud model and calculate the relative position relationship between the walls to obtain a first spatial connectivity graph; a first plane layout diagram is obtained by vectorizing the first spatial connectivity graph, and an image segmentation process is performed on the first plane layout diagram to obtain a second plane layout diagram; a shape feature vector and a spatial topology feature vector are extracted from the second plane layout diagram, and a support vector machine algorithm is used to train the shape feature vector and the spatial topology feature vector to obtain a room use classification model, and the room is labeled for office and residential use according to the classification model.

[0010] Specifically, a three-dimensional laser scanner is used to scan the interior of a building from multiple collection points and collect first point cloud data. A panoramic camera synchronously collects panoramic image data according to the position of the collection points. The first point cloud data is registered and spliced ​​to obtain second point cloud data. A projection transformation algorithm is used to perform geometric correction on the panoramic image data and register it with the second point cloud data to obtain a first texture point cloud model. According to the first texture point cloud model, the point cloud data is subjected to denoising and spatial clustering to obtain a second texture point cloud model. The indoor wall surface and floor surface are extracted using a regional growing algorithm. The wall thickness parameters and floor elevation parameters are obtained using a least squares plane fitting method. The relative position relationship between the walls is calculated through spatial topological analysis to obtain a first spatial connectivity graph. Based on the first spatial connectivity graph, a vectorization process is used to obtain a first plane layout diagram. The first plane layout diagram is segmented to obtain a second plane layout diagram. The wall contour is extracted from the second plane layout diagram and the door and window opening positions are identified. The room use area data is calculated using a polygon area calculation method. According to the second floor plan, the room shape feature vector is extracted, including the room length-to-width ratio, area ratio, and perimeter ratio. Based on the first spatial connectivity map, the spatial topological feature vector is extracted, including the number of adjacent rooms, door and window connectivity, and channel width. The support vector machine algorithm is used to establish a room use classification model. The classifier is trained by the shape feature vector and the topological feature vector. The classification threshold is set according to the minimum use area of ​​office space and the residential design specification, and the room is labeled for office, residential, and mixed use. The second spatial connectivity map is constructed based on the labeled results. The graph theory algorithm is used to analyze the spatial organization relationship of rooms with different uses, calculate the degree of aggregation of rooms with the same use, and determine the dominant use nature of each area of ​​the building. In the field of determining the use nature after the renovation of the interior of the building, when collecting point cloud data through a three-dimensional laser scanner, the scanner is set to a horizontal angular resolution of 0.02 degrees, a vertical angular resolution of 0.02 degrees, and a scanning distance of 0.5 meters to 30 meters. Multiple scanning stations are set inside the building, and the distance between adjacent stations is 3 meters to 5 meters to ensure that the overlap of the scanned data is above 40%. The image resolution collected by the panoramic camera is 8000×4000 pixels, the horizontal field of view of a single image is 360 degrees, and the vertical field of view is 180 degrees, ensuring that the image is free of distortion and chromatic aberration. The point cloud data registration adopts the iterative closest point algorithm. Each iteration step calculates the registration error of the two sets of point cloud data. When the error is less than 1 mm, the iteration is stopped to obtain accurately registered point cloud data. When the panoramic image is registered with the point cloud data, the image feature points and the point cloud feature points are extracted for matching, and the projection transformation matrix is ​​established to achieve accurate mapping of the image texture to the point cloud model. When extracting the wall surface and the floor surface, the region growing algorithm sets the normal vector angle threshold to 3 degrees and the plane fitting residual threshold to 2 mm, which can reliably extract the main plane components inside the building. The least squares method is used to fit the wall plane equation, ax+by+cz+d=0, where (a,b,c) is the plane normal vector and d is the distance from the plane to the origin.The wall thickness is obtained by calculating the distance between adjacent wall planes. The typical wall thickness is 120 mm to 240 mm. When extracting the room shape features, the room length-to-width ratio = maximum side length / minimum side length, area ratio = usable area / building area, and perimeter ratio = actual perimeter / minimum circumscribed rectangle perimeter are calculated. Taking an office area as an example, the room length-to-width ratio is 1.8, the area ratio is 0.85, and the perimeter ratio is 1.2, indicating that the room is a regular rectangle. In the spatial topology features, the number of adjacent rooms is 4, the door and window connectivity is 6, and the channel width is 1500 mm, which meets the requirements of office space design. The support vector machine classifier uses the radial basis kernel function, K(x,y)=exp(-γ||xy||²), where γ is the kernel function parameter and K(x,y) is the kernel function value, which represents the similarity between the two input vectors x and y in the feature space. The classification threshold is set to a usable area of ​​not less than 4 square meters per person for office space and a usable area of ​​not less than 9 square meters for residential bedrooms. A building is divided into office area (42%), residential area (35%) and mixed-use area (23%). Spatial aggregation analysis shows that the dominant use of the building is office.

[0011] Step S102: If the intelligent tax assistant determines that the use nature of the area is office or residential, a supervised machine learning algorithm is used to train the existing tax classification sample data to establish a tax classification model for determining the tax rate category corresponding to the office area and the residential area.

