Venue function diagnosis method based on spatiotemporal correlation dynamic atlas

Abstract

The application discloses a venue function diagnosis method based on a space-time correlation dynamic atlas, and belongs to the technical field of data processing and intelligent analysis. In view of the problems of split multi-source data, insufficient space behavior coupling and lack of quantitative constraints in existing venue diagnosis, the method obtains user interactive behavior data, user perception evaluation data and physical layout data of a target venue, generates a mixed demand category through confidence weighted fusion, constructs a dynamically updated space-time correlation dynamic atlas based on continuous behavior sequences and reachability constraints, quantitatively calculates a function redundancy index and a gap index, and finally uses multi-objective optimization to output a layout adjustment strategy under the constraints of space area and engineering budget, with the maximum demand matching and the minimum transformation cost as the target. The method can be used for health degree evaluation and scientific layout optimization of function configuration of various venues.

Description

Technical Field

[0001] This invention belongs to the field of data processing and intelligent analysis technology, specifically relating to a venue function diagnosis method based on spatiotemporal correlation dynamic graphs. Background Technology

[0002] In the diagnostic assessment of venue functional configuration and spatial layout, traditional methods typically rely on static functional zoning plans and a limited number of questionnaires, making it difficult to reflect the dynamic coupling relationship between user behavior and physical space during actual venue operation. Existing technologies often isolate functional requirements analysis, spatial accessibility assessment, and facility utilization efficiency calculation, lacking systematic collection and analysis of continuous user movement sequences within the venue. This results in the inability to quantify the correlation strength between functional points based on real behavioral data, making it difficult to objectively determine whether functional configuration is redundant or insufficient.

[0003] Meanwhile, existing solutions often rely on single questionnaire scores or simple visitor flow statistics to determine the demand attributes of each functional point, failing to fully utilize multi-source information from user interaction behavior data and perception evaluation data. This results in demand category determinations being limited by sample bias or data sparsity, making it difficult to reliably support subsequent diagnostic calculations. Due to the lack of effective modeling of spatial accessibility constraints between functional points, many evaluation methods treat functional points as independent units, ignoring the constraints of building physical structures (such as walls, partitions, and passageway distances) on users' actual movement paths. Consequently, the functional correlations derived from these methods often deviate from real-world usage scenarios.

[0004] Furthermore, existing technologies for identifying functional redundancy and gaps typically rely on fixed thresholds or empirical rules, failing to dynamically adjust based on venue operating hours, changes in visitor density, and evolving user needs. This results in poor timeliness of diagnostic results, making it difficult to reflect timely trends in venue functional status. In venue functional layout adjustment decisions, current methods largely depend on manual experience to formulate renovation plans, lacking quantitative optimization tools aimed at minimizing renovation costs and maximizing demand matching. Moreover, the generation of renovation strategies is often not effectively constrained by actual space area and project budgets, leading to insufficient feasibility of optimization solutions. These problems make it difficult for venue managers to scientifically and efficiently diagnose the health status of existing functional configurations and to provide reliable data support and strategic guidance for layout adjustments. Summary of the Invention

[0005] This invention provides a venue function diagnosis method based on a spatiotemporal dynamic map. It can integrate multi-source data such as user interaction behavior, perception evaluation and physical layout to construct and dynamically update a spatiotemporal dynamic map that reflects the spatial-behavioral coupling relationship. It can quantitatively assess the redundancy and gap of each functional point, and automatically generate layout adjustment strategies under the constraints of space area and engineering budget through multi-objective optimization, so as to achieve accurate diagnosis and scientific optimization of venue function configuration.

[0006] To achieve these objectives and other advantages of the present invention, a venue function diagnosis method based on spatiotemporal correlation dynamic graphs is provided, comprising: S1. Obtain user interaction behavior data, user perception evaluation data, and physical layout data of multiple functional points within the target venue; S2. Input the user interaction behavior data into the pre-trained demand classification model to obtain the objective demand category of each function point, and combine it with the category determined by the user perception evaluation data to generate the mixed demand category of each function point through confidence weighted fusion. S3. Based on the continuous behavioral sequence data generated by users during their movement within the venue, the transfer probability between functional points is mined, and the accessibility constraints between functional points are combined with the physical layout data to construct a spatiotemporal correlation dynamic graph reflecting the coupling relationship between venue space and behavior. The edge weights representing the correlation strength between functional points in the spatiotemporal correlation dynamic graph are dynamically updated according to the time decay function. S4. Based on the mixed demand categories and the dynamically updated spatiotemporal correlation dynamic map, the functional redundancy index and functional gap index of each functional point are quantitatively calculated. The functional gap index is determined by calculating the difference between the membership degree of the excitatory demand of the target functional point and the completeness of its associated complementary functions in the spatiotemporal correlation dynamic map. The completeness of the associated complementary functions is determined based on the joint determination of the spatial coupling degree and functional correlation degree of the functional point in the venue space. S5. With the optimization objectives of maximizing the matching degree of user needs and functions within the venue and minimizing the cost of physical space renovation, under the constraints of the area of ​​the venue that can be renovated and the project budget, a multi-objective optimization algorithm is used to search for the set of optimization strategy parameters for adjusting the layout of functional points, and the set of optimization strategy parameters is output.

[0007] Preferably, in step S1, obtaining user perception evaluation data specifically involves: Interactive touch terminals are deployed at the entrances and exits of various functional points in the venue to collect multi-dimensional perception score data of users on the current functional point before leaving the venue. The multi-dimensional perception score includes at least the functional importance dimension and the user experience dimension, and each dimension is quantified using a Likert scale. Furthermore, text comment data input by users on the interactive touch terminal is collected, and a sentiment analysis model deployed on the local server of the venue is used to extract the sentiment polarity of the text comment data in real time. The extracted sentiment tendency value is then aligned with the multidimensional perception score data after being timestamped, and used as a supplementary quantitative item for the user perception evaluation data of this function point.

[0008] Preferably, in step S2, a mixed requirement category for each functional point is generated by confidence-weighted fusion, specifically as follows: Obtain the first probability distribution corresponding to the objective demand category output by the demand classification model, and the second probability distribution obtained based on the user perception evaluation data mapping; The first time coverage is calculated based on the total number of samplings of user interaction behavior data at each functional point within the preset diagnostic period. The second sampling completeness is calculated based on the ratio of the number of valid questionnaires collected to the total number of questionnaires distributed for user perception evaluation data at each functional point. The first time coverage is then weighted and corrected based on the average daily passenger flow density at each functional point of the venue. The corrected first-time coverage and second-sample completeness are used as two original weights. First, the two are added together to obtain a total weight value. Then, each original weight is divided by this total weight value to obtain two normalized final weights. If the sum of the two is zero, they are directly assigned equal default weights. The normalized final weights are used as weighting coefficients to linearly weight the first probability distribution and the second probability distribution to obtain the fusion probability distribution of each function point belonging to different demand categories. The category corresponding to the highest probability is taken as the hybrid demand category.

[0009] Preferably, in step S3, the transition probability between the mining function points is specifically as follows: Get the continuous spatial location sequence of the user within a preset time window, collected by Bluetooth positioning base stations or radio frequency identification card readers deployed in the venue. Sort the multiple functional points visited by the user in the continuous spatial location sequence according to time to construct a continuous interaction sequence. Treat each functional point as a node, count the frequency of directed transfers from one functional point to another, and use the frequency of directed transfers as the initial weight of the directed edge. For data gaps in the interaction sequence caused by location signal obstruction or card reader blind spots within the venue, a data completion model based on a spatiotemporal graph convolutional network deployed on computing nodes at the venue edge is used. The completion model consists of stacked graph convolutional layers and temporal convolutional layers. The graph convolutional layers use an adjacency matrix constructed based on the venue's physical layout data to capture spatial dependencies between functional points. The temporal convolutional layers use gated linear units to capture long-range temporal dependencies of user activity sequences within the venue. The physical proximity distance between functional points and the historical venue passenger flow transfer patterns during the same period are used as prior constraints to regularize the completion results.

[0010] Preferably, the accessibility constraints between functional points based on the physical layout data specifically include: The physical layout data includes the building information model data of the venue. The three-dimensional spatial coordinates of each functional point are extracted from the building information model, and the spatial coordinate information of building components such as walls, partitions, columns, escalators and fire lanes is analyzed to construct a two-dimensional grid map marked with passable and impassable areas. Using the A* path search algorithm on the two-dimensional grid map, with the three-dimensional spatial coordinates of the functional points as the start and end nodes, the shortest physical reachable path and the corresponding actual walking distance between any two functional points in the venue are calculated. The cost function in the path search process includes a weighted sum of path length cost and passing pedestrian density cost. Directed edges between two functional points that have no physically reachable path are directly removed from the spatiotemporal dynamic graph, and the weights of edges whose actual walking distance exceeds a preset comfortable walking threshold are multiplied by a decay coefficient set according to distance segments, wherein the comfortable walking threshold is dynamically set according to the venue type and the average functional area spacing within the venue.

[0011] Preferably, in step S3, the dynamic updating of the edge weights representing the correlation strength between functional points in the spatiotemporal correlation dynamic graph based on the time decay function specifically involves: The time decay function adopts an asymmetric exponential decay model, and its decay rate parameter is dynamically switched according to the time period of the venue's operation. The time period includes the venue's pre-opening preparation period, normal operation period, and pre-closing clearing period. During normal operating hours, the operating hours are further divided into peak sub-periods and off-peak sub-periods according to the real-time passenger flow density of the venue. A first attenuation rate is used in the peak sub-periods and a second attenuation rate is used in the off-peak sub-periods. The first attenuation rate is less than the second attenuation rate. The weight value of each directed edge is updated in real time after being weighted and accumulated one by one according to the time difference between the timestamp of the corresponding transition behavior and the current diagnosis time, through the time decay function.

[0012] Preferably, in step S4, the calculation of the functional redundancy index is specifically as follows: Using the mixed demand category as a label, search for functional points with the same mixed demand category and whose edge weights are greater than a preset redundancy threshold in the spatiotemporal dynamic graph to form a candidate redundancy subgroup. The redundancy threshold is adaptively determined based on the average connection strength of each functional point in the venue and the venue space density coefficient. For any target functional point within the candidate redundant subgroup, the directed edge weights between it and all other functional points within the candidate redundant subgroup are extracted, a local directed adjacency matrix is ​​constructed, the row vectors of the local directed adjacency matrix are L1 normalized, the arithmetic mean of the normalized element values ​​is calculated, and the arithmetic mean is multiplied by the global clustering coefficient of the candidate redundant subgroup to obtain the functional redundancy index of the target functional point. The global clustering coefficient reflects the density of functional points within the subgroup in the venue space.

[0013] Preferably, in step S4, the method for determining the completeness of the associated complementary functions is as follows: In the spatiotemporal dynamic graph, the target functional point is used as the starting node, and a breadth-first search is performed along the direction of the directed edge to extract a set of directed neighborhood nodes of depth N, where N is a preset positive integer, and the value of N is related to the functional area level to which the target functional point belongs in the venue space. Nodes that have a pre-defined synergistic and complementary relationship with the exciting demand of the target functional point are selected from the neighborhood node set and marked as complementary nodes. The pre-defined synergistic and complementary relationship is predefined by the venue functional configuration knowledge base. Based on the physical proximity of each complementary node and the target functional point in the venue space and the graph edge weights between them, the spatial-functional joint support of each complementary node is calculated, and the joint support of all complementary nodes is accumulated and normalized to obtain the completeness of the associated complementary function.

[0014] Preferably, in step S5, the multi-objective optimization algorithm is a non-dominated sorting genetic algorithm with an elitist strategy, and its mutation operation is constrained by the spatiotemporal correlation dynamic graph, specifically: During the execution of the mutation operator in the genetic algorithm, the functional coupling degree between any two functional points is calculated based on the spatiotemporal correlation dynamic graph. The functional coupling degree is equal to the sum of the weights of the directed edges between the two functional points multiplied by the inverse ratio of their physical distance on the two-dimensional grid map of the venue. When the functional coupling between any two functional points is greater than a preset coupling threshold, a cooperative constraint is applied to the layout adjustment genes of these two functional points during the mutation operation: if the functional gap index of one functional point is higher than the preset gap threshold, the deletion gene site of the other functional point that has a strong coupling relationship with that functional point is locked in an inactive state to prohibit deletion; at the same time, high gap functional points are allowed to move towards their coupled functional points to improve spatial accessibility, but mutations that cause the distance between the two to increase due to reverse movement are prohibited; for the addition operation, the addition mutation is preferred only when the candidate position of the addition is within the preset neighborhood of the high gap functional point and the addition functional category has a cooperative complementary relationship with its excitatory demand.

