Tourism content recommendation system for analyzing favorites of tourism scenic spots

Abstract

The invention relates to the technical field of text travel data analysis, in particular to a text travel content recommendation system for analyzing preferences of text travel scenic spots, which comprises a culture gene bank construction and user interest modeling unit and a dynamic situation awareness and recommendation optimization unit. The data acquisition and preprocessing module extracts culture elements from a multi-channel text through optical character recognition and natural language processing, the culture gene extraction and fusion module constructs a knowledge graph and fuses similar elements to form a gene cluster, and the user interest model training module generates interest portraits based on user text travel platform data. In a dynamic situation awareness and recommendation optimization unit, a situation data acquisition and analysis module acquires and analyzes user situation data, a recommendation strategy generation and adjustment module generates a recommendation strategy according to a user state and interest by using a deep Q network algorithm, recommendation is optimized in combination with situation factors, and the accuracy of travel content recommendation is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cultural tourism data analysis, and in particular to a cultural tourism content recommendation system for analyzing preferences for cultural tourism attractions. Background Art

[0002] Cultural tourism data analysis is an important technology. As the cultural tourism industry is booming, users are faced with a huge amount of cultural tourism attractions and content information, from which it is difficult to quickly filter out options that suit their interests. Traditional cultural tourism recommendation systems are mainly based on simple user ratings, browsing history or popular lists, and have many limitations.

[0003] For example, when recommending attractions in a historical and cultural city, it is far from enough to judge solely based on whether the user has visited other attractions in the ancient city, because each attraction carries a unique cultural gene, such as local traditional handicrafts, historical legends, and architectural features. These deep cultural connotations are often overlooked in traditional recommendations. When users have a strong interest in specific cultural symbols or themes, the system cannot accurately identify and match these potential preferences, resulting in a large deviation between the recommended attractions or activities and the user's actual preferences, and cannot truly meet the user's demand for personalized, culturally rich cultural and tourism experiences. In order to solve this technical problem, we provide a cultural and tourism content recommendation system that analyzes preferences for cultural and tourism attractions. Summary of the invention

[0004] The purpose of the present invention is to provide a cultural and tourism content recommendation system that analyzes preferences for cultural and tourism attractions to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above objectives, a cultural tourism content recommendation system for analyzing preferences for cultural tourism attractions is provided, including a cultural gene library construction and user interest modeling unit and a dynamic context perception and recommendation optimization unit;

[0006] The cultural gene library construction and user interest modeling unit includes a data collection and preprocessing module, a cultural gene extraction and fusion module, and a user interest model training module, wherein:

[0007] The data collection and preprocessing module collects text data from multiple channels, converts and processes the text data using optical character recognition, analyzes the text data through natural language processing, extracts cultural elements, and establishes a candidate set of cultural elements, and then uses an image recognition algorithm to identify patterns in the text data and convert them into feature vectors;

[0008] The cultural gene extraction and fusion module constructs the extracted cultural elements into a knowledge graph based on the knowledge graph, extracts and annotates the emotional tendencies and cultural symbolic meanings in texts and images using sentiment analysis models and semiotic analysis methods, and fuses similar cultural elements in the knowledge graph through a semantic similarity calculation algorithm to form a cultural gene cluster with semantic associations and construct a cultural gene library;

[0009] The user interest model training module collects the user's recorded data on the cultural and tourism platform and constructs a user interest model. It uses a multi-layer perceptron to extract and classify the user behavior data, combines the attention mechanism, focuses on the user's attention to different cultural elements, and generates a personalized user interest portrait.

[0010] The dynamic context perception and recommendation optimization unit includes a context data collection and analysis module and a recommendation strategy generation and adjustment module, wherein:

[0011] The context data collection and analysis module collects the user's context data in real time, and performs spatial analysis on the geographic location data through the geographic information system to determine the type of scenic area where the user is located, uses time series analysis methods to analyze the influence of time and weather data on cultural and tourism activities, and uses social network analysis tools to mine user social relationships and group behavior characteristics to build a dynamic context information database;

[0012] The recommendation strategy generation and adjustment module designs a recommendation strategy generator based on a reinforcement learning algorithm, takes the current user's context information and user interest portrait as state input, recommends cultural and travel content as action output, and defines a reward function to generate the optimal recommendation strategy.

