Tourism preference modeling method and system based on user social behavior mining

Abstract

The invention provides a travel preference modeling method and system based on user social behavior mining, and relates to the technical field of travel preference modeling, and the method comprises the steps: executing the cross-domain social behavior data collection of a target user; performing multi-modal feature extraction on the cross-domain social behavior data, establishing a text emotion topic vector, an image scene recognition vector, a time sequence behavior feature vector and a friend influence weight matrix, and executing joint embedding analysis; constructing a preference state machine by using the initial embedded vector set; constructing a target user interest diffusion network, and executing relation-driven reinforcement of the state transition probability of the preference state machine by using the target user interest diffusion network; and generating the tourism preference of the target user according to the enhanced preference state machine, and visually presenting the tourism preference through a preference radar map. According to the method and the device, the technical problem that user travel preference modeling is not accurate enough in the prior art can be solved, and the technical effect of improving travel preference modeling accuracy is achieved.

Description

Technical Field

[0001] This application relates to the field of tourism preference modeling technology, and in particular to a tourism preference modeling method and system based on user social behavior mining. Background Technology

[0002] With the rapid development of social networks and mobile internet, users have generated a large amount of multi-dimensional behavioral data on various platforms, including text interactions, image sharing, location check-ins, and friend relationship chains. This data contains users' interests, preferences, and behavioral patterns, which are of great value for personalized recommendations, travel planning, and travel decisions.

[0003] Currently, existing tourism preference modeling methods mainly rely on single data types or explicit behavioral features, such as analyzing only users' browsing history, purchase records, or location check-in data, ignoring the implicit interests and emotional tendencies expressed by users in social interactions. Furthermore, traditional methods often lack a unified feature fusion mechanism when processing multimodal data, making it difficult to fully interact and map text, images, temporal behavioral information, and social relationship information, thus failing to form an accurate comprehensive preference representation.

[0004] In summary, existing technologies suffer from several technical problems. They rely excessively on single explicit behavioral data, lack unified fusion processing of multimodal data, and ignore implicit interests and emotional tendencies in social interactions. As a result, user travel preference modeling is not comprehensive and accurate enough, and cannot accurately reflect the dynamic changes in users' interests. This further affects the prediction effect of personalized recommendation systems, user experience, and the level of intelligence of travel services. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for modeling travel preferences based on user social behavior mining, in order to solve the technical problems in the existing technology that are not comprehensive and accurate in modeling user travel preferences due to over-reliance on single explicit behavioral data, lack of unified fusion processing of multimodal data, and neglect of implicit interests and emotional tendencies in social interactions. These problems fail to accurately reflect the dynamic changes in users' interests and further affect the prediction effect of personalized recommendation systems, user experience, and the level of intelligence of travel services.

[0006] In view of the above problems, this application provides a method and system for modeling travel preferences based on the mining of user social behavior.

[0007] Firstly, this application provides a tourism preference modeling method based on user social behavior mining, implemented through a tourism preference modeling system based on user social behavior mining, comprising: collecting cross-domain social behavior data of target users, the cross-domain social behavior data including tourism topic interaction data, non-tourism topic interaction data, friend relationship chains and interaction records, and location check-in data; extracting multimodal features from the cross-domain social behavior data, establishing text sentiment theme vectors, image scene recognition vectors, time-series behavioral feature vectors, and friend influence weight matrices, and performing joint embedding analysis to establish an initial embedding vector set; constructing a preference state machine including a long-term preference layer, a periodic interest layer, and a micro-trend amplification layer using the initial embedding vector set; constructing a target user interest diffusion network, the target user interest diffusion network updating the interest propagation path based on the friend influence weight matrix, and performing relationship-driven reinforcement of the state transition probabilities of the preference state machine using the target user interest diffusion network; generating target user tourism preferences based on the reinforced preference state machine, and visualizing them through a preference radar chart.

[0008] Preferably, the method for modeling tourism preferences based on user social behavior mining further includes: performing semantic segmentation and sentiment polarity recognition on the tourism topic interaction data and non-tourism topic interaction data to establish a mapped multi-dimensional topic distribution vector and sentiment intensity vector; extracting published image data from the tourism topic interaction data and non-tourism topic interaction data, performing scene recognition, object detection, and character composition analysis on the published image data, extracting geographic landmark tags and travel mode tags, using the extracted tags to perform cross-image multi-image aggregation analysis, and establishing an image scene recognition vector; performing time series encoding on the location check-in data under a multi-scale sliding window, extracting periodic travel patterns, travel time preferences, and sudden behavior intensity, and using Fourier time-frequency transform for periodic and non-periodic segmentation to establish a time-series behavior feature vector; performing graph structure analysis on the friend relationship chain, calculating friend interaction frequency, topic overlap rate, and number of joint trips, and establishing a friend influence weight matrix based on the calculation results.