[0015] Preferably, in step S5, the actual walking distance between any two functional points is pre-stored as a static spatial distance matrix and directly read during the iteration of the multi-objective optimization algorithm, prohibiting real-time invocation of the A* path search algorithm; the optimization strategy parameter set output by the multi-objective optimization algorithm needs to be verified by a digital twin simulation environment, and if the simulation fails, it will revert to the algorithm to search again until convergence; after the physical transformation is implemented, the demand classification model and the spatiotemporal correlation dynamic map are only updated offline in a quarterly or semi-annual cycle; between two offline updates, only the decay rate parameter of the time decay function is adjusted online, the map topology and the static spatial distance matrix remain unchanged, and the calculation range of functional coupling degree in the iteration is limited to the target functional point and the top K nodes in its neighborhood, where K is a preset positive integer and its value is between 10 and 15.

[0016] The present invention has at least the following beneficial effects: First, by constructing a unified diagnostic framework that integrates user interaction behavior data, perception evaluation data, and physical layout data, functional requirement analysis, spatial accessibility assessment, and behavioral pattern mining are incorporated into the same modeling system, effectively solving the problem of the difficulty in collaboratively utilizing multi-source heterogeneous data. At the data collection level, multi-dimensional perception scores are collected by deploying interactive touch terminals at functional point entrances and exits, and combined with real-time extraction of sentiment from text comments, enriching the dimensions and information content of perception data and avoiding scoring bias caused by retrospective recall. At the requirement fusion level, a confidence weighting coefficient dynamically calculated based on actual data collection is introduced. The fusion weight is adaptively adjusted according to the first-time coverage and the second-sampling completeness, effectively suppressing category judgment fluctuations caused by insufficient sampling of a certain data source in a specific time period or region, and improving the stability and robustness of the mixed requirement category output results.

[0017] Secondly, based on the continuous user behavior sequence, the probability of transfer between function points is mined, and a two-dimensional grid map of accessible areas is constructed by combining building information model data. The A* path search algorithm is used to calculate the shortest reachable path with path length and pedestrian density as the cost function. Unreachable edges are directly pruned, and edges exceeding the threshold are attenuated in segments according to distance, effectively purifying invalid connections in the spatiotemporal dynamic graph. This ensures that the edge weights of the spatiotemporal dynamic graph truly reflect the actual mobility probability of users under the constraints of the building's physical structure. Simultaneously, to address data gaps caused by missing positioning signals, a spatiotemporal graph convolutional network is used for completion. Graph convolutional layers capture spatial dependencies, and temporal convolutional layers use gated linear units to extract long-range temporal dependencies. Regularization is performed using physical channel proximity distance and historical passenger flow transfer patterns as prior constraints, ensuring statistical consistency of the data completion. Furthermore, an asymmetric exponential decay model is adopted, and the decay rate parameter is dynamically switched according to the operating period, allowing the edge weights to adaptively refresh with changes in venue operation rhythm and passenger flow density, always maintaining an accurate characterization of the current state's correlation strength.

[0018] Third, candidate redundant subgroups are selected using mixed demand categories as labels and a redundancy threshold adaptively determined based on average connection strength and spatial density coefficients. The redundancy index is calculated by multiplying the L1-normalized average of the row vectors in the local directed adjacency matrix with the global clustering coefficient of the subgroup. This simultaneously measures the connection strength contribution of the target functional point within the candidate redundant subgroup and the spatial compactness of the entire subgroup, making the redundancy determination spatially sensitive and adaptive. In terms of functional gap assessment, directed neighborhood nodes are extracted through breadth-first search, and nodes with predefined synergistic and complementary relationships are selected. The spatial-functional joint support is calculated by combining physical proximity and graph edge weights, and then accumulated and normalized. This ensures that the completeness assessment reflects both spatial reachability support and functional synergistic support dimensions, providing crucial assurance for the calculation quality of the functional gap index.

[0019] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the venue function diagnosis method based on spatiotemporal correlation dynamic graphs of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0022] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0023] like Figure 1 As shown, this invention provides a venue function diagnosis method based on spatiotemporal correlation dynamic graphs, constructing a complete closed-loop technical solution from data acquisition, demand fusion, spatiotemporal correlation dynamic graph modeling, quantitative diagnosis to optimization output. Specifically, it includes: S1: Acquisition of multi-source heterogeneous data: Collect three types of data from different sources and with complementary properties for each functional point in the target venue. These three types of data are user interaction behavior data, user perception evaluation data, and physical layout data of the venue.

[0024] User interaction data refers to automatically recorded user actions during their activities within the venue, including clicks on self-service terminals, passage records at turnstiles, browsing and selection on interactive screens, and signal interactions between users' mobile devices and Bluetooth beacons or Wi-Fi probes deployed within the venue. This data reflects users' objective usage of each function, encompassing multiple dimensions such as access frequency, dwell time, and depth of interaction, and serves as a crucial objective basis for inferring user needs.

[0025] User perception evaluation data comes from users' subjective feedback, reflecting their subjective perception of the functionality and its value. This method involves deploying interactive touch terminals at the entrances and exits of functional areas to collect users' immediate feedback before they leave, including multi-dimensional perception ratings and text comments. This real-time collection method ensures the timeliness and accuracy of the evaluation, effectively avoiding rating bias caused by retrospective recall. After being collected, text comments are processed in real-time by a sentiment analysis model to extract sentiment polarity values, which are then used as supplementary quantitative information in subsequent calculations.

[0026] The physical layout data originates from the venue's architectural design information, specifically including the spatial coordinates of each functional point and the spatial location information of building components such as walls, partitions, columns, escalators, and fire exits. This raw data will be used to construct a two-dimensional grid map marked with accessible and inaccessible areas, providing a foundation for subsequent accurate calculations of spatial accessibility. The introduction of physical layout data allows subsequent behavioral analysis to no longer treat functional points as isolated units, but rather to consider them comprehensively within the constraints of real architectural space.

[0027] The three types of data mentioned above are not used independently in subsequent steps, but are synergistically integrated through the unified framework of this invention. User interaction behavior data provides objective evidence of use, user perception evaluation data supplements subjective experience information, and physical layout data provides spatial constraints. Together, they support a comprehensive diagnosis of the venue's functions.

[0028] S2: Generation of Mixed Requirement Categories. After acquiring the above three types of data, step S2 involves effectively integrating objective behavioral data and subjective perception data to determine the mixed requirement category for each functional point, thereby providing a unified classification and labeling system for subsequent functional redundancy and gap diagnosis.

[0029] The demand classification model is a pre-trained neural network classifier or a decision tree-based ensemble learning model. Its input consists of user interaction behavior feature vectors collected from functional points. These feature vectors specifically include quantitative indicators such as access frequency, average dwell time, and operation completion rate per unit time. The model's output is the probability distribution of the functional point belonging to each preset demand category; this distribution is referred to as the first probability distribution in this method. The model is trained using supervised learning, with training data derived from historical behavior data of multiple venue functional points labeled with demand categories.

[0030] At the same time, user perception evaluation data is also mapped to a second probability distribution. This mapping can be implemented by combining the functional importance score and emotional polarity value from the multi-dimensional perception rating into a feature vector, and then using a pre-trained rating-to-category mapping model, such as a logistic regression model or a shallow neural network model, to convert this feature vector into a probability distribution indicating whether the functional point belongs to each demand category.

[0031] However, in actual venue operation, the availability of interactive behavior data and perceived evaluation data across different functional areas is often uneven. For example, some high-traffic functional areas may have abundant interactive behavior data, but the proportion of users actively completing evaluation questionnaires may be relatively low; conversely, other functional areas may exhibit the opposite. If a fixed weight is used for data fusion, insufficient sampling of a particular data source at a specific functional area can lead to inaccurate determinations of demand categories.

[0032] To address this, this invention introduces a confidence-weighted mechanism based on dynamic calculation of actual data sampling. The calculation of the first-time coverage is based on the total number of samplings of user interaction behavior data at each functional point within a preset diagnostic period. The more samplings, the more fully the data source characterizes the behavioral patterns of that functional point. Furthermore, since there are naturally significant differences in passenger flow at different functional points, directly comparing the number of samplings may not be objective enough. Therefore, this method also weights and corrects the first-time coverage based on the average daily passenger flow density of each functional point. The specific correction method is as follows: first, divide the total number of samplings by the average daily passenger flow of that functional point within the diagnostic period to obtain the number of samplings per unit passenger flow; then divide each number of samplings per unit passenger flow by the average value of this indicator for all functional points. The resulting ratio is the corrected first-time coverage. This means that when the number of samplings per unit passenger flow at a certain functional point is higher than the average level of the entire venue, the corrected coverage is greater than 1, indicating that the behavioral data collection at that functional point has achieved sufficient coverage relative to its passenger flow; otherwise, it is less than 1. Even a low absolute number of samples may not be high for a sparsely populated area, yet it may still achieve a high level of coverage; conversely, a densely populated area requires a larger sample size to achieve the same level of coverage. The second metric, sampling completeness, is calculated based on the ratio of the number of valid questionnaires returned for user perception evaluation data at each area to the total number of times the evaluation interface is triggered. The total number of times the evaluation interface is triggered refers to the total number of times users access the evaluation page via touchscreen terminals at the exit of that area, i.e., the total number of times the evaluation interface is presented to users. The number of valid questionnaires returned refers to the number of records where users actually submitted complete scale ratings and text comments. This indicator reflects the sufficiency and reliability of the subjective evaluation data; a higher ratio indicates better user cooperation and more representative collected perception evaluation data.

[0033] By using the corrected first-time coverage and second-sampling completeness as dynamic weighting coefficients, a linear weighted combination of the first and second probability distributions is performed to obtain the fused probability distribution of each functional point belonging to different demand categories. Finally, the category corresponding to the highest probability is taken as the mixed demand category for that functional point. This fusion method ensures that data sources with high data quality and sufficient coverage naturally receive higher weights during the fusion process, thereby effectively suppressing fluctuations in category determination caused by insufficient sampling of a particular data source on a specific functional point.

[0034] S3: Construction and Dynamic Update of the Spatiotemporal Relationship Dynamic Graph. This involves constructing a spatiotemporal relationship dynamic graph that simultaneously reflects the spatial accessibility between functional points and the correlation of user behavior transfers, and enabling it to dynamically evolve with changes in time and operational status. Essentially, the spatiotemporal relationship dynamic graph is a directed weighted graph structure, where each node represents a functional point within the venue, directed edges represent the behavioral transfer relationship of a user moving from one functional point to another, and the edge weights characterize the strength of this transfer relationship. The construction process of this spatiotemporal relationship dynamic graph consists of two stages: behavioral transfer probability mining and spatial accessibility constraints. These two stages work together to obtain a graph structure that accurately reflects the actual usage status of the venue.

[0035] In the behavior transition probability mining stage, this method utilizes Bluetooth positioning base stations or RFID readers deployed within the venue to collect the user's continuous spatial location sequence within a preset time window. Specifically, one option is to employ a Bluetooth triangulation method based on received signal strength indication, calculating the user's real-time coordinates on a two-dimensional plane by jointly measuring the signals emitted by the user's mobile device from three or more Bluetooth base stations. Another option is to deploy RFID readers at the entrances and exits of various functional points within the venue. Users wear uniquely identified RFID cards or wristbands that communicate with the readers in the near field, recording the sequence of functional points visited sequentially. After constructing a continuous interaction sequence by chronologically ordering the multiple functional points visited sequentially in the continuous spatial location sequence, each functional point is used as a node to count the frequency of directed transitions from one functional point to another. This frequency is directly used as the initial weight of the directed edges.

