Urban propagation scheme generation method and device based on multi-source data and hotspot feedback

Abstract

The application discloses a kind of urban propagation scheme generation method and device based on multi-source data and hot spot feedback, which comprises the following steps: collecting city five-dimensional heterogeneous data, historical time series data and cross-cultural literature, and outputting city propagation characteristic vector after pretreatment using factor analysis and entropy weight TOPSIS;Adaptive time series prediction model is constructed to extract time series potential, comprehensive score and strategy type label;According to country, the audience is layered to perform sentiment analysis and theme modeling, obtain interest preference, dynamic emotion and regional steady-state cultural characteristics;The above-mentioned characteristics are fused as condition vector, with strategy label as control, through template derivation, component combination and three paths of large model innovation to generate candidate scheme;Combined with theme semantic alignment degree and real-time hot topic evaluation sorting, output optimal scheme and enhance generation instruction document through knowledge base retrieval, and iteratively optimized with landing feedback closed loop.The application can efficiently and accurately generate optimal propagation scheme adapted to regional culture and hot topic.

Description

Technical Field

[0001] This invention belongs to the field of urban intelligent communication, specifically relating to a method and apparatus for generating urban communication schemes based on multi-source data and hotspot feedback. Background Technology

[0002] Currently, the academic community has conducted relevant theoretical explorations on the construction and communication strategies of the international communication of city image.

[0003] In terms of construction strategies, existing approaches mainly fall into two categories: one is to condense the unconscious, intrinsic abstract core, such as urban spirit, humanistic temperament, and values; the other is to shape conscious, external, concrete symbols, including landmark buildings, cultural and tourism features, folk customs, and other audiovisual elements. Regarding communication strategies, the research emphasizes the need to deeply grasp the patterns of international hot topic communication, integrate diverse channels and elements, and systematically present the city's image. Simultaneously, differentiated strategies need to be developed based on the cultural characteristics of different countries and regions to align with the target audience's acceptance habits and discourse expectations, and to explore cross-cultural issues with broad resonance. Through the precise matching of discourse strategies and media types, the precise generation of city image communication plans can be promoted.

[0004] The generation of urban communication plans relies on the support of massive, multimodal content in a communication resource database. Current practices typically rely on task keywords, resource attributes, or category tags for retrieval, but the results still contain a large amount of poorly matched material, requiring manual verification item by item, followed by a second round of screening based on communication objectives, regional culture, and trending topics. This process is not only labor-intensive but also struggles to coordinate multi-source heterogeneous data, assess communication potential in real time, and dynamically adapt to trending topics, resulting in insufficient accuracy, a lack of iteration mechanisms, and low screening efficiency. Summary of the Invention

[0005] This invention provides a method and apparatus for generating urban communication schemes based on multi-source data and hotspot feedback, in order to solve the problems of existing urban communication scheme generation and screening processes, such as reliance on manual retrieval and matching, insufficient accuracy, lack of iteration mechanism, and low screening efficiency.

[0006] In a first aspect, the present invention provides a method for generating urban dissemination schemes based on multi-source data and hotspot feedback, the method comprising the following steps:

[0007] Collect five-dimensional heterogeneous data from the target city, including search engines, media reports, online dissemination, inbound visitors, and multilingual social media comments, as well as historical dissemination time-series data and cross-cultural dissemination literature data from various countries and regions. Perform differentiated cleaning and normalization processing to output a standardized dataset.

[0008] Input the standardized dataset, use factor analysis and the TOPSIS algorithm based on entropy weight to perform high-dimensional feature dimensionality reduction and weight quantization, and output the urban propagation feature vector;

[0009] Input the historical propagation time series data and the city's propagation feature vector, construct an adaptive time series prediction model to extract future trend features, and combine them with its time series fluctuation features for quantitative evaluation, outputting a time series potential feature vector, a city's comprehensive propagation capability score, and a strategy type label;

[0010] The audience is segmented by country and region. The multilingual social media comment dimension data and cross-cultural communication literature data are input, sentiment analysis and topic modeling are performed, and the distribution of audience interest preferences, dynamic emotional feature vectors and regional steady-state cultural feature vectors are output for each region.

[0011] The city's communication feature vector, time-series potential feature vector, regional steady-state cultural feature vector, and audience interest and preference distribution are fused into a generation condition vector. The strategy type label is used as a control signal, and a set of regional candidate schemes with related communication strategy parameter vectors are generated in parallel through three paths: template derivation, component combination, and large model innovation.

[0012] Input the regional candidate scheme set, the dynamic emotion feature vector and the city comprehensive communication capability score, combine the topic semantic alignment degree and real-time hot spot status to perform differentiated evaluation and ranking, and output the optimal communication scheme for each region;

[0013] Input the optimal propagation scheme, and generate a corresponding scheme decision explanation document through knowledge base retrieval enhancement;

[0014] The feedback data from the landing site is used as incremental training samples and fed back into the adaptive time series prediction model, feature fusion weights, and scheme generation model to form an adaptive optimization closed loop.

[0015] Optionally, the differential cleaning and normalization process includes the following steps:

[0016] Max-min value normalization is performed on the numerical indicators in the five-dimensional heterogeneous data to eliminate dimensional differences;

[0017] The text data in the five-dimensional heterogeneous data is mapped into fixed-dimensional semantic vectors through a pre-trained multilingual model;

[0018] Missing values ​​in the five-dimensional heterogeneous data are filled using time series interpolation or mean.

[0019] Outliers in the five-dimensional heterogeneous data were identified using a box plot method and replaced with the median.

[0020] Optionally, the input of the standardized dataset, the use of factor analysis and the TOPSIS algorithm based on entropy weights for high-dimensional feature dimensionality reduction and weight quantization, and the output of the city propagation feature vector include the following steps:

[0021] The standardized dataset is subjected to the Kaiser-Meyer-Olkin test and the Bartlett test of sphericity. If the test passes, factor analysis is performed to extract common factors whose cumulative variance contribution rate reaches a preset threshold, and the dimensionality-reduced factor score matrix is ​​obtained.

[0022] An evaluation decision matrix is ​​constructed based on the factor score matrix, and the evaluation decision matrix is ​​then dimensionless to obtain a standardized decision matrix.

[0023] The information entropy of each evaluation indicator is calculated based on the standardized decision matrix, and the entropy weight of each evaluation indicator is calculated based on the information entropy.

[0024] Based on the standardized decision matrix, the positive ideal solution vector and the negative ideal solution vector are determined, and the Euclidean distance between each evaluation object and the positive ideal solution vector and the negative ideal solution vector are calculated respectively.

[0025] The relative proximity of each evaluation object is calculated based on the Euclidean distance, and the relative proximity is used as the weight quantification result. The factor score matrix and the weight quantification result are then concatenated to obtain the urban propagation feature vector.

[0026] Optionally, the formula for calculating the information entropy of each evaluation index is as follows:

[0027]

[0028] In the formula: Indicates the first Information entropy of each evaluation indicator Indicates the number of evaluation objects. Indicates the first The evaluation object is in the first The weighting values ​​under each evaluation indicator ;

[0029] The formulas for calculating the entropy weights of each evaluation index are as follows:

[0030]

[0031] In the formula: Indicates the first The entropy weight of each evaluation indicator, Indicates the number of evaluation indicators;

[0032] The formula for calculating the relative proximity is as follows:

[0033]

[0034] In the formula: Indicates the first The relative similarity of the evaluation objects Indicates the first The Euclidean distance between each evaluation object and the positive ideal solution vector. Indicates the first The Euclidean distance between each evaluation object and the negative ideal solution vector.