[0009] Preferably, the method for modeling travel preferences based on user social behavior mining further includes: performing semantic similarity attention analysis on multidimensional topic distribution vectors to establish a first-level attention factor; constructing a dual-channel attention for positive and negative emotions on the emotion intensity vector, calculating the positive emotion weight coefficient and the negative emotion weight coefficient respectively, constructing an emotion polarity response matrix, and establishing a second-level attention factor based on the emotion polarity response matrix; performing cross-channel cross-calculation using the first-level attention factor and the second-level attention factor to establish a topic-emotion joint modulation vector; obtaining the recent emotion fluctuation index of the target user, establishing a dynamic adjustment factor using the recent emotion fluctuation index, and performing multi-layer residual fusion on the topic-emotion joint modulation vector based on the dynamic adjustment factor to establish a modulated text emotion topic vector; and concatenating the modulated text emotion topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix, and then performing deep embedding mapping to establish the initial embedding vector set.

[0010] Preferably, the method for modeling travel preferences based on user social behavior mining further includes: performing principal component compression on the modulated text sentiment topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix; after removing low-contribution dimensions, establishing a first embedding subspace; constructing a multi-head graph convolutional network within the first embedding subspace; using the multi-head graph convolutional network to perform joint propagation of the friend relationship chain connection matrix and spliced ​​features to establish a second embedding subspace containing topological relationship awareness; and performing deep embedding mapping by fusing the global attention stream and local time series stream of the second embedding subspace using an adaptively gated dual-stream Transformer.

[0011] Preferably, the tourism preference modeling method based on user social behavior mining further includes: slicing the initial embedded vector set into multi-scale time windows to establish long-term feature subsets, periodic feature subsets, and short-term feature subsets respectively; performing latent variable Markov chain analysis on the long-term feature subset to extract the steady-state distribution of the long-term preference layer and establish an implicit preference migration matrix; performing seasonal and holiday interest pattern analysis on the periodic feature subset to establish a periodic interest layer; constructing a micro-trend amplifier in the short-term feature subset, the micro-trend amplifier being constructed based on a preset nearest neighbor time window, calculating the amplification factor under the constraint of sudden behavior intensity within the preset nearest neighbor time window, and establishing a micro-trend amplification layer after bias correction of the amplification factor; and fusing the long-term preference layer, periodic interest layer, and micro-trend amplification layer using a multi-layer residual coupling network to establish a preference state machine.

[0012] Preferably, the method for modeling travel preferences based on user social behavior mining further includes: establishing a psychological profile of the target user based on the time-series behavioral feature vector, wherein the psychological profile is constructed through risk-taking tendency features, independence features, and social features; and after constructing a modulating factor for travel preferences using the psychological profile, generating compensation for the target user's travel preferences based on the modulating factor.

[0013] Preferably, the method for modeling travel preferences based on user social behavior mining further includes: performing context-aware identification on the target user; when the context awareness of the target user triggers an event-driven mechanism, generating local driving instructions; the context-aware identification includes geolocation change identification and social topic concentration change identification; and adaptively updating the local parameters of the preference state machine according to the local driving instructions to complete the generation of the target user's travel preferences.

[0014] Preferably, the method for modeling travel preferences based on user social behavior mining further includes: performing synchronous analysis of the preference evolution of multiple friends based on the friend relationship chain and interaction records, and establishing a dynamic embedding of group preferences; and using the dynamic embedding of group preferences to perform target user interest diffusion network compensation management.

[0015] Preferably, the method for modeling travel preferences based on user social behavior mining further includes: establishing a self-updating mechanism, which is used to collect new data of target users at a preset period and use the results of the new data collection to manage the incremental update of the preference state machine.

[0016] Secondly, this application also provides a tourism preference modeling system based on user social behavior mining, used to execute a tourism preference modeling method based on user social behavior mining as described in the first aspect, including: a cross-domain social behavior data collection module, used to collect cross-domain social behavior data of target users, wherein the cross-domain social behavior data includes tourism topic interaction data, non-tourism topic interaction data, friend relationship chains and interaction records, and location check-in data; and an initial embedding vector set establishment module, used to extract multimodal features from the cross-domain social behavior data and establish text sentiment theme vectors, image scene recognition vectors, temporal behavior feature vectors, and friend influence vectors. The system comprises a weight matrix and performs joint embedding analysis to establish an initial embedding vector set; a preference state machine construction module, which uses the initial embedding vector set to construct a preference state machine including a long-term preference layer, a periodic interest layer, and a micro-trend amplification layer; a relationship-driven reinforcement execution module, which constructs a target user interest diffusion network, updates the interest propagation path based on the friend influence weight matrix, and performs relationship-driven reinforcement of the state transition probabilities of the preference state machine using the target user interest diffusion network; and a visualization module, which generates target user travel preferences based on the reinforced preference state machine and visualizes them through a preference radar chart.

