Multi-modal discriminative quality constrained urban community renewal demand dynamics classification method

Abstract

The application discloses a kind of multi-modal discriminant quality constraints urban community renewal demand power classification method.The method includes: obtaining the community level data of three modalities of building or image, resident opinion text, interest point, constructs each modality feature vector and standardization;After initial splicing of each modality subspace, first unsupervised clustering is executed to generate initial typing;Based on initial typing, estimate each modality discriminant contribution score;Calculate each modality quality score;Combine quality score and discriminant contribution score to construct adaptive fusion weight;Determine the minimum participation threshold of target modality and correct weight;Weighted fusion constructs fusion latent space, and executes second unsupervised clustering to output community renewal demand-power structure type.The application automatically identifies community renewal type through initial typing feedback, quality-discriminant joint weight and adaptive threshold, and improves classification stability and interpretability.

Description

Technical Field

[0001] This invention belongs to the fields of intelligent analysis of urban renewal, classification of urban communities, multimodal data fusion and unsupervised machine learning, and specifically relates to a method for classifying urban community renewal demand-dynamics based on joint modeling of modality discrimination contribution and quality constraints.

[0002] More specifically, this invention targets urban community renewal scenarios, using communities or neighborhoods as unified analysis units. It integrates multi-source heterogeneous data such as building or image modalities, resident opinion text modalities, and point of interest (POI) modalities. Through community-level multimodal observation signal construction, initial classification feedback, modal discrimination contribution estimation, modal quality constraints, adaptive fusion weight generation, adaptive determination of the minimum participation threshold of the target modality, and fusion latent space clustering, it achieves automatic identification of community renewal demand-dynamic structure types. Background Technology

[0003] Urban communities are crucial spatial units for residents' daily lives, public service provision, and grassroots governance. As urban development shifts from incremental expansion to improving existing infrastructure, the renewal of old communities, the construction of complete communities, and the optimization of community living circles have become important aspects of urban renewal. Unlike traditional renewal methods that primarily focus on physical space transformation, current community renewal places greater emphasis on the comprehensive coordination among residents' actual needs, public service provision, the quality of the built environment, and the conditions for implementation. Therefore, accurately identifying the structural differences in renewal needs and motivations among large-scale communities has become a critical technical challenge for refined urban renewal governance and decision support.

[0004] Existing research on community renewal largely focuses on aspects such as resident satisfaction, community governance, public services, built environment, and renewal performance. In recent years, researchers both domestically and internationally have begun to explore urban renewal and urban spatial identification issues using methods such as questionnaires, text mining, point-of-interest data, street view images, and multi-view learning, achieving promising results.

[0005] The above research provides an important foundation for the identification of community renewal needs, multi-source urban data analysis, and unsupervised classification. However, the following technical shortcomings still exist when dealing with the classification of community renewal needs-dynamic structures.

[0006] (1) Traditional methods for classifying and identifying the potential for community renewal often rely on manually constructed evaluation index systems, and use expert weighting, analytic hierarchy process (AHP), entropy weighting, or comprehensive scoring methods to rank or classify communities. While these methods have a certain degree of policy interpretability, the selection of indicators and the setting of weights are easily influenced by the researcher's experience, making it difficult to fully explore the implicit structural relationships within multi-source data such as resident opinion texts, points of interest (POIs), buildings, or images. At the same time, the comprehensive evaluation results are usually expressed as a single score or grade, making it difficult to reveal the combined differences between different communities in terms of resident needs, external service supply, and built environment conditions.

[0007] (2) Urban big data such as resident opinion texts, POI data and street view images can reflect residents' demands, service supply and physical environment status respectively, but the three types of data differ significantly in semantic level, spatial coverage, sparsity and data credibility. Existing studies often focus on a single data source or simple multi-source splicing, and have not yet fully addressed the differentiated role of different modalities in community renewal classification tasks.

[0008] (3) Existing multi-view clustering and multi-modal fusion methods can integrate multi-source feature representations of the same object under unsupervised conditions. However, when applied to community update classification, they are still prone to problems such as modality quality differences, modality redundancy, missing modalities, weak modality interference, and unbalanced modality contributions. If a direct splicing method is used, different modalities may cause imbalance in the fusion space due to differences in variance structure, sparsity, or noise level. If only modality quality is used for weighting, it can only answer whether the modality is reliable, but cannot answer whether the modality truly helps to distinguish community types.

[0009] (4) High data quality for a modality does not necessarily mean that the modality has a strong discriminative ability for community classification; on the contrary, although a modality may have a weak discriminative contribution under the current data conditions, it may be of great significance for the interpretation of community updates. For example, architectural or image modalities can characterize the physical built environment, but their discriminative contribution may be underestimated when the feature dimensions are few or the characterization is coarse. If fusion is based entirely on discriminative contribution, it may lead to a few strong discriminative modalities dominating the clustering results, weakening the integrity of multimodal analysis.

[0010] (5) Existing methods are insufficient in handling the minimum participation threshold of modalities and the addition of weak modalities. Setting a fixed minimum participation ratio of the target modality is based on human experience and is difficult to adapt to different datasets; when low-quality pseudo-modalities or random noise modalities are added to the model, there is a lack of a mechanism to automatically identify and suppress the influence of weak modalities, which may interfere with the final clustering structure. Summary of the Invention

[0011] The purpose of this invention is to provide a multimodal discriminative quality constraint-based classification method for urban community renewal demand dynamics, in order to solve the problems in the prior art such as strong subjectivity of manual weighting, easy modal imbalance caused by direct splicing of multimodal data, inability to reflect classification contribution by quality weighting alone, potential weakening of interpretive modalities by pure discriminative contribution fusion, insufficient automation of fixed thresholds, and insufficient robustness after weak modal access.