[0035] Optionally, the step of inputting the historical propagation time-series data and the city's propagation feature vector, constructing an adaptive time-series prediction model to extract future trend features, and combining them with its time-series fluctuation features for quantitative evaluation, and outputting a time-series potential feature vector, a city's comprehensive propagation capability score, and a strategy type label includes the following steps:

[0036] The historical propagation time series data is input into the adaptive time series prediction model. The adaptive time series prediction model adaptively selects a long short-term memory network or a Transformer encoder as the backbone network based on the urban development stage identifier in the urban propagation feature vector, and outputs a propagation trend prediction sequence for a future preset time window.

[0037] The growth slope, volatility variance, and peak potential coefficient are calculated based on the propagation trend prediction sequence. The growth slope, volatility variance, and peak potential coefficient are then concatenated to obtain the time series potential feature vector.

[0038] Based on the historical propagation time series data, the fluctuation coefficient and skewness coefficient are calculated to form a static measurement vector;

[0039] The city's comprehensive communication capability score is obtained by weighting and fusing the urban communication feature vector, the temporal potential feature vector, and the static measurement vector.

[0040] Based on the preset threshold range of the city's comprehensive communication capability score, the strategy type label is determined, which includes resource accumulation type label or fission explosion type label.

[0041] Optionally, the formula for calculating the peak potential coefficient is as follows:

[0042]

[0043] In the formula: Indicates the peak potential coefficient. This represents the maximum predicted value in the propagation trend prediction sequence. This represents the mean of the historical propagation time series data. This represents the standard deviation of the historical propagation time series data;

[0044] The formula for calculating the city's comprehensive communication capability score is as follows:

[0045]

[0046] In the formula: This indicates the city's overall communication capability score. This represents the urban propagation feature vector. This represents the time-series potential feature vector. Represents the static measure vector, Represents the L2 norm. The preset static feature weight coefficients, time series potential weight coefficients, and volatility measure weight coefficients are given. .

[0047] Optionally, the step of stratifying the audience by country / region, inputting the multilingual social media comment dimension data and the cross-cultural communication literature data, performing sentiment analysis and topic modeling, and outputting the distribution of audience interests and preferences, dynamic sentiment feature vectors, and regional steady-state cultural feature vectors for each region includes the following steps:

[0048] Language identification and translation alignment were performed on the multilingual social media comment dimension data to obtain a set of comment texts in a unified semantic space;

[0049] Fine-grained sentiment analysis is performed on the set of comment texts to obtain the sentiment polarity score and sentiment category probability distribution for each comment text;

[0050] Based on the emotional polarity score and the probability distribution of emotional categories, the mean emotional value, variance of emotional fluctuation, coefficient of emotional mutation and clustering degree of negative hot topics in each region are calculated, and the dynamic emotional feature vector is obtained by splicing them together.

[0051] Latent Dirichlet Allocation (LDA) topic modeling was performed on the cross-cultural communication literature data to extract high-frequency keywords and their probability distributions for each region, thereby obtaining the steady-state cultural feature vector of the region.

[0052] The frequency of audience interaction with different types of content in each region was statistically analyzed, and the audience interest preference distribution was obtained after normalization.

[0053] Optionally, the formula for calculating the clustering degree of negative trending topics is as follows:

[0054]

[0055] In the formula: Indicates the concentration of negative trending topics. Indicates the number of time slices within the time window. Indicates the first The percentage of negative emotional comments within a given time period. Indicates the time decay weight. The preset attenuation coefficient;

[0056] The formula for calculating the emotional abrupt change coefficient is as follows:

[0057]

[0058] In the formula: Indicates the emotional mutation coefficient. This represents the average sentiment in the current time slice. This represents the average mood of the previous time segment. The standard deviation of the historical sentiment mean.

[0059] Optionally, the step of fusing the city's communication feature vector, temporal potential feature vector, regional steady-state cultural feature vector, and audience interest and preference distribution into a generation condition vector, using the strategy type label as a control signal, and generating a regional candidate scheme set of related communication strategy parameter vectors in parallel through three paths: template derivation, component combination, and large model innovation, includes the following steps:

[0060] The urban communication feature vector, the temporal potential feature vector, the regional steady-state cultural feature vector, and the audience interest and preference distribution are spliced ​​and fused into a generation condition vector, and the strategy type label is encoded into a control signal vector.

[0061] Based on the template derivation path, the cosine similarity between the generated condition vector and the condition vectors of each template in the historical successful case template library is calculated. The best matching template that satisfies the strategy type constraint is selected for parameter replacement and local fine-tuning to generate the first candidate solution.

[0062] Based on the component combination path, the propagation target is decomposed into reusable components. With the goal of maximizing the combination score, a greedy strategy combined with a local search algorithm is used to select the optimal component combination and automatically splice it to generate a second candidate scheme.

[0063] Based on the aforementioned large-scale model innovation path, the generated condition vector and the control signal vector are decoded into natural language constraint instructions, driving the large language model to generate an innovative solution framework and generate a third candidate solution.

[0064] The first, second, and third candidate schemes are deduplicated and format-standardized to obtain the regional candidate scheme set, and each candidate scheme is bound to a propagation strategy parameter vector.

[0065] Optionally, the input of the regional candidate scheme set, the dynamic sentiment feature vector, and the city's comprehensive communication capability score, combined with the topic semantic alignment and real-time hotspot status for differentiated evaluation and ranking, and outputting the optimal communication scheme for each region includes the following steps:

[0066] Extract the topic distribution from the propagation strategy parameter vector of each candidate scheme in the regional candidate scheme set, extract the topic sentiment mean vector from the dynamic sentiment feature vector, calculate the cosine similarity between the topic distribution and the topic sentiment mean vector, and obtain the topic semantic alignment.

[0067] The current real-time hot topic status is determined based on the mean emotion value in the dynamic emotion feature vector and the clustering degree of negative hot topics. The real-time hot topic status includes a negative hot topic clustering state, a positive hot topic surge state, and a hot topic stable state.

[0068] The corresponding evaluation correction coefficient is determined based on the real-time hotspot status.

[0069] Based on the topic semantic alignment, the city's comprehensive communication capability score, and the evaluation correction coefficient, the adaptation score of each candidate scheme is calculated. The candidate schemes are then sorted in descending order of their adaptation scores, and the candidate scheme with the highest score is selected as the optimal communication scheme.

[0070] Optionally, the formula for calculating the adaptation score is as follows:

[0071]

[0072] In the formula: Indicates the fit score. Indicates the semantic alignment of the topics. This indicates the city's overall communication capability score. This represents the evaluation correction factor. , These are the preset topic alignment weights and city capability weights, respectively. Adjust the weights for preset hotspots. This indicates the coverage of the candidate solution with the distribution of audience interests and preferences;

[0073] The evaluation correction coefficient The segmentation and value selection rules are as follows:

[0074] When the real-time hot topic status is a state of negative hot topics clustering. ;

[0075] When the real-time hotspot status is a state of positive popularity surge. ;

[0076] When the real-time hot topic status is a stable hot topic state. ;

[0077] in, , These are the preset negative penalty coefficient and positive incentive coefficient, respectively. For negative trending topics, This represents the emotional mutation coefficient.

