Urban scenic spot recommendation management method and system based on artificial intelligence

Abstract

The invention relates to the technical field of scenic spot recommendation management, in particular to a city scenic spot recommendation management method and system based on artificial intelligence. The method comprises the following steps: receiving a scenic spot demand text input by a user, extracting a keyword data set in the scenic spot demand text, and performing word frequency evolution analysis to obtain keyword frequency evolution data; thirdly, deducing the playing tendency intention of the user according to the keyword frequency evolution data, obtaining tendency intention clustering data through clustering processing, and evaluating the adaptation degree of the urban scenic spots based on the data; and finally, extracting adaptive judgment logic, performing optimal recommendation strategy learning, outputting an adaptive recommendation strategy, and embedding the adaptive recommendation strategy into a terminal. The scenic spot recommendation management technology is optimized, so that the scenic spot recommendation management technology is more perfect.

Description

Technical Field

[0001] This invention relates to the field of scenic spot recommendation and management technology, and in particular to an artificial intelligence-based method and system for urban scenic spot recommendation and management. Background Technology

[0002] Traditional attraction recommendation systems typically rely on static user preferences or simple rules and rating mechanisms, making them ill-suited to rapidly changing tourist demands and dynamic market environments. Especially with the rapid development of big data and artificial intelligence technologies, improving the accuracy and personalization of attraction recommendations using advanced technologies has become a crucial issue for the tourism industry. By analyzing user needs, identifying user preferences, and optimizing recommendation results in real time, more accurate and personalized services can be provided. The application of artificial intelligence technologies, particularly Natural Language Processing (NLP), machine learning, and deep learning, enables attraction recommendation systems to extract valuable information from massive amounts of user data and attraction information, thereby better understanding users' travel preferences and needs.

[0003] In summary, traditional AI-based urban attraction recommendation and management methods suffer from poor accuracy in understanding user needs, resulting in low suitability of attraction recommendations. Summary of the Invention

[0004] Therefore, it is necessary to provide an artificial intelligence-based method and system for recommending and managing urban attractions in order to solve at least one of the aforementioned technical problems.

[0005] To achieve the above objectives, an artificial intelligence-based method for recommending and managing urban attractions is proposed, the method comprising the following steps: Step S1: Accept the scenic spot requirement text input by the user and determine the keyword dataset in the scenic spot requirement text; perform word frequency evolution analysis on the keyword dataset to obtain keyword frequency evolution data; Step S2: Based on the keyword frequency evolution data, deduce the user's travel intention and perform clustering processing to obtain intention clustering data; determine the suitability of city attractions based on the intention clustering data; Step S3: Extract the adaptation judgment logic from the city attraction adaptation degree, then learn the optimal recommendation strategy and output the adaptation recommendation strategy; embed the adaptation recommendation strategy into the terminal to execute city attraction recommendation management.

[0006] The present invention also provides an artificial intelligence-based urban attraction recommendation and management system for executing the artificial intelligence-based urban attraction recommendation and management method described above. The artificial intelligence-based urban attraction recommendation and management system includes: The data acquisition module is used to receive the scenic spot demand text input by the user and determine the keyword dataset in the scenic spot demand text; perform word frequency evolution analysis on the keyword dataset to obtain keyword frequency evolution data; The adaptation analysis module is used to infer users' travel intentions based on the keyword frequency evolution data, and perform clustering processing to obtain intention clustering data; and to determine the degree of adaptation of city attractions based on the intention clustering data. The recommendation strategy learning module is used to extract the adaptation judgment logic in the city attraction adaptation degree, and then learn the optimal recommendation strategy to output the adaptation recommendation strategy; the adaptation recommendation strategy is embedded into the terminal to perform city attraction recommendation management.

[0007] The beneficial effects of this invention lie in its ability to accurately capture users' interests and needs by accepting user-inputted text of attraction requests and extracting keyword datasets. This process, through word frequency evolution analysis, reveals the changing trends of keywords and the dynamic characteristics of user needs. This data processing method helps the system identify users' potential travel tendencies, thus providing valuable input for subsequent recommendations. Compared with traditional methods, this method can not only handle more complex text data but also track changes in user needs, ensuring that attraction recommendations are more aligned with users' current travel intentions. By deriving users' travel intentions based on keyword frequency evolution data and performing clustering, the system can categorize users with similar needs. The core value of this process lies in the ability to more accurately analyze the behavioral characteristics of different user groups through the application of clustering algorithms, revealing users' specific needs at different times and in different scenarios. This big data-based clustering analysis not only improves the accuracy of personalized recommendations but also enhances the adaptability of the recommendation system, enabling it to cope with more diverse user needs and a constantly changing market environment. In the process of extracting the judgment logic for the suitability of city attractions and learning the optimal recommendation strategy, the system can automatically optimize the recommendation strategy based on the user's intentions and the suitability of the attractions. The advantage of this process lies in its ability to adjust the recommendation algorithm in real time according to the needs of different users, ensuring a high degree of personalization and relevance in the recommendation results. By embedding adaptive recommendation strategies into the terminal, the recommendation system can achieve automated management of attraction recommendations, reducing the need for manual intervention and improving system response speed and user experience. Furthermore, continuous learning of the optimal recommendation strategy ensures the long-term effectiveness and adaptability of the recommendation results. Therefore, this invention is an optimization of a traditional AI-based urban attraction recommendation management method, solving the problem of poor accuracy in understanding user needs and resulting in low adaptability of attraction recommendations in traditional AI-based urban attraction recommendation management methods. It improves the accuracy of understanding user needs and enhances the adaptability of attraction recommendations. Attached Figure Description

[0008] Figure 1 A flowchart illustrating the steps of an AI-based urban attraction recommendation and management method; Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2. Detailed Implementation

[0009] Please see Figure 1 An artificial intelligence-based method for recommending and managing urban attractions, the method comprising the following steps: Step S1: Accept the scenic spot requirement text input by the user and determine the keyword dataset in the scenic spot requirement text; perform word frequency evolution analysis on the keyword dataset to obtain keyword frequency evolution data; In this embodiment of the invention, the input scenic spot demand text is obtained through a receiving device. After receiving the text, the device annotates each word segmentation unit in the sentence. During this process, the system performs preliminary word segmentation of the input text according to spaces and punctuation marks, and performs part-of-speech matching in conjunction with a pre-built tourism-specific dictionary. After annotation, key grammatical nodes such as nouns, verbs, and adjectives are extracted, and a lexical dependency graph is constructed. This graph is based on a dependency parsing algorithm, which determines the semantic association structure between words according to dependency edges such as subject-predicate, verb-object, attributive-head, and coordinate. Then, a keyword filtering operation based on a co-occurrence matrix is ​​performed on this structure. When the frequency of a certain word in all corpus windows exceeds the total number of words, the keyword filtering operation is performed. When a keyword has a correlation coefficient greater than 0.6 with a candidate subject or predicate node, it is identified as a keyword and stored in the keyword dataset. Then, the sliding window algorithm is used to perform word frequency trajectory analysis on the continuously input demand data in the time dimension. The window size is set to 50 records and the step size is 10 records. The frequency change curve of each keyword in the time series is calculated. The change rate is obtained by dividing the frequency difference between the two windows by the frequency of the previous window. Then, the noise is suppressed by the exponential smoothing algorithm with a smoothing coefficient set to 0.7. Finally, the keyword frequency evolution data containing timestamps, word frequency change rate and smoothing count is obtained, which provides the basic input for subsequent intent inference.

[0010] Step S2: Based on the keyword frequency evolution data, deduce the user's travel intention and perform clustering processing to obtain intention clustering data; determine the suitability of city attractions based on the intention clustering data; In this embodiment of the invention, based on hierarchical clustering algorithm and cosine similarity calculation method, the semantic structure between keyword time evolution vectors is hierarchically divided. The system first projects the keyword vectors into a 300-dimensional semantic vector space, then calculates the cosine similarity between any two keyword vectors, and the result is recorded as a semantic distance matrix. Next, an agglomerative hierarchical clustering algorithm is used to merge node groups with similarity greater than 0.8 in each iteration until the cluster structure is stable. Each cluster in this stage represents a potential user play intention. Subsequently, the system calculates the intra-cluster distance and inter-cluster distance to obtain a cluster density index, with the value limited to between 0 and 1. To prevent bias in the clustering results, an additional... The Davies-Bouldin index is used as an external validation parameter. By minimizing this index value, the optimal number of clusters is determined, generating clustering data with a tendency towards intent. Next, structured fields such as theme type, functional attributes, geographical coordinates, opening hours, and historical traffic flow of each attraction are extracted from the city's basic information database. The similarity score between the intent cluster and the attraction's theme keywords is calculated. The similarity is obtained using the weighted cosine similarity method, where the theme weight is set to 0.5, the functional attribute weight to 0.3, and the transportation convenience weight to 0.2. The attraction suitability score is generated by weighted summation of similarity scores, with the score limited to the range of 0 to 1. This score is the result data of the city's attraction suitability.