[0036] The venue inevitably faces location signal obstruction issues, such as attenuation of Bluetooth signals by large metal components or thick walls, and blind spots in card reader deployment. These factors can lead to data gaps in the collected interaction sequences. To address this issue, this method employs a data completion model based on a spatiotemporal graph convolutional network deployed at the venue's edge computing nodes. This completion model consists of stacked graph convolutional layers and temporal convolutional layers. The graph convolutional layers utilize an adjacency matrix constructed based on the venue's physical layout data to capture spatial dependencies. The closer two functional points are spatially, or the more direct and unobstructed the physical path between them, the larger the corresponding value in the adjacency matrix. The graph convolutional layer will then place greater emphasis on the characteristics of these spatially adjacent nodes when aggregating information. The temporal convolutional layers use gated linear units to capture the long-range temporal dependencies of user activity sequences within the venue, meaning that a user's behavioral tendencies at a given moment may be continuously influenced by functional points visited a long time ago. To ensure that the completion results do not deviate from actual physical possibilities, this method also uses the physical proximity distance between functional points and the historical passenger flow transfer patterns during the same period as prior constraints to regularize the completion results. Specifically, if the physical distance between the two functional points corresponding to the transfer relationship predicted by the completion model is too far, or if the completion result deviates significantly from the passenger flow transfer statistics of the same period in history, a corresponding penalty term is added to the loss function to encourage the model to output completion results that are more consistent with physical reality and historical patterns. For the cold start phase when a new venue is put into operation and there is no historical passenger flow transfer pattern data for the same period, the following alternative can be adopted: Based on the functional zoning plan preset by the venue operator, an initial passenger flow transfer probability matrix is ​​formed. That is, according to the logical correlation strength between functional areas, such as the expected high correlation between the catering area and the restrooms, and the expected low correlation between the exhibition hall and the equipment room, a prior probability value is set as a temporary substitute for regularization; after the venue's accumulated operating data reaches the preset diagnostic period, such as 30 days, the preset value is gradually replaced with the actual statistical value to achieve a smooth transition from preset priors to data-driven approaches.

[0037] In the spatial accessibility constraint stage, this method transforms Building Information Modeling (BIM) data into specific constraints on the spatiotemporal dynamic map structure. First, the three-dimensional spatial coordinates of each functional point are extracted from the BIM, and the spatial coordinate information of building components such as walls, partitions, columns, escalators, and fire exits is analyzed. Based on this, a two-dimensional grid map is constructed, marking accessible and inaccessible areas. This grid map divides the venue floor into several cells, each marked as accessible or inaccessible. Then, the A* path search algorithm is used on this grid map, with the three-dimensional spatial coordinates of the functional points as the start and end nodes, to calculate the shortest physical reachable path between any two functional points within the venue and the corresponding actual walking distance. The cost function of the A* path search algorithm includes a weighted sum of path length cost and pedestrian density cost. The former ensures that the searched path is as short as possible in physical distance, while the latter makes the algorithm tend to avoid areas with high real-time pedestrian flow, thus calculating an equivalent shortest path that better reflects the user's actual choices.

[0038] Based on this, directed edges between two functional points with no physically reachable paths are directly removed from the spatiotemporal dynamic graph, because any behavioral transfer statistics are meaningless when users cannot physically reach another point. For edges where the actual walking distance exceeds a preset comfortable walking threshold, their weight is multiplied by an attenuation coefficient set according to the relative proportion of the exceedance. To ensure logical consistency between the attenuation scale and the dynamic threshold, the segmentation is no longer based on fixed absolute distance values, but rather on the percentage of the exceedance relative to the threshold itself. Specifically, a relative proportion segmentation function is used: when the actual walking distance exceeds the comfortable walking threshold by less than or equal to 10% of the threshold itself, the attenuation coefficient is 0.85; when the exceedance is greater than 10% and less than or equal to 20%, the attenuation coefficient is 0.70; when the exceedance is greater than 20%, the attenuation coefficient is 0.50. Taking the preferred dynamic calculation method as an example, let the venue's dynamic comfortable walking threshold be T = 1.5 × D. avg D avgThis is the arithmetic mean of the Manhattan distances between all functional point pairs within the venue. For a large convention center with a distance T = 200 meters, the edge weight is multiplied by only 0.85 when the walking distance is between 200 and 220 meters, avoiding the unreasonable situation of being heavily penalized at 210 meters. If the walking distance exceeds 240 meters, it is multiplied by 0.50, reflecting a strict suppression of excessively long ineffective connections. Conversely, for a small community venue with a distance T = 80 meters, the weight is multiplied by 0.85 when the walking distance is between 80 and 88 meters, and enters the third tier of multiplication by 0.50 when the distance reaches 100 meters. This is consistent with the penalty for fixed absolute distances in this scenario, but with a more rigorous mathematical logic. Through this segmentation method based on relative proportions, the attenuation coefficient is ensured to adaptively scale with the actual spatial scale of the venue. Regardless of the venue type, edges exceeding the threshold by less than 10% enjoy the same relative tolerance, fundamentally solving the scale mismatch problem between fixed absolute distance segmentation and dynamic threshold setting.

[0039] After initially constructing the spatiotemporal dynamic graph, this method dynamically updates its edge weights over time. Specifically, an asymmetric exponential decay model is used, with the decay rate parameter dynamically switching based on the venue's operational time period. Time periods include at least the pre-opening preparation period, normal operation period, and pre-closing clearing period. Within the normal operation period, it is further divided into peak and off-peak sub-periods based on the venue's real-time visitor density. The peak sub-period uses a first decay rate, and the off-peak sub-period uses a second decay rate, with the first decay rate being less than the second. This is based on the principle that user behavior during peak hours is more representative and statistically stable, and should retain longer-term memory information; while off-peak hours have relatively larger random fluctuations, and should more quickly mitigate the impact of historical data. The weight values ​​of each directed edge are weighted and accumulated in real-time using a time decay function, based on the time difference between the timestamp of the corresponding transition behavior and the current diagnostic time. It is important to note that these real-time updated edge weights are only used to reflect the dynamic relational structure of the venue under its current operational status, serving as a reference for operational status monitoring and short-term diagnosis. For the multi-objective optimization decision involving functional point layout adjustment in step S5, the edge weights used are not directly taken from the real-time refresh value at a certain instant, but rather the average edge weights of the peak period (i.e., the several hours with the highest passenger flow density) within the diagnostic period are used as the optimization benchmark weights to ensure that the optimization scheme does not produce time-sensitive oscillations due to intraday passenger flow fluctuations, thus ensuring the stability of the decision results and the feasibility of the project.

[0040] S4: Quantitative Calculation of Functional Redundancy Index and Functional Gap Index. After obtaining the mixed demand categories for each functional point and the dynamically updated spatiotemporal correlation dynamic map, step S4 performs a quantitative diagnosis of the rationality of the configuration of each functional point in the venue, and calculates the two core quantitative indicators, namely the functional redundancy index and the functional gap index.

[0041] Functional redundancy refers to the existence of multiple functional points within a venue that provide the same or highly similar functions to meet user needs, and these functional points have a high correlation with user behavior transfer, indicating that they are interchangeable in actual user usage patterns. Moderate redundancy can improve fault tolerance and user experience, but excessive redundancy means a waste of space resources, thus requiring quantitative evaluation.

[0042] The calculation process for the functional redundancy index is as follows: Using mixed demand categories as classification labels, functional points with the same mixed demand category and whose edge weights are greater than a preset redundancy threshold are searched in the spatiotemporal dynamic graph, thus forming candidate redundant subgroups. This redundancy threshold is not a fixed value, but is adaptively determined based on the average connection strength of each functional point within the venue and the venue's spatial density coefficient. For example, in a venue with dense functional points and an average connection strength of 0.6, the redundancy threshold can be set to 0.5; while in a venue with sparse functional points and an average connection strength of only 0.3, the redundancy threshold is correspondingly lowered to 0.25. This adaptive design ensures that redundancy determination is comparable across venues of different densities.

[0043] For any target functional point within a candidate redundant subgroup, extract the weights of the directed edges between it and all other functional points within the subgroup, constructing a local directed adjacency matrix. The row vectors of this matrix are then L1 normalized (each element is divided by the sum of the absolute values ​​of all elements in that row), ensuring the comparability of the relative transfer strengths from each functional point to other functional points within the subgroup. The arithmetic mean of the normalized element values ​​is calculated, and this mean is multiplied by the global clustering coefficient of the candidate redundant subgroup. The resulting product is the functional redundancy index of the target functional point. The global clustering coefficient reflects the tightness of the connections between functional points within the subgroup; it can be calculated as the ratio of the actual number of directed edges within the subgroup to the total number of possible directed edges. A functional point with a high functional redundancy index indicates a strong transfer correlation with its peers, and the entire set of peers forms a tight cluster within the venue space. Users move frequently within this cluster, and the functional point's contribution to the overall cluster's connection strength is relatively large. This suggests that the functional point may be located in an over-densified functional area. To address the issue of excessively diluted normalized arithmetic mean and consequently low scores for true redundant nodes when the size of candidate redundant subgroups is too large (e.g., the number of function points within a subgroup exceeds 10), this method further specifies that when the size of the candidate redundant subgroup M > 10, the arithmetic mean is no longer used directly. Instead, a truncated harmonic mean is used. After removing the 30% of elements with the lowest normalized weights from the row vectors, the harmonic mean is calculated for the remaining elements and then multiplied by the global clustering coefficient. This approach retains the "size penalty" effect to suppress inflated scores in large clusters while ensuring that core redundant nodes with significant local connectivity strength in the cluster still receive sufficiently high redundancy scores, thus avoiding misjudgments in large-scale, reasonable clustering scenarios.

[0044] A functional gap refers to a situation where the exciting needs of a particular functional area within a venue lack sufficient and effective complementary functional support within the venue's existing functional configuration. Exciting needs refer to users' core expectations regarding the services provided by that functional area. For example, in a dining area, a user's exciting need might be to quickly obtain hot food; in a rest area, it might be to obtain a quiet and comfortable environment for relaxation. Complementary functions refer to other functional areas that can work in conjunction with the functional area to meet users' overall needs, such as restrooms near dining areas or drinking fountains near rest areas. A larger functional gap index means that when users visit that functional area, their associated needs for complementary functions cannot be effectively met, requiring users to travel long distances within the venue or making it impossible to find suitable complementary functions at all.

[0045] The calculation of the functional gap index first requires determining the membership degree of the excitatory demand of the target functional point. This membership degree has been pre-calculated and stored in step S2. Its physical meaning is the sum of the probability values ​​of all excitatory demand categories in the fusion probability distribution of the target functional point. If the mixed demand category of a functional point is itself excitatory, its membership degree is usually high; if it is another category but has some excitatory attributes, its quantitative contribution can also be retained through probability accumulation. Therefore, this step does not need to be calculated repeatedly, and the stored result of step S2 can be directly called. After obtaining the membership degree of the excitatory demand, with the target functional point as the starting node, a breadth-first search is performed in the spatiotemporal correlation dynamic graph along the direction of the directed edge to extract a set of directed neighborhood nodes of depth N. The value of N is related to the functional area hierarchy to which the target functional point belongs in the venue space: the core functional area located on the main axis of the venue can have an N value of 3, because its surrounding functional hierarchy is richer; while the auxiliary functional area located in the corner area can have an N value of 2. The division between core and auxiliary functional areas can be predefined by the venue management based on the functional zoning plan and entered into the system configuration; alternatively, the system can automatically determine this based on the shortest walking distance between functional points and the venue's geometric center or main entrances / exits. Areas with a distance less than one-third of the overall venue spatial scale are designated as core functional areas, and the remainder as auxiliary functional areas, allowing for automatic determination of the N value when manual annotation is unavailable. Regarding the dynamic determination of the N value, when expanding layer by layer from the target functional point, if there are no nodes in the directed neighborhood of the first layer (i.e., the out-degree is 0), the expansion stops and N is directly set to 1. This situation is considered a boundary condition where dynamic calculation cannot converge, and the system immediately reverts to a simplified determination method based on the functional area hierarchy. After the neighborhood node set is extracted, nodes with a pre-defined synergistic and complementary relationship between mixed demand categories and the excitatory demand of the target functional point are selected and marked as complementary nodes. This complementary relationship is predefined by the venue's functional configuration knowledge base, which pre-enters various common functional synergistic relationships, such as synergistic entries between the dining area and restrooms, and between the mother and baby room and accessible pathways. To address the challenges of knowledge base integration in new venues, common functional collaboration knowledge, such as catering and restroom pairings, can be imported as an initial framework. Then, the venue's BIM physical layout data can be introduced. The initial collaboration weights are adjusted based on the actual spatial reachability between functional points. Complementary relationships that are spatially adjacent and have smoother circulation are given higher confidence, while those severely obstructed by walls or floors have lower weights. This generates an initial knowledge base that conforms to specific spatial constraints. Subsequently, relying on a closed-loop self-evolution mechanism, collaboration items are automatically added or deleted and their weights dynamically adjusted based on the actual user migration frequency and perception scores after the optimization plan is implemented, completing a smooth transition from static predefined to data-driven evolution.To prevent system-level errors caused by the accidental deletion of critical collaborative relationships, such as "restroom" due to short-term data noise during the closed-loop self-evolution process, this method sets a minimum retention confidence threshold for collaborative relationships. The confidence weight of any preset collaborative item must not be lower than 50% of the initial skeleton weight. If it is lower than this threshold, a manual review process is triggered before the deletion operation can be performed, ensuring the safety boundary of the knowledge base evolution.