[0017] The technical solution provided in this application has at least the following technical effects or advantages: by achieving the technical goal of multimodal dynamic preference modeling based on users' cross-domain social behavior, it can capture users' long-term preferences, periodic interests and short-term micro-trends in real time, fully integrate text, images, time-series behavior and social relationship information, and combine group social influence to predict personalized travel preferences.

[0018] The above description is merely an overview of the technical solution of this application. To enable a clearer understanding of the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1This is a flowchart illustrating a tourism preference modeling method based on user social behavior mining, as proposed in this application.

[0021] Figure 2 This is a schematic diagram of the structure of a travel preference modeling system based on user social behavior mining, as proposed in this application.

[0022] Figure labeling: 1. Cross-domain social behavior data collection module; 2. Initial embedding vector set establishment module; 3. Preference state machine construction module; 4. Relationship-driven reinforcement execution module; 5. Visualization presentation module. Detailed Implementation

[0023] This application provides a method and system for modeling travel preferences based on user social behavior mining. It addresses the technical problems in existing technologies where over-reliance on single explicit behavioral data, lack of unified fusion processing of multimodal data, and neglect of implicit interests and sentiment tendencies in social interactions lead to incomplete and inaccurate user travel preference modeling. These shortcomings fail to accurately reflect dynamic changes in user interests, further impacting the prediction performance of personalized recommendation systems, user experience, and the level of intelligence in travel services. The application achieves the technical goal of multimodal dynamic preference modeling based on users' cross-domain social behavior, enabling real-time capture of long-term user preferences, cyclical interests, and short-term micro-trends. It fully integrates text, images, temporal behavioral information, and social relationship information, and combines group social influences to predict personalized travel preferences.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides a tourism preference modeling method based on user social behavior mining, which is applied to a tourism preference modeling system based on user social behavior mining, and specifically includes the following steps: S1: Perform cross-domain social behavior data collection for the target user. The cross-domain social behavior data includes tourism topic interaction data, non-tourism topic interaction data, friend relationship chain and interaction records, and location check-in data.

[0026] Specifically, the target users are those whose travel preferences are to be analyzed. The process involves collecting cross-domain social behavior data from multiple social platforms or different types of social scenarios to gather behavioral data relevant to the target users. "Cross-domain" refers to traversing different platforms or data types to ensure a more comprehensive information source, enabling a multi-faceted reconstruction of user behavior habits, rather than being limited to a single social environment.

[0027] Cross-domain social behavior data includes travel-related topic interaction data, non-travel-related topic interaction data, friend relationship chains and interaction records, and location check-in data. Travel-related topic interaction data refers to the target user's interactions on social networks involving travel themes, such as liking travel photos, commenting on travel articles, and participating in discussions related to travel, directly reflecting the target user's interest in travel. Non-travel-related topic interaction data refers to the target user's interactions in other areas on social networks, such as food, sports, music, or technology, reflecting the target user's overall interest characteristics. Friend relationship chains and interaction records describe the strength of connections and interactions between the target user and their social friends. Friend relationship chains show who is friends with whom, while interaction records further reflect the frequency and content type of interactions, such as the number of trips taken together and the frequency of mutual likes and comments. Location check-in data refers to the geolocation markers taken by the target user at different times and locations through social platforms. For example, checking in at a tourist attraction, restaurant, or hotel, or leaving location information during business trips and travels, not only reveals the target user's geographical activity trajectory but also reflects their travel frequency, frequently visited places, and potential travel interests.

[0028] S2: Perform multimodal feature extraction on the cross-domain social behavior data, establish text sentiment topic vectors, image scene recognition vectors, temporal behavior feature vectors and friend influence weight matrix, and perform joint embedding analysis to establish an initial embedding vector set.

[0029] Furthermore, this application also includes: performing semantic segmentation and sentiment polarity recognition on the tourism topic interaction data and non-tourism topic interaction data to establish a mapped multi-dimensional topic distribution vector and sentiment intensity vector; extracting published image data from the tourism topic interaction data and non-tourism topic interaction data, performing scene recognition, object detection, and character composition analysis on the published image data, extracting geographic landmark tags and travel mode tags, using the extracted tags to perform cross-image multi-image aggregation analysis, and establishing an image scene recognition vector; performing time series encoding on the location check-in data under a multi-scale sliding window, extracting periodic travel patterns, travel time preferences, and sudden behavior intensity, and using Fourier time-frequency transform for periodic and non-periodic segmentation to establish a time-series behavior feature vector; performing graph structure analysis on the friend relationship chain, calculating friend interaction frequency, topic overlap rate, and number of joint trips, and establishing a friend influence weight matrix based on the calculation results.