[0013] As a further improvement of the technical solution, in the data acquisition and preprocessing module, the specific method of converting and processing text data using optical character recognition is as follows:

[0014] Using an optical character recognition model based on deep learning, the optical character recognition model is constructed by combining a convolutional neural network and a recurrent neural network;

[0015] The convolutional neural network is used to extract local features of text images, and gradually reduce the image resolution and extract key features through convolution layers and pooling layers, and map the text image features to a low-dimensional space;

[0016] The recurrent neural network uses a long-term memory network to perform temporal processing on text image features to obtain contextual relationships between characters and convert the text image into a text character sequence.

[0017] As a further improvement of the technical solution, in the data collection and preprocessing module, the specific algorithm for analyzing text data through natural language processing and extracting cultural elements is:

[0018] Using a pre-trained language model based on named entity recognition, the text character sequence is input into the pre-trained language model, and the label probability distribution corresponding to each word is output. The probability of each word belonging to the cultural element category is calculated through a normalized exponential function, and the word is determined to be a cultural element based on a preset judgment threshold. Finally, the words determined to be cultural elements are extracted and a candidate set of cultural elements is constructed.

[0019] As a further improvement of the technical solution, in the cultural gene extraction and fusion module, the steps of constructing the extracted cultural elements into a knowledge graph based on the knowledge graph are as follows:

[0020] The graph database is used as the storage medium of the knowledge graph, the extracted cultural elements are used as nodes, and the semantic relationships between elements are used as edges to construct the initial graph structure;

[0021] A graph embedding algorithm is used to map nodes and edges into a low-dimensional vector space. For each node, a knowledge graph is constructed by learning its representation in the vector space.

[0022] As a further improvement of the technical solution, in the cultural gene extraction and fusion module, the specific method of fusing similar cultural elements in the knowledge graph through a semantic similarity calculation algorithm to form a cultural gene cluster with semantic association is as follows:

[0023] The cosine similarity algorithm is used to calculate the semantic similarity between nodes in the knowledge graph, and a similarity threshold is set. When the semantic similarity is greater than or equal to the similarity threshold, the two nodes and their associated edges are merged to form a cultural gene cluster, and the common cultural symbolic meanings and emotional tendencies are marked.

[0024] As a further improvement of the technical solution, in the user interest model training module, the specific algorithm for extracting and classifying the features of user behavior data using a multilayer perceptron is:

[0025] Construct a three-layer multilayer perceptron, where the first layer is the input layer, which is used to receive the user's behavior data on the cultural tourism platform and convert it into vector form input;

[0026] The second layer is a hidden layer, which uses the rectified linear unit as the activation function to perform nonlinear transformation on the input data and extract high-level features;

[0027] The third layer is the output layer, which is used to classify user behavior and output the user's preference probability for different cultural elements. The normalized exponential function is used to convert the neuron output into a probability distribution.

[0028] As a further improvement of the technical solution, in the situational data collection and analysis module, the specific steps of using the time series analysis method to analyze the influence of time and weather data on cultural and tourism activities are as follows:

[0029] For time data, we divide 24 hours a day into different time periods according to the activity level of cultural and tourism activities, count the frequency of users participating in various cultural and tourism activities in different time periods, and construct a time series.

[0030] For weather data, we divide weather conditions into three categories: sunny, rainy, and cloudy. We also count the frequency of users participating in cultural and tourism activities under different weather conditions to construct a weather sequence.

[0031] The autoregressive moving average model is used to model and forecast time series and weather series, and to predict the participation trend of cultural tourism activities in different time periods and weather conditions.

[0032] As a further improvement of the technical solution, in the recommendation strategy generation and adjustment module, the specific method of designing a recommendation strategy generator based on a reinforcement learning algorithm is as follows:

[0033] The recommendation strategy generator is constructed using the deep Q-network algorithm, which concatenates the current user's context information and user interest profile into a state vector and encodes the recommended cultural and tourism content into an action vector.

[0034] Construct a neural network with multiple fully connected layers as the Q network. The input layer receives the state vector, which is transformed by the hidden layer nonlinearly, and the output layer outputs the Q value estimate of each action.