[0046] For each marked complementary node, the joint spatial functional support of the complementary node is calculated based on its physical proximity to the target functional point in the venue space and the edge weights of the spatiotemporal dynamic graph between them. Physical proximity can be represented as the inverse normalized value of the Manhattan distance between the two on a two-dimensional grid map, while the edge weights of the spatiotemporal dynamic graph reflect the frequency of actual user movement between the two. Multiplying the two, the support of a complementary node for the target functional point depends both on its spatial proximity and on the frequency with which users utilize this complementary relationship in actual behavior. The completeness of the associated complementary functions is obtained by summing the joint support of all complementary nodes and then normalizing the results. Finally, the difference between the membership degree of the target functional point's excitement demand and the above completeness is used as the functional gap index. The larger the difference, the less sufficient the complementary support for the target functional point's excitement demand is in the current venue configuration. If the difference is negative, it indicates that the supply of complementary functions exceeds the excitement demand of the target functional point; in this case, the functional gap index is set to 0, indicating that there is no functional gap. The above truncation process ensures that the functional gap index remains in the non-negative range, facilitating the unification of dimensions and threshold settings in subsequent multi-objective optimization.

[0047] S5: Generation of layout adjustment strategy based on multi-objective optimization. Step S5 transforms the aforementioned diagnostic results into an optimization strategy that can be implemented in practice. Its goal is to search for the optimal functional point layout adjustment scheme through a multi-objective optimization algorithm under the dual constraints of the venue's modifiable space area and engineering budget.

[0048] There are two optimization objectives, which are somewhat mutually exclusive. The first objective is to maximize the match between user needs and functional configurations within the venue. This can be achieved by adding complementary functional configurations to fill gaps in the venue and adjusting the location of functional points to be closer to the target user group. The second objective is to minimize the cost of physical space renovation. This includes direct engineering costs such as wall demolition and alteration, partition relocation, equipment relocation, and electrical system upgrades, as well as indirect economic losses incurred due to venue shutdown or partial shutdown during construction.

[0049] This method employs a non-dominated sorting genetic algorithm with an elitist strategy as the multi-objective optimization solver. The algorithm's chromosome encoding scheme encodes the adjustment operation of each functional point within the venue into a binary vector, where each bit represents whether the corresponding functional point is deleted. A separate set of gene bits represents the target location coordinates of the newly added or moved functional point. During the execution of the mutation operator, this method introduces a spatiotemporal dynamic graph as a mutation constraint to ensure that the optimization scheme remains self-consistent in terms of functional coupling relationships.

[0050] The specific implementation is as follows: During the execution of the mutation operator in the genetic algorithm, the functional coupling degree between any two functional points is calculated based on the current spatiotemporal dynamic graph. This functional coupling degree is defined as the sum of the weights of the directed edges between the two functional points multiplied by the inverse ratio of their physical distance on the venue's two-dimensional grid map. The sum of the weights of the directed edges reflects the frequency of bidirectional movement between the two functional points, while the inverse ratio of physical distance ensures that the closer the two functional points are in space, the higher their functional coupling degree. This design allows the functional coupling degree to simultaneously characterize the two core dimensions of functional association strength and spatial proximity.

[0051] Regarding the physical distance between any two functional points, this method uses a static spatial distance matrix as the sole data source. This matrix is ​​constructed and stored once after the shortest path calculation for all functional points is completed in step S3. It is directly read and used during the genetic algorithm iteration process without calling the A* path search algorithm in real time. When a new functional point is generated by the mutation operation, a single A* search is performed on the new functional point and all existing functional points, and the result is incrementally written into the matrix. For the distance calculation of the new functional point, the expected physical environment after the modification of its location should be used as the evaluation benchmark. That is, the introduction of the new functional point is accompanied by the optimization and adjustment of the passability status within its surrounding preset influence range (such as a radius of 3-5 meters). The A* path search algorithm calculates the shortest reachable distance between it and existing functional points based on this expected environment, rather than being limited by the inaccessible area before the modification. This incremental writing operation belongs to the maintenance and update of the static matrix and is not included in the category of 'prohibiting real-time calls'. The calculation scope of functional coupling degree is limited to the target functional point affected by the mutation operation and its K neighboring nodes ranked in the spatiotemporal dynamic graph by out-degree and in-degree. K is a preset positive integer between 10 and 15, in order to reduce the computational complexity of a single mutation from O(N) to O(N). 2 The value is reduced to O(K·N), where N is the total number of functional points within the venue.

[0052] When the functional coupling between any two functional points exceeds a preset coupling threshold, a reverse locking constraint is applied to the binary deletion gene bits of these two functional points during the mutation operation. Specifically, if the mutation operation activates the deletion gene bit of one functional point, making that bit active and indicating the need to delete the functional point, the system automatically locks the deletion gene bit of the other functional point to an inactive state, meaning that the functional point cannot be deleted. The core logic of this constraint is that functional points with a strong complementary relationship to functional points whose functional gap index is higher than the preset gap threshold cannot be deleted simultaneously during the optimization process; otherwise, it will further worsen the existing functional gap. This gap threshold can be set by adding a standard deviation to the average gap index of all functional points in the venue. In practical applications, when the overall venue configuration is relatively complete and the gap index is generally low, this threshold is automatically at a lower level. Correspondingly, the criteria for judging high-gap functional points are more stringent, and the conditions for triggering the reverse locking constraint are more demanding, thus not excessively limiting the search space of the optimization algorithm. Conversely, when there are many functional gaps in the venue, the threshold automatically increases, and the conditions for triggering the constraint are correspondingly relaxed, giving the optimization algorithm more room for adjustment. Through the above constraint mechanism, this method ensures that the layout adjustment scheme searched by the multi-objective optimization algorithm will not sacrifice functional complementarity in exchange for space or cost optimization.

[0053] Finally, the algorithm outputs a complete set of optimization strategy parameters, which includes a list of functional points to be deleted, the types of functional points to be added and their recommended location coordinates, and the mapping relationship between the old and new locations of functional points that need to be repositioned. Venue managers can then formulate detailed renovation implementation plans based on this parameter set.

[0054] Compared to existing technologies, this solution integrates user interaction behavior, perception evaluation, and physical layout data at the data level. It enriches subjective evaluation dimensions through real-time rating collection and sentiment analysis, and employs a dynamic confidence-weighted fusion strategy to improve the stability of demand category determination and robustness under sparse data. At the graph construction level, it mines transfer probabilities based on continuous user behavior sequences, and integrates physical spatial constraints into graph edge weights using Building Information Modeling (BIM) and the A* path search algorithm, ensuring that association strength truly reflects actual movement possibilities. For data gaps, it uses a spatiotemporal graph convolutional network for completion, and regularization with physical proximity distance and historical passenger flow patterns ensures statistical consistency. An asymmetric exponential decay model dynamically adjusts the decay rate according to operating periods, allowing edge weights to be adaptively refreshed. At the diagnostic quantification level, it calculates a redundancy index by mixing demand category labels and adaptive redundancy thresholds, and integrating individual connection strength and group spatial density. It filters complementary nodes through breadth-first search and collaborative complementary relationships, and calculates completeness by comprehensively considering physical proximity and graph edge weights, ensuring the accuracy of gap assessment. At the optimization decision-making level, guided by the diagnostic results, a multi-objective genetic algorithm is used to search for layout strategies under space and budget constraints. The functional coupling degree reverse locking constraint ensures that complementary functions are not deleted at the same time, and outputs an optimization scheme that is feasible for engineering.

[0055] In another technical solution, the specific implementation method for obtaining user perception evaluation data in step S1 is further defined. Interactive touch terminals are deployed at the entrances and exits of each functional area in the venue to achieve real-time multi-dimensional perception score collection before users leave. These interactive touch terminals can be vertical all-in-one machines or wall-mounted tablets with touch displays. They are fixedly installed at the exit of each functional area to ensure that users can easily perform evaluation operations when they have completed their use of the functional area and are preparing to leave. The multi-dimensional perception score includes at least a functional importance dimension and a user experience dimension. The functional importance dimension is used to collect users' cognitive judgment of the importance of the functional area within the overall venue, while the user experience dimension is used to collect users' subjective feelings about the comfort of the environment, service efficiency, and the completeness of the facilities at that functional area. Each dimension uses a Likert scale to collect quantitative scores. The Likert scale is a commonly used psychological measurement tool, typically offering 5 or 7 levels for users to choose from. For example, "not very important" to "very important" corresponds to 1 to 5 points, or "very dissatisfied" to "very satisfied" corresponds to 1 to 7 points. Using a 5-point scale strikes a good balance between sufficient discrimination and user workload, thus being considered a preferred implementation method. By placing the evaluation terminal at the entrance and exit and guiding users to complete the evaluation before leaving, the temporal proximity of the evaluation data can be maximized, ensuring that users' impressions of the features remain clear and vivid. This avoids the memory decay and bias problems caused by users recalling questionnaires long after leaving, as in traditional methods. Furthermore, because the evaluation occurs on-site at the feature, users' understanding of the specific characteristics of the feature and their own usage experience is most accurate, resulting in highly authentic and valuable rating data.

[0056] In addition to structured scale ratings, the system also requires the collection of user-inputted text comment data on interactive touch terminals. Text comments are unstructured, free-form text information, allowing users to describe their specific feelings about features, offer suggestions for improvement, or point out existing problems in their own words. This free expression can capture rich information that structured scales cannot cover. The collected text comment data is transmitted to a sentiment analysis model deployed on a local server within the venue for real-time sentiment polarity extraction. The sentiment analysis model is a deep learning model based on natural language processing technology, typically using a pre-trained language model as the base encoder, such as a transformer-based bidirectional encoder representation model or its lightweight version. This model has been pre-trained and fine-tuned on a large amount of sentiment-polarized comment text data, accurately identifying whether the sentiment expressed in the input text is positive, negative, or neutral, and providing a corresponding quantitative score for the sentiment polarity. The reason for deploying the sentiment analysis model on a local server rather than in the cloud is to ensure the real-time nature of data processing and the security of user privacy, avoiding the privacy risks that might arise from uploading text data containing users' personal opinions to external networks. If the venue's local server has limited computing power and cannot support real-time inference of large-scale deep learning models, load reduction measures such as model quantization acceleration and lightweight student models after knowledge distillation can be adopted. Alternatively, an asynchronous processing mechanism can be used, where the evaluation interface can be closed after the text comment is submitted without waiting for the sentiment analysis results. The sentiment analysis is completed asynchronously in the background and then timestamped with the corresponding scale score to avoid affecting the user experience by waiting for the inference results on the terminal. The sentiment tendency value extracted by the sentiment analysis model is a continuous numerical value, with positive values ​​representing positive sentiment and negative values ​​representing negative sentiment. Its absolute value reflects the intensity of the sentiment. After the sentiment analysis is completed, the output sentiment tendency value and the multi-dimensional perception score data aligned with its timestamp are uniformly aggregated by the local server to the venue's data fusion center module, serving as the input data source for the second probability distribution mapping calculation in step S2.