[0030] Furthermore, this application also includes: performing semantic similarity attention analysis on multidimensional topic distribution vectors to establish a first-level attention factor; constructing a dual-channel attention for positive and negative emotions on the emotion intensity vector, calculating the positive emotion weight coefficient and the negative emotion weight coefficient respectively, constructing an emotion polarity response matrix, and establishing a second-level attention factor based on the emotion polarity response matrix; performing cross-channel cross-calculation using the first-level attention factor and the second-level attention factor to establish a topic-emotion joint modulation vector; obtaining the recent emotion fluctuation index of the target user, establishing a dynamic adjustment factor using the recent emotion fluctuation index, and performing multi-layer residual fusion on the topic-emotion joint modulation vector based on the dynamic adjustment factor to establish a modulated text emotion topic vector; concatenating the modulated text emotion topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix, and performing deep embedding mapping to establish the initial embedding vector set.

[0031] Furthermore, this application also includes: performing principal component compression on the modulated text sentiment topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix; after removing low-contribution dimensions, establishing a first embedding subspace; constructing a multi-head graph convolutional network within the first embedding subspace; using the multi-head graph convolutional network to perform joint propagation of the friend relationship chain connection matrix and spliced ​​features to establish a second embedding subspace containing topological relationship awareness; and performing deep embedding mapping by fusing the global attention stream and local time series stream of the second embedding subspace using an adaptively gated dual-stream Transformer.

[0032] Specifically, semantic segmentation and sentiment polarity recognition are performed on interactive data related to tourism topics and non-tourism topics. This means semantically dividing the target user's text content on social media platforms into tourism-related and tourism-unrelated segments. For example, a long comment is broken down into thematic fragments, and then positive, neutral, or negative sentiment is judged based on the content to extract the core topics in the text. Sentiment polarity recognition measures the target user's attitude towards the topics; for example, expressing liking for "island travel" is a positive emotion, while expressing complaints about "traffic congestion" is a negative emotion. Subsequently, the identified topics and sentiments are mapped into multi-dimensional topic distribution vectors and sentiment intensity vectors, which facilitates calculation and comparison.

[0033] Next, we extract posted image data from tourism-related and non-tourism-related interactive data. We further analyze the photos uploaded or shared by target users, performing scene recognition, object detection, and people composition analysis on the posted image data. Scene recognition identifies the background of the image by recognizing the overall environment, such as a beach, mountains, or city street scene; object detection identifies specific objects in the image, such as tents, tableware, or airplanes; and people composition analysis determines the number of people and their interaction relationships in the image. Based on the recognition results, we can extract geographical landmark tags, such as the Great Wall, and travel mode tags, such as independent travel, group tours, or self-driving tours. Then, by aggregating and analyzing multiple images using these tags, we can find the stable preferences of target users and establish corresponding image scene recognition vectors to represent their tourism interests.

[0034] Then, the location check-in data is encoded using a multi-scale sliding window time series method, segmenting the check-in data of target users at different time points to discover their travel patterns. Periodic travel patterns, travel time preferences, and the intensity of sudden behaviors are extracted. Periodic travel patterns refer to recurring travel habits, such as traveling abroad every summer; travel time preferences are the target user's preferred travel times, such as a higher probability of departing in the morning; and the intensity of sudden behaviors describes the frequency with which the target user suddenly takes a trip, such as suddenly traveling to different cities multiple times within a month. Fourier time-frequency transform is used to segment periodic and non-periodic travel, thus more accurately distinguishing between regular and random travel, and encoding this into a time-series behavioral feature vector.

[0035] Finally, a graph structure analysis of the friend relationship chain is performed, representing the relationships between the target user and their friends in graph form, where nodes represent people and edges represent friend relationships. Calculating friend interaction frequency measures the activity level of friend communication, calculating topic overlap rate reflects the similarity of topics friends focus on in social interactions, and calculating the number of trips taken together determines whether friends have frequent travel activities in real life. Based on the calculation results, a friend influence weight matrix can be established to record the potential influence of different friends on user behavior; for example, a friend who frequently travels with the user will have a higher weight.

[0036] Semantic similarity attention analysis is performed on multidimensional topic distribution vectors. This represents the multidimensional vectors representing different topic interests of the target user. By calculating semantic similarity, the topic that is closer to the user's core preferences can be determined, and the correlation between topics can be found. For example, the similarity between "island tourism" and "beach vacation" is higher than that between "island tourism" and "industrial exploration". This forms the first-level attention factor, making the modeling more accurate.

[0037] A dual-channel attention model for positive and negative emotions is constructed based on the emotional intensity vector. This involves dividing the target user's emotional expression into two independent channels: positive and negative emotions. The positive emotion weight coefficient represents the strength of the target user's positive attitude towards the topic, such as liking, anticipation, or appreciation, while the negative emotion weight coefficient represents the target user's negative attitude towards the topic, such as complaining, disgust, or disappointment. By calculating the weights separately, an emotional polarity response matrix can be constructed, simultaneously recording the strength of positive and negative emotions. A second-level attention factor can then be established to ensure that both positive and negative experiences are appropriately represented.