[0011] Step S3: Extract the adaptation judgment logic from the city attraction adaptation degree, then learn the optimal recommendation strategy and output the adaptation recommendation strategy; embed the adaptation recommendation strategy into the terminal to execute city attraction recommendation management; In this embodiment of the invention, a min-max normalization method is used to map all fit values ​​to the [0,1] interval, forming fit-normalized data. Then, a pattern analysis engine is activated to extract logical rules from the normalized data. The pattern analysis engine employs a decision tree-based rule extraction algorithm, using information gain ratio as the node splitting criterion. The information gain threshold is set to 0.02, and the tree depth is limited to 8. The system recursively splits the dataset to generate logical judgment linked lists, with each linked list representing a fit logic rule. For example, a high fit level is defined as when the topic similarity is greater than or equal to 0.7 and the traffic fit is greater than or equal to 0.6. The encoded logical structure is then passed to the reinforcement learning module. The reinforcement learning module uses the Q-learning algorithm to learn the optimal recommendation strategy for the adaptation judgment logic. During this learning process, the state space consists of combinations of scenic spot adaptation features, and the action space consists of two operations: "recommend" and "not recommend." The reward function is based on standardized user ratings, with a range of -1 to 1. The Q-value table is updated through multiple iterations, with a learning rate of 0.1 and a discount factor of 0.9. In each iteration, the system adjusts the optimal strategy value for the corresponding state according to the Bellman update criterion. When the convergence threshold reaches 1×10⁻⁶...- When it is t³, a stable optimal recommendation strategy vector is formed, and the output strategy parameters are embedded into the urban scenic spot recommendation execution terminal. After the terminal receives the new demand text, it directly reads the strategy vector for decision matching. According to the existing logical structure and reinforcement learning results, it comprehensively scores and ranks the adaptability of the candidate scenic spots, and outputs the final list of the recommended order of urban scenic spots.

[0012] Step S1 includes the following steps: Step S11: Receive the scenic spot demand text input by the user; Step S12: Mark the word nature in the scenic spot demand text, and determine the lexical relationship in the scenic spot demand text according to the word nature; Step S13: Determine the keyword data set in the scenic spot demand text according to the lexical relationship; Step S14: Perform word frequency evolution analysis on the keyword data set through the lexical relationship to obtain keyword frequency evolution data.

[0013] In the embodiment of the present invention, the scenic spot description data transmitted externally is received through the input interface. The length of the data buffer area of the input interface is set to 2048 characters, and the UTF-8 encoding method is used to standardize the characters of the input content. After the data enters the system, all non-text characters are first uniformly filtered. For example, carriage returns, line breaks, and tab characters are converted into spaces to ensure the continuity during subsequent text annotation. Then, the word segmentation operation based on the forward maximum matching algorithm is performed. This algorithm loads the tourism industry term dictionary with a scale of 50,000 records, and the matching window width is set to 6 characters. The algorithm starts from the left end of the input text and compares characters one by one, and preferentially matches the longest word for segmentation. During the processing, the stop word rule table is introduced for the continuously appearing place names, and auxiliary words such as "of", "very", "in", "to" are excluded. Finally, a labeled word segmentation sequence is generated, each entry is attached with the start and end character position indexes and the occurrence order number, and all entries are written into the structured array in the original text order, forming the standard data input unit of the scenic spot demand text.

[0014] In the part-of-speech tagging stage, a statistical tagging process based on the Hidden Markov Model (HMM) algorithm is performed on the standardized word sequences obtained in step S11. The training corpus uses a tourism industry descriptive corpus with a total of 1.5 million word samples. The tag set includes 10 categories of part-of-speech, including nouns, verbs, adjectives, adverbs, prepositions, numerals, locative words, place names, personal names, and organization names. The number of hidden states in the tagging model is set to 10, the emission probability is calculated using maximum likelihood estimation, and the transition probability matrix has a dimension of 10×10. The Viterbi algorithm is used to derive the optimal path for each state to obtain the most likely part-of-speech tag for each word. Based on the gender tags, dependency syntactic structure analysis is performed using an arc-normalized dependency tree construction algorithm. During the construction process, starting from the central verb of the syntactic root node, dependency edge relationships such as subject-predicate, verb-object, adverbial-head, attributive-head, and coordinate are established through tag constraints. The dependency weight is calculated by weighting the tag transition probability with the inverse of the distance between words, and the weight value is limited to between 0 and 1. When the weight of the relation edge is lower than 0.2, it is discarded. After the analysis is completed, a complete set of lexical relationships is formed in the text, and the parent and child node numbers, dependency types, and weight values ​​between words are recorded to support subsequent keyword extraction operations.

[0015] Using the vocabulary relation set constructed in step S12, a keyword candidate set is generated based on co-occurrence probability and information entropy. During the calculation, dependency edges are used as the analysis unit, and all adjacent word node pairs are traversed. Within a co-occurrence interval with an analysis window size of 5, the co-occurrence frequency of each word with other words is counted. Words with a co-occurrence frequency greater than 3 and a dependency relation weight greater than 0.4 are included in the candidate set. Subsequently, the information gain value of the candidate words is calculated based on the point mutual information algorithm. The information gain threshold is set to 0.1, and words below this threshold are removed. Next, a screening method based on text inverse document frequency is used to evaluate the distribution characteristics of candidate words in multiple demand text samples. This method defines a global word frequency upper limit of 1000 times and a lower limit of 3 times. Words within this range are retained. Then, they are sorted from high to low frequency, and the top 20% of words are selected to form a keyword dataset. In the keyword dataset, each record contains vocabulary content, part-of-speech type, dependency relation position index, word frequency, and timestamp. Through this set, the core information structure in the scenic spot demand text is transformed into analyzable data.

[0016] In the word frequency evolution analysis stage, time series statistical analysis is performed on the keyword dataset obtained in step S13. A time window mechanism is set to track keyword frequency changes, with each window length set to 100 text records and a window sliding step size of 20 records. The frequency of the same keyword appearing within consecutive time windows is incrementally counted, and the word frequency change rate is calculated. The change rate is defined as the word frequency of the subsequent window minus the word frequency of the previous window, then divided by the word frequency of the previous window. To reduce the interference of sudden fluctuations on the judgment, a weighted exponential smoothing algorithm is introduced for data smoothing, with a weight decay coefficient set to 0.6. This is used when calculating data for each time period. At that time, the statistical values ​​are updated using a single-pass weighted average method. In addition, in order to capture the degree of correlation between word frequency changes among different keywords, a word frequency covariance matrix is ​​constructed for all keyword pairs. The matrix dimension is equal to the square of the number of keywords. After standardization processing of the matrix, the correlation coefficient is calculated. When the correlation coefficient is greater than 0.7, it is considered that there is a frequency linkage relationship, and the linkage marker is recorded in the matrix. Finally, keyword frequency evolution data is formed. Its data structure includes keyword fields, time index, frequency change rate, smoothed frequency values ​​and correlation coefficient series, which are used to reflect the change trajectory and correlation structure of the semantic features of the entire tourism demand text in the time dimension.

[0017] refer to Figure 2 The aforementioned step S2 includes the following steps: Step S21: Analyze the semantic evolution structure in the keyword frequency evolution data; In this embodiment of the invention, the keyword frequency evolution data generated in step S14 is read in a structured manner. The data fields include keyword index, time series number, smoothed word frequency, rate of change, and covariance coefficient. To establish a semantic evolution structure, a two-layer association algorithm based on semantic similarity and time dependence is adopted. First, the semantic distance between any two keywords is calculated within the same time window. The semantic distance is implemented using the average word vector difference method, represented by a 300-dimensional vector. Each keyword corresponds to a fixed set of semantic vectors, which are derived from a set of word vectors trained on a preset tourism corpus. If the cosine similarity between two keyword vectors is greater than 0.75, a connection edge is recorded in the semantic relationship matrix, and the similarity value is used as the edge weight. Then, in the time dimension... The frequency change direction sign of the same keyword in two consecutive time windows is calculated. When the change direction is the same and the amplitude difference is less than 2%, an evolutionary connection is generated. This connection represents the continuous evolutionary relationship of the semantic topic. Then, the keyword relationship matrix in all time periods is stacked in chronological order to form a three-dimensional tensor of temporal semantic relationship. Path tracing operation based on graph decomposition is performed on this tensor. The number of keyword nodes is defined as n, the number of time layers is t, and the threshold for the tracing path length is t-2. The average edge weight and node frequency stability index are calculated for each semantic path. Paths with an index lower than 0.3 are removed. Finally, a semantic evolution network is generated in the three-dimensional structure of the tensor. The number of network edges is controlled within 1.5 times the number of nodes. This network is the semantic evolution structure of the keyword frequency evolution data.