[0012] The technological breakthrough of this invention lies in establishing a multimodal adaptive fusion mechanism consisting of "initial classification feedback—discriminative contribution estimation—quality constraint fusion—adaptive modality participation preservation—weak modality suppression." This mechanism does not simply input text, POI, and building features into a conventional clustering model. Instead, it first uses the feedback from the initial community classification structure to estimate the actual contribution of each modality to the classification. Then, it constructs weights by jointly using these contributions and modality quality scores, and maintains the reasonable participation of the target modality through an adaptive tau threshold, thereby achieving the identification of community renewal needs and dynamic structure types.

[0013] According to one aspect of the present invention, a method for classifying urban community renewal demand dynamics under multimodal discriminative quality constraints is provided, comprising: acquiring community-level data of at least two modalities of a target urban community, wherein the modalities include architectural or image modalities, resident opinion text modalities, and point-of-interest modalities; performing community-level aggregation and feature representation on the data of each modality respectively, constructing community-level feature vectors corresponding to each modality, and preprocessing the community-level feature vectors to obtain standardized subspace representations of each modality; initially concatenating the subspace representations of each modality, performing dimensionality reduction processing, and then performing a first unsupervised clustering. Generate an initial community classification structure; based on the initial community classification structure, calculate the discriminant contribution score of each modality to the community classification; calculate the modality quality score of each modality, which is used to characterize the data availability, integrity, reliability, and stability of the corresponding modality; construct adaptive fusion weights for each modality based on the modality quality score and discriminant contribution score, and perform weighted fusion of the subspace representations of each modality according to the adaptive fusion weights to construct a fusion latent space; perform a second unsupervised clustering in the fusion latent space to output the community update requirement-dynamic structure type.

[0014] As a further technical solution, the discriminant contribution score is calculated as follows: for any modality, based on the inter-class differences between different initial categories under the initial community classification structure and the intra-class fluctuations within the same initial category, the ratio of the inter-class differences to the intra-class fluctuations is calculated and used as the discriminant contribution score of the modality.

[0015] As a further technical solution, the adaptive fusion weights are constructed as follows: the modality quality score of each modality is multiplied by the discrimination contribution score to obtain the quality-discrimination joint score of each modality; the quality-discrimination joint score of each modality is divided by the sum of the quality-discrimination joint scores of all modalities to obtain the adaptive fusion weight of each modality.

[0016] As a further technical solution, an adaptive determination step for the minimum participation threshold of the target modality is also included: for a preset target modality, a threshold search interval is determined based on the number of modalities participating in the fusion; within the search interval, weight correction, fusion latent space construction, and unsupervised clustering are performed on candidate thresholds, and a comprehensive evaluation function is constructed based on clustering effectiveness index, clustering stability index, minimum participation index of the target modality, and class balance index; the candidate threshold with the optimal comprehensive evaluation function is selected as the minimum participation threshold of the target modality; and the adaptive fusion weights are corrected according to the minimum participation threshold of the target modality.

[0017] As a further technical solution, the target modality is a building or image modality.

[0018] As a further technical solution, the community-level feature vector of the resident opinion text modality includes the number of basic requests, the number of improvement requests, the number of upgrade requests, the total number of requests, text availability variables, and text confidence; the community-level feature vector of the building or image modality includes one or more of the following: building quality score, building floor structure, street view image recognition score, and community image recognition score; the community-level feature vector of the point of interest modality includes the total number of points of interest, the number of major categories of points of interest, the number of subcategories of points of interest, the density of public service facilities, and the density of commercial service facilities.

[0019] As a further technical solution, it also includes modality access and weak modality suppression steps: when adding a new candidate modality, it is determined that the new candidate modality meets the preset access conditions. The preset access conditions include being able to map to a unified community analysis unit, being able to form a community-level feature representation, being able to calculate a modality quality score, and being able to estimate a discrimination contribution score based on the initial community classification structure. After access, if the modality quality score and / or discrimination contribution score of the new candidate modality is lower than a preset threshold, the adaptive fusion weight of the new candidate modality is reduced.

[0020] As a further technical solution, the upper bound of the threshold search interval... Determined based on the number M of modes participating in the fusion. ,in This is the upper bound of the search ratio.

[0021] According to one aspect of the present invention, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to implement the multimodal discriminant quality constraint-based urban community renewal demand dynamics classification method when executing the instructions.

[0022] According to one aspect of the present invention, a multimodal discriminant quality constraint-based urban community renewal demand dynamics classification system is provided to implement the multimodal discriminant quality constraint-based urban community renewal demand dynamics classification method, comprising: a data acquisition module for acquiring community-level data of at least two modalities of a target urban community, wherein the modalities include architectural or image modalities, resident opinion text modalities, and point-of-interest modalities; an observation signal construction module for performing community-level aggregation and feature representation on the data of each modality respectively, constructing community-level feature vectors corresponding to each modality; a feature processing module for performing missing value processing, variable transformation, and standardization processing on the community-level feature vectors to obtain standardized subspace representations of each modality; and an initial classification module for initially concatenating the subspace representations of each modality. After dimensionality reduction, a first unsupervised clustering is performed to generate an initial community classification structure. A discriminant contribution estimation module calculates the discriminant contribution score of each modality to the community classification based on the initial community classification structure. A quality assessment module calculates the modality quality score for each modality, which characterizes the data availability, integrity, reliability, and stability of the corresponding modality. An adaptive fusion module constructs adaptive fusion weights for each modality based on its modality quality score and discriminant contribution score. A fusion latent space module performs weighted fusion of the subspace representations of each modality based on the adaptive fusion weights to construct a fusion latent space. A final clustering module performs a second unsupervised clustering in the fusion latent space and outputs the community update requirement-dynamic structure type.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: First, it reduces the subjectivity of manual weighting. This invention does not pre-set fixed modal weights or manually pre-determine community types. Instead, it automatically generates fusion weights through initial classification feedback, modal quality assessment, and modal discrimination contribution estimation.