[0078] Optionally, the step of inputting the optimal propagation scheme and generating a corresponding scheme decision description document through knowledge base retrieval includes the following steps:

[0079] Construct an urban communication knowledge base, which includes a multi-dimensional communication indicator library, a comprehensive publicity and operation case library, a communication effect evaluation model library, and an external communication experience and rule library;

[0080] The joint retrieval query vector is formed by concatenating the dissemination strategy parameter vector of the optimal dissemination scheme, the urban dissemination feature vector, the audience interest and preference distribution of the corresponding region, the dynamic emotional feature vector, the regional steady-state cultural feature vector, and the strategy type label.

[0081] Based on the joint retrieval query vector, feature similarity retrieval is performed in the urban communication knowledge base to match quantitative support data, similar implementation cases and reasoning logic.

[0082] The system automatically renders a pre-defined standardized report template and outputs a solution decision description document that includes the core content of the solution, decision-making logic, expected audience acceptance, risk warnings, and implementation suggestions.

[0083] Optionally, the step of using the landing feedback data as incremental training samples to feed back into the adaptive time-series prediction model, feature fusion weights, and scheme generation model to form an adaptive optimization closed loop includes the following steps:

[0084] Collect data on actual dynamic changes in sentiment, shifts in audience interests and preferences, and deviations in key communication indicators in each region after the implementation of the current round of plans at fixed intervals.

[0085] A loss function is constructed based on the actual dynamic emotional changes, audience interest and preference shifts, and key communication indicator deviations.

[0086] The incremental training data is fed back to the adaptive temporal prediction model to update network parameters, the feature fusion weight adjustment factor analysis and entropy weight parameters, and the embedding layer and attention mechanism parameters of the conditional encoder in the scheme generation model for fine-tuning.

[0087] The update is paused when the loss decreases by less than the preset convergence threshold after two consecutive iterations to prevent overfitting.

[0088] Secondly, the present invention also provides an apparatus for generating urban communication schemes based on multi-source data and hotspot feedback, the apparatus comprising:

[0089] The data preprocessing module is used to collect five-dimensional heterogeneous data and historical dissemination time series data of the target city, and simultaneously collect cross-cultural dissemination literature data of various countries and regions. It performs differential cleaning and normalization processing on the five-dimensional heterogeneous data and outputs a standardized dataset.

[0090] The urban propagation feature modeling module is used to input the standardized dataset, perform high-dimensional feature dimensionality reduction and weight quantization using factor analysis and the TOPSIS algorithm based on entropy weight, and output the urban propagation feature vector.

[0091] The temporal potential assessment module is used to input the historical propagation time series data and the city's propagation feature vector, construct an adaptive temporal prediction model to extract future trend features, and perform quantitative assessment in combination with its temporal fluctuation features, outputting a temporal potential feature vector, a city's comprehensive propagation capability score, and a strategy type label;

[0092] The hot topic sentiment calculation module is used to segment the audience by country and region. It takes the multilingual social media comment dimension data and the cross-cultural communication literature data as input, performs sentiment analysis and topic modeling, and outputs the audience interest preference distribution, dynamic emotional feature vector and regional steady-state cultural feature vector for each region.

[0093] The candidate scheme generation module is used to integrate the urban dissemination feature vector, the temporal potential feature vector, the regional steady-state cultural feature vector, and the audience interest and preference distribution into a generation condition vector. Using the strategy type label as the control signal, it generates a regional candidate scheme set with associated dissemination strategy parameter vectors in parallel through three paths: template derivation, component combination, and large model innovation.

[0094] The scheme evaluation and ranking module is used to take into account the regional candidate scheme set, the dynamic emotion feature vector and the city's comprehensive communication capability score, and to perform differentiated evaluation and ranking based on the topic semantic alignment degree and real-time hot spot status, and output the optimal communication scheme for each region.

[0095] The decision explanation document generation module is used to take the optimal propagation scheme as input and generate a corresponding decision explanation document through knowledge base retrieval and enhancement.

[0096] The iterative optimization control module is used to feed the landing feedback data back as incremental training samples to the adaptive time series prediction model, feature fusion weight and scheme generation model to form an adaptive optimization closed loop.

[0097] The beneficial effects of this invention are:

[0098] This study collects five-dimensional heterogeneous data, historical dissemination time-series data, and cross-cultural dissemination literature data of the target city. After differential cleaning and normalization, a standardized dataset is obtained. Factor analysis and the TOPSIS algorithm with entropy weight are used for dimensionality reduction and quantization, outputting a city dissemination feature vector. An adaptive time-series prediction model extracts trend features, outputting a time-series potential feature vector, a city comprehensive dissemination capability score, and strategy type labels. Layered audience sentiment analysis and topic modeling are conducted, outputting audience interest preferences, dynamic emotional characteristics, and regional stable cultural characteristics. Multi-dimensional features are integrated into a generation condition vector, with strategy type labels as control signals, to generate regional candidate scheme sets in parallel through three paths. The optimal dissemination scheme is obtained by combining topic semantic alignment and real-time hot topic status evaluation and ranking. A scheme decision-making explanation document is generated through knowledge base retrieval. An adaptive optimization closed loop is formed through implementation feedback. Based on the above methods, a city dissemination endowment profile can be constructed through multi-source data modeling, proactively mining dissemination potential and automatically matching strategy types, generating differentiated candidate schemes in parallel, dynamically completing accurate screening and automatic document generation in conjunction with hot topics, and achieving continuous optimization across the entire chain through feedback, outputting an optimal dissemination scheme adapted to regional culture and hot topics. Compared to manual production models, this approach significantly improves the accuracy and responsiveness of urban international communication plans, reduces the human resource costs of cross-cultural analysis and hot topic control, and builds a self-optimizing long-term communication operation loop. Attached Figure Description

[0099] Figure 1 This is a flowchart illustrating the urban communication scheme generation method based on multi-source data and hotspot feedback in this invention.

[0100] Figure 2 This is a schematic diagram of the urban communication scheme generation device based on multi-source data and hotspot feedback in this invention. Detailed Implementation

[0101] This invention discloses a method for generating urban communication schemes based on multi-source data and hotspot feedback.

[0102] Referring to Figure 1, the method for generating urban communication schemes based on multi-source data and hotspot feedback specifically includes the following steps:

[0103] S1. Collect five-dimensional heterogeneous data from the target city, including search engines, media reports, online dissemination, inbound visitors, and multilingual social media comments, as well as historical dissemination time-series data and cross-cultural communication literature data from various countries and regions. Perform differential cleaning and normalization processing to output a standardized dataset.

[0104] Specifically, this step is used to complete the compliant collection and unified standardization of five types of dissemination data across the entire domain, stable cultural documents from different countries and regions, and long-term historical time-series data. Data collection follows the principle of prioritizing mainstream domestic platforms, supplementing with publicly available international data, and using official statistical data as the basis, thus avoiding data compliance risks throughout the process.