[0057] Furthermore, the extracted sentiment tendency values ​​are time-stamp aligned with the multidimensional perception score data and used as supplementary quantification items in the user perception evaluation data for that feature. Timestamp alignment means that when each text comment is input into the sentiment analysis model, the system records the collection time of the comment and, after sentiment analysis, associates the output sentiment tendency value with the multidimensional perception score data submitted by the same user at the same time, ensuring that both belong to a complete record of the same evaluation behavior. This timestamp alignment operation is the foundation of data fusion; it guarantees the consistency of sentiment tendency values ​​and scale scores in the time dimension, allowing them to be used together as complete perception evaluation data for the same user and the same feature in subsequent steps. By adding sentiment tendency values ​​as supplementary quantification items to the user perception evaluation data, the structured data, which originally only contained numerical scores, receives information enhancement from free text. The subtle emotional information contained in the text comments is transformed into quantifiable numerical features, enriching the information dimensions of the perception evaluation data and providing more comprehensive data support for subsequent demand category determination.

[0058] By deploying interactive touch terminals at the entrances and exits of functional points and guiding users to complete multi-dimensional perception ratings before leaving, the time lag and memory bias problems caused by post-event recall in traditional questionnaires are effectively solved, enabling the collected evaluation data to truly reflect the user's immediate feelings at the moment of using the functional point. Simultaneously, by introducing sentiment analysis from text comments and aligning sentiment tendency values ​​as supplementary quantitative items with the scale scores using timestamps, unstructured text information is transformed into structured numerical features that can participate in subsequent quantitative calculations. This not only enriches the information dimensions of the perception evaluation data but also makes subjective evaluation data objectively quantifiable and computable, providing a more comprehensive and reliable data foundation for the demand category fusion and determination in step S2.

[0059] In another technical solution, the specific implementation method of generating mixed demand categories for each functional point through confidence-weighted fusion in step S2 is further defined. Two probability distributions are obtained and fused using a dynamic weighting strategy. The first probability distribution originates from the output of a demand classification model. When user interaction behavior data is input into the model, the model outputs a probability vector reflecting the likelihood of a functional point belonging to various preset demand categories after forward computation. The values ​​of each element in this probability vector are between 0 and 1, and the sum is 1. The second probability distribution is obtained based on the mapping of user perception evaluation data. Specifically, the mapping method can be to combine the multidimensional scale scores and sentiment tendency values ​​in the perception evaluation data into a feature vector, and then convert it into the corresponding probability distribution through a pre-trained classification mapping model. These two probability distributions come from two different information sources: objective behavioral data and subjective perception data. They each reflect different aspects of the functional point demand categories. Using either source alone cannot obtain a comprehensive and accurate demand judgment result; only by effectively fusing the two can they mutually verify and complement each other.

[0060] To achieve a reasonable integration, the calculation method for the dynamic confidence weighting coefficient was further defined. The first-time coverage is calculated based on the total number of samplings of user interaction behavior data from each functional point within a preset diagnostic period. The diagnostic period can be set according to the venue's operational characteristics, such as 30 or 90 days, or divided by quarter. A 30-day diagnostic period scheme achieves a good balance between data accumulation timeliness and statistical sufficiency, and is therefore considered a preferred implementation method. The higher the total number of samplings, the more sufficient the behavioral data accumulation for that functional point, and the more reliable the output of the demand classification model trained based on this data. To eliminate the impact of differences in visitor flow among different functional points on the coverage evaluation, the first-time coverage also needs to be weighted and corrected based on the average daily visitor flow density of each functional point in the venue. The correction method can be the ratio of the total number of samplings to the ratio of the average daily visitor flow of that functional point within the diagnostic period to the average daily visitor flow of all functional points. The second sampling completeness is calculated based on the ratio between the number of valid questionnaires collected for user perception evaluation data at each function point and the total number of times the evaluation interface is triggered. The total number of times the evaluation interface is triggered refers to the total number of times users enter the evaluation page on the touch terminal at the exit of that function point. This ratio reflects the quality and reliability of the subjective evaluation data. The higher the ratio, the better the user cooperation and the more representative the collected perception evaluation data is.

[0061] The corrected first-time coverage and second-sample completeness are used as two original weights. To ensure that the fused result still constitutes a valid probability distribution, before formal weighting, the two original weights are first added together to obtain a total weight. Then, each original weight is divided by this total weight to obtain two normalized final weights, the sum of which is always equal to one. In the extreme case where no user interaction data or user perception evaluation data is collected for a certain function during the diagnostic period, resulting in both original weights being zero, the two final weights are directly set to be equal, i.e., each accounting for half. Subsequently, the two normalized final weights are multiplied by the probability values ​​corresponding to each category in the first and second probability distributions, respectively. Then, the weighted results of the corresponding categories in both distributions are added together to obtain the fused probability distribution. Since the sum of the two normalized weights is always one, and both the first and second probability distributions are valid probability distributions, the fused result still satisfies the basic properties of probability distributions, i.e., each probability value is non-negative and the sum is one. This ensures that the subsequent operation of selecting the category corresponding to the highest probability is mathematically valid and rigorous. The core advantage of this dynamic weighting method lies in the fact that the weighting coefficients are adaptively determined based on the actual data collection situation of each functional point. If the interaction behavior data of a functional point is sufficiently sampled but the perception evaluation data recovery rate is low, the first probability distribution will receive a greater weight in the fusion, and vice versa. This effectively suppresses the category judgment bias caused by insufficient sampling of a certain data source on a specific functional point. Finally, the category corresponding to the highest probability in the fusion probability distribution is taken as the mixed requirement category of that functional point. This method ensures that the determination results of requirement categories among different functional points are comparable and consistent.

[0062] While generating the mixed demand categories for each functional point, it is also necessary to pre-determine the membership degree of the excitement-type demand for each functional point for subsequent functional gap index calculation. The specific determination method is as follows: A venue demand category system is pre-established, dividing the demand categories into three levels: basic demand, expected demand, and excitement demand. Excitement-type demand refers to demand categories that users do not explicitly expect, but whose fulfillment would significantly increase user surprise and satisfaction. Examples in venue scenarios include: immersive interactive experience areas in exhibition halls, VIP viewing lounges in sports stadiums, and self-service disinfection book cabinets in libraries. In the confidence-weighted fusion process of step S2, in addition to ultimately selecting the category corresponding to the highest probability as the mixed demand category, the fusion probability value of the functional point belonging to each demand category in the aforementioned demand category system is also retained. From this fusion probability distribution, the sum of the probability values ​​of the functional point belonging to all excitement-type demand categories is extracted, which is the excitement-type demand membership degree of the functional point. This membership degree takes a continuous value between 0 and 1; a higher value indicates that the functional point is perceived by the user to have a stronger excitement attribute.

[0063] By introducing a confidence-weighted coefficient dynamically calculated based on actual data sampling, the determination of demand categories can adaptively adjust the fusion weights according to the available quality of each functional point and data source, effectively overcoming the limitations of fixed-weight fusion methods in scenarios with uneven data distribution. When interactive behavior data for a certain functional point is relatively abundant but perceptual evaluation data is relatively sparse, objective behavioral data dominates the fusion process; conversely, when perceptual evaluation data is abundant but behavioral data is limited, subjective evaluation data receives a higher weight contribution. This dynamic adaptive mechanism significantly improves the stability and robustness of the mixed demand category output results, avoiding fluctuations in category determination caused by insufficient sampling of a single data source at a specific functional point, making the final output demand category labels more reliable and trustworthy.

[0064] In another technical solution, the specific implementation method for mining the transfer probability between functional points in step S3 is further defined. Its core lies in collecting the user's continuous spatial location sequence and statistically analyzing the directed transfer frequency through positioning infrastructure deployed within the venue. Bluetooth positioning base stations are low-power wireless signal transmitting devices that are evenly deployed throughout the venue, continuously emitting beacon signals containing unique identifiers. After receiving these signals, users' smartphones or other mobile devices can measure the received signal strength of each base station and use triangulation algorithms to calculate their own two-dimensional coordinates within a large spatial area. Radio frequency identification (RFID) readers are devices that automatically identify targets carrying electronic tags and obtain their unique identification codes through radio frequency signals. They are typically deployed at the entrances and exits of functional points. When a user's RFID card or wristband passes near the reader, the reader records the user's unique identifier and the timestamp of the passage. Within a preset time window, such as 7 or 14 days, a longer time window accumulates richer transfer data but reduces timeliness. Using 7 days as a sampling period achieves a good balance between statistical sufficiency and timeliness, thus it can be considered a preferred time window setting. A continuous interaction sequence can be constructed by arranging the access points in the collected continuous spatial location sequence of users in chronological order. Then, using each functional point as a node, the frequency of directed transfers from one functional point to another is counted. This frequency is directly used as the initial weight of the directed edges, reflecting the degree to which users recognize the correlation between different functional points in their actual behavior.

[0065] This technical solution also fully considers the unavoidable problem of incomplete data collection in the venue environment and proposes a data completion model based on spatiotemporal graph convolutional networks. In actual venue operation, obstacles such as large metal components, reinforced concrete walls, and elevator shafts in the building structure can shield and attenuate Bluetooth signals, or the limited density of card reader deployment can lead to blind spots in certain areas, resulting in data gaps in the collected location sequences, meaning that users lack continuous location records within a certain time period. To address this issue, this method uses a data completion model based on spatiotemporal graph convolutional networks deployed on edge computing nodes in the venue. The reason for deploying the model on edge computing nodes rather than in the central cloud is that data completion requires low latency to meet real-time requirements, and the original location data involves user privacy; processing at the edge reduces privacy risks associated with data transmission. This completion model consists of stacked graph convolutional layers and temporal convolutional layers, and is a deep learning architecture specifically designed for processing spatiotemporal graph structured data. Graph convolutional layers use an adjacency matrix constructed based on the physical layout data of the venue to capture spatial dependencies. This adjacency matrix reflects the proximity and connectivity of different functional points in the physical space. When two functional points are close to each other or have a direct connection, the value of the corresponding position in the adjacency matrix is ​​larger. This allows the graph convolutional layer to pay more attention to the information contribution of these spatially adjacent nodes when updating the feature representation of each node.

[0066] The temporal convolutional layer employs gated linear units (LRUs) to capture the long-range temporal dependencies of user activity sequences within the venue. A LRU is a time-series modeling component that selectively transmits relevant information within a sequence through a gating mechanism. This effectively captures behavioral dependencies spanning long time intervals; for example, a user's early visit to a particular feature after entering the venue may have a lasting impact on their subsequent activity choices over a longer period. This long-range dependency modeling capability is crucial for accurately understanding and predicting the complete behavioral trajectory of users within the venue. To ensure that the missing data imputed by the model does not deviate from actual physical and social patterns, this method also uses the physical proximity distance between feature points and historical visitor flow patterns during the same period as prior constraints to regularize the imputation results. The physical proximity distance constraint ensures that the imputed transfer paths are spatially reasonable; that is, the actual walking distance between two feature points should not exceed a certain reasonable range. This range can be estimated based on the total area of ​​the venue and the density of feature points. For example, in a venue with a total area of ​​10,000 square meters and 50 feature points, the reasonable distance range can be set to no more than 200 meters. Constraints based on historical passenger flow patterns during the same period ensure that the completed flow relationships remain consistent with historical statistical patterns for similar time periods, avoiding anomalous completion results that significantly deviate from historical norms. For the cold start phase of a new venue's operation, where historical passenger flow patterns for the same period are unavailable, the following alternative approach can be adopted: An initial passenger flow probability matrix is ​​formed based on the venue operator's pre-defined functional zoning plan. This involves setting prior probability values ​​based on the logical correlation strength between functional areas (e.g., a high expected correlation between dining areas and restrooms, and a low expected correlation between exhibition halls and equipment rooms), serving as temporary substitutes for regularization. Once the venue's accumulated operational data reaches a pre-defined diagnostic period (e.g., 30 days), the pre-defined values ​​are gradually replaced with actual statistical values, achieving a smooth transition from pre-defined priors to data-driven approaches. By regularizing under these two prior constraints, the missing data in the model output conforms to both the objective limitations of physical space and the historical operational patterns of the venue. Once the complete behavioral sequence data is filled in, it is transmitted back from the edge computing node to the data fusion center module of the local server via the venue's internal LAN. Here, it is combined with the perception evaluation data output by the sentiment analysis module and used together for the construction of the spatiotemporal dynamic graph and the calculation of edge weights in step S3.