[0038] By using first-level and second-level attention factors for cross-channel calculation, the correlation between topics and emotional attitudes are combined to establish a topic-emotion joint modulation vector. This vector not only represents the topics that the target user likes, but also reflects the strength of their attitude towards the topics. For example, if the target user has a moderate level of interest in the topic of "self-driving tour", but their emotional expression is extremely positive, then the overall result will give it a larger weight in the preference model.

[0039] The system obtains the recent sentiment fluctuation index of target users, which quantifies their emotional stability or volatility by monitoring the magnitude of changes in their emotional expression over a period of time. For example, if a person's comments over the past 30 days frequently alternate between positive and negative emotions, it indicates a high sentiment fluctuation index. A dynamic adjustment factor is then established, allowing the model to reduce the weight of preferences when emotions are unstable, avoiding bias caused by short-term fluctuations. Subsequently, multi-layer residual fusion is used to repeatedly combine and refine the dynamic adjustment effect with the obtained topic-sentiment joint modulation vector, ultimately yielding a more accurate modulated text sentiment topic vector that dynamically reflects the target users' true travel interests.

[0040] The modulated text sentiment theme vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix are concatenated to unify and integrate text, image, time, and social features.

[0041] Then, principal component compression is performed on the modulated text sentiment topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix to reduce the dimensionality of high-dimensional features from different sources. Principal component compression simplifies the data by retaining the components that best explain the overall differences, while eliminating dimensions with lower contributions to avoid information redundancy and computational burden, thus establishing the first embedding subspace.

[0042] A multi-head graph convolutional network is constructed within the first embedding subspace. Graph neural network methods are used to further model friend relationships and feature information in the dimensionality-reduced space. Multi-heading captures feature relationships from multiple different perspectives simultaneously, with each head representing an adjacency propagation pattern. This allows for joint propagation of the connection matrix of the friend relationship chain and the concatenated user features, meaning it analyzes not only the user's own features but also the topological structure and interactive influence between friend groups. The second embedding subspace, which then incorporates not only user feature information but also topological relationship awareness, enables the model to understand the user's position and relationship strength within the social network.

[0043] The adaptive-gated dual-stream Transformer fuses the global attention stream and the local time-series stream in the second embedding subspace, representing the simultaneous consideration of two different information streams during the deep learning phase. The global attention stream captures long-range dependencies between overall user features through the Transformer structure, such as a user's comprehensive behavioral patterns across different times and scenarios. The local time-series stream focuses on capturing dynamic changes within short time windows, such as someone's frequent check-ins and image uploads over the past 7 days. The adaptive gating mechanism dynamically adjusts the fusion ratio based on the weights of different inputs, ensuring that the model neither ignores global patterns nor loses local details. After dual-stream fusion, deep embedding mapping is performed, resulting in a high-quality embedding representation that integrates text, images, temporal sequences, and social structures. The resulting initial embedding vector set forms the foundational data structure for the entire travel preference modeling, containing a comprehensive representation of the user's multimodal features.

[0044] S3: Construct a preference state machine including a long-term preference layer, a periodic interest layer, and a micro-trend amplification layer using the initial embedding vector set.

[0045] Furthermore, this application also includes: slicing the initial embedded vector set into multi-scale time windows to establish long-term feature subsets, periodic feature subsets, and short-term feature subsets respectively; performing latent variable Markov chain analysis on the long-term feature subset to extract the steady-state distribution of the long-term preference layer and establish an implicit preference migration matrix; performing seasonal and holiday interest pattern analysis on the periodic feature subset to establish a periodic interest layer; constructing a micro-trend amplifier in the short-term feature subset, the micro-trend amplifier being constructed based on a preset nearest neighbor time window, calculating the amplification factor under the constraint of sudden behavior intensity within the preset nearest neighbor time window, and establishing a micro-trend amplification layer after bias correction of the amplification factor; and fusing the long-term preference layer, periodic interest layer, and micro-trend amplification layer using a multi-layer residual coupling network to establish a preference state machine.

[0046] Specifically, the initial embedded vector set is sliced ​​into multi-scale time windows, that is, the feature representation of a user at different times is divided into segments of different scales according to time to observe different patterns of behavioral change. Then, long-term feature subsets, periodic feature subsets, and short-term feature subsets are established. The long-term feature subset focuses on the target user's stable preferences over a longer period, such as the target user consistently liking a certain type of music over two years; the periodic feature subset emphasizes the target user's recurring interests in fixed periods, such as searching for travel-related information every summer; and the short-term feature subset reflects recent, immediate behavior, such as frequently browsing a certain brand of clothing in the last seven days.