[0012] The standard deviation method is used to identify outliers in the tax classification sample data, and the mean filling method is used to supplement the missing values ​​based on the outlier identification results to obtain the first sample data; the minimum and maximum value standardization processing is performed on the first sample data, and the standardization processing result is converted into a one-hot encoding to obtain the third sample data; the variance selection method is used to remove low-discrimination features from the third sample data, and the principal component analysis method is used to extract the principal components whose cumulative contribution rate reaches a specified threshold from the data after removing the low-discrimination features to obtain the sixth sample data; a random forest classifier is established based on the sixth sample data, and the probability value of the tax rate category is calculated by the random forest classifier. If the probability value is greater than the specified threshold, the corresponding tax rate category is adopted; if the probability value is less than the specified threshold, the tax rate category is determined by a probability weighting method.

[0013] Specifically, the tax classification sample data is processed for outliers through data cleaning, the triple standard deviation method is used to identify outliers, the mean filling method is used to supplement missing values ​​to obtain the first sample data, the numerical features in the first sample data are processed by minimum and maximum value standardization to obtain the second sample data, and the categorical features are converted to one-hot encoding to obtain the third sample data. The Pearson correlation coefficient is calculated based on the third sample data, and the redundant features with a correlation coefficient greater than 0.9 are removed to obtain the fourth sample data. The variance selection method is used to remove low-discrimination features with a variance less than 0.1 to obtain the fifth sample data. For the fifth sample data, key features such as building area, years of use, and regional value are extracted, and the feature contribution rate is calculated using the principal component analysis method. The principal component with a cumulative contribution rate of 90% is selected to obtain the sixth sample data. A random forest classifier is established based on the sixth sample data, and the number of decision trees is set to 100, the maximum depth of the tree is 10, and the minimum number of leaf node samples is 5. The classifier is trained using the five-fold cross validation method to obtain the first tax classification model. According to the first tax classification model, the newly added areas are predicted and the probability values ​​of each tax rate category are calculated. If the maximum probability value is greater than 0.8, the corresponding tax rate category is adopted. If the maximum probability value is less than 0.8, the tax rate category is determined by probability weighting. For mixed-use areas, the office area ratio and residential area ratio are calculated respectively, and the tax rates corresponding to different uses are weighted averaged using the area ratio to obtain the final tax rate. In the process of tax classification sample data processing, the building area data of the real estate shows an obvious outlier distribution. The mean of the building area is calculated to be 120 square meters and the standard deviation is 30 square meters. The three-times standard deviation method is used to identify records with a building area greater than 210 square meters or less than 30 square meters as outliers. There are 15% missing values ​​in the value characteristics of the area. The mean of this feature is calculated to be 12,000 yuan / square meter, and the missing data is filled with this mean. Correlation analysis was performed on 11 original features, including the age of real estate, house structure type, and decoration grade. By calculating the Pearson correlation coefficient, it was found that the correlation coefficient between the age of use and the newness rate of the house was 0.92, indicating that the two features have a strong linear correlation, and the age of use feature was retained. The house structure types include reinforced concrete, brick-concrete, and brick-wood, which are converted into three binary features using unique hot encoding. In the feature selection process, the variance value of each feature was calculated, and it was found that the variance of the house orientation feature was 0.08, which was less than the threshold value of 0.1, indicating that the feature had a low degree of discrimination for the sample. The principal component analysis method was used to reduce the dimensionality of the remaining features. The feature contribution rate of the first principal component was 45%, the second principal component was 28%, and the third principal component was 17%. The cumulative contribution rate reached 90%, and these three principal components were selected as classification features. The training process of the random forest classifier uses five-fold cross-validation, that is, the sample data is randomly divided into 5 parts, 4 parts are selected as training sets each time, and 1 part is selected as a validation set.By setting different parameter combinations for training, when the number of decision trees is 100, the maximum depth is 10, and the minimum number of leaf node samples is 5, the classification accuracy reaches the optimal value of 92%. For a newly added area with a building area of ​​150 square meters, the probability of the office tax rate category is 0.85, and the probability of the residential tax rate category is 0.15. Since the maximum probability value of 0.85 is greater than the threshold value of 0.8, it is determined that the office tax rate is applicable to this area. For another mixed-use area with a building area of ​​200 square meters, the office area is 120 square meters, the residential area is 80 square meters, the office area accounts for 0.6, the residential area accounts for 0.4, the office tax rate is 12%, and the residential tax rate is 4%. The weighted average calculation results in a comprehensive tax rate of 8.8% for this area.

[0014] Step S103: If the intelligent tax assistant determines that the use nature of the area is mixed, the mixed nature area is subdivided into several sub-areas through a cluster analysis algorithm, and the sub-areas are determined to be office or residential in nature based on the area proportion and use characteristics of the sub-areas.

[0015] The basic feature data of door and window positions, partition walls, and channel connections in mixed-nature areas are collected, and the maximum and minimum value normalization method is used to obtain the first feature vector; the Euclidean distance is calculated based on the first feature vector to obtain the first similarity matrix, and the density clustering algorithm is used to obtain the first sub-area division result; the room connectivity relationship is extracted based on the first sub-area division result, and the second similarity matrix is ​​obtained by calculating the channel width and connectivity between adjacent rooms, and the graph cut algorithm is used to obtain the second sub-area division result; the building area data is extracted based on the second sub-area division result, the crowd density data is collected through infrared sensors, and the hierarchical clustering algorithm is used to obtain office clusters and residential clusters, and the sub-area usage nature is determined according to the area weight and cluster feature weight.