[0018] Step S22: Derive the user's play intention based on the semantic evolution structure; In this embodiment of the invention, during the stage of inferring user play intentions based on semantic evolution structure, the semantic evolution network generated in step S21 is read, and the backbone structure is extracted using a semantic tree transformation algorithm. Based on the minimum spanning tree principle, the algorithm selects the path with the largest edge weight as the starting point of the backbone path using the Prim algorithm, expanding adjacent nodes layer by layer until 60% of the total number of nodes is covered. This part is defined as the semantic backbone layer. Then, the nodes in the backbone layer are decomposed hierarchically according to the time window index order. Each set of nodes within three consecutive time windows is divided into a semantic hierarchical unit. The average node frequency, similarity concentration, and time span coefficient of each unit are calculated. Hierarchical units with a frequency standard deviation below 0.05 and a time span of less than or equal to three windows are marked as stable semantic layers with high concentration. Nodes with a weight of 0.8 are defined as core topic points in the backbone layer. Next, if the decrease in edge weight in adjacent paths between core topic points exceeds 0.2, a semantic topic shift is considered. A topic environment index value is generated at this shift node, representing the weight strength of the semantic migration direction, with a value range of 0 to 1. Based on this, nodes in the backbone path are divided into several semantic partitions according to the degree of clustering of core topic points. Each partition represents a potential user interest trend. Further analysis is conducted on verb-type nodes within each partition, such as "play," "visit," "take photos," and "experience." By calculating the time frequency increment of verb nodes, an action tendency index is determined. This index is used to determine which type of play activity the user prefers, thereby deriving user play tendency intention data from the entire semantic evolution structure.

[0019] Step S23: Cluster the user's play intentions to obtain intention clustering data; In this embodiment of the invention, during the clustering process of user play intentions, the user play intention data obtained in step S22 is input. The data fields include topic type number, semantic partition center coordinates, action tendency degree, and time stability index. For clustering analysis, an Euclidean distance clustering algorithm based on K-means++ initialization is adopted, with the initial number of cluster centers set to 8, the maximum number of iterations to 10, and the convergence error threshold to 1×10⁻⁶. -4 Clustering operations are performed on the two-dimensional feature space composed of the semantic partition center coordinate vector and the action tendency vector. In each iteration, the squared distance from each sample point to each cluster center is recalculated and the center position is updated. After the iteration, 8 clusters are output. Each cluster contains a group of samples with similar play type tendencies. To reduce the interference of outliers, the internal squared error of each cluster is calculated. Clusters with an overall average error exceeding 0.2 times the standard deviation are further divided. After the division, the final tendency intention clustering data is obtained. Each cluster record includes the intention cluster number, the semantic vector of the cluster center, the statistics of the members inside the cluster, and the average action tendency. Through this clustering result, the play intentions of different users can be distinguished and expressed.

[0020] Step S24: Obtain basic information about the current city's attractions; In this embodiment of the invention, during the stage of acquiring basic information of current urban attractions, a publicly available urban tourism data repository is accessed and the structured information of all attractions in the current city is read. The fields include attraction number, name, theme tag, geographic coordinates, site area, average daily visitor volume, historical traffic speed, holiday visitor flow fluctuation index, entry restriction time, landscape type, and equipment completeness value. The total number of records is 1200. Each data is uniformly parsed in JSON format and stored in memory to form a basic information table. To achieve standardized comparison, zero-mean normalization is performed on numerical fields, with a normalization range of [0,1]. The Euclidean distance from the city center point to the attraction is calculated for the latitude and longitude data in the geographic coordinate information, with the unit being kilometers. The distance field is then linearly mapped and scaled to ensure that different dimensional features have a unified analysis scale. Text fields such as theme tags are converted into binary matrices using one-hot encoding. The information table is supplemented with the average monthly climate temperature data and the number of surrounding traffic types for each attraction over the past 5 years to describe the background features of the attraction. This basic information serves as the input dataset for subsequent matching calculations.

[0021] Step S25: Determine the suitability of urban attractions based on the basic information and clustering data of current urban attractions.

[0022] In this embodiment of the invention, during the stage of determining the suitability of urban attractions based on the basic information and intention clustering data of current urban attractions, the basic information of attractions obtained in step S24 and the intention clustering data obtained in step S23 are aligned in terms of features. An attraction attribute matrix and an intention feature matrix are constructed in the feature vector space. The attraction attribute matrix includes a theme category vector, a transportation convenience vector, a distance vector, and a passenger flow stability index. The intention feature matrix includes a semantic center vector, an action tendency vector, and a theme relevance coefficient. First, the cosine similarity between each attraction theme category vector and the intention semantic center vector is calculated. The similarity threshold is set to 0.6, and only results greater than 0.6 are retained. The matching pairs of this value are then used to calculate the correlation coefficient between transportation convenience and action tendency in the transportation dimension. The Pearson correlation method is used to calculate the value in the interval [-1,1], and the absolute value is taken as the transportation matching degree. The weighting coefficient is set to 0.4. At the same time, the theme matching degree weighting coefficient is calculated to 0.6. Then, the theme matching degree and transportation matching degree of each attraction are weighted and summed to form a preliminary fit score. To enhance stability, the score is normalized by mean and variance and then weighted and averaged. Finally, the attraction fit degree evaluation data is output. The data fields include attraction number, theme matching degree, transportation matching degree and comprehensive fit index. This index is used for subsequent recommendation judgment and priority ranking.

[0023] Step S21 includes the following steps: Step S211: Analyze the semantic hierarchical relationship in the keyword frequency evolution data; determine the logical connection nodes of the keyword frequency evolution data based on the semantic hierarchical relationship; Step S212: Based on the logical connection nodes and the semantic hierarchical relationship, evaluate the keyword KL divergence difference of the keyword frequency evolution data, measure the difference in keyword frequency distribution in the context structure, and obtain keyword KL divergence difference data. Step S213: Based on the keyword KL divergence difference data, perform demand hotspot migration structure analysis on the keyword frequency evolution data to obtain the demand hotspot migration structure; Step S214: Analyze the semantic evolution structure in the keyword frequency evolution data based on the demand hotspot migration structure.

[0024] In this embodiment of the invention, the received keyword frequency evolution data is preprocessed. Regular expression-based text cleaning removes non-linguistic characters, URLs, and punctuation marks. A word segmentation dictionary is used for precise segmentation. Then, part-of-speech tagging is performed on the segmented terms. The tagging uses a rule-based prefix-suffix dictionary and word frequency table parallel matching method, with a word frequency threshold of 5. Terms below the threshold are classified as low-frequency candidates. Next, a word co-occurrence matrix is ​​constructed using a sliding window method. The window length is set to 5 terms, and the step size is 1. Any word pair within the same window is counted as a co-occurrence. The co-occurrence count matrix is ​​row-normalized to obtain a conditional probability distribution. Then, semantic hypernym / hypernym relationship analysis is performed. First, the point mutual information value of word pairs is calculated from the co-occurrence conditional probability distribution, with a point mutual information threshold of 1.5 to distinguish strongly semantically related word pairs. Then, the hypernym / hypernym relationship of word pairs that meet the point mutual information threshold is determined. Hypernym / hypernym candidates are... The judgment criteria are based on the occurrence rate of general terms in the context being higher than that of sub-terms. The threshold for the general occurrence rate is set to 0.6. At the same time, it is checked whether the part-of-speech combination of words conforms to the hierarchical structure of noun-noun or noun-phrase combination. If it does, the word pair is marked as hierarchical relationship. All marked hierarchical relationships are put into a directed semantic graph. The nodes of the directed semantic graph are terms, and the edges are hierarchical pointing relationships. The edge weights are taken from the normalized co-occurrence conditional probability. Then, based on the directed semantic graph, the shortest path algorithm based on path length is used to identify logical connection nodes. The criterion for logical connection nodes is that the sum of the in-degree and out-degree of the node is in the top 10 percentile of the node degree distribution of the graph and the frequency of the node as an intermediate node on the hierarchical path is higher than 50 times. Such nodes are identified as logical connection nodes of keyword frequency evolution data. The entire process outputs the marked semantic hierarchical relationship and logical connection node list for subsequent steps.