[0067] By deploying infrastructure such as Bluetooth positioning base stations or RFID readers within the venue to collect continuous spatial location sequences of users, a systematic record of users' complete activity trajectories within the venue was achieved, providing a reliable data foundation for the quantitative statistics of transfer probabilities between functional points. Addressing the data gaps caused by unavoidable signal obstruction and reader blind spots in the venue environment, a filling model based on a spatiotemporal graph convolutional network was introduced. This model captures spatial dependencies through graph convolutional layers, extracts long-range temporal dependencies using gated linear units in temporal convolutional layers, and uses physical proximity distances and historical passenger flow transfer patterns as prior constraints for regularization, achieving high-quality completion of missing data. This architectural design effectively ensures the integrity and statistical consistency of behavioral sequence data, making the extracted transfer probabilities between functional points more realistic and reliable, providing accurate behavioral data support for the construction of a spatiotemporal dynamic graph.

[0068] In another technical solution, the specific implementation method of implementing accessibility constraints between functional points in step S3 by combining physical layout data is further defined. Spatial information is extracted from the Building Information Model (BIM) and a two-dimensional grid map marking accessible areas is constructed. A BIM is a data file that digitally represents the physical and functional characteristics of a building, containing precise three-dimensional spatial coordinates and geometric dimensions of each building component within the venue. This method extracts the three-dimensional spatial coordinates of each functional point from the model and simultaneously analyzes the spatial coordinate information of building components such as walls, partitions, columns, escalators, and fire escape routes. Walls are the main obstacles dividing space; partitions are non-load-bearing partitions that flexibly divide interior areas; columns are vertical support components that bear structural loads; escalators are important vertical transportation facilities connecting different floors; and fire escape routes are critical evacuation routes in emergencies. The spatial information of these components collectively determines which areas within the venue are accessible to people and which areas are insurmountable obstacles. After obtaining the above information, this method divides the venue's planar area into several grid units according to a certain spatial resolution, and labels each grid unit as a passable or impassable area based on the location information of each building component, thereby constructing a complete two-dimensional grid map. The choice of grid resolution needs to strike a balance between computational accuracy and efficiency. For example, a grid size of 0.5 meters by 0.5 meters can be selected. This size can accurately reflect the width of the passage and the boundaries of obstacles, and will not cause excessive computational load for subsequent path search due to too many grids. Therefore, it can be considered a preferred implementation method.

[0069] After the two-dimensional grid map is constructed, the A* path search algorithm is further used to calculate the shortest physically reachable path and corresponding actual walking distance between any two functional points within the venue. The A* path search algorithm is a heuristic search algorithm widely used in path planning. It evaluates the sum of the actual cost from the starting node to the current node and the heuristically estimated cost from the current node to the target node for each node to be expanded during the search process, prioritizing the expansion of the node with the lowest overall cost, thus efficiently finding the optimal path from the starting point to the destination. In this application scenario, the starting and ending nodes of the path search are the corresponding positions of the three-dimensional spatial coordinates of the two functional points on the two-dimensional grid map. The cost function in the algorithm search process includes a weighted sum of path length cost and pedestrian density cost. The pedestrian density cost is based on pedestrian density data, which can be flexibly selected from real-time collected values ​​or estimated values ​​based on historical statistics, depending on the actual computing power configuration of the computing nodes at the venue edge, to achieve a balance between path planning accuracy and system processing load. The path length cost is the sum of the number of grid cells traversed by the search path, reflecting the actual walking distance. The pedestrian density cost is the sum of the pedestrian density values ​​of each grid cell traversed by the search path at the current moment, reflecting the congestion level of the path. By introducing the pedestrian density cost into the cost function, the shortest physically reachable path searched by the algorithm is not solely based on the shortest path of geometric distance, but rather an effective shortest path that comprehensively considers both distance and walking comfort. This better aligns with the decision-making logic of users actually choosing walking routes within the venue. The weight ratio between the path length cost and the pedestrian density cost can be adjusted according to the actual operational strategy of the venue. For example, the weight of the path length cost can be set to 0.7, and the weight of the pedestrian density cost can be set to 0.3. This ratio achieves a good balance in most conventional venues and can therefore be used as a preferred set of reference weight values.

[0070] By extracting detailed spatial component information from Building Information Modeling (BIM) and constructing a two-dimensional grid map marking accessible areas, abstract architectural design data is transformed into calculable spatial constraints. Based on this, the A* path search algorithm is used to calculate the shortest physically reachable path using a weighted cost function of path length and pedestrian density. This ensures that spatial accessibility assessment considers both the physical limitations of the building structure and the comfort of passage under real-time operational conditions, effectively addressing the problem of traditional methods treating functional points as independent units while ignoring the constraints of the building's physical structure on actual user movement. By pruning inaccessible edges and segmenting and attenuating out-of-range edges, the structure of the spatiotemporal dynamic graph is effectively purified. The retained edge weights more realistically reflect the actual mobility possibilities of users under the physical spatial constraints of the building, laying a solid spatial constraint foundation for subsequent graph-based diagnostic calculations.

[0071] In another technical solution, the specific implementation method of dynamically updating the edge weights of the spatiotemporal correlation dynamic graph based on the time decay function in step S3 is further defined. An asymmetric exponential decay model is adopted, and its decay rate parameter is dynamically switched according to the venue's operating time period. The asymmetric exponential decay model is a time series weighted model. Its basic characteristic is that it assigns different weights to events occurring at different time points. Events closer to the current time have higher weights, and events farther away from the current time have lower weights, and the weight decays over time following an exponential function curve. This model is called asymmetric because its decay rate can be asymmetric in different directions of the time axis; that is, the decay in the forward time direction from the past to the present can have different characteristics than the correlation processing in the reverse time direction. In the application scenario of this method, the decay rate parameter is not a fixed constant but is dynamically switched according to the venue's operating time period. This is because the reliability and reference value of user behavior data vary significantly in different time periods. The specific time period types include the venue's pre-opening preparation period, normal operation period, and pre-closing clearing period. During the pre-opening preparation period, the venue has not yet officially received visitors, and almost no user behavior data is generated during this time. Even the small amount of data mainly consists of staff activity records, which are not representative of ordinary user behavior patterns. Therefore, the limited historical data generated during this period should be rapidly attenuated to eliminate its interference with the current diagnosis. During the pre-closing clearing period, users have gradually left the venue, and their behavior patterns differ fundamentally from those during normal operation. The data during this period also requires even faster attenuation processing. Specifically, the attenuation rate parameter for the pre-closing clearing period is set to 0.25 to ensure that special behavioral data related to leaving during the clearing period is sufficiently diluted before entering the next normal operation cycle, and does not substantially affect the statistical analysis of behavioral correlation strength during normal operation.

[0072] During normal operating hours, the operating period is further divided into peak and off-peak sub-periods based on the venue's real-time visitor density, with different decay rate parameters applied to each. Real-time visitor density can be obtained through various methods, such as gate counting systems at each entrance / exit, active device counts in Bluetooth positioning systems, or crowd density analysis modules in video surveillance systems. Its quantification is typically expressed as people per square meter. For example, a population density greater than 0.8 per square meter can be considered a peak sub-period, less than 0.3 per square meter a off-peak sub-period, and values ​​in between are considered medium-period and included in the off-peak sub-period for decay processing. During peak sub-periods, the number of users in the venue is high, and the frequency of use and behavior shifts at various functional points are also high. The resulting behavioral data is large in volume and has high statistical stability and representativeness, accurately reflecting user behavior patterns under normal operating conditions. Therefore, a smaller decay rate is used to extend the historical data retention period; for example, the decay rate parameter can be set to 0.05. During off-peak sub-periods, the number of users is relatively sparse, the amount of behavioral data is small and the random fluctuations are large. Individual abnormal behaviors have a more significant impact on the statistical results. Therefore, a larger decay rate is adopted to make historical data fade more quickly. For example, the decay rate parameter can be set to 0.15. Through this differentiated decay strategy driven by passenger flow density, the edge weights of the spatiotemporal correlation dynamic graph can be adaptively adjusted according to the actual rhythm of venue operation.

[0073] This technical solution also specifies the refresh method for the weight values ​​of each directed edge. Specifically, it refreshes the weights in real-time by weighting and accumulating the weights of each edge's corresponding transfer behavior based on the time difference between the timestamp of the transfer and the current diagnosis time, using a time decay function. Each user transfer behavior associated with each directed edge is accompanied by a timestamp, recording the precise time of the transfer. When refreshing the spatiotemporal dynamic graph, the system iterates through all historical transfer records on all directed edges, calculates the time difference between the occurrence time and the current diagnosis time for each transfer record, and then substitutes the decay rate parameter determined by the current time period type and passenger flow density into the asymmetric exponential decay function to calculate the weight contribution value of the transfer at the current time. Finally, the weight contribution values ​​of all transfer records are summed to obtain the new weight of the directed edge after the refresh. This weighted accumulation refresh method ensures that each historical transfer record receives an appropriate weight contribution according to its occurrence time and current operating status, keeping the edge weights of the entire spatiotemporal dynamic graph in the latest state adapted to the current venue operating status. It should be noted that the real-time refreshed edge weights are mainly used to reflect the dynamic relational structure of the venue under its current operating status, serving as a reference for operational status monitoring and short-term diagnosis. For the multi-objective optimization decision involving functional point layout adjustments in step S5, the edge weights used are not directly taken from the real-time refresh values ​​at a specific instant, but rather the average edge weights of the peak periods (i.e., the several hours with the highest passenger flow density) within the diagnostic period are used as the optimization benchmark weights. This ensures that the optimization scheme does not experience time-sensitive oscillations due to intraday passenger flow fluctuations, guaranteeing the stability of the decision results and the feasibility of the project.

[0074] By introducing an asymmetric exponential decay model and dynamically switching the decay rate parameter according to the venue's operating time and real-time visitor density, the problem of traditional static maps failing to reflect the impact of changes in venue operating status on the strength of functional associations is effectively solved. A smaller decay rate is used during peak hours to maintain a longer memory period for statistically representative behavioral data, while a larger decay rate is used during off-peak hours to quickly fade data with greater random fluctuations. This time-adaptive decay strategy ensures that the edge weights of the spatiotemporal dynamic map accurately reflect the true association structure of the venue under its current operating status, providing a timely data foundation for subsequent map-based diagnostic calculations.

[0075] In another technical solution, the specific calculation method of the functional redundancy index in step S4 is further defined. Candidate redundant subgroups are searched in the spatiotemporal dynamic graph using mixed demand categories as labels, and subgroup members are filtered by an adaptively determined redundancy threshold. Using mixed demand categories as labels means that only functional points with the same demand category label are considered during the search process, because the definition of functional redundancy presupposes that multiple functional points provide the same or highly similar demand-satisfying functions; only functional points of the same category have a functional substitution relationship. The specific method of searching in the spatiotemporal dynamic graph is to traverse all pairs of functional points with the same mixed demand category and check whether the weight of the directed edge between them is greater than the preset redundancy threshold. If it is greater than the threshold, the two functional points are included in the same candidate redundant subgroup. The value of the redundancy threshold needs to be adaptively determined based on the actual situation of the venue, rather than using a fixed empirical value, because the density of functional points and user activity vary greatly among different venues. The determination of the redundancy threshold includes the average connection strength of each functional point within the venue and the venue's spatial density coefficient. Average connectivity strength is the average weight of directed edges between all functional points within a venue, reflecting the tightness of the overall functional connections within the venue. Venue spatial density coefficient is the ratio of the number of functional points to the total area of ​​the venue, reflecting the density of functional points within the venue space. For example, in a venue with high functional point density and high average connectivity strength, the redundancy threshold can be set to a relatively high value, such as 0.65; while in a venue with sparse functional points and low average connectivity strength, the redundancy threshold should be lowered accordingly, such as to 0.35.