[0035] During the training process, the agent selects actions, environmental feedback rewards and the next state according to the current state, and uses the experience replay mechanism to store the current state selected actions, environmental feedback rewards and the next state in the replay buffer, randomly samples a batch of data from the buffer, and updates the Q network parameters according to the Q learning update formula.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] In a cultural and tourism content recommendation system that analyzes preferences for cultural and tourism attractions, the data collection and preprocessing module uses optical character recognition and natural language processing technology to extract cultural elements from text data collected from multiple channels, laying the foundation for subsequent analysis. The cultural gene extraction and fusion module is built based on the knowledge graph, integrating similar cultural elements to form gene clusters, and deeply exploring cultural connotations and associations. The user interest model training module uses a multi-layer perceptron combined with an attention mechanism. Based on the user's recorded data on the cultural and tourism platform, it accurately focuses on the user's attention to different cultural elements, generates personalized interest portraits, deeply analyzes the cultural essence of cultural and tourism attractions, and accurately identifies users' potential cultural interest preferences, so that the recommended content is closely aligned with user needs, thereby improving the accuracy of cultural and tourism content recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is an overall block diagram of the present invention.

[0039] The meaning of each number in the figure is:

[0040] 1. Cultural gene library construction and user interest modeling unit; 11. Data collection and preprocessing module; 12. Cultural gene extraction and fusion module; 13. User interest model training module; 2. Dynamic situational awareness and recommendation optimization unit; 21. Situational data collection and analysis module; 22. Recommendation strategy generation and adjustment module. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The present invention provides a cultural tourism content recommendation system for analyzing the preferences of cultural tourism attractions. Figure 1 As shown, it includes cultural gene library construction and user interest modeling unit 1 and dynamic situation perception and recommendation optimization unit 2;

[0043] The cultural gene library construction and user interest modeling unit 1 includes a data collection and preprocessing module 11, a cultural gene extraction and fusion module 12 and a user interest model training module 13, wherein:

[0044] The data collection and preprocessing module 11 collects text data from multiple channels, converts and processes the text data using optical character recognition, analyzes the text data through natural language processing, extracts cultural elements, and establishes a candidate set of cultural elements, and then uses an image recognition algorithm to identify patterns in the text data and convert them into feature vectors.

[0045] In the data collection and preprocessing module 11, the specific method of converting and processing text data using optical character recognition is as follows:

[0046] Determine the size of the input text image and construct the first convolution layer. Perform a convolution operation on the input text image to obtain a feature map. Then add a pooling layer and use maximum pooling to pool the feature map to obtain a pooled feature map. Repeat the above convolution and pooling operations to gradually reduce the image resolution and extract more abstract key features, mapping the text image features to a low-dimensional space.

[0047] The feature map output by the last pooling layer of the convolutional neural network is expanded into a feature sequence in row or column order. The long short-term memory network consists of an input gate, an output gate, a forget gate, and a memory unit. For each element in the input sequence, the input gate, forget gate, output gate, and memory unit updates are calculated.

[0048] The entire feature sequence is processed by the above-mentioned long short-term memory network unit to capture the contextual relationship between characters. The final output hidden state sequence is converted into a text character sequence through a fully connected layer. During the training process, a large amount of annotated text image data is used to optimize the model parameters by minimizing the cross entropy loss function between the predicted character sequence and the actual character sequence, so that the model can accurately convert the text image into a text character sequence and realize the optical character recognition function. Through the above steps, the optical character recognition model based on deep learning can effectively extract the features of the text image and convert it into a text character sequence, providing a basis for subsequent cultural element extraction and other operations.

[0049] In the data collection and preprocessing module 11, the text data is analyzed through natural language processing, and the specific algorithm for extracting cultural elements is as follows:

[0050] Select a pre-trained language model based on named entity recognition. Pre-trained language models are models trained on large-scale text data. These models have learned rich language knowledge and semantic representations, and can provide strong support for subsequent named entity recognition. Input the text character sequence obtained through optical character recognition into the selected pre-trained language model. The pre-trained language model will perform pre-processing operations such as word segmentation and encoding on the input text, and convert the text into a vector representation suitable for model processing.

[0051] The pre-trained language model analyzes each word in the input text according to its internal neural network structure and training parameters, and outputs its corresponding label probability distribution. For a simple named entity recognition task, the label categories include "people", "places", "cultural activities", "cultural items", categories related to cultural elements, and other non-cultural element categories. For each word's label probability distribution, the probability of it belonging to the cultural element category is calculated using a normalized exponential function.