[0024] Second, it improves the adaptability of multimodal fusion to classification objectives. This invention not only considers modality quality, but also estimates the modality discrimination contribution based on the initial community classification structure, so that the modality weights simultaneously reflect "whether the modality is reliable" and "whether the modality helps to distinguish categories".

[0025] Third, it improves cluster separation and type structure clarity. By estimating modal discriminative contribution feedback, this invention can enhance the influence of modalities that have a real distinguishing effect on community typing and reduce the interference of weak-quality or weak-discriminative modalities on the fusion space.

[0026] Fourth, it maintains the integrity of the multimodal method. This invention uses an adaptive determination mechanism for the minimum participation threshold of the target modality to ensure that modalities such as buildings or images, which have important interpretative significance but whose current discriminative contribution may be low, maintain reasonable participation, thus avoiding the classification results from degenerating into a near-monomodal or near-bimodal method dominated by a few strong discriminative modalities.

[0027] Fifth, improve the automation level of the parameter determination process. This invention transforms the fixed minimum participation ratio into an adaptive TAU optimization problem, where the threshold is jointly determined by clustering effectiveness, stability, target modality participation, and class balance, thereby avoiding setting a single empirical value.

[0028] Sixth, improve the robustness of the method to modal changes. This invention sets modal access conditions and a weak modality suppression mechanism, ensuring that newly added modes only receive higher fusion weights when they possess high quality and strong discriminative contributions; for low-quality, low-discrimination, or random noise modes, the system can... , and The combined constraints reduce its impact.

[0029] Seventh, enhance the interpretability of community renewal classification results. This invention outputs not a single comprehensive score, but rather a community renewal demand-motivation structure type, which can be used for community renewal type identification, renewal priority assessment, renewal strategy zoning, and urban governance decision support. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a general technical roadmap for the urban community renewal demand dynamics classification method with multimodal discriminative quality constraints, as described in this invention. Figure 2 This is a flowchart illustrating the construction process of community-level multimodal observation signals according to an embodiment of the present invention. Figure 3 This is a diagram illustrating the modal discrimination contribution estimation mechanism based on initial classification feedback in an embodiment of the present invention. Figure 4This is a diagram illustrating the adaptive determination mechanism of the quality-discrimination joint weight and the minimum participation threshold of the target mode in an embodiment of the present invention. Figure 5 This diagram illustrates the modal access conditions and weak mode suppression mechanism in an embodiment of the present invention. Figure 6 This is a structural diagram of the urban community renewal demand-motivation classification system module according to an embodiment of the present invention; Figure 7 This is a graph showing the changes in the adaptive TAU optimization comprehensive score and clustering index in an embodiment of the present invention. Figure 8 This is a graph showing the average adaptive weight variation of each mode under different tau values ​​according to an embodiment of the present invention. Figure 9 This is the average fusion weight map after adding pseudo-noise modes in an embodiment of the present invention; Figure 10 This is a heatmap illustrating the final community update requirements and power types in this embodiment of the invention. Figure 11 This is an explanatory diagram of the final eight categories of community update needs and motivation types in this embodiment of the invention. Detailed Implementation

[0032] This invention primarily addresses the following technical problems: First, how to fuse heterogeneous modalities such as buildings or images, resident opinion texts, and POIs without relying on manually fixed weights; second, how to distinguish between modality quality and modality discriminative ability, so that the fusion weights reflect both data reliability and classification contribution; third, how to avoid excessively suppressing building or image modalities that have important interpretative significance but currently have weak discriminative contributions; fourth, how to transform the fixed modality participation ratio into data-driven adaptive parameters; and fifth, how to reduce the interference of invalid modalities on clustering results when modality combinations change or weak information modalities are added.

[0033] Based on the aforementioned problems, a new urban community renewal demand-motivation classification method is needed. This method should be able to automatically determine the actual contribution of different modalities to community classification without relying on manual fixed weighting or pre-setting community types. It should also consider modality quality and discriminative ability, ensure the reasonable participation of target modalities with important explanatory significance, and maintain the stability and robustness of classification results when modality combinations change or weak modalities are added.

[0034] It should be noted that: The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0037] This embodiment uses urban community renewal classification as an application scenario, taking communities or neighborhoods as unified analysis units. It integrates building or image modalities, resident opinion text modalities, and Points of Interest (POI) modalities to construct community-level multimodal observation signals (i.e., constructing community-level feature vectors corresponding to each modality). Through initial classification feedback, modality discrimination contribution estimation, quality-discrimination joint weight construction, adaptive determination of the minimum participation threshold for the target modality, and fusion latent space clustering, it achieves the identification of community renewal demand-dynamic structure types. It should be understood that this embodiment is only used to illustrate specific implementations of the present invention and is not intended to limit the scope of protection of the present invention.

[0038] Step 1: Acquisition of community-level multimodal data.

[0039] Using communities or neighborhoods as the unified analysis unit, basic community or neighborhood data, building or image data, resident opinion text data, and Point of Interest (POI) data are acquired within the study area. Basic community or neighborhood data is used to determine the unified analysis unit; building or image data is used to characterize the physical built environment of the community; resident opinion text data is used to extract residents' explicit renovation needs; and POI data is used to characterize the service supply and functional structure around the community.