[0076] After identifying candidate redundant subgroups, the algorithm further refines the process by extracting the directed edge weights between any target functional point within a subgroup and all other functional points within that subgroup, constructing a local directed adjacency matrix. In this matrix, rows correspond to the target functional point, and columns correspond to other functional points within the subgroup. Each element in the matrix represents the weight of the directed edge from the target functional point to the corresponding column functional point. Then, the row vectors of this local directed adjacency matrix undergo L1 normalization. L1 normalization divides each element value in a row vector by the sum of the absolute values ​​of all elements in that row, ensuring that the normalized row vector contains elements between zero and one, with a total sum of one. The purpose of this normalization is to eliminate differences in the absolute values ​​of edge weights between different functional points due to variations in passenger flow, making the relative transfer intensities of functional points within subgroups of various sizes comparable.

[0077] After normalizing the row vectors, the arithmetic mean of the normalized element values ​​is calculated. It's important to note that this arithmetic mean calculation after normalization introduces a "scale penalty" effect—that is, the larger the candidate redundant subgroup, the lower the average contribution of individual functional points. For example, the average is 1 / 3 for a scale of 3 and 1 / 5 for a scale of 5. The beneficial effect of this design is that without this normalization averaging mechanism, nodes in large clusters of similar functions would generally receive higher redundancy scores due to the overall group size advantage, leading to many functional points that are not truly redundant being misjudged as redundant. In actual venue operation, large clusters of similar functions often reflect reasonable functional aggregation; the truly redundant ones are those functional points with extremely high local connectivity strength within the cluster and significant substitution relationships in user behavior. This design transforms the judgment criterion from "absolute strength" to "relative density" through arithmetic averaging, ensuring that high redundancy indices are only awarded to functional points that occupy a core position in strongly connected local structures. This effectively suppresses the systematically inflated individual scores in large clusters and improves the accuracy of redundancy identification. To address the issue of excessively diluted normalized arithmetic mean leading to undervalued scores for true redundant nodes when candidate redundant subgroups are too large (e.g., M > 10 function points within the subgroup), this method further specifies that when M > 10, the arithmetic mean is no longer used directly. Instead, a truncated harmonic mean is used—after removing the 30% of elements with the lowest normalized weights from the row vectors, the harmonic mean is calculated for the remaining elements and then multiplied by the global clustering coefficient. This approach retains the "size penalty" effect to suppress inflated scores in large clusters while ensuring that core redundant nodes with significant local connectivity strength in the cluster still receive sufficiently high redundancy scores, avoiding misjudgments in large-scale, reasonable clustering scenarios.

[0078] Based on this, the arithmetic mean is multiplied by the global clustering coefficient of the candidate redundant subgroups. The resulting product is the functional redundancy index of the target functional point. The global clustering coefficient reflects the density of all functional points within the candidate redundant subgroup in the venue space. It can be calculated as the ratio of the actual number of directed edges within the subgroup to the total number of possible directed edges within the subgroup. The closer this ratio is to one, the more complete the connections between functional points within the subgroup, and the more compact the entire subgroup forms a functional cluster in the venue space. Multiplying the normalized mean of the row vectors by the global clustering coefficient ensures that the final redundancy index reflects two dimensions of information: the mean represents the contribution of the target functional point to the connection strength relative to other functional points within the subgroup, and the global clustering coefficient represents the spatial density of the entire subgroup. The product ensures that a high functional redundancy index is obtained only when a functional point has a high relative connection strength within its subgroup and the overall spatial density of the subgroup is high, thus guaranteeing that redundancy determination has both individual and group-level distinguishing capabilities.

[0079] In another technical solution, the method for determining the completeness of the complementary functions in step S4 is further specified. A breadth-first search is used to extract the set of directed neighborhood nodes of the target functional point from the spatiotemporal dynamic graph. Breadth-first search is a graph traversal algorithm that starts from the starting node and expands the search outwards layer by layer according to the number of hops from the starting node. First, it searches all nodes within one hop, then all nodes within two hops, and so on. In this method, the target functional point is used as the starting node, and breadth-first search is performed along the direction of the directed edges. That is, only other functional points reachable from the target functional point along the directed edges are considered. This is because the direction in which users flow from the target functional point to other functional points reflects the actual direction of their behavior after visiting the target functional point, which has direct practical significance for determining which functional points may serve as complementary functional supports for the target functional point. The search depth N is a preset positive integer that determines the size of the neighborhood range. The value of N is dynamically determined based on the local functional association distribution density of the target functional point in the venue space. Specifically, taking the target functional point as the center, the system expands outward layer by layer in the spatiotemporal dynamic graph according to the number of hops of directed edges. The number of functional points contained in each layer and the average directed edge weight between all functional points in that layer and the target functional point are statistically analyzed. Starting from the first layer, the growth rate of the number of newly added functional points is calculated layer by layer. When the average growth rate of the number of newly added functional points is lower than a preset growth rate threshold after expanding two consecutive layers (preferably 15%), or when the average directed edge weight between functional points and the target functional point in each layer decays to less than 20% of the average edge weight of the entire venue, it indicates that the effective functional connections around the target functional point have become sparse. Continuing to expand will introduce a large number of weak connections or even invalid nodes. In this case, the number of expansion layers is used as the N value. Regarding the boundary conditions for dynamically determining the N value, when expanding layer by layer with the target functional point as the center, if there are no nodes in the directed neighborhood of the first layer (i.e., the directed out-degree of the functional point is zero), the expansion stops and N is directly set to 1. This situation is considered a boundary condition where the dynamic calculation cannot converge, and the system immediately reverts to a simplified judgment method based on the functional area hierarchy. If the dynamic calculation fails to converge stably due to data sparsity during the diagnostic period, a simplified determination method based on functional zone hierarchy will be implemented: the venue management will predefine core and auxiliary functional zones according to the functional zoning plan and enter the system configuration; in the absence of manual annotation, the normalized distance between the target functional point and the geometric center of the venue will be used as a reference, and areas with a distance less than one-third of the overall spatial scale of the venue will be defaulted to the core zone (N=3), while the rest will be auxiliary zones (N=2). This simplified method will only be used when the dynamic calculation fails to converge.

[0080] After extracting the directed neighborhood node set, the system further refines this set by selecting nodes that have a pre-defined synergistic and complementary relationship with the exciting needs of the target functional points, based on mixed demand categories. These nodes are then marked as complementary nodes. A synergistic and complementary relationship refers to the mutual cooperation and complementarity between two functional points at the user demand satisfaction level. In other words, when a user visits one functional point, they often generate a need for another. For example, after visiting a dining area, a user typically needs restrooms or a rest area; after visiting a ticket office, a user typically needs the ticket verification entrance. These synergistic and complementary relationships are predefined by a venue functional configuration knowledge base. This knowledge base is a structured database built based on in-depth analysis of venue types and functional configurations, and it establishes functional synergistic relationship mapping tables for different types of venues. For example, in museum-type venues, the knowledge base can define synergistic and complementary relationships between exhibition halls and information desks, and between souvenir shops and packaging service counters; in stadium-type venues, it can define synergistic and complementary relationships between spectator stands and food and beverage vending points, and between changing rooms and shower facilities. By querying the knowledge base for the ideal complementary functional categories corresponding to the exciting requirements of the target functional points, and then matching them with the mixed requirement categories of each node in the neighboring node set, the selection of complementary nodes can be completed. To prevent system-level errors caused by the accidental deletion of key collaborative relationships (such as "restroom") due to short-term data noise during the closed-loop self-evolution process, this method sets a minimum retention confidence threshold for collaborative relationships. The confidence weight of any preset collaborative entry must not be lower than 50% of the initial skeleton weight. If it is lower than this threshold, a manual review process is triggered before the deletion operation can be performed, ensuring the safety boundary of the knowledge base evolution.

[0081] After marking all complementary nodes, the spatial functional joint support of each complementary node is calculated based on its physical proximity to the target functional point within the venue space and the spatiotemporal relational dynamic graph edge weights between them. The joint support of all complementary nodes is then summed and normalized to obtain the completeness of the associated complementary functions. Physical proximity can be represented by the reciprocal of the actual walking distance between the two on the two-dimensional grid map; the closer the distance, the higher the physical proximity. The spatiotemporal relational dynamic graph edge weights reflect the actual frequency of user movement between the two; a higher frequency indicates a greater tendency for users to use both in conjunction. Multiplying the physical proximity by the graph edge weights yields the spatial functional joint support of the complementary node. This calculation considers both the spatial accessibility of the complementary function and the strength of the behavioral association. After calculating the joint support of all complementary nodes, all values ​​are summed and then normalized. The normalization benchmark can be the theoretically maximum possible support, such as the sum of all ideal complementary functions around the target functional point with the highest support under an ideal configuration. The normalized completeness value ranges from zero to one. The closer it is to one, the more complete the existing complementary functions in the venue are, and the closer it is to zero, the less complete the complementary functions are.

[0082] In another technical solution, the multi-objective optimization algorithm in step S5 employs a non-dominated sorting genetic algorithm with an elitist strategy, introducing a collaborative constraint mechanism based on functional coupling degree in its mutation operator. Functional coupling degree is defined as the sum of the weights of the directed edges between two functional points multiplied by the inverse ratio of their physical distance, simultaneously representing the strength of functional association and spatial proximity. The physical distance is directly read from a pre-stored static spatial distance matrix, prohibiting real-time calls to the A* path search algorithm during iteration. This matrix is ​​constructed and stored once in step S3, and only when a new functional point is added is a single A* search performed, with the result incrementally written. This incremental operation is not included in the "prohibition of real-time calls" category; after writing, the distance value is frozen until the next offline calibration cycle. The calculation range of functional coupling degree is limited to the target functional point and its top K out-degree / in-degree neighboring nodes (K is 10 to 15), reducing the complexity of a single mutation from O(N0) to O(N1). 2 ) decreased to O(K·N).

[0083] When the coupling degree between any two functional points exceeds an adaptively determined coupling threshold, three types of collaborative constraints are applied: In deletion operations, if the gap index of a functional point is higher than a preset gap threshold and it is strongly coupled with another functional point, the latter's deleted gene site is locked in an inactive state (forced retention); in movement operations, functional points with high gaps are allowed to move towards coupled functional points to improve accessibility, while reverse movement is prohibited; in addition operations, the mutation is retained only if the addition location is within a preset neighborhood of a functional point with high gaps and the addition category has a collaborative complementary relationship with its excitatory demand. These constraints limit the search space to a feasible region that conforms to the spatial behavior coupling rules.

[0084] The optimized strategy parameter set must be verified in a closed-loop digital twin simulation environment. Specifically, the following steps are taken: using the current spatiotemporal dynamic graph as the initial state, the link prediction model is used to simulate the redistribution of graph edge weights after the implementation of the optimized strategy parameter set, and the redundancy index and gap index of each functional point are recalculated based on the redistributed graph. If the diagnostic indicators after the simulation update meet the preset convergence conditions, the current optimized strategy parameter set is confirmed as the final output. If the convergence conditions are not met, the deviation information of this simulation is fed back to the multi-objective optimization algorithm as a taboo constraint to drive the algorithm to correct the strategy parameters in the new round of search. The corrected parameter set then re-enters the digital twin simulation verification, forming an iterative closed loop of "optimization-simulation-feedback". To ensure the engineering convergence under extreme data conditions, a maximum number of backtracking iterations is preset (e.g., but not limited to 50 times). If the number of backtracking iterations exceeds this threshold, the iteration loop is forcibly terminated, and the feasible solution with the highest comprehensive score of the digital twin simulation evaluation index in the current Pareto front is selected as the compromise optimized strategy parameter set output. At the same time, a manual review alarm is triggered to the venue management terminal. The final output strategy parameter set includes a list of function points to be deleted, the types and recommended locations of newly added function points, and the mapping relationship between the old and new locations of function points whose locations need to be adjusted. This allows managers to formulate a hierarchical update architecture after physical modifications to the construction blueprint: between two offline calibrations, only the decay rate parameter of the time decay function is adjusted online, while the map topology and static spatial distance matrix remain unchanged. After the project is completed and normal operation resumes beyond the calibration cycle, the demand classification model and the spatiotemporal dynamic map are updated offline incrementally on a quarterly or semi-annual basis. The demand classification model uses an elastic weight consolidation method for incremental learning, employing the Fisher information matrix as a parameter importance measure, and applying a secondary penalty to the update magnitude to maintain the classification ability for old samples.