[0025] In the keyword KL divergence difference assessment stage based on logical connection nodes and semantic hierarchical relationships, the logical node table and the directed graph structure of semantic hierarchical relationships output in step S211 are first extracted. Contextual word frequency distribution information is extracted for the time window of each logical connection node. The relative distribution ratio of word frequencies of the hierarchical and subordinate nodes within the window and their adjacent windows is statistically analyzed. The time window length is set to 10 records, and the step size is 2 records. The frequency of keyword occurrences is counted within each window, and the distribution difference between adjacent windows is calculated. Subsequently, a divergence analysis operation based on logarithmic probability difference is performed. The word frequency distribution of the hierarchical node is used as the baseline distribution, and the distribution of the subordinate node is used as the comparison distribution. The difference between the two is determined by accumulating the logarithmic differences. The divergence result is set to a value range of 0 to 1. The closer the divergence value is to 1, the more significant the semantic shift between contexts. In the experimental parameters, a 1×10⁻⁶ value is used for the data distribution with a probability of 0. -6 The smoothing correction value is used to avoid calculation errors caused by abnormal numerical fluctuations. After the calculation is completed, a record entry containing window index, hypernym identifier, hyponym identifier, divergence value and timestamp is generated for each logical node. Then, the divergence values ​​of all nodes are integrated into the keyword KL divergence difference dataset. The aggregation operation is performed on the dataset to calculate the average divergence value and standard deviation, so as to provide a quantitative basis for subsequent demand hotspot migration structure analysis.

[0026] In the stage of demand hotspot migration structure analysis based on keyword KL divergence difference data of keyword frequency evolution data, the keyword KL divergence difference data obtained in step S212 is first read, and the divergence threshold is defined as 0.5. Nodes with a divergence value higher than this value are regarded as semantic hotspot change nodes. Then, a time series index table is established, and all hotspot change nodes are arranged in ascending order of time. The time difference between adjacent nodes is calculated. If the time difference is less than 3 windows and there is the same superordinate word between the nodes, it is considered to belong to a continuous hotspot chain. This chain is recorded as the initial migration path. In order to identify the direction of hotspot migration, a first-order difference operation is performed on the time series of each migration path. When the word frequency change rate changes from positive to negative or from negative to positive, the corresponding time point is marked with a migration direction indicator. The direction flag value is +1 to indicate topic diffusion and -1 to indicate topic convergence. Then, all migration paths are traversed over the entire time range. For paths with a migration time span of more than 5 windows, length truncation is performed, retaining only the data within the first 4 windows to maintain sample balance. Next, the distribution of the number of hotspot paths within each time window is counted. If the number of hotspot paths in a certain window exceeds 30% of the total number of paths, that time window is set as a hotspot concentration segment. The average divergence difference is calculated for the interconnections between the concentration segments to generate a migration intensity matrix. The matrix dimension is the square of the number of hotspot segments. A depth-first traversal is performed on the matrix to extract the migration chain with the highest frequency, and the data is organized into the required hotspot migration structure data. This data includes the migration start point, end point, direction code, average divergence value, and span value. It's important to explain that performing a first-order difference operation on the time series of a migration path is equivalent to calculating the "amplitude of change" between two adjacent time windows. This is the difference between the word frequency change rate at the later time point and the word frequency change rate at the previous time point. This directly reflects the "increase or decrease trend" of frequency changes, rather than just a static value. In simple terms, it's not just about "how many times the word is mentioned now," but "whether the rate of increase in the number of mentions is accelerating or slowing down." When the word frequency change rate turns from positive to negative, it means that the keyword's frequency was rising in the previous few time periods (positive change), but then it started to decline (negative change), indicating that the attention is decreasing, the semantic focus is decreasing, and the theme is gradually converging. When the word frequency change rate turns from negative to positive, it means that the keyword, which was previously in a state of declining attention, is now on an upward trend, indicating that users are starting to focus on this theme again in a new stage, and the scope and influence of the theme's semantics are expanding, which is a diffusion process.

[0027] In the stage of parsing the semantic evolution structure of keyword frequency evolution data based on the demand hotspot migration structure, the demand hotspot migration structure data generated in step S213 is read. This data structure contains migration start words, end words, and direction information. To establish the semantic evolution structure, a topological mapping of the hotspot migration chain needs to be performed on the feature space. First, a keyword index matrix is ​​created, with rows corresponding to keyword numbers and columns corresponding to time window numbers. The matrix elements store their frequency normalization values. Then, semantic connectivity calculation is performed based on the migration chain defined in the demand hotspot migration structure. For each migration path, the semantic vector similarity between the start word and the end word is calculated. A 300-dimensional word vector representation is used, and the cosine similarity value is calculated. The similarity value is multiplied by the migration direction code to generate a weighted semantic transfer value. When the weighted result is greater than 0.6, it indicates that the topic evolution is stable. In this case, the starting word and ending word are connected by an undirected edge and weighted as the transition value; otherwise, a directed edge with convergence direction is established. This process is repeated until all transition structures are mapped, resulting in a semantic evolution network containing keyword nodes, transition directions, and connection weights. To extract the evolutionary backbone structure, a shortest path search algorithm is executed on the generated network, using the node with the highest semantic concentration as the root node, and gradually expanding outwards. For each layer traversed, the difference between the out-degree and in-degree of the node is calculated. If the out-degree is greater than the in-degree, it is marked as a semantic expansion node; otherwise, it is marked as a semantic continuation node. Finally, the semantic evolution structure data composed of nodes, edges, and direction labels is output. Its fields include node number, level index, connection weight, direction code, and evolution depth, which are used to represent the global semantic change process of keyword frequency evolution data.

[0028] Step S22 includes the following steps: Step S221: Perform semantic tree hierarchical analysis on the semantic evolution structure to obtain the semantic evolution tree hierarchy; Step S222: Calculate the depth and branching degree of each level in the semantic evolution tree hierarchy; evaluate the backbone semantic structure in the semantic evolution tree hierarchy based on the depth and branching degree; Step S223: Decompose the semantic evolution coherence process of the main semantic structure to obtain the semantic evolution coherence data of the main semantic structure; Step S224: Construct a focused topic background for the main semantic structure to obtain the focused topic background of the main structure; Step S225: Based on the main semantic evolution coherence data and the main focused theme background, infer the user's play intention.

[0029] In this embodiment of the invention, during the semantic tree hierarchical analysis stage of the semantic evolution structure, the semantic evolution structure data output in step S214 is first read. This data includes node numbers, time indices, hierarchical indices, connection weights, and direction codes. During the analysis, a semantic evolution tree is constructed based on the connection weights and time indices between nodes. A weight-first breadth-first traversal algorithm is used as the tree construction method. Starting from the root node with the largest weight, the tree expands in the direction of increasing time indices. In each traversal, nodes with connection weights not less than 0.4 are added first, while nodes with weights less than 0.2 are not added. Nodes are not included in the tree structure. The maximum width of the tree is set to 15 nodes, and the maximum depth is limited to 8 levels. After traversal, a complete semantic evolution tree is obtained. The parent-child node relationship of the tree clearly indicates the evolution path of semantics from the upper-level concept to the lower-level subdivision, which is the basis for subsequent hierarchical analysis calculations. Then, all nodes are recorded hierarchically according to the node time index to form a semantic evolution tree hierarchical data table. Each record in the table includes the layer number, node ID, parent node ID, connection weight and direction code, and identifies the number of nodes in each layer for subsequent depth and branch degree analysis.

[0030] In the stage of calculating the depth and branching degree of each level in the semantic evolution tree hierarchy, the semantic evolution tree hierarchy data table generated in step S221 is first traversed and statistically analyzed. The root layer depth is defined as 1. For each level, the depth value is determined by recursively adding 1 according to the parent-child link direction. When calculating the branching degree, the number of child nodes of each parent node is defined as its branching degree. The branching degree of all nodes is recorded and the average branching degree index of each level is obtained. Then, nodes whose depth level and branching degree index differ by more than twice the standard deviation are considered as abnormal branch nodes and are excluded. After that, the hierarchy expansion curve is plotted based on the hierarchy depth distribution and the expansion rate index value is calculated. This value is obtained by weighting the difference in average branching degree between adjacent layers and the total number of node connections. When the expansion rate is stable in the range of 0.1 to 0.2, the segment is determined to be the trunk segment. The set of nodes containing 3 consecutive layers in the segment is defined as the trunk semantic structure layer. Through this calculation and screening of hierarchy features, the core main path of the semantic evolution tree is extracted, laying the data structure foundation for subsequent semantic coherence decomposition.