[0040] In this embodiment, the building or image modality can be represented by building quality rating, building floor structure, street view image recognition results, or community image recognition results; the resident opinion text modality can be represented by the number of basic requests, the number of improvement requests, the number of upgrade requests, the total number of requests, text confidence, and text availability variables; the POI modality can be represented by indicators such as the total number of POIs, the number of POI major categories, the number of POI subcategories, and the density of public service facilities.

[0041] See Figure 1 The overall technical approach of this invention is not simply to directly splice multimodal data and then cluster it. Instead, it addresses the problems of modality imbalance, quality differences, discriminant contribution differences, insufficient participation of target modalities, and weak modal interference in community-updated multimodal data by constructing an adaptive fusion clustering mechanism.

[0042] Step 2: Construction and preprocessing of community-level multimodal observation signals.

[0043] This step involves community-level aggregation and feature representation of each modal data, forming modal feature vectors for each community (i.e., community-level multimodal observation signals). The demand side is primarily expressed through resident opinion text modalities, including basic demands, improvement demands, upgrade demands, overall demand strength, text availability, and text confidence. The driving force or condition side is primarily expressed through POI modalities and architectural or image modalities, where POI modalities represent service supply and functional structure, and architectural or image modalities represent the physical built environment status.

[0044] See Figure 2 For resident opinion text data, batch inference or natural language processing models can be used for structured extraction to identify whether the text contains specific community information, community name, administrative district or street information, demand level, demand theme, and confidence level. Valid text records are aggregated by community to form community-level text indicators such as the number of basic demands, the number of improvement demands, the number of upgrade demands, the total number of demands, and the average confidence level.

[0045] For POI data, statistical aggregation is performed according to a certain service area around the community or neighborhood to form features such as the total number of POIs, the number of major POI categories, and the number of subcategories of POIs. For building or image modalities, building ratings, building forms, floor structures, or image recognition features that can characterize the physical built environment of the community are extracted.

[0046] Subsequently, each modal feature is mapped to the same community-level feature matrix using a unified community identifier. The community identifier field, used for data association, indexing, and result write-back, is not included as a modeling feature in clustering. Count features are transformed using a log1p transform, continuous features are standardized, and a text availability variable is constructed based on whether valid records exist in the text.

[0047] Furthermore, the following multimodal feature cleaning and standardization processes are performed: missing value processing, outlier processing, compression transformation of count variables, and scale unification processing are applied to each modal feature. For the resident opinion text modality, a text availability variable is constructed; for count features, log1p transformation is used to reduce the influence of long-tail distribution; for continuous features, standardization methods are used to map different modal features to a comparable scale.

[0048] Step 3: Initial community subtype structure generation.

[0049] After constructing the community-level multimodal observation signals, the features of each modality are first initially stitched together, then standardized and dimensionality reduced before unsupervised clustering is performed to generate an initial community classification structure. This initial community classification structure is not the final classification result, but serves as a feedback basis for subsequent estimation of the discrimination contribution of each modality.

[0050] In one implementation, Principal Component Analysis (PCA) is used to reduce the dimensionality of the initial spliced ​​features, and the KMeans algorithm is used to generate initial cluster labels. This initial genotyping structure provides a computable category reference, enabling subsequent determination of whether different modalities truly have a classifying function under this category structure.

[0051] Step 4: Estimation of modal discrimination contribution based on initial fractal feedback.

[0052] See Figure 3 Based on the initial community classification structure, the inter-class differences of each modality between different initial cluster categories and the intra-class fluctuations within the same initial category are calculated to obtain the discrimination contribution score of each modality.

[0053] For the m-th mode, its discrimination contribution score can be expressed as: , in, This represents the discrimination contribution score of the m-th mode. This represents the inter-class difference measure between different initial cluster categories for the m-th modality under the initial community classification structure. This represents the intra-cluster variation measure of the m-th mode within the same initial cluster category, with epsilon being a tiny positive number to prevent the denominator from being zero.

[0054] When a certain modality varies significantly across different community types and is more concentrated within the same type, the corresponding modality... A high value indicates that the modality has a strong discriminative contribution to community typing. Therefore, this invention transforms the modality weighting from traditional static splicing or manual weighting to a discriminative contribution estimation based on feedback from the initial classification structure.

[0055] Step 5: Calculate the modal mass fraction.

[0056] Calculate the modal mass fraction for each mode. The aforementioned Used to characterize the usability, completeness, credibility, and stability of the corresponding modality within the target community. Text modality... It can be determined by whether the text is available (i.e., text availability), text confidence, and the number of valid texts; architectural or image modalities It can be determined by the completeness of building data, the confidence level of image recognition, or the validity of building ratings; POI modalities It can be determined by the coverage of POI data, category completeness, and spatial matching validity.

[0057] Step 6: Construct the joint weight of quality-discrimination contribution.

[0058] Based on modal mass score Modal discrimination contribution score Construct adaptive fusion weights for each modality: , in, This represents the adaptive fusion weights for the m-th modality. This represents the mass fraction of the m-th mode. This represents the discrimination contribution score of the m-th mode. This represents the sum of the products of the quality scores and the discrimination contribution scores of all participating modes. In this way, the mode weights are no longer determined solely by mode quality, but rather by a combination of mode reliability and mode classification contribution.

[0059] See Figure 4This mechanism enables modal weights to simultaneously reflect "whether the modality is reliable" and "whether the modality helps to distinguish categories," thereby avoiding the problem that classification contribution cannot be reflected when only modal quality is used for weighting, and also avoiding the problem of uncontrollable contributions of different modalities when directly splicing them together.