[0052] According to the preset judgment threshold, the probability of each word belonging to the cultural element category is judged. If the probability of belonging to the cultural element category is greater than or equal to the judgment threshold, the word is judged to be a cultural element. Otherwise, it is judged to be a non-cultural element. The words judged to be cultural elements are extracted to form a candidate set of cultural elements. This candidate set will serve as the basis for further analysis and processing, so as to more deeply explore the cultural connotations of cultural and tourism attractions and the interests and preferences of users, and provide data support for accurate cultural and tourism content recommendations.

[0053] Through the above steps, the named entity recognition method based on the pre-trained language model can be used to effectively extract cultural elements from text data, providing a key data foundation for the construction of the cultural gene library and user interest modeling of the entire cultural and tourism content recommendation system, and improving the system's understanding and application capabilities of the cultural and tourism field, thereby achieving more accurate and personalized recommendation services.

[0054] The cultural gene extraction and fusion module 12 constructs the extracted cultural elements into a knowledge graph based on the knowledge graph. It uses sentiment analysis models and semiotic analysis methods to extract and annotate the emotional tendencies and cultural symbolic meanings in texts and images. It uses a semantic similarity calculation algorithm to fuse similar cultural elements in the knowledge graph to form a cultural gene cluster with semantic associations and build a cultural gene library.

[0055] In the cultural gene extraction and fusion module 12, the steps of constructing the extracted cultural elements into a knowledge graph based on the knowledge graph are as follows:

[0056] The cultural elements extracted by the data collection and preprocessing module 11 are taken as nodes. For example, if cultural elements such as "Forbidden City", "Great Wall" and "Peking Opera" are extracted, corresponding nodes are created in the graph database. Each node can contain some attribute information, such as the name, type and related descriptive text of the cultural element.

[0057] Determine the semantic relationship between elements as an edge. For example, there is a semantic relationship of "belongs to" between "Forbidden City" and "Ancient Chinese Royal Buildings". Create a directed edge between the "Forbidden City" node and the "Ancient Chinese Royal Buildings" node, and mark the edge type as "belongs to". For "Peking Opera" and "Traditional Opera", create similar semantic relationship edges. In this way, all cultural elements and their semantic relationships are constructed into an initial graph structure and stored in the graph database.

[0058] Through the graph embedding algorithm, random walks are performed on the constructed initial graph structure. The length and number of walks are set. Starting from each node, neighbor nodes are selected for walks according to a certain probability distribution to generate a series of node sequences with fixed length. Multiple random walks are performed on each node in the graph to obtain a large number of node sequences.

[0059] The generated node sequence is input into the preset natural language processing model. The goal of the natural language processing model is to predict the nodes around a given node. For each node in the node sequence, the node is taken as the center and the nodes in the window before and after it are predicted. The vector representation of the node is learned by maximizing the predicted probability. Let the node The vector representation is , its neighbor nodes The vector representation is , then the objective function of the natural language processing model can be expressed as:

[0060] ;

[0061] in ; is the total number of nodes in the graph, Represented as the length of the node sequence, Represents a window size parameter, which is used to determine the range of surrounding elements centered on a certain element in the calculation. By continuously optimizing this objective function and using optimization algorithms such as gradient descent to update the vector representation of the node, nodes with similar semantic relationships are closer in the low-dimensional vector space.

[0062] After learning the graph embedding algorithm, each node is represented by a vector in a low-dimensional vector space. These nodes and their vector representations, together with the edge information, together form the final knowledge graph. In subsequent operations, these vector representations can be used to calculate semantic similarity and integrate cultural gene clusters, further exploring the deep relationship between cultural elements and providing richer knowledge support for cultural and tourism content recommendations.

[0063] Through the above steps, the extracted cultural elements can be effectively constructed into a knowledge graph, laying the foundation for the system to analyze the cultural connotations of cultural and tourism attractions and accurately match user interests.

[0064] In the cultural gene extraction and fusion module 12, the specific method of fusing similar cultural elements in the knowledge graph through the semantic similarity calculation algorithm to form a cultural gene cluster with semantic association is as follows:

[0065] First, obtain the low-dimensional vector representation corresponding to each node from the previously constructed knowledge graph. Assume that the node The vector is ,node The vector is ,in is the dimension of the vectors, which are learned by the graph embedding algorithm.

[0066] Calculate nodes according to the cosine similarity formula and nodes The semantic similarity between , based on the knowledge and practical application experience in the field of culture and tourism, set a similarity threshold For each pair of nodes in the knowledge graph, their semantic similarity is calculated and compared with the set threshold Make a comparison.