[0031] In the semantic evolution coherence decomposition stage of the main semantic structure, the main semantic structure layer determined in step S222 is invoked. Within this layer, node sequences are extracted in chronological order, and semantic similarity, time interval, and directional consistency between adjacent nodes are calculated. Semantic similarity is represented by the cosine similarity of 300-dimensional word vectors, with a similarity threshold set to 0.65. The time interval is obtained by the difference between the node time indices. Directional consistency is assigned a value of 1 or -1 based on whether the direction codes are the same or opposite. Subsequently, a coherence index is calculated for each pair of adjacent nodes, and the semantic similarity and directional consistency are combined. The consistency, the inverse of the time interval, and the weighted average are calculated with weights of 0.5, 0.3, and 0.2, respectively. After calculation, a coherence distribution curve is formed on the continuous node sequence. The length of the curve is equal to the number of backbone nodes minus 1. Segments with a coherence of less than 0.3 are broken. The backbone semantic path is divided into several coherent segments, and the average similarity and time span of each segment are calculated to generate backbone semantic evolution coherence data. Each record contains the start ID, end ID, coherence value, segment length, and corresponding time period to characterize the temporal continuity of the semantic backbone evolution process.

[0032] In the stage of constructing the focused topic background for the main semantic structure, the main semantic structure layer and the main semantic evolution coherence data obtained in step S223 are read. The node group with the highest coherence value in the main path is selected as the candidate topic core group, and the number of candidate groups is set to 5. Then, word frequency resampling operation is performed on the nodes in these core groups to count the standardized frequency of keywords in adjacent time periods. The frequency curve is smoothed by Gaussian filtering with a filtering window width of 3 time units. After smoothing, the peak position is identified, and the time node where the peak occurs is taken as the topic focus point. To construct the focused topic background, the superordinate semantic words and subordinate dependent words corresponding to each focus node are extracted to form the topic context. The frequency of their occurrence relationship is calculated, with a weight ratio of 0.7 for superordinate words and 0.3 for subordinate words. The context word vectors are weighted and averaged to generate the topic background expression vector. This vector has a dimension of 300 and includes the semantic direction and the coordinates of the topic distribution center. Finally, the topic focus, relevant time interval, background vector and weight set are combined to form the main focused topic background dataset.

[0033] In the stage of inferring user play intentions based on the semantic evolution coherence data and the thematic background of the main framework, the results data of steps S223 and S224 are read. The time span index in the coherence data is aligned with the focus time point in the thematic background data. If the time overlap ratio is greater than 0.5, a play intention line is established in the semantic evolution framework. Verb keywords, such as "visit," "stroll," "experience," and "take photos," are extracted from the node set within the intention line. The verb density index is calculated by statistically analyzing the frequency of these action nodes and their concentration in the time series. The value is set to 0.2. When the verb density is higher than the threshold, the behavioral driving factor of the intent line is defined as strong behavioral orientation; otherwise, it is marked as weak behavioral orientation. Then, the interest domain label is determined by combining the higher-level semantic label in the background of the focused topic and combined with the behavioral driving type to generate the play intention code. The encoding structure is a concatenation of interest type code and behavioral orientation code. Each code represents a specific user play tendency. Finally, the user play tendency intent data table is output, which includes play intention code, associated topic ID, behavioral orientation type, time span and focus intensity index, for subsequent tendency intent clustering and attraction adaptation analysis.

[0034] Step S25 includes the following steps: Step S251: Extract the game background theme, historical traffic conditions, and geographical coordinates from the basic information; Step S252: Select scene adaptability based on the theme of the tour background according to the clustering data of the tendency intention, and obtain the scenic spot background adaptability selection data; Step S253: Extract travel time tolerance from the tendency intention clustering data; perform spatiotemporal traffic accessibility matching on historical traffic conditions and geographic coordinates based on the travel time tolerance to obtain scenic spot traffic adaptation data; Step S254: Perform bivariate coupling analysis based on the scenic spot background adaptation selection data and scenic spot traffic adaptation data to obtain bivariate adaptation coupling data; Step S255: Determine the degree of adaptation of urban attractions based on the bivariate adaptation coupling data.

[0035] In this embodiment of the invention, during the stage of extracting the basic information such as the theme of the visit, historical traffic conditions, and geographical coordinates, the basic information records of the attractions are first retrieved from a publicly available urban attraction database. Each record includes the attraction number, a set of thematic keywords, opening hours, geographical coordinates (longitude and latitude), historical traffic speed sequence, and surrounding traffic network nodes. Thematic keywords are extracted using a regular expression matching algorithm, filtering out auxiliary words containing non-noun parts of speech, and retaining only descriptive words to form the theme of the visit. Subsequently, the historical traffic conditions are standardized by time series, and the traffic flow records for five consecutive years are divided into hourly units. The location is divided into 8760 time points. The average hourly traffic speed is calculated and a time distribution vector is generated. To ensure data stability, sudden outliers are corrected using median smoothing. The outlier detection threshold is the mean plus or minus three times the standard deviation. The historical traffic state vector is obtained. Then, the geographic coordinates are processed using coordinate system one. The latitude and longitude are transformed into rectangular coordinates by spherical projection. The coordinate unit is kilometers, and the precision is controlled to three decimal places. The extracted travel background theme, the smoothed historical traffic state data, and the standardized geographic coordinates are combined to form the output dataset of step S251.

[0036] In the stage of selecting scene adaptability for the game background theme based on the tendency intention clustering data, the tendency intention clustering data obtained in step S23 is first imported. Each record contains the intention cluster number, semantic center vector, interest domain encoding, and behavior orientation type. The game background theme text sequence extracted in step S251 is compared with the theme keywords, and each keyword is segmented and encoded, converting each word into a 300-dimensional word vector. Then, a game background semantic matrix is ​​constructed, where rows represent keywords and columns represent vector dimensions. Next, the cosine similarity matrix between the game background semantic matrix and the tendency intention cluster center vector is calculated. The result ranges from 0 to 1. Then, attention weighting is performed, and weights are assigned according to the behavior orientation type parameter. The weight parameter is defined as 0.6 for strong behavior orientation and 0.4 for weak behavior orientation. The weights are applied to the similarity matrix and weighted summation is performed to obtain the scene fit score. All attractions are sorted and the mean and standard deviation of the scores are calculated. Records with scores higher than 0.5 times the standard deviation of the mean are retained as high-fit samples, generating attraction background fit selection data. Each data point includes attraction number, semantic similarity value, behavior orientation weight value and total fit score, which are used for subsequent bivariate analysis matching.

[0037] In the stage of extracting travel time tolerance from the intention clustering data and performing spatiotemporal traffic accessibility matching, user travel time preference parameters recorded in the intention clustering data are read. These parameters include the shortest tolerance time, the longest acceptable travel time, and the average travel speed preference, in minutes and kilometers per hour. These parameters are converted into traffic time threshold ranges, with the upper and lower limits corresponding to the shortest and longest tolerance times, respectively. Then, the historical traffic state vector and geographic coordinate data generated in step S251 are used to calculate the shortest path travel time between various attractions within the city. The Dijkstra algorithm is used, with historical average travel speed as the edge weight. The results are output in minutes. Then, the path time is compared with the travel time tolerance range. If the median time falls within the range, it is rated as reachable. If it is higher than the upper limit, it is marked as exceeding the tolerance range. If it is lower than the lower limit, it is marked as highly convenient. The scoring range of 0 to 1 is mapped proportionally to form a spatiotemporal traffic accessibility score matrix. The geographic coordinate distance factor is multiplied by a decay coefficient of 0.9 to compensate for the spatial influence, thereby generating scenic spot traffic adaptation data. Each record contains scenic spot number, accessibility score, travel time preference index and distance decay value.