[0060] Step 7: Adaptive determination of the minimum participation threshold for the target modality. For target modalities that have significant interpretative value but whose discrimination contribution may be low under the current feature conditions, a minimum participation threshold tau is set. In this embodiment, the target modality is an architectural or image modality, and its minimum participation constraint is expressed as: , in, tau represents the fusion weight of the target mode, and tau represents the minimum participation threshold of the target mode.

[0061] Unlike fixed empirical thresholds, tau in this invention is not pre-fixed as a single constant, but is determined through adaptive refinement search. Specifically, the search interval [0, tau_max] is determined based on the number of modalities participating in the fusion, where: , M represents the number of modes participating in the fusion, and rho represents the threshold search upper bound ratio coefficient, used to control the maximum search range of the minimum participation threshold and avoid excessive forced participation of the target mode. Here, ρ is greater than 0 and less than or equal to 1. In one implementation, rho is set to 0.6; when the number of modes participating in the fusion is 3, = 0.6 × 1 / 3 = 0.20.

[0062] After determining the search interval, a coarse search, a local refined search, and a further refined search are performed within that interval. For each candidate tau, modality weight adjustment, latent space construction, and unsupervised clustering are re-executed. A comprehensive evaluation function is constructed based on clustering effectiveness, clustering stability, the minimum participation index of the target modality, and class balance. The tau with the optimal comprehensive evaluation function is selected as the minimum participation threshold of the target modality.

[0063] In this embodiment, the comprehensive evaluation function can comprehensively consider the silhouette coefficient, Calinski-Harabasz index, Davies-Bouldin index, cluster stability ARI, minimum participation of the target mode, and class balance. See also Figure 7 Under the current data conditions, the minimum participation threshold for the optimal target mode obtained by adaptive search is: = 0.04, the optimal number of clusters is 8, the cumulative explained variance of PCA is 0.9525, silhouette is 0.471678, stability_ari is 0.984624, and cluster_balance is 0.902059. These results indicate that the minimum participation threshold for the target modality can be determined jointly by the data structure and the clustering objective, rather than being fixed to a certain empirical value manually.

[0064] See Figure 8 As tau increases, the average fusion weight of the building or image modality gradually increases, while the average fusion weight of the text modality and POI modality are adjusted accordingly. When tau is too high, it will over-interfere with the original adaptive weight structure. Therefore, tau is not necessarily better the larger it is, but should be determined by clustering effectiveness, stability, and modality participation.

[0065] The adaptive optimization of tau can also be determined by Bayesian optimization, sequential model optimization, black box optimization, or soft constraint penalty term optimization.

[0066] Step 8: Construction of multimodal fusion latent space and final clustering.

[0067] After determining the adaptive fusion weights for each modality and the minimum participation threshold for the target modality, the representations of each modal subspace are weighted and fused to form a joint representation of community update demand and motivation. For community i, its fused representation can be expressed as: , in, Let i represent the community i in the m-th modal subspace. Let represent the final fusion weight of the m-th mode, and M represent the number of modes participating in the fusion. This represents the square root operation.

[0068] Subsequently, unsupervised clustering is performed in the fusion latent space to obtain the community update demand-dynamic structure type. The output includes the community type label, the number and proportion of each type of community, the feature profile of each type in modalities such as buildings or images, resident text and POIs, and the corresponding demand-dynamic structure explanation.

[0069] See Figure 10 In this embodiment, the final community renewal demand-motivation type presents differentiated characteristic profiles across dimensions such as building rating, textual appeals, total number of POIs, number of POI categories, and text availability. This demonstrates that the output of this invention is not a single comprehensive score, but rather an interpretable community renewal demand-motivation structure type.

[0070] See further Figure 11 In this embodiment, based on the clustering results of the fusion potential space after adaptive tau optimization, eight types of community renewal demand-driving structure are finally formed, namely: high supply and low expression type, general supply and low expression type, high improvement demand-multiple supply support type, conventional improvement demand type, outstanding building quality and low expression type, low supply and low expression type, upgrade-oriented high demand type and built environment shortcomings and low expression type.

[0071] Among them, the high-supply, low-expression type is characterized by a high total number and functional categories of POIs but a low level of explicit resident demands, indicating that the external service supply of this type of community is relatively abundant and the pressure for short-term explicit renewal is relatively weak; the general-supply, low-expression type is characterized by low textual demands and a general level of POI supply, making it suitable as a target for routine maintenance and dynamic monitoring; the high-improvement-multi-supplier-supported type is characterized by high basic and improvement demands, and a certain service supply foundation, making it suitable for quality improvement and facility enhancement; the routine improvement demand type is characterized by stable but relatively moderate resident improvement needs, making it suitable for inclusion in the gradual community renewal sequence.

[0072] The "high-quality building - low-expression" type is characterized by high architectural or image modal scores but low explicit resident demands, indicating that the focus of renovation may not be on the building itself, but on improving service facilities and living convenience. The "low-supply - low-expression" type is characterized by low textual demands and weak POI supply, which may indicate a potential blind spot in renovation identification, requiring further on-site investigation to determine the true needs. The "upgrade-oriented - high-demand" type is characterized by high demands for upgrades and overall needs, indicating that this type of community has a strong need for quality improvement and can be identified as a key renovation target. The "built environment shortcomings - low-expression" type is characterized by low architectural or image modal scores and insufficient explicit resident expression, which is suitable for further analysis through community surveys, facility improvement, and governance intervention.

[0073] The above eight categories of results are clustering interpretations obtained based on data from a specific study area in this embodiment, and are not intended to limit the present invention to output a fixed number of types. The core of the present invention lies in automatically identifying interpretable community update demand-dynamic structure types through multimodal adaptive fusion clustering; under different city conditions, different data scales, or different modal combinations, the final number of types can be determined comprehensively based on the clustering effectiveness, stability, and interpretability.