[0016] Specifically, the first basic feature data such as the location of doors and windows, partition walls, channel connections, water and electricity layout, lighting and ventilation conditions, and equipment and facility configuration are collected for the mixed nature area, and the first basic feature data is normalized by the maximum and minimum value normalization method to obtain the first feature vector. The Euclidean distance is calculated based on the first feature vector to obtain the first similarity matrix, and the density clustering algorithm is used to subdivide the mixed nature area, setting the clustering radius to 0.5 and the minimum number of sampling points to 4 to obtain the first sub-area division result. Based on the first sub-area division result, the room connectivity relationship is extracted, the channel width and connection mode between adjacent rooms are calculated to obtain the second similarity matrix, and the graph cut algorithm is used to optimize the boundaries of the sub-areas that do not meet the connectivity requirements to obtain the second sub-area division result. For the second sub-area division result, the second basic feature data such as building area, partition type, and functional attributes are extracted, the crowd density data is collected through infrared sensors, and the equipment usage frequency data is extracted based on the video surveillance data to obtain the second feature vector. Based on the second eigenvector, the hierarchical clustering algorithm is used to classify the sub-areas. The number of clusters is set to 2, and the ward minimum variance method is used to obtain the office clusters and residential clusters. The area data of each sub-area is counted by the polygon area calculation method, and the total area ratio of the office area and the total area ratio of the residential area are calculated. The area weight of 0.7 and the cluster feature weight of 0.3 are used to determine the use nature of the sub-area. In the subdivision process of mixed-nature areas, the basic feature data contains multiple dimensions: the location of doors and windows is represented by coordinate points, such as (x1=350, y1=280) represents the upper left corner of the door; the partition wall is represented by a line segment, such as (x1=350, y1=280, x2=350, y2=580) represents a wall; the channel connection is recorded by the adjacency matrix, the value 1 indicates direct connection, and the value 0 indicates disconnection; the water and electricity layout records the spatial distribution density of water supply points and power sockets. These raw data are normalized, and the maximum and minimum value normalization method is used to map all eigenvalues ​​to the [0,1] interval. Taking a mixed area as an example, the door and window position coordinates are normalized from (350, 280) to (0.58, 0.47), the wall length is normalized from 300 cm to 0.6, and the channel width is normalized from 150 cm to 0.5. In density clustering, setting the cluster radius to 0.5 means that points with a Euclidean distance less than 0.5 are regarded as neighborhood points, and the minimum number of sampling points is 4, which means that at least 4 neighborhood points are required to form a cluster core. The calculation results show that the first sub-area division result contains 5 sub-areas, of which the largest sub-area area is 280 square meters and the smallest sub-area area is 85 square meters. The room connectivity analysis shows that sub-area 1 and sub-area 2 are connected by a corridor with a width of 180 cm, sub-area 2 and sub-area 3 are connected by a door opening with a width of 90 cm, and sub-area 4 and sub-area 5 are connected by a channel with a width of 150 cm.Based on these connectivity features, the graph cut algorithm is used to optimize the boundary, cut off the connection relationship with a channel width less than 120 cm, and re-divide it into 4 sub-areas. In the extraction of functional attribute features, the infrared sensor records the changes in the density of people flow within 24 hours. The office area has a higher density of people flow during working hours from 9:00 to 18:00, and the residential area has a double peak in the morning and evening. Video surveillance data shows that the frequency of office equipment use reaches a peak during working hours, while the frequency of living facilities use is higher during non-working hours. When hierarchical clustering is used, the inter-class distance is calculated by the ward minimum variance method. When the number of clusters is 2, the inter-class variance reaches the maximum value. The total area of ​​the office area is 320 square meters, accounting for 0.64; the total area of ​​the residential area is 180 square meters, accounting for 0.36. Combined with the clustering feature value, the comprehensive score is calculated according to the area weight of 0.7 and the feature weight of 0.3. The sub-area greater than 0.55 is determined as office nature, and the sub-area less than 0.45 is determined as residential nature.

[0017] Step S104, through the case reasoning algorithm, retrieve cases similar to the sub-region situation from the existing tax case database, compare and analyze relevant attributes, including area area, usage function, and personnel occupancy, predict the tax rate tendency of the area, and form preliminary tax recommendations.

[0018] According to the tax case database, the first attribute data of building area, usage function, and traffic convenience are extracted, and the feature vector is obtained through the minimum and maximum value normalization method and the one-hot encoding method; the cosine similarity method is used to retrieve cases with similarity greater than a threshold from the tax case database according to the feature vector to obtain a first case set; for the first case set, the second attribute data of regional area ratio, space utilization rate, and personnel density are extracted, and the principal component whose cumulative contribution rate meets the requirements is selected through the principal component analysis method to obtain the second case set; based on the second case set, a decision tree model is constructed, and the information gain ratio is used to select the optimal split feature extraction decision path, and the tax rate prediction model is obtained through multi-layer perceptron training.