[0038] In the bivariate coupling analysis stage based on the selection data for scenic spot background adaptation and the scenic spot traffic adaptation data, the two datasets obtained in steps S252 and S253 are read. Data with the same scenic spot number are aligned by key value to establish a bivariate comparison table, where variable one is the background adaptation score and variable two is the traffic adaptation score. Dimensionless processing is performed on the two variables, and all values ​​are mapped to the interval [0,1] through max-min normalization. After normalization, the data is linearly weighted and synthesized. The weight coefficient of the background adaptation data is set to 0.6, and the weight coefficient of the traffic adaptation data is set to 0.4. Feature quantization embedding is performed on the weighted data. The theme tag strength and scene semantic vector of the scenic spot background semantic features are vector-concatenated with the travel distance features and travel time indicators in the scenic spot traffic adaptation data to form a bivariate feature matrix with a matrix dimension of M×4, where M is the number of samples. Then, the eigenvalues ​​of the principal components of the matrix are calculated, and the coupling strength coefficient is calculated by the eigenvector projection method with a threshold range of 0 to 1. The coupling strength coefficients of all scenic spot samples are saved as bivariate adaptation coupling data for final adaptation degree evaluation.

[0039] In the stage of determining the suitability of urban attractions based on bivariate adaptive coupling data, the bivariate adaptive coupling data obtained in step S254 is read, and the coupling strength coefficients of all attractions are aggregated and calculated. The mean, standard deviation, and maximum and minimum value ranges are calculated. A grading threshold is set with the mean as the center. When the coupling value is greater than the mean plus 0.5 times the standard deviation, it is defined as a high-fit layer; when it is greater than the mean minus 0.5 times the standard deviation and less than the mean plus 0.5 times the standard deviation, it is defined as a medium-fit layer; and when it is lower than the mean minus 0.5 times the standard deviation, it is defined as a low-fit layer. Each attraction is assigned to the corresponding level, and the final fit index is further weighted and corrected by combining the theme similarity ratio. The weighting ratio is 0.7 for theme matching degree and 0.3 for traffic coupling value. The final calculated comprehensive fit index ranges from 0 to 1. The index value and the attraction number are recorded together in the fit results and used for judging and ranking the recommendation priority of urban attractions.

[0040] Step S252 includes the following steps: The intention clustering data is subjected to multi-level intention decoupling and vectorization encoding to obtain intention-level feature vector groups; The aforementioned travel background theme is decomposed into scenic spot scene elements to obtain scenic spot scene decomposition elements; Attention mechanism matching is performed on scenic spot scene decomposition elements based on intent-level feature vector groups to obtain intent-scene association matching data. Perform feature cross-projection processing on the intent scene association matching data to generate intent scene matching features; Scene adaptation is selected based on the intent scene matching features to obtain scenic spot background adaptation selection data.

[0041] In this embodiment of the invention, during the multi-level intent decoupling and vectorization encoding process of the intent clustering data, the intent clustering data obtained in step S23 is first imported into the computational processing system. The data includes intent cluster numbers, interest topic codes, semantic center vectors, behavioral guidance labels, and time weight values. Each record is divided into three analysis units: interest layer, behavioral layer, and time layer. The field data of each layer is processed independently. The interest layer uses the word embedding vector averaging method to calculate the topic center representation. The 300-dimensional vectors of all relevant keywords are averaged and normalized to obtain the interest topic vector. The behavioral layer converts the behavioral guidance labels into one-hot codes, and the vector length is... The vector is 10-dimensional, with each dimension corresponding to a specific action category, such as sightseeing, shooting, resting, and experiencing. The time weight values ​​of the time layer are normalized according to the time window number, with a standard deviation of 0.25 and a mean of 0 for amplitude adjustment, so that the weights of different time periods participate in the calculation on the same scale. Then, a layered concatenation method is used to sequentially concatenate the vectors of the interest layer, behavior layer, and time layer into a composite vector. Each intent cluster outputs one composite feature record. Finally, an intent-level feature vector group is constructed, where each vector has a uniform length of 310 dimensions, including three parts: topic representation, behavior distribution, and time factor, which serve as the high-dimensional input basis for subsequent scene matching processing.

[0042] In the stage of decomposing the scenic spot scene elements of the main theme of the tour, the main theme text of the tour generated in step S251 is extracted. First, a word segmentation system based on the bidirectional maximum matching algorithm is used to parse the word groups. The window width is set to 5 characters. Nouns and adjectives with a frequency higher than the threshold of 0.03 in the text are extracted to generate a preliminary set of scene words. Then, through part-of-speech filtering, three categories of words are retained: location features, activity features, and atmosphere features. Among them, location features include natural words such as mountain, lake, and park; activity features include verbs such as hiking, sightseeing, and taking pictures; and atmosphere features are... The surrounding features include descriptive words such as tranquil, lively, and romantic. Then, the weight of each word in the entire text is calculated. The weight value is defined as the word frequency divided by the total number of words and then multiplied by the inverse document frequency. Finally, the top 30% of words are selected by weight ranking to form a set of key elements of the scenic spot scene. Each element is semantically classified, numbered, and standardized and vectorized, and converted into a 300-dimensional floating-point array. The category label and word weight factor are recorded at the same time. The output result is the set of decomposed elements of the scenic spot scene, where each record corresponds to an element name, category number, semantic vector, and weight value.

[0043] In the attention mechanism matching stage of scenic spot scene decomposition elements based on intent-level feature vector groups, the generated intent-level feature vector groups and the obtained scenic spot scene decomposition element set are first input into the matching process. The dot product correlation value of each intent vector and scene element vector is calculated, and the correlation threshold is set to 0.5. Weighted normalization is performed on all correlation values. The weights are determined by the time layer factor in the intent vector. The weight range of 0 to 1 represents the perceived intensity of the intent to the time window. After normalization, an attention distribution matrix is ​​established. The matrix rows correspond to intent vectors, the columns correspond to scene elements, and the matrix values ​​represent the intensity of attention allocation. Then, through a weighted aggregation operation, the weight of each row in the matrix is ​​multiplied by the corresponding scene vector and summed to obtain the response vector of the intent to the scene. This response vector reflects the attention ratio of a certain intent in all scenic spot background elements. Then, the density value of the response vectors of all intents in the semantic space is calculated. Intents with a density higher than 0.6 are identified as the main attention intents. The intent-scene association matching data is output, and its fields include intent ID, scene element ID, correlation value, attention coefficient, and weighted response vector.

[0044] In the stage of feature cross-projection processing of intent scene association matching data, the output intent scene association matching data is called, and a linear orthogonal transformation is performed on the response vector of each record, mapping the 300-dimensional vector to two independent feature subspaces. Subspace 1 represents the topic similarity axis, and subspace 2 represents the behavior orientation axis. The projection basis is defined as the mean direction of the interest part and the behavior part in the intent-level feature vector group. The projection operation adopts the principal axis projection method, and the projection value is calculated for each response vector to obtain two sets of feature components. Then, the cross mutual information index between topic features and behavior features is calculated. Based on the mutual information result, the features of the two subspaces are merged into a set of cross feature vectors with a length of 60 dimensions, where the first 30 dimensions represent the topic semantic distribution and the last 30 dimensions represent the behavior tendency distribution. In order to prevent the bias caused by the difference in feature scale, the output vector is normalized to zero mean, so that the overall distribution is balanced. The output result is used as the intent scene matching feature data, and its fields include intent number, topic component, and behavior component.

[0045] In the stage of scene adaptation selection based on intent scene matching features, the output intent scene matching feature data is read, and the scene adaptation score is calculated based on the theme component and behavior component of each scenic spot. The weight of the theme dimension is defined as 0.55 and the weight of the behavior dimension is 0.45. Weighted summation is performed on all intent samples to obtain the comprehensive matching value range of 0 to 1 for each scenic spot. Then, all matching value sequences are summarized by scenic spot number, and the mean and standard deviation of each scenic spot are calculated. The Z-score standardization method is used to transform the data into comparable values. Then, scenic spots with a standardized score greater than 0.6 are identified as high-matching objects. The identification is then mapped to the original intent type to form an adaptation assignment mapping. The mapping information of all high-matching scenic spots constitutes the scenic spot background adaptation selection dataset. The data fields include scenic spot number, comprehensive matching value, theme intensity value, behavior relevance and sample size statistics, providing a high-accuracy input data source for the subsequent bivariate coupling analysis in step S252.

[0046] Step S254 includes the following steps: The background adaptation selection data and the traffic adaptation data of the scenic spots are processed to be dimensionless, and the data are mapped to a preset closed interval [0, 1] to obtain background adaptation normalized data and traffic adaptation normalized data respectively. Feature quantization and embedding processing is performed on the background adaptation normalized data and the traffic adaptation normalized data. The theme semantic features and scene association strength features in the scenic spot background adaptation selection data are vectorized and rearranged with the spatiotemporal accessibility features and traffic tolerance mapping features in the scenic spot traffic adaptation data to construct bivariate feature coupled data. Perform coupling strength inference processing on the bivariate feature coupling data to obtain bivariate coupling strength data; The bivariate coupling strength data is subjected to multidimensional aggregation processing to obtain the bivariate adaptive coupling data.