[0105] First, an independent data collection channel was established to collect data from search engine popularity, online dissemination across the entire internet, authoritative media reports, inbound visitor data, multilingual social media comments, historical dissemination time-series data, and original data from cross-cultural communication documents from various countries and regions. This process was used to construct the original dataset, which is mathematically expressed as follows:

[0106]

[0107] In the formula: Represents the original dataset across the entire domain; This is the raw data for search engine popularity; To disseminate the original data across the entire network; Original data for authoritative media reports; For raw data of inbound visitors; Raw data for multilingual social media comments; This is the original data for historical dissemination timelines; This data comprises original documents on cross-cultural communication across countries and regions. The search engine popularity data primarily collects search frequency of city and related cultural tourism keywords, platform PV and UV metrics, and temporal fluctuation characteristics of popularity across all time periods. Popularity growth rate, popularity volatility, and peak popularity coefficient are calculated using daily, weekly, and monthly statistical granularities, and multi-year time series data are concatenated to construct a popularity feature matrix. Search volume is normalized to its maximum and minimum values. The standardized search popularity index calculation formula is as follows:

[0108]

[0109] In the formula: This represents the original search popularity index value; This represents the minimum value of the sample within the statistical period for this indicator. This represents the maximum value of the sample within the statistical period for this indicator. This is the search popularity index output after normalization.

[0110] The data collected includes information from the entire network, short video exposure, total reposts, average dissemination depth, and the scale of cross-regional coverage nodes. Preferably, by analyzing user behavior sequences, invalid samples such as paid commenters and automated click fraud are identified and removed. After removal, the overall dissemination index is calculated, and the network dissemination diffusion coefficient D is defined.

[0111]

[0112] In the formula: Represents the network propagation and diffusion coefficient; Represents the total number of content forwarded across the entire network; This represents the average depth of content dissemination. Authoritative media reports include data from central media, local media, and overseas partner media. Weighted statistics are applied based on the media's hierarchical authority, including total reporting volume, the proportion of positive and negative sentiment, monthly increase in media volume, and the issue's persistence coefficient. The data also summarizes the media's positive index and negative risk index, thus vectorizing the issue distribution.

[0113] Inbound visitor data is obtained from monthly sampling data from the cultural and tourism and statistics departments, including statistics on the number of inbound visitors, the geographical distribution of visitors, the frequency of mobile device use, and the average length of stay. Calculate the standardized visitor activity index:

[0114]

[0115] In the formula: , The normalized balance coefficient is T, representing the total number of days in a single statistical period. Location entropy is used to quantify the concentration of customer origin by region. Multilingual social media comments are collected in Chinese, English, Japanese, Korean, Arabic, and other languages, with the language set dynamically adjusted to reflect the target city's international customer origin structure. Cross-cultural communication literature extracts regional cultural preferences, communication taboos, and long-term discourse characteristics. Historical communication time series aggregate all indicators daily and weekly to generate continuous volume time series samples. Overseas data sources utilize compliant third-party APIs and public snapshots for dual-track collection, with all user data aggregated anonymously. Data cleaning rules include: linear interpolation for missing time series data, and filling static indicators with the median of the same dimension. Criteria for removing outlier samples; text deduplication of all text; numerical mapping. Intervals; all types of text are semantically encoded using multilingual-BERT, outputting a standardized dataset. It should be noted that all statistical thresholds and normalization coefficients in this step are preferred values ​​in this embodiment. Those skilled in the art can adaptively adjust them according to the data distribution of the target city and the data source type, without limiting the scope of protection of this invention.

[0116] S2. Input the standardized dataset, use factor analysis and the TOPSIS algorithm based on entropy weight to perform high-dimensional feature dimensionality reduction and weight quantization, and output the urban propagation feature vector.

[0117] Specifically, based on standardized datasets Constructing a high-dimensional feature matrix Rows represent time-series samples, and columns represent individual propagation indicators. First, all normalized numerical values ​​and coded text are integrated. Factor analysis is used for dimensionality reduction, and common factors with a variance contribution rate greater than 0.80 are selected. The factor decomposition formula is:

[0118]

[0119] In the formula: Represents the original high-dimensional feature matrix; Representative factor loading matrix; Represents the common factor score matrix; This represents a special factor matrix, characterizing the residual portion of the original features that cannot be explained by the common factors. The higher the variance contribution rate, the stronger the explanatory power of the common factor for the original indicators. 0.80 is an optimal empirical threshold that balances information retention and dimensionality reduction efficiency.

[0120] Initial weights are obtained based on the proportion of factor variance. Then, the information entropy is used to correct the weights to eliminate subjective bias and obtain the optimal feature weights. .

[0121] Formula for calculating the information entropy of an indicator:

[0122]

[0123] Formula for calculating the objective weight of indicators:

[0124]

[0125] In the formula: For the first Information entropy of the evaluation indicators; Calculate the coefficients for information entropy; The number of representative evaluation indicators; Representing the The first evaluation object Dimensionless values ​​for each indicator; Representing the The evaluation object is in the first The weighting values ​​under each evaluation indicator; For the first The entropy weight of each evaluation indicator;

[0126] Obtain the objective weights of each indicator Subsequently, the dimensionality-reduced indicators and their corresponding objective weights are input into the TOPSIS model. Using each time window as the evaluation object, the relative closeness of each window to the positive and negative ideal solutions is calculated, thereby compressing the multi-dimensional propagation indicators into a unified-scale comprehensive score. Weighted fusion yields a unified-dimensional urban propagation feature vector:

[0127]

[0128] In the formula: A unified representation of the five static communication attributes of urban search, media, network, foreign affairs, and hot topics is used as a unified benchmark input for subsequent time-series modeling and scheme generation; This represents the total number of common factors. For the first Common factor scores; For the first The objective weights corresponding to the common factors.

[0129] S3. Input the historical propagation time series data and the city propagation feature vector, construct an adaptive time series prediction model to extract future trend features, and combine the time series fluctuation features extracted based on the historical propagation time series data for quantitative evaluation, and output the time series potential feature vector, the city's comprehensive propagation capability score and the strategy type label.

[0130] Specifically, time-series samples, statistically analyzed monthly for 3 to 5 consecutive years, are extracted from a standardized dataset and input into an adaptive time-series prediction model to perform potential mining. The model engineering adaptation rules are as follows: a 4-layer, 8-head Transformer time-series encoder is used when the historical time-series data storage duration is greater than or equal to 3 years; otherwise, an LSTM network is used. These two architectures are only for engineering implementation adaptation and do not change the core technology protection scheme of this invention. The model fits the multi-year popularity change curve to predict the spread popularity sequence for the next 12 to 24 months. Extract three types of features: the slope of popularity growth, annual fluctuation variance, and peak season potential, and output a time-series potential feature vector. :

[0131]

[0132]

[0133] In the formula: The slope of the heat increase; To predict the length of the time window; The historical time series sample value at time t; for Predict the sequence value at any given time. For a complete sequence of future transmission trend predictions; This serves as a historical time series baseline reference value. To predict the variance of series fluctuations; This represents the peak potential coefficient.