[0074] Step 9: Modal access conditions and weak modal robustness verification.

[0075] See Figure 5 This invention is not limited to only three modalities: buildings or images, resident opinion texts, and Points of Interest (POIs). For new modalities, as long as the following conditions are met, they can be included in the framework of this invention as candidate modalities: First, it can be mapped to a unified community analysis unit; Second, it can form community-level feature representations; Third, it can calculate modal mass fraction. ; Fourth, it can estimate modality discrimination contribution scores based on the initial community fractal structure. .

[0076] For low-quality, low-discrimination, or random noise modes, by , and The joint constraints reduce the impact of its final fusion.

[0077] To verify the robustness under modal variation conditions, this embodiment further includes modal subset experiments and pseudo-noise modal experiments. Modal subset experiments include combinations of text modality with POI modality, text modality with architecture or image modality, POI modality with architecture or image modality, and complete combinations of text modality, POI modality, and architecture or image modality. These experiments are used to verify that the method can operate under different available modal combinations and, based on… and Adaptively adjust the fusion weights.

[0078] Furthermore, a random pseudo-noise mode is constructed as a new mode to be integrated into the fusion framework. Experimental results show that after adding the pseudo-noise mode, the discrimination contribution score of this mode is improved. The average fusion weight is 0.00104518. The value was 0.015088, significantly lower than the text, POI, and building or image modalities. See also Figure 9 The average fusion weight of the pseudo-noise mode is significantly reduced, indicating that the present invention can suppress the weight of newly added modes with low discrimination and low information content, and has a certain robustness to weak modes.

[0079] It should be noted that the above experiments do not imply that arbitrarily adding a modality will necessarily improve the clustering effect, but rather illustrate that the applicability of this invention does not depend on a fixed modality name, but on whether the modality can meet the access conditions such as community-level mapping, feature representation, quality score calculation, and discriminative contribution estimation; for weak quality or weak discriminative modalities, this invention can... , and The joint constraints reduce their interference with the final clustering results.

[0080] This invention also provides a multimodal classification system for urban community renewal demand dynamics with quality constraints, which can be used for urban renewal classification, renewal priority identification, and planning governance decision support. This system can be implemented as a software system, deployed on a server or local computing device, or integrated into an electronic device or computer-readable storage medium. Figure 6 The modular structure of the system in this embodiment is shown.

[0081] The system includes: a data acquisition module, an observation signal construction module, a feature processing module, a quality assessment module, an initial classification module, a discriminant contribution estimation module, an adaptive fusion module, a threshold optimization module, a fusion latent space module, a final clustering module, a result interpretation module, and a robustness verification module. Each module corresponds one-to-one with the steps in the method embodiment, as detailed below: The data acquisition module is used to acquire community-level data in at least two modalities from the target city community. The acquired modalities include architectural or image modalities, resident opinion text modalities, and point-of-interest (POI) modalities. Among them, architectural or image modalities are used to characterize the physical built environment status of the community; resident opinion text modalities are used to characterize residents' explicit renewal needs; and POI modalities are used to characterize the service supply and functional structure around the community.

[0082] The observation signal construction module is used to perform community-level aggregation and feature representation of data from each modality, constructing community-level feature vectors corresponding to each modality. Specifically, for resident opinion text data, this module can use batch inference or natural language processing models for structured extraction, identifying community information, demand levels, demand themes, and confidence levels in the text. Valid text records are aggregated by community to form community-level text indicators such as the number of basic demands, improvement demands, upgrade demands, total demands, and average confidence levels. For point-of-interest (POI) data, this module performs statistical aggregation according to a certain service area around the community or neighborhood, forming features such as the total number of POIs, the number of major POI categories, and the number of subcategories of POIs. For building or image data, this module extracts building ratings, building forms, floor structures, or image recognition features that characterize the physical built environment of the community. Subsequently, the features of each modality are mapped to the same community-level feature matrix through a unified community identifier.

[0083] The feature processing module is used to handle missing values, transform variables, and standardize community-level feature vectors to obtain standardized representations of each modal subspace. Specifically, this module excludes community identifier fields used for data association, indexing, and result write-back from clustering as modeling features; performs logarithmic transformation on count features to reduce the influence of long-tail distribution; standardizes continuous features; and constructs a text availability variable based on whether there are valid records in the text.

[0084] The quality assessment module is used to calculate the modality quality score for each modality. The modality quality score characterizes the data availability, completeness, reliability, and stability of the corresponding modality. Specifically, the quality score for the text modality is determined by the text availability variable, text recognition confidence, and the number of valid texts; the quality score for the building or image modality is determined by the building data completeness, image recognition confidence, or building rating validity; and the quality score for the point of interest (POI) modality is determined by the POI data coverage, category completeness, and spatial matching validity.

[0085] The initial classification module is used to initially concatenate the representations of each modality subspace, perform dimensionality reduction, and then perform the first unsupervised clustering to generate an initial community classification structure. This initial classification structure is not the final classification result, but serves as a feedback basis for subsequent estimation of the discriminative contribution of each modality. In one implementation, this module uses principal component analysis for dimensionality reduction and the KMeans algorithm to generate initial cluster labels.

[0086] The discriminant contribution estimation module calculates the discriminant contribution score of each modality to community typing based on the initial community typing structure. For any modality, this module calculates the ratio of the inter-class difference to the intra-class fluctuation based on the inter-class difference of the modality's features across different initial categories and the intra-class fluctuation within the same initial category. This ratio is used as the discriminant contribution score for that modality. When a modality shows significant differences across different community types and is more concentrated within the same type, the discriminant contribution score for that modality is higher, indicating that the modality has a strong discriminant contribution to community typing.