[0047] In this embodiment of the invention, the scenic spot background adaptation selection data and scenic spot traffic adaptation data generated in steps S252 and S253 are called, and the numerical fields are extracted respectively. The former includes the comprehensive matching value, theme intensity value and scene association strength, and the latter includes the accessibility score, travel distance attenuation value and travel time preference index. The total number of samples is defined as N. Each field value is mapped to the interval [0,1] using the minimum-maximum normalization method. During the calculation, the minimum and maximum values ​​are calculated for each field. The original value is subtracted from the minimum value and then divided by the difference in the numerical range, which realizes the dimensionless standard conversion. In order to maintain data consistency, uniform precision control is performed on all mapped values, and 4 decimal places are retained. At the same time, the distribution test is performed on the normalized results. The test method is Kolmogorov-Smirnov statistics. If the skewness coefficient of the field deviating from the normal distribution exceeds 0.2, the square root transformation method is used to balance the distribution. Finally, two normalized result data are output, named background adaptation normalized data and traffic adaptation normalized data respectively, and stored in a unified data structure for quantitative embedding processing.

[0048] In the feature quantization and embedding process of background-adapted normalized data and traffic-adapted normalized data, the weight coefficient of the background-adapted normalized data is set to 0.6, and the weight coefficient of the traffic-adapted normalized data is set to 0.4. Based on these two, a feature synthesis matrix is ​​established. The matrix contains multi-dimensional feature concatenation records corresponding to each scenic spot. The background feature part includes theme semantic features and scene association strength features, and the traffic feature part includes spatiotemporal accessibility features and traffic tolerance mapping features. In the vectorization and rearrangement process, the theme semantic features are first subjected to word vector averaging transformation. The 300-dimensional theme word vectors are averaged and reduced to 30-dimensional principal component representation. Then, the scene association strength is converted into a 10-dimensional measurable index by weighted average according to the node adjacency degree. This index is concatenated with the traffic features to form a 40-dimensional composite feature vector. Each vector is assigned a scenic spot number index for easy subsequent retrieval. All vectors are arranged by number to form bivariate feature coupling data, where rows represent sample scenic spots and columns represent the joint feature space of background and traffic.

[0049] In the stage of performing coupling strength inference processing on bivariate feature coupling data, the bivariate feature coupling data is input into the vector strength calculation process. Linear correlation detection is performed on the 40-dimensional vector of each sample scenic spot. The Pearson correlation coefficient matrix between each dimension feature is calculated. The matrix dimension is 40×40. Then, the elements with an absolute value greater than 0.6 along the diagonal are counted as strongly correlated feature pairs. For each sample, the product of the number of strongly correlated feature pairs and the average correlation is calculated to generate the initial value of coupling strength. To improve stability, the initial value is processed by three moving averages with a window length of 5 samples. Then, the least squares fitting curve is used to calculate the smoothing trend. The slope of the curve represents the rate of change of coupling strength. The smoothed coupling value and the rate value are combined into a coupling strength coefficient to identify each scenic spot. Finally, bivariate coupling strength data is generated, which includes fields such as scenic spot number, coupling coefficient, strong correlation count and rate of change, to quantify the degree of aggregation of interaction between the background and traffic dimensions.

[0050] In the multidimensional aggregation process of the bivariate coupling strength data, the bivariate coupling strength data calculated in the previous step is first grouped and aggregated according to the scenic spot number. After averaging the data within the group, global statistical features are calculated, including mean, standard deviation, and range. At the same time, a multidimensional projection matrix is ​​constructed, with the scenic spot coupling coefficient, background normalization value, traffic normalization value, and rate of change as the four principal axes. Principal component analysis is performed to extract the two most important feature axes, and the cumulative contribution rate is set to 90% as the cutoff standard. In this two-dimensional coordinate space, the coupling similarity of each pair of scenic spots is calculated using Euclidean distance. Samples with similarity greater than 0.7 are grouped into the same cluster unit. The center vector of each unit is counted and its weighted average strength is calculated. The multidimensional aggregation result is output with the mean and cluster radius as variables. All units are integrated to form the final bivariate fit coupling data. This data consists of scenic spot number, coupling center coordinates, principal axis feature values, and final composite strength, which is used to support the quantitative determination of the overall fit of urban scenic spots in step S255.

[0051] Step S3 includes the following steps: Step S31: Normalize the adaptation degree of urban attractions to obtain normalized adaptation degree data; Step S32: Extract the adaptation judgment logic from the adaptation degree normalization data; Step S33: Use the Q-learning reinforcement learning algorithm in artificial intelligence technology to learn the optimal recommendation strategy for the adaptation judgment logic and output the adaptation recommendation strategy; Step S34: Embed the adaptation recommendation strategy into the terminal to perform city attraction recommendation management.

[0052] In this embodiment of the invention, during the normalization process of the urban attraction suitability level, the attraction suitability data output in step S255 is first imported into the data standardization process. This data structure includes attraction number, comprehensive suitability index, theme matching degree, transportation matching degree, and aggregation weight parameters. To achieve consistent calculations across different feature scales, the comprehensive suitability index is used as the central variable for dimensionless standardization transformation. The minimum and maximum values ​​of each data field are calculated, and these values ​​are used as boundaries to map the original values ​​to the interval [0,1], ensuring that all samples are on a uniform scale. For comparison, a normal distribution fit test is performed on the mapped distribution. Fields with a skewness exceeding 0.15 are smoothed using logarithmic smoothing to adjust for extreme values. To prevent the accumulation of calculation errors, all values ​​are retained to four decimal places. Weight control parameters are introduced during the normalization process, with the weight of theme matching degree set to 0.7 and the weight of traffic matching degree set to 0.3. The weighted results of the two are then normalized again to obtain the final normalized data of the fit degree. The fields include the attraction number, normalized fit value, adjusted theme proportion coefficient, and traffic correction coefficient.

[0053] In the adaptation judgment logic stage of extracting adaptation-normalized data, the adaptation-normalized data generated in step S31 is read. Using the comprehensive adaptation value as the target variable, the topic proportion coefficient and traffic correction coefficient are used as input variables to perform rule extraction analysis based on a decision tree. The information gain criterion is used as the node splitting standard, with a threshold set to 0.03. The maximum tree depth is 10 levels, and each branch contains at least 15 samples. When constructing the tree, the data is first divided into three equal intervals corresponding to high, medium, and low categories based on the comprehensive adaptation value. The information gain value is calculated and recorded for each splitting node. The splitting condition is to stop splitting when the information gain is below a threshold. After the tree is built, all leaf node path rules are extracted. Each rule consists of multiple condition thresholds. For example, when the topic proportion coefficient is ≥0.65 and the traffic correction coefficient is ≥0.5, the adaptation level is high. The extracted conditions are linked into a list and encoded into a rule set. Then, the co-occurrence frequency of each condition is calculated through correlation matrix analysis. Condition pairs with an occurrence frequency greater than 0.4 are defined as stable logical association pairs. The dataset is then consolidated to form an adaptation judgment logic dataset, which includes rule number, logical condition string, output level weight, and support index.

[0054] In the optimal recommendation strategy learning stage of the adaptation judgment logic using the Q-learning reinforcement learning algorithm, the adaptation judgment logic dataset obtained in step S32 is first input into the reinforcement learning process. The state space is defined as the adaptation level distribution of each attraction at different time periods, with the number of states set to 200. The action space is defined as whether to recommend the attraction, with values ​​in binary form 0 or 1. The initial Q-table size is the number of states multiplied by the number of actions, i.e., 200 × 2. All Q values ​​are initialized to 0. The reward function is standardized to the interval [-1, 1] based on user satisfaction ratings. During the learning process, the learning rate α is set to 0.15, the discount factor γ to 0.85, and the exploration rate ε to 0.1. In each iteration, an action is randomly selected and the Q value is updated. The update method is based on the Bellman equation to calculate the current reward and the maximum expected future return. The average reward change is statistically analyzed every 100 iterations. If the change is less than 1 × 10⁻⁶, the reward is considered lost. - ³ indicates convergence. The total number of stable iterations is approximately 5000, ultimately generating an optimal recommendation policy table. The table records the optimal Q-value and corresponding expected return when performing recommended or non-recommended actions in each state. The action with the maximum value is regarded as the policy output item, and the output result is defined as the adaptive recommendation policy, which includes the state index, action type, and optimal action Q-value.