[0047] Next, a latent variable Markov chain analysis is performed on the long-term feature subset, which uses a probabilistic model to infer the potential shift patterns of users' long-term preferences. A latent variable Markov chain is a model that assumes transition probabilities between states. It can extract the steady-state distribution of long-term preference layers from user behavior sequences. For example, in the long run, users may have a 70% probability of favoring electronic products and a 30% probability of favoring sporting goods. The transition relationships between different preferences are recorded using an implicit preference transition matrix; for example, the probability of shifting from sporting goods preference to electronic products is 10%. This allows us to reveal the evolution of long-term preferences between different states.

[0048] Then, seasonal and holiday interest pattern analysis is performed on the cyclical feature subset to identify regular changes in user interests at specific time points. Seasonal interest patterns might manifest as increased demand for down jackets in winter, while holiday interest patterns might show a higher frequency of gift purchases during the Spring Festival. By establishing a cyclical interest layer through analysis, it is possible to capture the fluctuations in target users' preferences over natural time cycles and specific time periods.

[0049] Furthermore, a micro-trend amplifier is constructed within the short-term feature subset to amplify subtle changes in recent behavior. The micro-trend amplifier is calculated based on a preset nearest-neighbor time window, such as detecting whether users have suddenly increased their browsing frequency for a certain type of product within the last three days. If a target user's clicks on new products increase fivefold in a short period, the signal is amplified by an amplification factor, while bias corrections are used to avoid overstating accidental behavior. This micro-trend amplification layer can more sensitively capture short-term, explosive preferences.

[0050] Finally, a multi-layer residual coupling network is used to fuse the long-term preference layer, the periodic interest layer, and the micro-trend amplification layer. This involves using a deep network to complement and superimpose features from different time scales, avoiding the limitations of single-level information. Residual coupling means preserving the original features and superimposing them with new features during the fusion process, thereby improving the stability and expressive power of the model. Furthermore, a preference state machine is established, capable of dynamically describing the overall state of the target user across long-term stable preferences, periodic interests, and short-term trend changes.

[0051] S4: Construct a target user interest diffusion network. The target user interest diffusion network updates the interest propagation path based on the friend influence weight matrix. The target user interest diffusion network is used to perform relationship-driven reinforcement of the state transition probability of the preference state machine.

[0052] Furthermore, this application also includes: performing synchronous analysis of the preference evolution of multiple friends based on the friend relationship chain and interaction records, and establishing a dynamic embedding of group preferences; and using the dynamic embedding of group preferences to perform target user interest diffusion network compensation management.

[0053] Specifically, the system performs synchronous analysis of the preference evolution of multiple friends based on friend relationship chains and interaction records. This involves comprehensively considering the strength of relationships and interactive behaviors between the target user and different friends, such as chat frequency, number of likes, and shared topics. Friend relationship chains refer to the social network structure formed between the target user and their friends, while interaction records are traces of actual communication. Through synchronous analysis, it is possible to observe how the interests and preferences of multiple friends change collectively over a period of time. For example, over the past six months, a group of friends may have gradually shifted their focus from movies to sports events. This allows for the establishment of a dynamic embedding of group preferences, representing the dynamic change process in a vectorized manner, enabling the system to capture and quantify the interest migration trends of the entire friend group.

[0054] Next, dynamic embedding of group preferences is used for target user interest diffusion network compensation management, which involves introducing group influence factors when modeling individual user interests. An interest diffusion network represents the process of interest spreading within social relationships; for example, when five friends simultaneously start paying attention to a certain type of tourist attraction, the target user is more likely to be influenced and develop an interest. Compensation management involves using dynamic embedding of group preferences to correct and supplement insufficient or biased interest information about the target user. For instance, if the target user has no significant recent behavioral data, but their friends show high interest in a new phone, the diffusion compensation mechanism can predict that the target user is also likely to be interested in that phone.

[0055] A target user interest diffusion network is an interest propagation structure built around a target user as the core node, incorporating their friend group and social relationships. It simulates the diffusion process of interests within a social network. The network is updated based on a friend influence weight matrix, a mathematical representation of the degree of influence between friends, where the values ​​reflect the contribution of different friends to changes in the user's interests. For example, if a friend has traveled with the target user 20 times in the past year and interacts frequently, their weight may be significantly higher than that of a casual acquaintance. By updating the interest propagation path, the strength and priority of interest transmission from a friend to the target user can be dynamically adjusted.

[0056] Next, a relationship-driven reinforcement of the state transition probabilities of the preference state machine is performed using the target user interest diffusion network. This means enhancing the predictive power of the state machine through group relationships. A preference state machine is a modeling tool used to depict the process of a target user switching from one interest preference state to another. The state transition probability is the likelihood that a target user will migrate from one preference to another under certain conditions. Relationship-driven reinforcement emphasizes that the probability correction stems from the influence of social relationships. For example, if eight friends have recently shifted their focus from travel to digital products, the target user's state machine will be strengthened in terms of the migration probability from travel to digital products, improving prediction accuracy.