[0087] The adaptive fusion module constructs adaptive fusion weights for each modality based on its modality quality score and discrimination contribution score. Specifically, this module multiplies the quality score of each modality by its discrimination contribution score to obtain the joint quality-discrimination score for each modality; it then divides this joint quality-discrimination score by the sum of the joint quality-discrimination scores of all modalities to obtain the adaptive fusion weight for each modality. Through this method, the modality weights are no longer determined solely by modality quality, but rather by a combination of modality reliability and classification contribution.

[0088] A threshold optimization module is used to determine the minimum participation threshold for the target modality. For a preset target modality (in this embodiment, architecture or image modality), this module determines a threshold search interval based on the number of modalities participating in the fusion. The upper bound of the threshold search interval is determined according to the number of modalities participating in the fusion and controlled by a search upper bound proportional coefficient. Within the search interval, weight correction, fusion latent space construction, and unsupervised clustering are performed on candidate thresholds. A comprehensive evaluation function is constructed based on clustering effectiveness, clustering stability, minimum participation of the target modality, and class balance. The candidate threshold with the optimal comprehensive evaluation function is selected as the minimum participation threshold for the target modality. The adaptive fusion module is also used to correct the adaptive fusion weights according to this threshold, ensuring that the fusion weight of the target modality is not lower than the threshold. In one embodiment, the upper bound of the threshold search interval is determined according to the number of modalities participating in the fusion, specifically by multiplying the search upper bound proportional coefficient by the reciprocal of the number of modalities participating in the fusion. In one embodiment, the search upper bound proportional coefficient is set to 0.6.

[0089] The fusion latent space module is used to perform weighted fusion of the representations of each modal subspace based on the adaptive fusion weights after threshold correction, and to construct the fusion latent space.

[0090] The final clustering module is used to perform a second unsupervised clustering in the fusion potential space and output the community update requirement-dynamic structure type.

[0091] The results interpretation module outputs feature profiles of each type of community across modalities such as buildings or images, resident text, and points of interest, along with corresponding demand-dynamic structure explanations. The output includes the community type label, the number and proportion of communities of each type, and the corresponding demand-dynamic structure explanation for each type.

[0092] The robustness verification module is used to perform modality subset experiments and / or pseudo-noise modality experiments to verify the classification stability of the system under different modality combinations. This module can add new candidate modalities and verify whether they meet preset access conditions. For low-quality, low-discrimination, or random noise modalities, the system reduces their final fusion impact through joint constraints of quality score, discrimination contribution score, and fusion weight, ensuring the system's robustness to modality changes.

[0093] When the system is working, the data acquisition module acquires at least two modalities of data from the target city community; the observation signal construction module performs community-level aggregation and feature representation of each modal data, and constructs community-level feature vectors corresponding to each modality; the feature processing module completes missing data processing, variable transformation and standardization, and obtains the representation of each modal subspace.

[0094] The quality assessment module and the initial classification module are executed in parallel or sequentially. The initial classification module initially assembles and clusters the representations of each modality subspace to generate an initial community classification structure; the discriminant contribution estimation module calculates the discriminant contribution score for each modality based on this classification structure. The quality assessment module calculates the quality score for each modality. The adaptive fusion module constructs adaptive fusion weights based on the quality score and the discriminant contribution score. The threshold optimization module determines the minimum participation threshold for the target modality, and the adaptive fusion module adjusts the weights accordingly.

[0095] The latent space fusion module performs weighted fusion of the representations of each modal subspace based on the corrected weights; the final clustering module performs clustering in the latent space fusion and outputs the community update requirement-dynamic structure type; the result interpretation module outputs the type profile and interpretation results.

[0096] The system in this embodiment can be implemented as a software system and deployed on a server or local computing device. In one embodiment, the system is integrated into an electronic device, and the functions of the above-mentioned modules are implemented by a processor executing a computer program stored in memory. In another embodiment, the system is implemented as a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the functions of the above-mentioned modules. The system output results can be used for community update classification, update priority ranking, and planning governance decision support.

[0097] This invention also provides an electronic device, including a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the urban community renewal demand dynamics classification method with multimodal discrimination quality constraints as described in any of the above method embodiments. This electronic device can be a server, a personal computer, an edge computing device, or a dedicated computing device.

[0098] In one embodiment, the electronic device further includes a communication interface for acquiring community-level data in at least two modalities of the target urban community, the modalities including building or image modalities, resident opinion text modalities, and point-of-interest modalities.

[0099] In one embodiment, the electronic device further includes a display device for visually displaying community update needs - power structure types and characteristic profiles of each type of community.

[0100] The output of the electronic device is used for urban community renewal classification, renewal priority ranking, and planning and governance decision support.