[0019] Specifically, the first attribute data such as building area, use function, traffic convenience, decoration grade, equipment and facilities, and personnel capacity are extracted from the tax case database, the numerical attributes are normalized by the minimum and maximum value standardization method, and the categorical attributes are converted by the one-hot encoding method to obtain the feature vector. A distance measurement matrix is ​​constructed based on the feature vector, and similar cases are retrieved from the tax case database using the cosine similarity method. The similarity threshold is set to 0.8, and cases with similarity greater than the threshold are selected to form the first case set. Second attribute data such as regional area ratio, space utilization, and personnel density are extracted from the first case set. The principal component analysis method is used to reduce the dimension of the second attribute data, and the principal component with a cumulative contribution rate of 90% is selected to obtain the second case set. A decision tree is constructed based on the second case set, and the maximum depth of the tree is set to 5, the minimum number of leaf node samples is set to 10, and the information gain ratio is used to select the optimal split feature. The decision path is extracted to obtain the first rule set. The rule confidence and support are calculated based on the first rule set, and the confidence threshold is set to 0.7 and the support threshold is set to 0.1. The rules that meet the threshold conditions are screened to form the second rule set. The second rule set was trained using a multilayer perceptron, with the number of hidden layer nodes set to 10, the learning rate set to 0.01, and the number of training rounds set to 1000, to obtain a tax rate prediction model. Based on the tax rate prediction model, the probability distribution of sub-regions under different tax rate categories was calculated, and the tax rate prediction value was obtained using the probability weighting method. The prediction value was corrected for deviations by combining the regional use attribute, area ratio attribute, and use intensity attribute. The original attribute data obtained from the tax case database contains multiple dimensions. The building area value ranges from 50 to 1000 square meters, which is converted to the interval of 0 to 1 through minimum and maximum value standardization; the use function includes three types: office, commercial, and residential, which are converted into three binary features using one-hot encoding; the traffic convenience is represented by the distance from the subway station, which is recorded as 1 within 300 meters, 0.6 from 300 to 800 meters, and 0.2 above 800 meters. In the case retrieval process, cosine similarity is used to calculate the similarity between cases. The formula for cosine similarity is cos(θ)=(A·B) / (||A||·||B||), where A and B are the eigenvectors of the two cases. For an office area with a building area of ​​280 square meters and 450 meters away from the subway station, the eigenvector (0.28,1,0,0,0.6) is obtained after standardization. The similarity with all cases in the database is calculated, and 52 cases with a similarity greater than 0.8 are selected. In the principal component analysis process, the eigenvalues ​​of the feature covariance matrix are calculated to be 2.8, 1.5, 0.9, 0.5, and 0.3, respectively. The corresponding eigenvectors constitute the principal components. The contribution rate of the first principal component is 47%, the second principal component is 25%, and the third principal component is 18%. The cumulative contribution rate reaches 90%, and these three principal components are retained. The decision tree construction process uses the information gain ratio criterion to select split features. Information gain ratio = information gain / feature entropy.Taking the building area as an example, the samples are divided into two groups according to the area: less than 200 square meters and greater than or equal to 200 square meters. The entropy value before splitting is calculated to be 1.8, the entropy value after splitting is 0.9, the information gain is 0.9, the characteristic entropy is 0.8, and the information gain ratio is 1.125. The formula for calculating the confidence of the rule is confidence = number of samples that meet the rule / number of samples in the conditional part of the rule. Taking the rule "if the building area is greater than 200 square meters and is less than 500 meters away from the subway station, the tax rate is 12%" as an example, the total number of samples that meet the conditions is 40, of which 32 samples have a tax rate of 12%, and the confidence is calculated to be 0.8. The support calculation formula is support = number of samples that meet the rule / total number of samples, and the support of this rule is 0.15. During the training process of the multilayer perceptron, the back propagation algorithm is used to optimize the network parameters, and the loss function uses the cross entropy function. For the input samples, the output values ​​of neurons in each layer are calculated through forward propagation. The ReLU activation function is used for the hidden layer: f(x)=max(0,x). The output layer uses the Softmax function to convert the output into a probability distribution. After 1000 rounds of training, the prediction accuracy on the validation set reached 88%. In the process of correcting the prediction results, the weight of the use attribute is set to 0.4, the weight of the area ratio is set to 0.4, and the weight of the use intensity is set to 0.2. The probability of the office tax rate in a mixed-use area is predicted to be 0.65, and the probability of the residential tax rate is 0.35. Considering that the office area ratio in this area is 0.7, the tax rate prediction value is finally adjusted to the office tax rate.

[0020] Step S105, using a logistic regression algorithm, analyze the influence weight of each area and tax rate on the comprehensive tax rate, calculate the comprehensive tax rate of the office area and the residential area using a weighted average method, and determine the proportion coefficient of each area in the tax amount calculation.

[0021] The original attribute data of the building area is collected, and the original attribute data is processed by the minimum and maximum value standardization method to obtain the first standardized data; the mean variance of the area data series and the tax rate data series is calculated according to the first standardized data, and the data series is processed by the centralization method to obtain the second standardized data; a logistic regression predictor is constructed based on the second standardized data, and the predictor parameters are optimized by the stochastic gradient descent method to obtain the area weight value and the tax rate weight value; a linear combination equation is constructed according to the area weight value and the tax rate weight value, and the parameters of the equation are calculated by the least squares method to obtain the influence coefficient, the influence coefficient is normalized by the sigmoid function to obtain the regional contribution, and the tax amount calculation ratio coefficient is calculated by the additive combination method.