[0134] Calculate the historical time series fluctuation coefficient based on all historical time series data. With time series skewness coefficient :

[0135]

[0136]

[0137] In the formula: The standard deviation of historical propagation time series data; This is the mean of historical propagation time series data; This refers to the historical time series fluctuation coefficient; The skewness coefficient for historical time series data is used to characterize the degree of asymmetry in the distribution of time series data. It is a static measure vector.

[0138] Belonging to two sets of feature spaces, this invention sets two possible implementation methods. The main implementation method vertically concatenates the two types of vectors to obtain an extended feature vector. The data is then fed into the TOPSIS model to calculate the overall score. A simplified engineering approach is used, employing a linear weighted approximation. The weighting coefficients are calculated using the following formula:

[0139]

[0140] In the formula: Assess the city's overall communication capabilities. For vector fusion, dynamic weight coefficients are used; while retaining the optimal values ​​for each segment: hour , hour , hour Temporal skewness The time weight is reduced by 0.1, as mentioned above. The segmentation threshold, weighted value, and skewness determination threshold are all preferred parameters in this embodiment, and can be adaptively adjusted according to the historical propagation time sequence distribution of the city.

[0141] Scoring based on comprehensive communication capabilities Automatic tagging strategy for dividing intervals, with preset dynamic thresholds. It is obtained from historical sample clustering or quantile statistics, and is not a fixed constant. Mark resource accumulation strategy, when The marked fission trigger strategy will , one-hot encoded The output is sent to S5 as a constraint for scheme generation.

[0142] S4. Stratify the audience by country and region, input the multilingual social media comment dimension data and cross-cultural communication literature data, perform sentiment analysis and topic modeling, and output the audience interest preference distribution, dynamic emotional feature vector and regional steady-state cultural feature vector for each region.

[0143] Specifically, cross-cultural documents and multilingual social media comments are grouped and processed according to country and region. For cross-cultural communication documents, multilingual-BERT encoding is used to generate regional steady-state cultural feature vectors. It carries regional culture and disseminates taboo-related stable attributes. The device automatically counts the effective sample size of monolingual comments. When the social media sample size is lower than a preset statistical significance threshold, it uniformly converts the comments to general English text. This threshold is determined based on the central limit theorem and the representativeness requirement of the samples. The device reads the confidence score from the machine-translated output text and corrects the original sentiment score using a confidence compensation formula. The compensation calculation formula is as follows:

[0144]

[0145] In the formula: The translation confidence compensation coefficient; The benchmark compensation coefficient; Confidence score of machine-translated output text; baseline compensation coefficient The compensation coefficient is lowered when the confidence level is below 0.60, and approaches 1 when the confidence level is above 0.90. 0.60, 0.90 and 1.00 are the preferred values ​​in this embodiment, and can be dynamically adjusted according to the performance of the translation model.

[0146] The revised text is fed into a fine-grained sentiment model, and LDA topic modeling is performed. After softmax normalization, an audience interest preference distribution vector is generated. Aggregate all revised sentiment results according to a fixed weekly time window, and calculate the mean sentiment score for each theme sequentially. Variance of mood fluctuation Emotional mutation coefficient Negative Hot Topics Concentration Concatenate to generate regional dynamic emotion feature vectors The regional steady-state vector and the audience interest distribution are spliced ​​together to form a complete audience fusion feature, which is then supplied to the S5 solution generation and S6 hot topic scoring process.

[0147] S5. The city's communication feature vector, time-series potential feature vector, regional steady-state cultural feature vector, and audience interest and preference distribution are fused into a generation condition vector. The strategy type label is used as a control signal. The regional candidate scheme set of related communication strategy parameter vectors is generated in parallel through three paths: template derivation, component combination, and large model innovation.

[0148] Specifically, the S2 city dissemination vector, S3 time-series potential vector, S4 regional steady-state vector, and audience interest distribution are vertically concatenated to construct a region-specific generation condition vector:

[0149]

[0150] Convert the S3 policy tag into a one-hot encoded vector. As an additional control signal, three generation paths—template derivation, component combination, and large model innovation—are initiated in parallel. The output scheme of each path is bound to a structured propagation strategy parameter vector. Template derivation path calculation Cosine similarity with the regional historical template library is used to select matching strategy tag templates for parameter replacement and localization fine-tuning; the component combination path optimizes for maximizing matching scores, heuristically combining theme, channel, and budget components; the large model innovation path uses a Transformer architecture pre-trained language model to... , Decode into prompts with regional cultural constraints; set preset thematic dictionaries A validity check mechanism is in place if the model output topic is not in the preset dictionary. If the result is determined to be a model illusion, the system automatically reverts to the component combination path for regeneration. The three output schemes are deduplicated using a semantic similarity threshold of 0.85. 0.85 is the preferred value in this embodiment and can be adjusted according to the scheme's discriminative power. After denormalization, a candidate scheme set corresponding to the country / region is formed and sent to the S6 sorting stage.

[0151] S6. Input the regional candidate scheme set, the dynamic emotion feature vector and the city comprehensive communication capability score, combine the topic semantic alignment degree and real-time hot spot status to perform differentiated evaluation and ranking, and output the optimal communication scheme for each region.

[0152] Specifically, this step pre-constructs a complete hot topic correction rule base, extracts a set of topics for each candidate solution, and calculates the average semantic alignment with the mean of each topic within the dynamic sentiment vector. The formula for calculating the topic semantic alignment is as follows:

[0153]

[0154] In the formula: For the first Thematic semantic alignment of candidate solutions; For the first The complete set of dissemination topics corresponding to the candidate solutions. For the first The total number of topics disseminated in the candidate schemes; Themes in dynamic emotion feature vectors The corresponding thematic sentiment average.

[0155] Based on dynamic sentiment vector analysis, three types of trending scenarios are analyzed, and correction coefficients are configured differently. Threshold for determining the concentration of negative hot topics Threshold for determining the concentration of positive hotspots Threshold for judging emotional fluctuations All are preferred embodiments and can be adaptively adjusted. Negative aggregation scenario. Offline gathering-related plans will be subject to negative adjustments, while public communication-related plans will be subject to positive adjustments. In cases of extreme negative overall impact, large-scale offline event plans will be directly filtered out; positive-trend scenarios will be addressed. Furthermore, most average values ​​increased, with only the highly matched hot topic solutions showing a positive correction; the correction coefficient for hot topics remained stable at 1. The formula for calculating the overall suitability score of a single solution is as follows:

[0156]

[0157] In the formula: For the first The overall fit score of the candidate solutions; Adjustment coefficients for differentiated configurations. Weighting for topic semantic alignment; Weighting of the city's comprehensive communication capabilities score; Weighting of regional audience interest coverage; For the first Thematic semantic alignment of candidate solutions; Assess the city's overall communication capabilities. The representative scheme's coverage of the regional audience's interest distribution; scores are truncated to... The interval is sorted by score in descending order, and the optimal number of solutions are selected and sent to S7 to generate the accompanying document.

[0158] S7. Input the optimal propagation scheme and generate a corresponding scheme decision description document by retrieving and enhancing the knowledge base.