[0067] if , then the node and nodes and its associated edges, for example, if the node Represents "Forbidden City", node represents "Summer Palace". They have high similarity in architectural style and historical and cultural background, meet the threshold condition, and when merged, create a new cultural gene cluster node , integrate the attribute information of nodes and nodes, and mark the common cultural symbolic meaning and emotional tendency. and nodes The connected edges are adjusted to point to the new cultural gene cluster node , in order to maintain the structural integrity of the knowledge graph.

[0068] Repeat the above judgment and merging operations for all node pairs in the knowledge graph until all possible node pairs are traversed, and finally form multiple cultural gene clusters with semantic associations. These cultural gene clusters can more effectively summarize and organize the cultural elements in the field of culture and tourism, and provide a more valuable knowledge basis for subsequent user interest modeling and recommendation strategy generation.

[0069] Through the above steps, the cosine similarity algorithm can be used to accurately integrate similar cultural elements in the knowledge graph, construct cultural gene clusters, and enhance the system's understanding and application capabilities of the cultural connotation of cultural tourism, thereby better serving the core functions of the cultural tourism content recommendation system.

[0070] The user interest model training module 13 collects the user's recorded data on the cultural and tourism platform and constructs a user interest model. It uses a multi-layer perceptron to extract and classify the user behavior data, combines the attention mechanism, focuses on the user's attention to different cultural elements, and generates a personalized user interest portrait.

[0071] In the user interest model training module 13, the specific algorithm for extracting features and classifying user behavior data using a multi-layer perceptron is:

[0072] Collect user behavior data from the cultural and tourism platform, such as user browsing history, likes, comments, purchases and other behavior information, clean and preprocess these data, remove invalid data and noise data, convert the processed user behavior data into vector form, and construct the input layer of the multi-layer perceptron. The number of neurons is equal to the dimension of the user behavior data vector. The function of the input layer is to receive the user behavior data vector and pass it to the next layer.

[0073] For the hidden layer, assume that neurons, initialize the weight matrix of the hidden layer , whose shape is , and the bias vector , whose shape is The weights can be initialized randomly or based on a specific distribution.

[0074] The hidden layer performs a linear transformation on the input data and calculates , get a The intermediate result vector of dimension , then, yes By applying the rectified linear unit activation function for nonlinear transformation, the output vector of the hidden layer is obtained. The model can learn the nonlinear characteristics in the data and enhance the expressive power of the model.

[0075] Initialize the weight matrix of the output layer , the output layer performs a linear transformation on the output of the hidden layer and calculates , we get a result vector, and use the normalized exponential function to convert the neuron output of the output layer into a probability distribution. The formula of the normalized exponential function is: ;in Indicates the user's The preference probability of a cultural element, That is the final output probability vector of users’ preference for different cultural elements.

[0076] The cross entropy loss function is used to measure the difference between the user preference probability predicted by the model and the actual preference. Assume that for a user, the actual preference label is , then the cross entropy loss function for: ; Use a large number of user behavior data samples to train the multilayer perceptron. In each training iteration, a batch of user behavior data is input into the model and the output preference probability vector is calculated. , and then calculate the loss value according to the above loss function.

[0077] The gradient of the loss function to the weights and biases of each layer of the model is calculated through the back-propagation algorithm, and the weights and biases are updated using optimization algorithms such as gradient descent to minimize the loss function. By continuously adjusting hyperparameters such as the learning rate and the number of training rounds, the model is continuously optimized on the training data, improving the accuracy of classifying user behaviors and the ability to model user interests, and ultimately generating accurate personalized user interest portraits, providing strong support for cultural and tourism content recommendations.

[0078] Through the above steps, the three-layer multi-layer perceptron can be used to effectively extract and classify user behavior data, accurately build a user interest model, and achieve precise analysis and modeling of user cultural and tourism interests.

[0079] The dynamic context perception and recommendation optimization unit 2 includes a context data collection and analysis module 21 and a recommendation strategy generation and adjustment module 22, wherein:

[0080] The situational data collection and analysis module 21 collects the user's situational data in real time, and performs spatial analysis on the geographic location data through the geographic information system to determine the type of scenic area where the user is located, uses time series analysis methods to analyze the impact of time and weather data on cultural and tourism activities, and uses social network analysis tools to mine user social relationships and group behavior characteristics to build a dynamic situational information database.