[0022] Specifically, original attribute data such as building area, usable area, actual utilization rate, traffic convenience, decoration grade, equipment and facilities are collected for office areas and residential areas, and the minimum and maximum value standardization method is used to normalize the original attribute data to obtain the first standardized data. The mean and variance of the area data series and the tax rate data series are calculated according to the first standardized data, and the data series are processed by the centralization method to obtain the second standardized data. A logistic regression predictor is constructed based on the second standardized data, and the learning rate is set to 0.01 and the number of iterations is 1000. The predictor parameters are optimized by the random gradient descent method to obtain the area weight value and the tax rate weight value. A linear combination equation is constructed according to the area weight value and the tax rate weight value, and the equation parameters are calculated by the least squares method to obtain the area influence coefficient and the tax rate influence coefficient. The comprehensive influence value of each region is calculated based on the area influence coefficient and the tax rate influence coefficient, and the influence value is normalized by the sigmoid function to obtain the regional contribution. The area proportion weight and the tax rate proportion weight are set according to the regional contribution, and the comprehensive proportion coefficient of each region is calculated by the additive combination method. The final tax calculation proportion coefficient is obtained by normalization. In the calculation of tax rates for mixed office and residential areas, the original attribute data showed obvious numerical differences. The office area has a construction area of ​​280 square meters, a usable area of ​​252 square meters, and an actual utilization rate of 0.9; the residential area has a construction area of ​​160 square meters, a usable area of ​​144 square meters, and an actual utilization rate of 0.85. The minimum and maximum standardization method is used for normalization, and the conversion formula is x'=(x-min) / (max-min). After processing, the area characteristic value of the office area is 0.7, and the area characteristic value of the residential area is 0.4. The standardized data is centered, the mean μ and standard deviation σ of the data sequence are calculated, and x'=(x-μ) / σ is used for conversion. Taking the area data as an example, the office area deviates from the mean by 2.1 standard deviations, and the residential area deviates from the mean by 1.2 standard deviations. The converted data is more suitable for logistic regression modeling. The logistic regression predictor uses the stochastic gradient descent method to optimize parameters, and the loss function is the cross entropy function, L=-Σ(ylog(p)+(1-y)log(1-p)), where y is the actual label and p is the predicted probability. After 1000 iterations of optimization, the area weight of the office area converges to 0.65, and the tax rate weight converges to 0.35; the area weight of the residential area converges to 0.55, and the tax rate weight converges to 0.45. The linear combination equation is in the form of ax+by+c, where x is the area feature and y is the tax rate feature. The parameter values ​​are calculated by the least squares method, and the equation for the office area is 0.65x+0.35y+0.2, and the equation for the residential area is 0.55x+0.45y+0.15. Substituting the standardized eigenvalues ​​into the equation, the comprehensive impact value of the office area is calculated to be 0.72, and that of the residential area is 0.58.The sigmoid function σ(x)=1 / (1+e^(-x)) is used to normalize the impact value and map it to the (0,1) interval. The normalized impact value of the office area is 0.67, and that of the residential area is 0.54. Based on these values, the weights are set, with the area share weight being 0.6 and the tax rate share weight being 0.4. The additive combination w1x1+w2x2 is used to calculate the comprehensive share coefficient. The final calculation results show that the office area's share coefficient in the tax calculation is 0.63, taking into account its larger area share and higher utilization rate; the residential area's share coefficient is 0.37, reflecting its smaller area size and relatively lower usage intensity. This weighted calculation method based on multidimensional features reasonably balances the contribution of different use areas in the tax allocation.

[0023] Step S106, comprehensively considers the area, tax rate and proportion coefficient of each area, adopts dynamic programming algorithm, establishes tax calculation model, obtains the tax payable of the building, and evaluates the impact of changes in regional area and tax rate on the tax amount through sensitivity analysis method, and forms tax optimization plan through intelligent tax assistant.

[0024] A tax state transfer equation f(i, j) including the area number i and the area value j is constructed, and the state transfer equation is calculated by a dynamic programming algorithm to obtain a first optimal solution; an objective function z is constructed based on the first optimal solution, and the objective function includes the product sum of the area adjustment amount xi and the tax rate adjustment amount yi, and the second optimal solution is obtained by using the simplex method; for the second optimal solution, the partial derivative of the area change with respect to the tax amount is calculated to obtain an area sensitivity matrix, and the partial derivative of the tax rate change with respect to the tax amount is calculated to obtain a tax rate sensitivity matrix; a group of constraint equations is constructed based on the area sensitivity matrix and the tax rate sensitivity matrix, and the optimal adjustment plan is solved by the Lagrange multiplier method, and the optimal adjustment plan is iteratively optimized by the gradient descent method to obtain a third optimal solution.