[0159] Specifically, a city-specific, comprehensive communication knowledge base will be built, divided into four sub-bases: a multi-dimensional indicator database, a comprehensive implementation case database, a communication effect evaluation database, and an external operation rule database. All stored content will adopt [the relevant standards / principles]. Uniform dimensional feature encoding ensures complete alignment of the vector space. The retrieval query vector is constructed by concatenating the optimal solution strategy parameters, city dissemination vector, regional audience fusion vector, and strategy tags.

[0160]

[0161] The FAISS approximate nearest neighbor retrieval method is adopted, and the lower similarity threshold of 0.65 is the preferred adjustable parameter in this embodiment. When the similarity between the recall vector and the query vector is lower than this threshold, the knowledge base lacks effective supporting data, and the device automatically switches to a general preset template to ensure the basic usability of the report. The retrieval simultaneously recalls the top 10 benchmarking indicators, the top 5 similar implementation cases, and the top 3 adapted dissemination rules. The retrieval data is automatically filled by retrieving the standardized government report template. The report has six fixed modules, namely: scheme overview, multi-dimensional data evidence, similar implementation case reference, implementation operation parameters, hot spot risk warning, and medium- and long-term effect prediction. The document is output in both Word and PDF formats simultaneously, and only minor manual adjustments are needed for application and review of urban cultural tourism and external publicity projects.

[0162] S8. The landing feedback data is used as incremental training samples and fed back to the condition encoder of the adaptive time series prediction model, feature fusion weight and scheme generation model to form an adaptive optimization closed loop.

[0163] Specifically, the device employs a dual-cycle scheduling mechanism of weekly light iterations and monthly full reruns. After each cycle, it automatically collects incremental feedback data generated by the implementation of solutions in each region. This incremental data includes regional dynamic sentiment shifts. Audience interest and preference shift Time-series heat prediction residual The four types of information—the implementation effect of the solution, the matching error, etc.—are integrated to construct a global loss function as the model optimization objective. The formula for calculating the loss function is as follows:

[0164]

[0165] For temporal and generative deep learning networks, a PEFT parameter fine-tuning strategy is adopted to freeze the underlying general semantic and temporal extraction layers and only fine-tune the top-level adapter and layer normalization parameters.

[0166] The optimizer selected is an adaptive moment estimation optimizer, and a base learning rate is set. To avoid drastic parameter oscillations, the parameters are decayed in each iteration. For factor analysis and TOPSIS entropy weight statistics modules, a fixed sliding window mechanism of 12 months is used to continuously update the covariance matrix and indicator information entropy, eliminating the need for full backtracking of historical samples over many years, thus balancing update efficiency and feature timeliness. For the hot topic correction rule base, the error is automatically fine-tuned based on the actual negative clustering and hot topic leap sample errors in the current round. , The correction factor is fixed and requires no manual configuration. After each round of parameter updates, the difference in loss reduction between two consecutive rounds is calculated. Preset convergence threshold , For this embodiment, adjustable parameters are preferred. If the model converges, terminate the current round of incremental training and retain the optimal set of model parameters. If it does not converge, continue backpropagation updates, setting a maximum of 20 iterations to prevent infinite loops. After model convergence, simultaneously refresh the entire knowledge base, storing all candidate solutions, the optimal solution, and deployment hotspot trajectories in the case library. Recalculate the profile vectors of all solutions in the library in batches, and simultaneously update the policy judgment thresholds for each region. The entire process forms a complete adaptive closed loop of generation, evaluation, implementation, feedback, and iteration, which can continuously adapt to the dynamic changes of audience culture, trending topics across the internet, and urban communication stages in different countries and regions.

[0167] A device for generating urban communication schemes based on multi-source data and hotspot feedback, comprising:

[0168] The data preprocessing module is used to collect five-dimensional heterogeneous data and historical dissemination time series data of the target city, and simultaneously collect cross-cultural dissemination literature data of various countries and regions. It performs differential cleaning and normalization processing on the five-dimensional heterogeneous data and outputs a standardized dataset.

[0169] The urban propagation feature modeling module is used to input the standardized dataset, and to perform high-dimensional feature dimensionality reduction and weight quantization using factor analysis and the TOPSIS algorithm based on entropy weight, and output the urban propagation feature vector.

[0170] The temporal potential assessment module is used to input the historical propagation time-series data and the city's propagation feature vector, construct an adaptive temporal prediction model to extract future trend features, and perform quantitative assessment in combination with its temporal fluctuation features, outputting a temporal potential feature vector, a city's comprehensive propagation capability score, and a strategy type label.

[0171] The hot topic sentiment calculation module is used to segment the audience by country and region. It takes as input the multilingual social media comment dimension data and the cross-cultural communication literature data, performs sentiment analysis and topic modeling, and outputs the audience interest preference distribution, dynamic emotional feature vector and regional steady-state cultural feature vector for each region.

[0172] The candidate scheme generation module is used to integrate the city's communication feature vector, time-series potential feature vector, regional steady-state cultural feature vector, and audience interest and preference distribution into a generation condition vector. Using the strategy type label as a control signal, it generates a regional candidate scheme set with associated communication strategy parameter vectors in parallel through three paths: template derivation path, component combination path, and large model innovation path.

[0173] The scheme evaluation and ranking module is used to take into account the regional candidate scheme set, the dynamic emotion feature vector and the city's comprehensive communication capability score, and to perform differentiated evaluation and ranking based on the topic semantic alignment and real-time hotspot status, and output the optimal communication scheme for each region.

[0174] The decision description document generation module is used to input the optimal propagation scheme and generate a corresponding decision description document by retrieving and enhancing the scheme from the knowledge base.

[0175] The iterative optimization control module is used to feed the landing feedback data back as incremental training samples to the adaptive time series prediction model, feature fusion weight and scheme generation model to form an adaptive optimization closed loop.

[0176] This application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a method for generating urban dissemination schemes based on multi-source data and hotspot feedback.

[0177] The terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server. The terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0178] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0179] The memory can be an internal storage unit of the terminal device, such as a hard disk or RAM of the terminal device, or an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) equipped on the terminal device. Furthermore, the memory can be a combination of internal storage units and external storage devices of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0180] In this terminal device, the urban propagation scheme generation method based on multi-source data and hotspot feedback in the above embodiment is stored in the memory of the terminal device and loaded and executed on the processor of the terminal device for convenient use.

[0181] This application also discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it employs a method for generating urban communication schemes based on multi-source data and hotspot feedback as described in the above embodiments.