[0081] In the situational data collection and analysis module 21, the specific steps of using the time series analysis method to analyze the influence of time and weather data on cultural and tourism activities are as follows:

[0082] According to the experience and historical data of the cultural and tourism industry, the 24 hours of a day are divided into different time periods. The number of times users participate in various cultural and tourism activities in different time periods is obtained from the historical data or relevant statistical information of the cultural and tourism platform. Participate in activities The number of times is The total number of users is , then calculate the user participation activity during this period Frequency For each time period and each cultural and tourism activity, the corresponding frequency is calculated to construct a time series ,in Indicates different time periods. Indicates different types of cultural and tourism activities.

[0083] The weather conditions are divided into three categories: sunny, rainy, and cloudy. The number of times users participate in various cultural and tourism activities under different weather conditions is also counted from historical data. Let the weather be Participate in activities The number of times is , It means sunny. It means rain. indicates negative, the total number of users is , then calculate the user participation activity under this weather Frequency .

[0084] For each type of weather and each cultural and tourism activity, the corresponding frequency is calculated to construct a weather sequence ,in Indicates different weather categories. Indicates different types of cultural and tourism activities.

[0085] For time series and weather sequence , respectively select the autoregressive moving average model, the form of the autoregressive moving average model is ;in is the time series at time The value of is the autoregressive coefficient, is the moving average coefficient, is a white noise sequence. By analyzing the time series and weather series, we can determine value, and according to Determine the parameters of the model .

[0086] For the model with determined parameters, the maximum likelihood estimation method is used to estimate the parameters of the model. and , for time series , let the estimated parameters be and , then the established autoregressive moving average model is , similarly for the weather series , the established autoregressive moving average model is , test the established model. When the model passes the test, it can be used for prediction. For time series, suppose you want to predict the future period Internal user participation activities Frequency , substitute the known historical data into the time series model, and gradually calculate the predicted value. Similarly, for the weather series, predict the frequency of future cultural and tourism activities under different weather conditions. ,Through these prediction values, the participation trends of cultural ,tourism activities in different time periods and weather conditions can be analyzed, ,providing a basis for the generation of recommendation strategies.

[0087] Through the above steps, we can effectively use time series analysis methods and autoregressive moving average models to analyze the impact of time and weather data on cultural and tourism activities, and make predictions, providing important contextual information support for the cultural and tourism content recommendation system.

[0088] The recommendation strategy generation and adjustment module 22 designs a recommendation strategy generator based on the reinforcement learning algorithm, takes the current user's context information and user interest portrait as state input, and the recommended cultural and travel content as action output, and defines a reward function to generate the optimal recommendation strategy.

[0089] In the recommendation strategy generation and adjustment module 22, the specific method of designing a recommendation strategy generator based on the reinforcement learning algorithm is as follows:

[0090] The current user's context information is obtained from the context data collection and analysis module 21, including geographic location information, time information, weather information, and user social relationships and group behavior characteristics obtained through social network analysis. At the same time, the user interest portrait is obtained from the user interest model training module 13, which includes information such as the user's preference for different cultural elements.

[0091] These context information and user interest profiles are spliced ​​together to form a state vector of fixed length , determine the set of cultural and tourism content that the system can recommend, assuming that it contains Different cultural and tourism contents are encoded to form an action vector , construct a neural network containing multiple fully connected layers as the Q network, and the input layer has neurons, which receive the state vector .

[0092] The first hidden layer can have neurons, using the rectified linear unit as the activation function, performing nonlinear transformation on the input data and calculating the output of the first hidden layer , and then activated by the rectified linear unit to obtain , you can add one or more hidden layers, such as the second hidden layer has neurons, calculate the output of the second hidden layer , after activation of the rectified linear unit, we get , the output layer has neurons, which output the Q-value estimate for each action, , calculate the output of the output layer .

[0093] In each training step, the agent is trained based on the current state vector , the Q value estimate of each action is calculated through the Q network Then choose actions with probability Randomly choose an action with probability Select the action with the largest Q value. Suppose the selected action is The system will recommend corresponding cultural and travel content to users, and users will generate certain feedback. The reward function is defined based on user feedback and actual business rules. , if the user accepts the recommendation and makes a reservation or positive review, a higher positive reward is given; if the user ignores the recommendation, a smaller negative reward is given. At the same time, the system observes the next state , which is constructed in the same way as the current state vector, and is a concatenation of contextual information and user interest portraits that are updated according to the user's new behavior and changes in the environment.