[0025] Specifically, the tax state transfer equation f(i,j)=max{f(i-1,j),f(i-1,j-vi)+wi} is constructed based on the area data, tax rate data and proportion coefficient of each region, where i represents the region number, j represents the area value, vi represents the region area, and wi represents the region tax amount. The first optimal solution is calculated using the dynamic programming algorithm. For the first optimal solution, the area change range and tax rate change range are set as constraints, and the objective function z=Σ(xi*yi) is constructed, where xi represents the region area adjustment amount, and yi represents the region tax rate adjustment amount. The second optimal solution is obtained by solving the simplex method. The sensitivity Si calculation formula based on the second optimal solution is: , calculate the partial derivative of the area change with respect to the tax amount to obtain the area sensitivity matrix, and calculate the partial derivative of the tax rate change with respect to the tax amount to obtain the tax rate sensitivity matrix. According to the area sensitivity matrix and the tax rate sensitivity matrix, a set of constraint equations is constructed, and the Lagrange multiplier method is used to solve the optimal adjustment scheme to obtain the area adjustment coefficient and tax rate adjustment coefficient of each region. Based on the adjustment coefficient, the area adjustment amount and tax rate adjustment amount of each region are calculated, and the adjustment step is set to 0.01. The gradient descent method is used to iteratively optimize the adjustment amount to obtain the third optimal solution. For the third optimal solution, a tax burden impact assessment function g(x)=Σ(αixi+βiyi) is constructed, where αi and βi represent the impact weights of the area and tax rate, respectively. The impact value of each adjustment amount on the overall tax burden is calculated by function derivation. When solving the tax amount optimization problem in dynamic programming, taking the office area and residential area of ​​a building as an example, the office area is 280 square meters, the tax rate is 12%, and the proportion coefficient is 0.65; the residential area is 160 square meters, the tax rate is 4%, and the proportion coefficient is 0.35. Construct the state transfer equation f(i,j), where i represents the current area number, j represents the current area value, vi represents the area of ​​area i, and wi represents the tax contribution of area i. The state matrix is ​​obtained through state recursive calculation. In the matrix, f(1,280)=21.84 represents the optimal tax amount when only the office area is considered, and f(2,440)=27.44 represents the optimal tax amount when all areas are considered. Set the area change range to ±20 square meters and the tax rate change range to ±2 percentage points, and construct a linear programming model z=Σ(xi*yi), where z represents the total tax amount, xi represents the area adjustment amount, and yi represents the tax rate adjustment amount. Take the partial derivative of the objective function to calculate the sensitivity. Taking the office area as an example, the area sensitivity It means that every increase of 1 square meter in area will lead to an increase of 7800 yuan in tax amount, and the tax rate sensitivity It means that every 1 percentage point increase in the tax rate will lead to an increase of 23,400 yuan in tax. The area sensitivity of the residential area is 0.014, and the tax rate sensitivity is 0.64, which shows that the impact of the tax rate adjustment on the tax amount is more significant. Based on the sensitivity analysis results, the Lagrangian function L(x,y,λ)=z(x,y)+λ(g(x,y)-c) is constructed, where g(x,y) represents the constraint condition, c represents the constraint boundary, and z(x,y) is the original objective function. The optimal adjustment scheme is obtained by solving the Lagrangian equation. The office area is reduced by 15 square meters, and the tax rate is reduced by 1.2 percentage points; the residential area is increased by 12 square meters, and the tax rate is increased by 0.8 percentage points. The gradient descent method is used to optimize the adjustment scheme, and the learning rate is set to 0.01. When the gradient value is less than 0.001, the iteration is stopped. After 78 rounds of iterations, the final solution is obtained. The office area is adjusted to 267 square meters, and the tax rate is adjusted to 10.9%; the residential area is adjusted to 170 square meters, and the tax rate is adjusted to 4.7%. An impact assessment function g(x)=Σ(αixi+βiyi) was constructed, where αi and βi were 0.6 and 0.4 respectively. The impact value of area adjustment on tax burden was calculated to be -0.86, the impact value of tax rate adjustment was -1.24, and the comprehensive impact value was -2.1, indicating that the optimization scheme has achieved the optimization of the tax burden structure while keeping the total tax amount basically stable.

[0026] Step S107, generating an intelligent tax recommendation report in the intelligent tax assistant according to the comprehensive tax rate, tax payable and tax optimization plan of the building.

[0027] An association rule mining algorithm is used to calculate the comprehensive tax rate index of buildings, and a first association rule set is obtained according to the support and confidence between the indicators; hierarchical clustering is performed on the first association rule set, and the rules are grouped by the cluster quantity parameter to obtain a first report framework, which includes background information, data analysis, optimization suggestions and conclusion summary; keyword groups are extracted according to the first report framework, and the semantic similarity is calculated by the word vector embedding method, and the cosine distance method is used for semantic clustering to obtain a second report framework; for the second report framework, cases are retrieved from the historical report library, and cases within the similarity threshold are selected by the text similarity calculation method to obtain a standardized report template, which includes a professional terminology dictionary and a numerical indicator rule library.

[0028] Specifically, data indicators are collected for the comprehensive tax rate, tax payable and tax optimization plan of the building, and the support and confidence between indicators are calculated by association rule mining algorithm. The minimum support threshold is set to 0.3 and the minimum confidence threshold is set to 0.7 to obtain the first association rule set. The report title system is constructed based on the first association rule set, and the hierarchical clustering method is used to group the rules. The number of clusters is set to 4, corresponding to the background information, data analysis, optimization suggestions and conclusion summary in the report, respectively, to obtain the first report framework. According to the first report framework, key words and syntactic structures are extracted, and the semantic similarity is calculated by word vector embedding method. The semantics is clustered by cosine distance method to obtain the second report framework. Based on the second report framework, similar cases are retrieved from the historical report library, and the text similarity calculation method is used to select cases with similarity greater than 0.8 as reference templates to obtain the first report template. A professional terminology dictionary and numerical indicator rule library are constructed for the first report template, and the string matching method is used to identify professional terms in the text. The regular expression method is used to extract numerical indicators to obtain the second report template. Based on the second report template, the standardized text content is generated, and the text verification rules are used to check the numerical accuracy, unit labeling, and format specifications. The structured report content is formed through the natural language generation method. In the process of generating the intelligent tax recommendation report, the association rule mining algorithm is used to analyze the data indicators, and multiple groups of association rules are found. When the office area is larger than 200 square meters, the support for the comprehensive tax rate exceeding 8% is 0.35, and the confidence is 0.82; when the residential area is larger than 150 square meters, the support for the tax rate optimization space greater than 1 percentage point is 0.42, and the confidence is 0.75. The report framework is constructed by the hierarchical clustering method, and the content of similar topics is clustered under the same title. The data analysis part includes three sub-topics: area distribution, tax rate structure, and space utilization. The optimization suggestion part includes three sub-topics: area adjustment suggestions, tax rate optimization measures, and comprehensive benefit evaluation. Background information and conclusion summary are respectively used as independent topics. The semantics of the report are vectorized, and the word vector embedding method is used to calculate the semantic similarity. Taking "tax rate optimization" as an example, the similarity with "tax burden adjustment" is 0.92, the similarity with "tax planning" is 0.88, and the similarity with "tax amount calculation" is 0.75. Based on the semantic similarity clustering results, similar concepts are expressed in a unified manner to form a standardized professional terminology system. When retrieving similar cases from the historical report library, the cosine distance method is used to calculate the text similarity. For a new report text, its similarity distribution with historical cases is 0.85 for case 1, 0.82 for case 2, and 0.76 for case 3. Case 1 and case 2 with similarities greater than 0.8 are selected as reference templates to extract the expression methods and format specifications.Define standard terminology expressions in the professional term dictionary, such as "building area" is uniformly expressed as "taxable area", "tax rate" is uniformly expressed as "applicable tax rate", and "optimization plan" is uniformly expressed as "tax burden optimization suggestion". The numerical indicator rules stipulate that the area is retained as an integer, the tax rate is retained to 1 decimal place, the amount is retained to 2 decimal places, and the proportion is retained to 1 place before the percentage sign. The text standardization check includes multiple dimensions. The numerical accuracy check found that the area data "280.45 square meters" should be corrected to "280 square meters"; the unit marking check found that the percentage sign was missing after "12.5%"; the format standardization check found that the title level was inconsistent, and the third-level title "3.2.1 Tax rate structure analysis" should be adjusted to "3.2 Tax rate structure analysis". Through the natural language generation method, the standardized content is organized into a structured report to form a tax recommendation report with rigorous logic and standardized expression.