[0055] In the stage of embedding the adaptive recommendation strategy into the terminal to perform city attraction recommendation management, the adaptive recommendation strategy obtained in step S33 is imported into the terminal execution system. The strategy table is stored as a structured data file with fields including state index, recommended action, and optimal benefit value. When the terminal receives new user input text during runtime, it first generates the current state vector based on the keyword extraction in step S1 and the intent clustering results in step S2. The state vector is matched with the corresponding state index number according to the adaptation judgment rule. The optimal action item is found by calling the strategy table through the index. If the optimal action is a recommendation, it is entered into the attraction candidate queue for sorting; otherwise, it is marked as a reserved item. Then, the Q value of all attractions with recommended actions is taken as the final scoring basis. The scoring results are arranged in descending order to form a recommendation order table. When the recommendation is displayed, the comprehensive adaptation index is used as the explicit sorting parameter. At the same time, the feedback data of each recommendation is saved and written back to the cache for subsequent strategy retraining and parameter correction. Throughout the process, data transmission is encapsulated in a standardized JSON structure, thereby realizing a complete closed loop of terminal execution management and recommendation decision-making.

[0056] The present invention also provides an artificial intelligence-based urban attraction recommendation and management system for executing the artificial intelligence-based urban attraction recommendation and management method described above. The artificial intelligence-based urban attraction recommendation and management system includes: The data acquisition module is used to receive the scenic spot demand text input by the user and determine the keyword dataset in the scenic spot demand text; perform word frequency evolution analysis on the keyword dataset to obtain keyword frequency evolution data; The adaptation analysis module is used to infer users' travel intentions based on the keyword frequency evolution data, and perform clustering processing to obtain intention clustering data; and to determine the degree of adaptation of city attractions based on the intention clustering data. The recommendation strategy learning module is used to extract the adaptation judgment logic in the city attraction adaptation degree, and then learn the optimal recommendation strategy to output the adaptation recommendation strategy; the adaptation recommendation strategy is embedded into the terminal to perform city attraction recommendation management.

[0057] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for recommending and managing urban attractions based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Accept the scenic spot requirement text input by the user and determine the keyword dataset in the scenic spot requirement text; perform word frequency evolution analysis on the keyword dataset to obtain keyword frequency evolution data; Step S2: Based on the keyword frequency evolution data, deduce the user's travel intention and perform clustering processing to obtain intention clustering data; determine the suitability of city attractions based on the intention clustering data; Step S3: Extract the adaptation judgment logic from the city attractions' suitability, then learn the optimal recommendation strategy and output the suitability recommendation strategy; The adaptive recommendation strategy is embedded into the terminal to perform urban attraction recommendation management.

2. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Accept the user's input of attraction request text; Step S12: Mark the parts of speech in the scenic spot requirement text, and determine the lexical relationships in the scenic spot requirement text based on the parts of speech; Step S13: Determine the keyword dataset in the scenic spot demand text based on the lexical relationships; Step S14: Perform word frequency evolution analysis on the keyword dataset through the lexical relationships to obtain keyword frequency evolution data.

3. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Analyze the semantic evolution structure in the keyword frequency evolution data; Step S22: Derive the user's play intention based on the semantic evolution structure; Step S23: Cluster the user's play intentions to obtain intention clustering data; Step S24: Obtain basic information about the current city's attractions; Step S25: Determine the suitability of urban attractions based on the basic information and clustering data of current urban attractions.

4. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 3, characterized in that, Step S21 includes the following steps: Step S211: Analyze the semantic hierarchical relationship in the keyword frequency evolution data; determine the logical connection nodes of the keyword frequency evolution data based on the semantic hierarchical relationship; Step S212: Based on the logical connection nodes and the semantic hierarchical relationship, evaluate the keyword KL divergence difference of the keyword frequency evolution data, measure the difference in keyword frequency distribution in the context structure, and obtain keyword KL divergence difference data. Step S213: Based on the keyword KL divergence difference data, perform demand hotspot migration structure analysis on the keyword frequency evolution data to obtain the demand hotspot migration structure; Step S214: Analyze the semantic evolution structure in the keyword frequency evolution data based on the demand hotspot migration structure.

5. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 4, characterized in that, Step S22 includes the following steps: Step S221: Perform semantic tree hierarchical analysis on the semantic evolution structure to obtain the semantic evolution tree hierarchy; Step S222: Calculate the depth and branching degree of each level in the semantic evolution tree hierarchy; evaluate the backbone semantic structure in the semantic evolution tree hierarchy based on the depth and branching degree; Step S223: Decompose the semantic evolution coherence process of the main semantic structure to obtain the semantic evolution coherence data of the main semantic structure; Step S224: Construct a focused topic background for the main semantic structure to obtain the focused topic background of the main structure; Step S225: Based on the main semantic evolution coherence data and the main focused theme background, infer the user's play intention.

6. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 3, characterized in that, Step S25 includes the following steps: Step S251: Extract the game background theme, historical traffic conditions, and geographical coordinates from the basic information; Step S252: Select scene adaptability based on the theme of the tour background according to the clustering data of the tendency intention, and obtain the scenic spot background adaptability selection data; Step S253: Extract travel time tolerance from the tendency intention clustering data; perform spatiotemporal traffic accessibility matching on historical traffic conditions and geographic coordinates based on the travel time tolerance to obtain scenic spot traffic adaptation data; Step S254: Perform bivariate coupling analysis based on the scenic spot background adaptation selection data and scenic spot traffic adaptation data to obtain bivariate adaptation coupling data; Step S255: Determine the degree of adaptation of urban attractions based on the bivariate adaptation coupling data.

7. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 6, characterized in that, Step S252 includes the following steps: The intention clustering data is subjected to multi-level intention decoupling and vectorization encoding to obtain intention-level feature vector groups; The aforementioned travel background theme is decomposed into scenic spot scene elements to obtain scenic spot scene decomposition elements; Attention mechanism matching is performed on scenic spot scene decomposition elements based on intent-level feature vector groups to obtain intent-scene association matching data. Perform feature cross-projection processing on the intent scene association matching data to generate intent scene matching features; Scene adaptation is selected based on the intent scene matching features to obtain scenic spot background adaptation selection data.

8. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 6, characterized in that, Step S254 includes the following steps: The background adaptation selection data and the traffic adaptation data of the scenic spots are processed to be dimensionless, and the data are mapped to a preset closed interval [0, 1] to obtain background adaptation normalized data and traffic adaptation normalized data respectively. Feature quantization and embedding processing is performed on the background adaptation normalized data and the traffic adaptation normalized data. The theme semantic features and scene association strength features in the scenic spot background adaptation selection data are vectorized and rearranged with the spatiotemporal accessibility features and traffic tolerance mapping features in the scenic spot traffic adaptation data to construct bivariate feature coupled data. Perform coupling strength inference processing on the bivariate feature coupling data to obtain bivariate coupling strength data; The bivariate coupling strength data is subjected to multidimensional aggregation processing to obtain the bivariate adaptive coupling data.

9. The urban scenic spot recommendation and management method based on artificial intelligence according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Normalize the adaptation degree of urban attractions to obtain normalized adaptation degree data; Step S32: Extract the adaptation judgment logic from the adaptation degree normalization data; Step S33: Use the Q-learning reinforcement learning algorithm in artificial intelligence technology to learn the optimal recommendation strategy for the adaptation judgment logic and output the adaptation recommendation strategy; Step S34: Embed the adaptation recommendation strategy into the terminal to perform city attraction recommendation management.

10. An artificial intelligence-based urban attraction recommendation and management system, characterized in that, For executing the AI-based urban attraction recommendation and management method as described in claim 1, the AI-based urban attraction recommendation and management system includes: The data acquisition module is used to receive the scenic spot demand text input by the user and determine the keyword dataset in the scenic spot demand text; perform word frequency evolution analysis on the keyword dataset to obtain keyword frequency evolution data; The adaptation analysis module is used to infer users' travel intentions based on the keyword frequency evolution data, and perform clustering processing to obtain intention clustering data; and to determine the degree of adaptation of city attractions based on the intention clustering data. The recommendation strategy learning module is used to extract the adaptation judgment logic in the city attraction adaptation degree, and then learn the optimal recommendation strategy to output the adaptation recommendation strategy; the adaptation recommendation strategy is embedded into the terminal to perform city attraction recommendation management.

Images