[0057] S5: Generate target user travel preferences based on the enhanced preference state machine and visualize them through a preference radar chart.

[0058] Furthermore, this application also includes: establishing a psychological profile of the target user based on the time-series behavioral feature vector, wherein the psychological profile is constructed through risk-taking tendency features, independence features, and social features; and constructing a modulating factor for travel preferences using the psychological profile, and then performing travel preference generation compensation based on the modulating factor.

[0059] Furthermore, this application also includes: performing context-aware recognition on the target user; when the target user's context awareness triggers an event-driven mechanism, generating local driving instructions; the context-aware recognition includes geolocation change recognition and social topic concentration change recognition; and adaptively executing local parameter updates of the preference state machine according to the local driving instructions to complete the generation of the target user's travel preferences.

[0060] Furthermore, this application also includes: establishing a self-updating mechanism, which is used to perform new data collection of target users at a preset period and to use the new data collection results to perform incremental update management of the preference state machine.

[0061] Specifically, a psychological profile of the target user is built based on temporal behavioral feature vectors. This involves analyzing the behavioral patterns of the target user over time to extract patterns that reflect their psychological characteristics. Temporal behavioral feature vectors are tools that numerically represent factors such as the target user's travel frequency, travel time distribution, and intensity of sudden behavioral changes. The psychological profile, on the other hand, depicts the target user's psychological tendencies through behavioral characteristics, such as whether they like novelty, rely on group decision-making, or frequently engage in proactive social interaction. The resulting psychological profile can then serve as the basis for personalized interest modeling of the target user.

[0062] Next, the psychological profile is constructed using risk-taking tendency, independence, and social characteristics. Risk-taking tendency measures whether the target user is inclined to try high-risk or novel activities, such as whether they are willing to travel to remote areas or participate in extreme sports. Independence refers to whether the target user is more autonomous in decision-making and preferences, manifested in making choices independently without relying on a group, such as traveling abroad alone. Social characteristics reflect the target user's level of activity in interacting with others, such as whether they enjoy traveling in groups or frequently participate in gatherings during their travels.

[0063] Then, a modulating factor for travel preferences is constructed using psychological profiles. This factor represents a regulatory parameter generated based on psychological characteristics to modify the user's travel preference model. For example, if the target user has a high risk appetite and high independence, the modulating factor will increase the system's recommendation weight for adventure and independent travel content.

[0064] Finally, compensating for target user travel preferences based on modulation factors means that when target user behavioral data is insufficient or noisy, preferences can be supplemented or corrected through the moderating effect of mental profiling. For example, if a user's recent behavioral data indicates that they mainly focus on short-distance urban travel, but their mental profiling shows that their adventurous tendencies have been rising continuously over the past six months, then recommendations for outdoor adventure categories will be compensatorily increased in preference prediction.

[0065] Furthermore, context-aware identification is performed on target users, combining their current environment and real-time state to make judgments. Context-aware identification includes geolocation change identification and social topic concentration change identification. Geolocation change identification refers to monitoring changes in user location information, such as a target user moving from home to the airport, which may indicate that a trip is about to begin. Social topic concentration change identification refers to analyzing the target user's topic interaction on social platforms. If the participation or attention to a certain topic increases rapidly in a short period of time, it indicates that a change in the concentration of social topics has occurred. For example, the number of discussions about a tourist destination increases from 100 to 500 within 3 days.

[0066] Next, when the target user's context awareness triggers the event-driven mechanism, local driving instructions are generated. The event-driven mechanism refers to automatically activating the corresponding response when a key condition is detected, such as the target user's location suddenly appearing in a scenic area, generating travel-related recommendations. Local driving instructions are specific operation commands used to quickly correct the target user's preference model in specific scenarios, such as immediately increasing the weight of content related to that topic after detecting a sudden change in social topics.

[0067] Then, based on local driving instructions, the local parameters of the preference state machine are adaptively updated, that is, the parts of the user preference model closely related to the current context are dynamically adjusted. A preference state machine is a modeling tool that describes how a target user's interests change over time. Local parameter updates flexibly modify detailed features without disrupting the overall trend, such as increasing the probability of interest in a particular travel destination in the short term. Finally, through local updates, the target user's travel preferences can be generated; that is, when combined with the context, the target user's preferences become more personalized and timely.

[0068] Furthermore, establishing a self-updating mechanism means automatically maintaining and updating the user preference model. A self-updating mechanism is an automated management method used to ensure that the preference state machine continuously reflects the latest behavior and interest changes of the target users.