[0029] It will be apparent to those skilled in the art that the present application is not limited to the details of the exemplary embodiments described above, and that the present application can be implemented in other specific forms without departing from the spirit or essential features of the present application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present application. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An intelligent tax assistant interaction method, characterized in that: The method comprises: Through laser scanning and panoramic cameras, the spatial layout diagram and area data of each area after the renovation of the internal partitions of the building are obtained, and image segmentation and target detection are performed on the spatial layout diagram to extract the geometric features and spatial relationships of each area. Combined with the building design specifications and tax classification standards, the use nature of each area is determined, including office, residential and mixed use; If the intelligent tax assistant determines that the area is used for office or residential purposes, it uses a supervised machine learning algorithm to train the existing tax classification sample data and establish a tax classification model to determine the tax rate category corresponding to the office area and the residential area; If the intelligent tax assistant determines that the use nature of the area is mixed, the mixed nature area will be subdivided into several sub-areas through the cluster analysis algorithm, and the sub-areas will be determined to be office or residential based on the area proportion and use characteristics of the sub-areas; Through case-based reasoning algorithms, similar cases to the sub-region are retrieved from the existing tax case database, and relevant attributes, including area, usage function, and occupancy, are compared and analyzed to predict the tax rate tendency of the region and form preliminary tax recommendations; The logistic regression algorithm is used to analyze the influence of the area and tax rate of each region on the comprehensive tax rate, and the weighted average method is used to calculate the comprehensive tax rate of office and residential areas, and the proportion coefficient of each region in the tax calculation is determined; Based on the area, tax rate and proportion coefficient of each area, a dynamic programming algorithm is used to establish a tax calculation model to obtain the tax payable for the building. At the same time, the sensitivity analysis method is used to evaluate the impact of changes in regional area and tax rate on the tax amount, and a tax optimization plan is formed through the intelligent tax assistant; Generate an intelligent tax recommendation report in the Intelligent Tax Assistant based on the building’s comprehensive tax rate, tax payable, and tax optimization plan.

2. The method according to claim 1, characterized in that The laser scanning and panoramic camera are used to obtain the spatial layout diagram and area data of each area after the internal partition transformation of the building, perform image segmentation and target detection on the spatial layout diagram, extract the geometric features and spatial relationships of each area, and combine the building design specifications and tax classification standards to determine the use nature of each area, which includes office, residential and mixed use, including: A three-dimensional laser scanner is used to obtain first point cloud data inside the building, and a first texture point cloud model is obtained by registering the first point cloud data with the panoramic image data; Performing spatial clustering processing on the first texture point cloud model to obtain a second texture point cloud model, extracting wall surfaces using a region growing algorithm based on the second texture point cloud model and calculating the relative position relationship between walls to obtain a first spatial connectivity graph; Performing vectorization processing on the first spatial connectivity graph to obtain a first plane layout graph, and performing image segmentation processing on the first plane layout graph to obtain a second plane layout graph; Shape feature vectors and spatial topology feature vectors are extracted from the second plan layout diagram, and the shape feature vectors and spatial topology feature vectors are trained using a support vector machine algorithm to obtain a room use classification model, and the rooms are labeled for office and residential use according to the classification model.

3. The method according to claim 1, characterized in that If the intelligent tax assistant determines that the use nature of the area is office or residential, a supervised machine learning algorithm is used to train the existing tax classification sample data to establish a tax classification model for determining the tax rate category corresponding to the office area and the residential area, including: The standard deviation method is used to identify outliers in the tax classification sample data, and the mean filling method is used to supplement the missing values ​​according to the outlier identification results to obtain the first sample data; Perform minimum and maximum value normalization processing on the first sample data, and perform one-hot encoding conversion on the normalization processing result to obtain third sample data; A variance selection method is used to remove low-discrimination features from the third sample data, and a principal component analysis method is used to extract the principal component whose cumulative contribution rate reaches a specified threshold value from the data after the low-discrimination features are removed to obtain a sixth sample data; A random forest classifier is established based on the sixth sample data, and the probability value of the tax rate category is calculated by the random forest classifier. If the probability value is greater than a specified threshold, the corresponding tax rate category is adopted; if the probability value is less than the specified threshold, the tax rate category is determined by a probability weighted method.