[0085] The long-term calibration mechanism synchronously feeds back and optimizes the variation constraint rules: The system monitors the actual effect of three types of operations in each calibration cycle. For new operations, the collaborative confidence weight is adjusted based on the transfer frequency and perception score; for movement operations, the allowable movement range is adjusted based on the gap index change and transfer attenuation; for deletion operations, the subsequent utilization rate of the function points retained due to reverse locking is monitored. If it is continuously lower than the minimum threshold, it is fed back to the redundancy threshold for adaptive calculation, releasing inefficient space while maintaining the complementary relationship.

[0086] The aforementioned layered update architecture enables a long-term evolutionary cycle of data collection, diagnosis, simulation pre-verification, optimization implementation, and effect monitoring. The graphs, models, and constraint rules are gradually optimized with the calibration cycle, and the accuracy of diagnosis and the effectiveness of strategies are gradually improved with the operation time.

[0087] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0088] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A venue function diagnosis method based on spatiotemporal correlation dynamic graph, characterized in that, include: S1. Obtain user interaction behavior data, user perception evaluation data, and physical layout data of multiple functional points within the target venue; S2. Input the user interaction behavior data into the pre-trained demand classification model to obtain the objective demand category of each function point, and combine it with the category determined by the user perception evaluation data to generate the mixed demand category of each function point through confidence weighted fusion. S3. Based on the continuous behavioral sequence data generated by users during their movement within the venue, the transfer probability between functional points is mined, and the accessibility constraints between functional points are combined with the physical layout data to construct a spatiotemporal correlation dynamic graph reflecting the coupling relationship between venue space and behavior. The edge weights representing the correlation strength between functional points in the spatiotemporal correlation dynamic graph are dynamically updated according to the time decay function. S4. Based on the mixed demand categories and the dynamically updated spatiotemporal correlation dynamic map, the functional redundancy index and functional gap index of each functional point are quantitatively calculated. The functional gap index is determined by calculating the difference between the membership degree of the excitatory demand of the target functional point and the completeness of its associated complementary functions in the spatiotemporal correlation dynamic map. The completeness of the associated complementary functions is determined based on the joint determination of the spatial coupling degree and functional correlation degree of the functional point in the venue space. S5. With the optimization objectives of maximizing the matching degree of user needs and functions within the venue and minimizing the cost of physical space renovation, under the constraints of the area of ​​the venue that can be renovated and the project budget, a multi-objective optimization algorithm is used to search for the set of optimization strategy parameters for adjusting the layout of functional points, and the set of optimization strategy parameters is output.

2. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 1, characterized in that, In step S1, user perception evaluation data is obtained, specifically as follows: Interactive touch terminals are deployed at the entrances and exits of various functional points in the venue to collect multi-dimensional perception score data of users on the current functional point before leaving the venue. The multi-dimensional perception score includes at least the functional importance dimension and the user experience dimension, and each dimension is quantified using a Likert scale. Furthermore, text comment data input by users on the interactive touch terminal is collected, and a sentiment analysis model deployed on the local server of the venue is used to extract the sentiment polarity of the text comment data in real time. The extracted sentiment tendency value is then aligned with the multidimensional perception score data after being timestamped, and used as a supplementary quantitative item for the user perception evaluation data of this function point.

3. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 1, characterized in that, In step S2, a mixed requirement category for each functional point is generated through confidence-weighted fusion, specifically as follows: Obtain the first probability distribution corresponding to the objective demand category output by the demand classification model, and the second probability distribution obtained based on the user perception evaluation data mapping; The first time coverage is calculated based on the total number of samplings of user interaction behavior data at each functional point within the preset diagnostic period. The second sampling completeness is calculated based on the ratio of the number of valid questionnaires collected to the total number of questionnaires distributed for user perception evaluation data at each functional point. The first time coverage is then weighted and corrected based on the average daily passenger flow density at each functional point of the venue. The corrected first-time coverage and second-sample completeness are used as two original weights. First, the two are added together to obtain a total weight value. Then, each original weight is divided by this total weight value to obtain two normalized final weights. If the sum of the two is zero, then the two are directly assigned equal default weights; Using the normalized final weights as weighting coefficients, the first probability distribution and the second probability distribution are linearly weighted and combined to obtain the fusion probability distribution of each function point belonging to different demand categories, and the category corresponding to the highest probability is taken as the hybrid demand category.

4. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 1, characterized in that, In step S3, the transition probability between the mining function points is specifically as follows: Get the continuous spatial location sequence of the user within a preset time window, collected by Bluetooth positioning base stations or radio frequency identification card readers deployed in the venue. Sort the multiple functional points visited by the user in the continuous spatial location sequence according to time to construct a continuous interaction sequence. Treat each functional point as a node, count the frequency of directed transfers from one functional point to another, and use the frequency of directed transfers as the initial weight of the directed edge. For data gaps in the interaction sequence caused by location signal obstruction or card reader blind spots within the venue, a data completion model based on a spatiotemporal graph convolutional network deployed on computing nodes at the venue edge is used. The completion model consists of stacked graph convolutional layers and temporal convolutional layers. The graph convolutional layers use an adjacency matrix constructed based on the venue's physical layout data to capture spatial dependencies between functional points. The temporal convolutional layers use gated linear units to capture long-range temporal dependencies of user activity sequences within the venue. The physical proximity distance between functional points and the historical venue passenger flow transfer patterns during the same period are used as prior constraints to regularize the completion results.

5. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 4, characterized in that, The accessibility constraints between functional points based on the physical layout data are specifically as follows: The physical layout data includes the building information model data of the venue. The three-dimensional spatial coordinates of each functional point are extracted from the building information model, and the spatial coordinate information of building components such as walls, partitions, columns, escalators and fire lanes is analyzed to construct a two-dimensional grid map marked with passable and impassable areas. Using the A* path search algorithm on the two-dimensional grid map, with the three-dimensional spatial coordinates of functional points as the start and end nodes, the shortest physical reachable path and the corresponding actual walking distance between any two functional points in the venue are calculated. The cost function in the path search process includes a weighted sum of path length cost and passing pedestrian density cost. The passing pedestrian density cost uses the estimated passenger flow density of each passage area in the corresponding time period as the input parameter. The estimated passenger flow density includes at least one of the following: an estimated value determined based on the historical average value of the same period or an estimated value determined based on the operator's preset time period benchmark table. Directed edges between two functional points that have no physically reachable path are directly removed from the spatiotemporal dynamic graph, and the weights of edges whose actual walking distance exceeds a preset comfortable walking threshold are multiplied by a decay coefficient set according to distance segments, wherein the comfortable walking threshold is dynamically set according to the venue type and the average functional area spacing within the venue.

6. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 1, characterized in that, In step S3, the edge weights representing the correlation strength between functional points in the spatiotemporal correlation dynamic graph are dynamically updated according to the time decay function, specifically as follows: The time decay function adopts an asymmetric exponential decay model, and its decay rate parameter is dynamically switched according to the time period of the venue's operation. The time period includes the venue's pre-opening preparation period, normal operation period, and pre-closing clearing period. During normal operating hours, the operating hours are further divided into peak sub-periods and off-peak sub-periods according to the real-time passenger flow density of the venue. A first attenuation rate is used in the peak sub-periods and a second attenuation rate is used in the off-peak sub-periods. The first attenuation rate is less than the second attenuation rate. The weight value of each directed edge is updated in real time after being weighted and accumulated one by one according to the time difference between the timestamp of the corresponding transition behavior and the current diagnosis time, through the time decay function.

7. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 1, characterized in that, In step S4, the calculation of the functional redundancy index is specifically as follows: Using the mixed demand category as a label, search for functional points with the same mixed demand category and whose edge weights are greater than a preset redundancy threshold in the spatiotemporal dynamic graph to form a candidate redundancy subgroup. The redundancy threshold is adaptively determined based on the average connection strength of each functional point in the venue and the venue space density coefficient. For any target functional point within the candidate redundant subgroup, the directed edge weights between it and all other functional points within the candidate redundant subgroup are extracted, a local directed adjacency matrix is ​​constructed, the row vectors of the local directed adjacency matrix are L1 normalized, the arithmetic mean of the normalized element values ​​is calculated, and the arithmetic mean is multiplied by the global clustering coefficient of the candidate redundant subgroup to obtain the functional redundancy index of the target functional point. The global clustering coefficient reflects the density of functional points within the subgroup in the venue space.

8. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 1, characterized in that, In step S4, the method for determining the completeness of the associated complementary functions is as follows: In the spatiotemporal dynamic graph, the target functional point is used as the starting node, and a breadth-first search is performed along the direction of the directed edge to extract a set of directed neighborhood nodes of depth N, where N is a preset positive integer, and the value of N is related to the functional area level to which the target functional point belongs in the venue space. Nodes that have a pre-defined synergistic and complementary relationship with the exciting demand of the target functional point are selected from the neighborhood node set and marked as complementary nodes. The pre-defined synergistic and complementary relationship is predefined by the venue functional configuration knowledge base. Based on the physical proximity of each complementary node and the target functional point in the venue space and the graph edge weights between them, the spatial-functional joint support of each complementary node is calculated, and the joint support of all complementary nodes is accumulated and normalized to obtain the completeness of the associated complementary function.

9. The venue function diagnosis method based on spatiotemporal correlation dynamic graph as described in claim 1, characterized in that, In step S5, the multi-objective optimization algorithm is a non-dominated sorting genetic algorithm with an elitist strategy, and its mutation operation is constrained by the spatiotemporal correlation dynamic graph, specifically: During the execution of the mutation operator in the genetic algorithm, the functional coupling degree between any two functional points is calculated based on the spatiotemporal correlation dynamic graph. The functional coupling degree is equal to the sum of the weights of the directed edges between the two functional points multiplied by the inverse ratio of their physical distance on the two-dimensional grid map of the venue. When the functional coupling between any two functional points is greater than a preset coupling threshold, a cooperative constraint is applied to the layout adjustment genes of these two functional points during the mutation operation: if the functional gap index of one functional point is higher than the preset gap threshold, the deletion gene site of the other functional point that has a strong coupling relationship with that functional point is locked in an inactive state to prohibit deletion; at the same time, high gap functional points are allowed to move towards their coupled functional points to improve spatial accessibility, but mutations that cause the distance between the two to increase due to reverse movement are prohibited; for the addition operation, the addition mutation is preferred only when the candidate position of the addition is within the preset neighborhood of the high gap functional point and the addition functional category has a cooperative complementary relationship with its excitatory demand.

10. The venue function diagnosis method based on spatiotemporal correlation dynamic map as described in any one of claims 1 to 9, characterized in that, In step S5, the actual walking distance between any two functional points is pre-stored as a static spatial distance matrix and directly read during the iteration of the multi-objective optimization algorithm. Real-time invocation of the A* path search algorithm is prohibited. The optimization strategy parameter set output by the multi-objective optimization algorithm needs to be verified by a digital twin simulation environment. If the digital twin simulation fails, the algorithm will revert to searching again until convergence. After the physical transformation is implemented, the demand classification model and the spatiotemporal correlation dynamic map will only be updated offline on a quarterly or semi-annual basis. Between two offline updates, only the decay rate parameter of the time decay function will be adjusted online. The map topology and the static spatial distance matrix will remain unchanged. The calculation range of the functional coupling degree during the iteration is limited to the target functional point and the top K nodes in its neighborhood, where K is a preset positive integer and its value is between 10 and 15.

Images