[0101] In summary, this invention provides a method and system for classifying urban community renewal demand-dynamics based on joint modeling of modal discriminative contribution and quality constraints. The method uses communities or sub-communities as unified analysis units, acquiring multimodal data such as building or image modalities, resident opinion text modalities, and Point of Interest (POI) modalities to construct community-level multimodal observation signals. It cleans, aggregates, characterizes, and standardizes the data for each modality, calculating modal quality scores. An initial community classification structure is generated using the initial clustering results, and the discriminative contribution of each modality to community classification is estimated based on this initial classification structure. Adaptive fusion weights are constructed based on the modal quality scores and modal discriminative contribution scores. For target modalities with significant interpretative value, such as buildings or images, a minimum participation threshold is determined through an adaptive τ-threshold optimization mechanism. Finally, a fusion latent space is constructed, and unsupervised clustering is performed to output the community renewal demand-dynamic structure type. This invention, through initial classification feedback, quality-discrimination joint weights, adaptive target modality participation preservation, and weak modality suppression mechanisms, can achieve automatic identification of multimodal community update types without the need for manual weight presets or manual labeling of community types. It also improves the separation, stability, multimodal completeness, and interpretability of classification results, and can be used for urban community update classification, update priority identification, and planning and governance decision support.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal classification method for urban community renewal demand dynamics with quality constraints, characterized in that, include: Acquire community-level data in at least two modalities from the target city community, including architectural or image modalities, resident opinion text modalities, and point-of-interest modalities; Community-level aggregation and feature representation are performed on each modality data to construct community-level feature vectors corresponding to each modality. The community-level feature vectors are then preprocessed to obtain standardized subspace representations of each modality. The representations of each modal subspace are initially spliced ​​together, and after dimensionality reduction, the first unsupervised clustering is performed to generate the initial community classification structure. Based on the initial community classification structure, the discriminant contribution score of each modality to the community classification is calculated. The discriminant contribution score is calculated as follows: for any modality, based on the inter-class differences of the modality features between different initial categories under the initial community classification structure, and the intra-class fluctuations within the same initial category, the ratio of the inter-class differences to the intra-class fluctuations is calculated as the discriminant contribution score of the modality. Calculate the modal quality score for each modality, which is used to characterize the data availability, integrity, reliability, and stability of the corresponding modality; Based on the modality quality score and discrimination contribution score of each modality, an adaptive fusion weight is constructed for each modality. The subspace representations of each modality are then weighted and fused according to the adaptive fusion weight to construct a fusion latent space. The adaptive fusion weight is constructed as follows: the modality quality score and discrimination contribution score of each modality are multiplied to obtain the quality-discrimination joint score of each modality; the quality-discrimination joint score of each modality is divided by the sum of the quality-discrimination joint scores of all modalities to obtain the adaptive fusion weight of each modality. Perform a second unsupervised clustering in the fusion potential space to output the community update requirement-dynamic structure type.

2. The urban community renewal demand dynamics classification method with multimodal discriminant quality constraints according to claim 1, characterized in that, It also includes an adaptive determination step for the minimum participation threshold of the target mode: For a preset target modality, a threshold search interval is determined based on the number of modalities participating in the fusion. Within the search interval, weight correction, fusion potential space construction, and unsupervised clustering are performed on the candidate thresholds. A comprehensive evaluation function is constructed based on the clustering effectiveness index, clustering stability index, minimum participation index of the target modality, and class balance index. The candidate threshold that is optimal in the comprehensive evaluation function is selected as the minimum participation threshold of the target mode; the adaptive fusion weights are then corrected based on the minimum participation threshold of the target mode.

3. The urban community renewal demand dynamics classification method with multimodal discriminant quality constraints according to claim 2, characterized in that, The target modality is a building or image modality.

4. The urban community renewal demand dynamics classification method with multimodal discriminant quality constraints according to claim 1, characterized in that, The community-level feature vector of the resident opinion text modality includes the number of basic requests, the number of improvement requests, the number of upgrade requests, the total number of requests, text usability variables, and text confidence; the community-level feature vector of the building or image modality includes one or more of the following: building quality score, building floor structure, street view image recognition score, and community image recognition score; the community-level feature vector of the point of interest modality includes the total number of points of interest, the number of major categories of points of interest, the number of subcategories of points of interest, the density of public service facilities, and the density of commercial service facilities.

5. The urban community renewal demand dynamics classification method with multimodal discriminant quality constraints according to claim 1, characterized in that, It also includes modal access and weak mode suppression steps: When adding a new candidate modality, it is determined that the new candidate modality meets the preset access conditions. The preset access conditions include being able to map to a unified community analysis unit, being able to form a community-level feature representation, being able to calculate the modality quality score, and being able to estimate the discrimination contribution score based on the initial community classification structure. After access, if the modality quality score and / or discrimination contribution score of the newly added candidate modality are lower than the preset threshold, the adaptive fusion weight of the newly added candidate modality will be reduced.

6. The urban community renewal demand dynamics classification method with multimodal discriminant quality constraints according to claim 2, characterized in that, The upper bound of the threshold search interval Determined based on the number M of modes participating in the fusion. ,in This is the upper bound of the search ratio.

7. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor is configured to implement, when executing the instructions, the urban community renewal demand dynamics classification method with multimodal discriminative quality constraints as described in any one of claims 1 to 6.

8. A multimodal discriminant quality-constrained urban community renewal demand dynamics classification system, used to implement the multimodal discriminant quality-constrained urban community renewal demand dynamics classification method as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to acquire community-level data in at least two modalities from the target city community, including architectural or image modalities, resident opinion text modalities, and point-of-interest modalities. The observation signal construction module is used to perform community-level aggregation and feature representation on each modality data, and construct the community-level feature vector corresponding to each modality; The feature processing module is used to process the community-level feature vectors by handling missing values, transforming variables, and standardizing them to obtain standardized representations of each modal subspace. The initial classification module is used to initially splice together the representations of each modal subspace, perform the first unsupervised clustering after dimensionality reduction, and generate the initial community classification structure. The discriminant contribution estimation module is used to calculate the discriminant contribution score of each mode to the community classification based on the initial community classification structure. The quality assessment module is used to calculate the modality quality score for each modality, and the modality quality score is used to characterize the data availability, integrity, reliability and stability of the corresponding modality; An adaptive fusion module is used to construct adaptive fusion weights for each mode based on the mode quality score and the discrimination contribution score of each mode; The latent space fusion module is used to perform weighted fusion of the representations of each modal subspace based on the adaptive fusion weights to construct the latent space fusion ... The final clustering module is used to perform a second unsupervised clustering in the fusion potential space and output the community update requirement-dynamic structure type.

Images