4. The method according to claim 1, characterized in that: If the intelligent tax assistant determines that the use nature of the area is mixed, the mixed nature area is subdivided into several sub-areas through a cluster analysis algorithm, and the sub-areas are determined to be office or residential based on the area proportion and use characteristics of the sub-areas, including: The first eigenvector is obtained by collecting the basic feature data of the door and window positions, partition walls and channel connections in the mixed nature area and using the maximum and minimum value normalization method; Calculating the Euclidean distance according to the first eigenvector to obtain a first similarity matrix, and using a density clustering algorithm to obtain a first sub-region division result; Extracting the room connectivity relationship based on the first sub-region division result, obtaining a second similarity matrix by calculating the channel width and connectivity between adjacent rooms, and obtaining the second sub-region division result using a graph cut algorithm; The building area data is extracted according to the second sub-area division result, the crowd density data is collected through infrared sensors, and the hierarchical clustering algorithm is used to obtain office clusters and residential clusters. The use nature of the sub-area is determined according to the area weight and cluster feature weight.

5. The method according to claim 1, characterized in that: The case-based reasoning algorithm retrieves cases similar to the sub-region from the existing tax case database, compares and analyzes relevant attributes, including area, usage function, and occupancy, predicts the tax rate tendency of the area, and forms preliminary tax recommendations, including: The first attribute data of building area, use function and traffic convenience are extracted from the tax case database, and the feature vector is obtained through the minimum and maximum value standardization method and the one-hot encoding method; Retrieving cases with similarity greater than a threshold from a tax case database according to the feature vector using a cosine similarity method to obtain a first case set; Extracting the second attribute data of regional area ratio, space utilization rate and personnel density from the first case set, and selecting the principal component with the cumulative contribution rate meeting the requirements by the principal component analysis method to obtain the second case set; A decision tree model was constructed based on the second case set, the information gain ratio was used to select the optimal split feature extraction decision path, and the tax rate prediction model was obtained through multi-layer perceptron training.

6. The method according to claim 1, characterized in that The logistic regression algorithm is used to analyze the weight of the impact of each area and tax rate on the comprehensive tax rate, and the weighted average method is used to calculate the comprehensive tax rate of the office area and the residential area, and the proportion coefficient of each area in the tax calculation is determined, including: Collecting original attribute data of the building area, and processing the original attribute data by a minimum-maximum value normalization method to obtain first standardized data; Calculating the mean variance of the area data sequence and the tax rate data sequence according to the first standardized data, and processing the data sequence using a centralization method to obtain second standardized data; Constructing a logistic regression predictor based on the second standardized data, and optimizing the predictor parameters by using a stochastic gradient descent method to obtain an area weight value and a tax rate weight value; A linear combination equation is constructed based on the area weight value and the tax rate weight value, the least squares method is used to calculate the parameters of the equation to obtain the influence coefficient, the influence coefficient is normalized by the sigmoid function to obtain the regional contribution, and the additive combination method is used to calculate the tax amount calculation ratio coefficient.

7. The method according to claim 1, characterized in that The tax calculation model is established by integrating the area, tax rate and proportion coefficient of each area, using a dynamic programming algorithm to obtain the tax payable for the building. At the same time, the sensitivity analysis method is used to evaluate the impact of changes in regional area and tax rate on the tax amount, and a tax optimization plan is formed through the intelligent tax assistant, including: Constructing a tax state transfer equation f(i, j) including the area number i and the area value j, wherein the state transfer equation is calculated using a dynamic programming algorithm to obtain a first optimal solution; Constructing an objective function z according to the first optimal solution, the objective function includes the product sum of the regional area adjustment amount xi and the tax rate adjustment amount yi, and using the simplex method to solve to obtain the second optimal solution; For the second optimal solution, the partial derivative of the area change with respect to the tax amount is calculated to obtain an area sensitivity matrix, and the partial derivative of the tax rate change with respect to the tax amount is calculated to obtain a tax rate sensitivity matrix; A set of constraint equations is constructed according to the area sensitivity matrix and the tax rate sensitivity matrix, an optimal adjustment scheme is solved by using the Lagrange multiplier method, and the optimal adjustment scheme is iteratively optimized by using the gradient descent method to obtain a third optimal solution.

8. The method according to claim 1, characterized in that According to the building's comprehensive tax rate, tax payable and tax optimization plan, an intelligent tax recommendation report is generated in the intelligent tax assistant, including: The association rule mining algorithm is used to calculate the comprehensive tax rate index of buildings, and the first association rule set is obtained according to the support and confidence between the indicators; Performing hierarchical clustering processing on the first association rule set, grouping the rules by cluster quantity parameters to obtain a first report framework, wherein the first report framework includes background information, data analysis and optimization suggestions, and conclusion summary; Extract key words according to the first report framework, calculate semantic similarity by word vector embedding method, and use cosine distance method to cluster semantics to obtain the second report framework; For the second report framework, cases are retrieved from the historical report library, and cases within a similarity threshold are selected through a text similarity calculation method to obtain a standardized report template, wherein the report template includes a professional terminology dictionary and a numerical indicator rule library.

Images