[0182] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0183] The method for generating a city propagation scheme based on multi-source data and hotspot feedback, as described in the above embodiments, is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0184] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0185] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for generating urban communication schemes based on multi-source data and hotspot feedback, characterized in that, Includes the following steps: Collect five-dimensional heterogeneous data from the target city, including search engines, media reports, online dissemination, inbound visitors, and multilingual social media comments, as well as historical dissemination time-series data and cross-cultural dissemination literature data from various countries and regions. Perform differentiated cleaning and normalization processing to output a standardized dataset. Input the standardized dataset, use factor analysis and the TOPSIS algorithm based on entropy weight to perform high-dimensional feature dimensionality reduction and weight quantization, and output the urban propagation feature vector; Input the historical propagation time series data and the city's propagation feature vector, construct an adaptive time series prediction model to extract future trend features, and combine the time series fluctuation features extracted based on the historical propagation time series data for quantitative evaluation, and output the time series potential feature vector, the city's comprehensive propagation capability score and the strategy type label; The audience is segmented by country and region. The multilingual social media comment dimension data and cross-cultural communication literature data are input, sentiment analysis and topic modeling are performed, and the distribution of audience interest preferences, dynamic emotional feature vectors and regional steady-state cultural feature vectors are output for each region. The city's communication feature vector, time-series potential feature vector, regional steady-state cultural feature vector, and audience interest and preference distribution are fused into a generation condition vector. The strategy type label is used as a control signal, and a set of regional candidate schemes with related communication strategy parameter vectors are generated in parallel through three paths: template derivation, component combination, and large model innovation. Input the regional candidate scheme set, the dynamic emotion feature vector and the city comprehensive communication capability score, combine the topic semantic alignment degree and real-time hot spot status to perform differentiated evaluation and ranking, and output the optimal communication scheme for each region; Input the optimal propagation scheme, and generate a corresponding scheme decision explanation document through knowledge base retrieval enhancement; The feedback data from the landing site is used as incremental training samples and fed back into the adaptive time series prediction model, feature fusion weights and scheme generation model to form an adaptive optimization closed loop.

2. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The differential cleaning and normalization process includes the following steps: Max-min value normalization is performed on the numerical indicators in the five-dimensional heterogeneous data to eliminate dimensional differences; The text data in the five-dimensional heterogeneous data is mapped into fixed-dimensional semantic vectors through a pre-trained multilingual model; Missing values ​​in the five-dimensional heterogeneous data are filled using time series interpolation or mean. Outliers in the five-dimensional heterogeneous data were identified using a box plot method and replaced with the median.

3. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The input standardized dataset is subjected to factor analysis and the TOPSIS algorithm based on entropy weights for high-dimensional feature dimensionality reduction and weight quantization. The output city propagation feature vector includes the following steps: The standardized dataset is subjected to the Kaiser-Meyer-Olkin test and the Bartlett test of sphericity. If the test passes, factor analysis is performed to extract common factors whose cumulative variance contribution rate reaches a preset threshold, and the dimensionality-reduced factor score matrix is ​​obtained. An evaluation decision matrix is ​​constructed based on the factor score matrix, and the evaluation decision matrix is ​​then dimensionless to obtain a standardized decision matrix. The information entropy of each evaluation indicator is calculated based on the standardized decision matrix, and the entropy weight of each evaluation indicator is calculated based on the information entropy. Based on the standardized decision matrix, the positive ideal solution vector and the negative ideal solution vector are determined, and the Euclidean distance between each evaluation object and the positive ideal solution vector and the negative ideal solution vector are calculated respectively. The relative proximity of each evaluation object is calculated based on the Euclidean distance, and the relative proximity is used as the weight quantification result. The factor score matrix and the weight quantification result are then concatenated to obtain the urban propagation feature vector.

4. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 3, characterized in that, The formulas for calculating the information entropy of each evaluation indicator are as follows: In the formula: Indicates the first Information entropy of each evaluation indicator Indicates the number of evaluation objects. Indicates the first The evaluation object is in the first The weighting values ​​under each evaluation indicator ; The formulas for calculating the entropy weights of each evaluation index are as follows: In the formula: Indicates the first The entropy weight of each evaluation indicator, Indicates the number of evaluation indicators; The formula for calculating the relative proximity is as follows: In the formula: Indicates the first The relative similarity of the evaluation objects Indicates the first The Euclidean distance between each evaluation object and the positive ideal solution vector. Indicates the first The Euclidean distance between each evaluation object and the negative ideal solution vector.

5. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The process of inputting the historical propagation time-series data and the city's propagation feature vector, constructing an adaptive time-series prediction model to extract future trend features, and combining these with its time-series fluctuation features for quantitative evaluation, outputting a time-series potential feature vector, a city's comprehensive propagation capability score, and a strategy type label includes the following steps: The historical propagation time series data is input into the adaptive time series prediction model. The adaptive time series prediction model adaptively selects a long short-term memory network or a Transformer encoder as the backbone network based on the urban development stage identifier in the urban propagation feature vector, and outputs a propagation trend prediction sequence for a future preset time window. The growth slope, volatility variance, and peak potential coefficient are calculated based on the propagation trend prediction sequence. The growth slope, volatility variance, and peak potential coefficient are then concatenated to obtain the time series potential feature vector. Based on the historical propagation time series data, the fluctuation coefficient and skewness coefficient are calculated to form a static measurement vector; The city's comprehensive communication capability score is obtained by weighting and fusing the urban communication feature vector, the temporal potential feature vector, and the static measurement vector. Based on the preset threshold range of the city's comprehensive communication capability score, the strategy type label is determined, which includes resource accumulation type label or fission explosion type label.

6. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 5, characterized in that, The formula for calculating the peak potential coefficient is as follows: In the formula: Indicates the peak potential coefficient. This represents the maximum predicted value in the propagation trend prediction sequence. This represents the mean of the historical propagation time series data. This represents the standard deviation of the historical propagation time series data; The formula for calculating the city's comprehensive communication capability score is as follows: In the formula: This indicates the city's overall communication capability score. This represents the urban propagation feature vector. This represents the time-series potential feature vector. Represents the static measure vector, Describing the L2 norm, The preset static feature weight coefficients, time series potential weight coefficients, and volatility measure weight coefficients are given. .

7. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The process of segmenting the audience by country / region, inputting the multilingual social media comment dimension data and the cross-cultural communication literature data, performing sentiment analysis and topic modeling, and outputting the audience interest preference distribution, dynamic sentiment feature vector, and regional steady-state cultural feature vector for each region includes the following steps: Language identification and translation alignment were performed on the multilingual social media comment dimension data to obtain a set of comment texts in a unified semantic space; Fine-grained sentiment analysis is performed on the set of comment texts to obtain the sentiment polarity score and sentiment category probability distribution for each comment text; Based on the emotional polarity score and the probability distribution of emotional categories, the mean emotional value, variance of emotional fluctuation, coefficient of emotional mutation, and clustering degree of negative hot topics in each region are calculated, and the dynamic emotional feature vector is obtained by splicing them together. Latent Dirichlet Allocation (LDA) topic modeling was performed on the cross-cultural communication literature data to extract high-frequency keywords and their probability distributions for each region, thereby obtaining the steady-state cultural feature vector of the region. The frequency of audience interaction with different types of content in each region was statistically analyzed, and the audience interest preference distribution was obtained after normalization.

8. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 7, characterized in that, The formula for calculating the clustering degree of negative trending topics is as follows: In the formula: Indicates the concentration of negative trending topics. Indicates the number of time slices within the time window. Indicates the first The percentage of negative emotional comments within a given time period. Indicates the time decay weight. The preset attenuation coefficient; The formula for calculating the emotional abrupt change coefficient is as follows: In the formula: Indicates the emotional mutation coefficient. This represents the average sentiment in the current time slice. This represents the average mood of the previous time segment. The standard deviation of the historical sentiment mean.

9. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The process of fusing the urban communication feature vector, temporal potential feature vector, regional steady-state cultural feature vector, and audience interest and preference distribution into a generation condition vector, using the strategy type label as a control signal, and generating a regional candidate scheme set of related communication strategy parameter vectors in parallel through three paths—template derivation, component combination, and large model innovation—includes the following steps: The urban communication feature vector, the temporal potential feature vector, the regional steady-state cultural feature vector, and the audience interest and preference distribution are spliced ​​and fused into a generation condition vector, and the strategy type label is encoded into a control signal vector. Based on the template derivation path, the cosine similarity between the generated condition vector and the condition vectors of each template in the historical successful case template library is calculated. The best matching template that satisfies the strategy type constraint is selected for parameter replacement and local fine-tuning to generate the first candidate solution. Based on the component combination path, the propagation target is decomposed into reusable components. With the goal of maximizing the combination score, a greedy strategy combined with a local search algorithm is used to select the optimal component combination and automatically splice it to generate a second candidate scheme. Based on the aforementioned large-scale model innovation path, the generated condition vector and the control signal vector are decoded into natural language constraint instructions, driving the large language model to generate an innovative solution framework and generate a third candidate solution. The first, second, and third candidate schemes are deduplicated and format-standardized to obtain the regional candidate scheme set, and each candidate scheme is bound to a propagation strategy parameter vector.

10. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The process of inputting the regional candidate solution set, the dynamic sentiment feature vector, and the city's comprehensive communication capability score, combined with the topic semantic alignment and real-time hot topic status for differentiated evaluation and ranking, and outputting the optimal communication solution for each region includes the following steps: Extract the topic distribution from the propagation strategy parameter vector of each candidate scheme in the regional candidate scheme set, extract the topic sentiment mean vector from the dynamic sentiment feature vector, calculate the cosine similarity between the topic distribution and the topic sentiment mean vector, and obtain the topic semantic alignment. The current real-time hot topic status is determined based on the mean emotion value in the dynamic emotion feature vector and the clustering degree of negative hot topics. The real-time hot topic status includes a negative hot topic clustering state, a positive hot topic surge state, and a hot topic stable state. The corresponding evaluation correction coefficient is determined based on the real-time hotspot status. Based on the topic semantic alignment, the city's comprehensive communication capability score, and the evaluation correction coefficient, the adaptation score of each candidate scheme is calculated. The candidate schemes are then sorted in descending order of their adaptation scores, and the candidate scheme with the highest score is selected as the optimal communication scheme.

11. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 10, characterized in that, The formula for calculating the adaptation score is as follows: In the formula: Indicates the fit score. Indicates the semantic alignment of the topics. This indicates the city's overall communication capability score. This represents the evaluation correction factor. , These are the preset topic alignment weights and city capability weights, respectively. Adjust the weights for preset hotspots. This indicates the coverage of the candidate solution with the distribution of audience interests and preferences; The evaluation correction coefficient The segmentation and value selection rules are as follows: When the real-time trending topic status is a state of negative trending topics clustering. ; When the real-time hotspot status is a state of positive popularity surge. ; When the real-time hot topic status is a stable hot topic state. ; in, , These are the preset negative penalty coefficient and positive incentive coefficient, respectively. For negative trending topics, This represents the emotional mutation coefficient.

12. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The process of inputting the optimal propagation scheme and generating a corresponding scheme decision description document through knowledge base retrieval and enhancement includes the following steps: Construct an urban communication knowledge base, which includes a multi-dimensional communication indicator library, a comprehensive publicity and operation case library, a communication effect evaluation model library, and an external communication experience and rule library; The joint retrieval query vector is formed by concatenating the dissemination strategy parameter vector of the optimal dissemination scheme, the urban dissemination feature vector, the audience interest and preference distribution of the corresponding region, the dynamic emotional feature vector, the regional steady-state cultural feature vector, and the strategy type label. Based on the joint retrieval query vector, feature similarity retrieval is performed in the urban communication knowledge base to match quantitative support data, similar implementation cases and reasoning logic. The system automatically renders a pre-defined standardized report template and outputs a solution decision description document that includes the core content of the solution, decision-making logic, expected audience acceptance, risk warnings, and implementation suggestions.

13. The method for generating urban communication schemes based on multi-source data and hotspot feedback according to claim 1, characterized in that, The step of feeding the landing feedback data back to the adaptive time-series prediction model, feature fusion weights, and scheme generation model as incremental training samples to form an adaptive optimization closed loop includes the following steps: Collect data on actual dynamic changes in sentiment, shifts in audience interests and preferences, and deviations in key communication indicators in each region after the implementation of the current round of plans at fixed intervals. A loss function is constructed based on the actual dynamic emotional changes, audience interest and preference shifts, and key communication indicator deviations. The incremental training data is fed back to the adaptive temporal prediction model to update network parameters, the feature fusion weight adjustment factor analysis and entropy weight parameters, and the embedding layer and attention mechanism parameters of the conditional encoder in the scheme generation model for fine-tuning. The update is paused when the loss decreases by less than the preset convergence threshold after two consecutive iterations to prevent overfitting.

14. A device for generating urban communication schemes based on multi-source data and hotspot feedback, characterized in that, The device includes: The data preprocessing module is used to collect five-dimensional heterogeneous data and historical dissemination time series data of the target city, and simultaneously collect cross-cultural dissemination literature data of various countries and regions. It performs differential cleaning and normalization processing on the five-dimensional heterogeneous data and outputs a standardized dataset. The urban propagation feature modeling module is used to input the standardized dataset, perform high-dimensional feature dimensionality reduction and weight quantization using factor analysis and the TOPSIS algorithm based on entropy weight, and output the urban propagation feature vector. The temporal potential assessment module is used to input the historical propagation time series data and the city's propagation feature vector, construct an adaptive temporal prediction model to extract future trend features, and perform quantitative assessment in combination with its temporal fluctuation features, outputting a temporal potential feature vector, a city's comprehensive propagation capability score, and a strategy type label; The hot topic sentiment calculation module is used to segment the audience by country and region. It takes the multilingual social media comment dimension data and the cross-cultural communication literature data as input, performs sentiment analysis and topic modeling, and outputs the audience interest preference distribution, dynamic emotional feature vector and regional steady-state cultural feature vector for each region. The candidate scheme generation module is used to integrate the urban dissemination feature vector, the temporal potential feature vector, the regional steady-state cultural feature vector, and the audience interest and preference distribution into a generation condition vector. Using the strategy type label as the control signal, it generates a regional candidate scheme set with associated dissemination strategy parameter vectors in parallel through three paths: template derivation, component combination, and large model innovation. The scheme evaluation and ranking module is used to take into account the regional candidate scheme set, the dynamic emotion feature vector and the city's comprehensive communication capability score, and to perform differentiated evaluation and ranking based on the topic semantic alignment degree and real-time hot spot status, and output the optimal communication scheme for each region. The decision explanation document generation module is used to take the optimal propagation scheme as input and generate a corresponding decision explanation document through knowledge base retrieval and enhancement. The iterative optimization control module is used to feed the landing feedback data back as incremental training samples to the adaptive time series prediction model, feature fusion weight and scheme generation model to form an adaptive optimization closed loop.

15. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it implements the urban communication scheme generation method based on multi-source data and hotspot feedback as described in any one of claims 1 to 13.

16. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it implements the urban communication scheme generation method based on multi-source data and hotspot feedback as described in any one of claims 1 to 13.

Images