[0094] The current state , Selected Action , Rewards and the next state Form an experience tuple , and stored in the playback buffer. The playback buffer has a certain capacity limit. When the buffer is full, the first-in-first-out strategy can be used to replace the old experience.

[0095] Randomly sample a batch of data from the playback buffer. Assume that the number of samples is , and get a batch of experience tuples. The expression of the experience tuple is , for each sampled experience tuple, according to The update formula of learning updates the Q network parameters. The update formula of Q learning is:

[0096] ;

[0097] in is the learning rate, which controls the step size of parameter updates, is a discount factor that weighs the importance of future rewards and current rewards, Represents the next state, Represents the next state The corresponding action, is in the next state All possible actions Maximum Value, in the actual update, for the sampled experience tuple , calculate the target Q value ,in, Indicates the immediate status. Indicates the immediate action corresponding to the immediate state. Expressed as an immediate reward, the parameters of the Q network are then updated by minimizing the mean square error loss function. The expression for minimizing the mean square error loss function is ; To update the parameters of the Q network, you can use optimization algorithms such as stochastic gradient descent, calculate the gradient of the weights and biases of the Q network according to the loss function, and update them. Continuously iterate the training so that the Q network can learn the optimal recommendation strategy, that is, it can accurately estimate the Q value of each recommended action according to different user states, so as to select the most appropriate cultural and tourism content for recommendation, thereby improving the performance of the recommendation system and user satisfaction.

[0098] In the present invention, the data collection and preprocessing module 11 extracts cultural elements from multi-channel texts through optical character recognition and natural language processing, the cultural gene extraction and fusion module 12 constructs a knowledge graph and fuses similar elements to form a gene cluster, the user interest model training module 13 generates an interest portrait based on the user's cultural and travel platform data, and in the dynamic situation perception and recommendation optimization unit 2, the situation data collection and analysis module 21 collects and analyzes user situation data, and the recommendation strategy generation and adjustment module 22 uses a deep Q network algorithm to generate a recommendation strategy based on user status and interests, and optimizes the recommendation in combination with situational factors, thereby effectively improving the accuracy of cultural and travel content recommendations.

[0099] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions, characterized in that: It includes the cultural gene library construction and user interest modeling unit (1) and the dynamic context perception and recommendation optimization unit (2); The cultural gene library construction and user interest modeling unit (1) comprises a data collection and preprocessing module (11), a cultural gene extraction and fusion module (12) and a user interest model training module (13), wherein: The data collection and preprocessing module (11) collects text data from multiple channels, converts and processes the text data using optical character recognition, analyzes the text data through natural language processing, extracts cultural elements, and establishes a candidate set of cultural elements, and then uses an image recognition algorithm to identify patterns in the text data and convert them into feature vectors; The cultural gene extraction and fusion module (12) constructs the extracted cultural elements into a knowledge graph based on the knowledge graph, extracts and annotates the emotional tendencies and cultural symbolic meanings in texts and images using a sentiment analysis model and a semiotic analysis method, and fuses similar cultural elements in the knowledge graph using a semantic similarity calculation algorithm to form a cultural gene cluster with semantic associations and construct a cultural gene library; The user interest model training module (13) collects the user's record data on the cultural tourism platform and constructs a user interest model. It uses a multi-layer perceptron to extract and classify the user behavior data, combines the attention mechanism, focuses on the user's attention to different cultural elements, and generates a personalized user interest portrait. The dynamic context perception and recommendation optimization unit (2) comprises a context data collection and analysis module (21) and a recommendation strategy generation and adjustment module (22), wherein: The context data collection and analysis module (21) collects the context data of the user in real time, and performs spatial analysis on the geographical location data through a geographic information system to determine the type of scenic area where the user is located, uses a time series analysis method to analyze the influence of time and weather data on cultural and tourism activities, and uses a social network analysis tool to mine the social relationships and group behavior characteristics of users, and build a dynamic context information database; The recommendation strategy generation and adjustment module (22) designs a recommendation strategy generator based on a reinforcement learning algorithm, takes the current user's context information and user interest profile as state input, and the recommended cultural and travel content as action output, and defines a reward function to generate an optimal recommendation strategy.

2. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions according to claim 1, characterized in that: In the data acquisition and preprocessing module (11), the specific method of converting and processing text data using optical character recognition is as follows: Using an optical character recognition model based on deep learning, the optical character recognition model is constructed by combining a convolutional neural network and a recurrent neural network; The convolutional neural network is used to extract local features of text images, and gradually reduce the image resolution and extract key features through convolution layers and pooling layers, and map the text image features to a low-dimensional space; The recurrent neural network uses a long-term memory network to perform temporal processing on text image features to obtain contextual relationships between characters and convert the text image into a text character sequence.

3. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions according to claim 2, characterized in that: In the data collection and preprocessing module (11), the specific algorithm for analyzing text data through natural language processing and extracting cultural elements is: Using a pre-trained language model based on named entity recognition, the text character sequence is input into the pre-trained language model, and the label probability distribution corresponding to each word is output. The probability of each word belonging to the cultural element category is calculated through a normalized exponential function, and the word is determined to be a cultural element based on a preset judgment threshold. Finally, the words determined to be cultural elements are extracted and a candidate set of cultural elements is constructed.

4. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions according to claim 3, characterized in that: In the cultural gene extraction and fusion module (12), the steps of constructing the extracted cultural elements into a knowledge graph based on the knowledge graph are as follows: The graph database is used as the storage medium of the knowledge graph, the extracted cultural elements are used as nodes, and the semantic relationships between elements are used as edges to construct the initial graph structure; A graph embedding algorithm is used to map nodes and edges into a low-dimensional vector space. For each node, a knowledge graph is constructed by learning its representation in the vector space.

5. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions according to claim 4, characterized in that: In the cultural gene extraction and fusion module (12), the specific method of fusing similar cultural elements in the knowledge graph through a semantic similarity calculation algorithm to form a cultural gene cluster with semantic association is as follows: The cosine similarity algorithm is used to calculate the semantic similarity between nodes in the knowledge graph, and a similarity threshold is set. When the semantic similarity is greater than or equal to the similarity threshold, the two nodes and their associated edges are merged to form a cultural gene cluster, and the common cultural symbolic meanings and emotional tendencies are marked.

6. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions according to claim 5, characterized in that: In the user interest model training module (13), the specific algorithm for extracting features and classifying user behavior data using a multi-layer perceptron is: Construct a three-layer multilayer perceptron, where the first layer is the input layer, which is used to receive the user's behavior data on the cultural tourism platform and convert it into vector form input; The second layer is a hidden layer, which uses the rectified linear unit as the activation function to perform nonlinear transformation on the input data and extract high-level features; The third layer is the output layer, which is used to classify user behavior and output the user's preference probability for different cultural elements. The normalized exponential function is used to convert the neuron output into a probability distribution.

7. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions according to claim 6, characterized in that: In the situational data collection and analysis module (21), the specific steps of using the time series analysis method to analyze the influence of time and weather data on cultural and tourism activities are as follows: For time data, we divide 24 hours a day into different time periods according to the activity level of cultural and tourism activities, count the frequency of users participating in various cultural and tourism activities in different time periods, and construct a time series. For weather data, we divide weather conditions into three categories: sunny, rainy, and cloudy. We also count the frequency of users participating in cultural and tourism activities under different weather conditions to construct a weather sequence. The autoregressive moving average model is used to model and forecast time series and weather series, and to predict the participation trend of cultural tourism activities in different time periods and weather conditions.

8. A cultural and tourism content recommendation system for analyzing preferences for cultural and tourism attractions according to claim 7, characterized in that: In the recommendation strategy generation and adjustment module (22), the specific method for designing a recommendation strategy generator based on a reinforcement learning algorithm is as follows: The recommendation strategy generator is constructed using the deep Q-network algorithm, which concatenates the current user's context information and user interest profile into a state vector and encodes the recommended cultural and tourism content into an action vector. Construct a neural network with multiple fully connected layers as the Q network. The input layer receives the state vector, which is transformed by the hidden layer nonlinearly, and the output layer outputs the Q value estimate of each action. During the training process, the agent selects actions, environmental feedback rewards and the next state according to the current state, and uses the experience replay mechanism to store the current state selected actions, environmental feedback rewards and the next state in the replay buffer, randomly samples a batch of data from the buffer, and updates the Q network parameters according to the Q learning update formula.

Images