[0069] Next, the self-updating mechanism performs new data collection on the target user at preset intervals. This means acquiring the user's recent behavioral data at fixed time intervals, such as collecting the user's social interactions, location check-ins, and content browsing data every 7 days. New data collection refers to the latest information obtained based on existing data, which could be newly posted pictures, topics participated in, or new travel records.

[0070] Then, the incremental update management of the preference state machine is performed using the newly collected data. That is, after obtaining the latest data, local or incremental corrections are made based on the new information. The preference state machine is a model that represents the changes in user interests over time, and incremental update management can quickly adjust the state transition probabilities and preference strengths.

[0071] In summary, the tourism preference modeling method based on user social behavior mining provided in this application has the following technical effects: by achieving the technical goal of multimodal dynamic preference modeling based on user cross-domain social behavior, it can capture users' long-term preferences, periodic interests and short-term micro-trends in real time, fully integrate text, image, time-series behavior and social relationship information, and combine group social influence to predict personalized tourism preferences.

[0072] Example 2: Based on the same inventive concept as the tourism preference modeling method based on user social behavior mining in the foregoing examples, this application also provides a tourism preference modeling system based on user social behavior mining. Please refer to the appendix. Figure 2The system includes: a cross-domain social behavior data collection module 1, used to collect cross-domain social behavior data of target users, including tourism topic interaction data, non-tourism topic interaction data, friend relationship chains and interaction records, and location check-in data; an initial embedding vector set establishment module 2, used to extract multimodal features from the cross-domain social behavior data, establish text sentiment theme vectors, image scene recognition vectors, time-series behavior feature vectors, and friend influence weight matrices, and perform joint embedding analysis to establish an initial embedding vector set; a preference state machine construction module 3, used to construct a preference state machine including a long-term preference layer, a periodic interest layer, and a micro-trend amplification layer using the initial embedding vector set; a relationship-driven reinforcement execution module 4, used to construct a target user interest diffusion network, which updates the interest propagation path based on the friend influence weight matrix, and uses the target user interest diffusion network to perform relationship-driven reinforcement of the state transition probability of the preference state machine; and a visualization presentation module 5, used to generate target user tourism preferences based on the reinforced preference state machine and visualize them through a preference radar chart.

[0073] Furthermore, the tourism preference modeling system based on user social behavior mining is also used for: performing semantic segmentation and sentiment polarity recognition on the tourism topic interaction data and non-tourism topic interaction data to establish a mapped multi-dimensional topic distribution vector and sentiment intensity vector; extracting published image data from the tourism topic interaction data and non-tourism topic interaction data, performing scene recognition, object detection, and character composition analysis on the published image data, extracting geographic landmark tags and travel mode tags, using the extracted tags to perform cross-image multi-image aggregation analysis, and establishing an image scene recognition vector; performing time series encoding on the location check-in data under a multi-scale sliding window, extracting periodic travel patterns, travel time preferences, and sudden behavior intensity, and using Fourier time-frequency transform for periodic and non-periodic segmentation to establish a time-series behavior feature vector; performing graph structure analysis on the friend relationship chain, calculating friend interaction frequency, topic overlap rate, and number of joint trips, and establishing a friend influence weight matrix based on the calculation results.

[0074] Furthermore, the aforementioned tourism preference modeling system based on user social behavior mining is also used for: performing semantic similarity attention analysis on multi-dimensional topic distribution vectors to establish a first-level attention factor; constructing a dual-channel attention for positive and negative emotions on the emotion intensity vector, calculating the positive emotion weight coefficient and the negative emotion weight coefficient respectively, constructing an emotion polarity response matrix, and establishing a second-level attention factor based on the emotion polarity response matrix; performing cross-channel cross-calculation using the first-level attention factor and the second-level attention factor to establish a topic-emotion joint modulation vector; obtaining the recent emotion fluctuation index of the target user, establishing a dynamic adjustment factor using the recent emotion fluctuation index, and performing multi-layer residual fusion on the topic-emotion joint modulation vector based on the dynamic adjustment factor to establish a modulated text emotion topic vector; and concatenating the modulated text emotion topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix, and then performing deep embedding mapping to establish the initial embedding vector set.

[0075] Furthermore, the tourism preference modeling system based on user social behavior mining is also used for: performing principal component compression on the modulated text sentiment topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix; establishing a first embedding subspace after removing low-contribution dimensions; constructing a multi-head graph convolutional network in the first embedding subspace; using the multi-head graph convolutional network to perform joint propagation of the friend relationship chain connection matrix and spliced ​​features to establish a second embedding subspace containing topological relationship awareness; and performing deep embedding mapping by fusing the global attention stream and local time series stream of the second embedding subspace using an adaptively gated dual-stream Transformer.

[0076] Furthermore, the tourism preference modeling system based on user social behavior mining is also used for: multi-scale time window slicing of the initial embedded vector set to establish long-term feature subsets, periodic feature subsets, and short-term feature subsets respectively; performing latent variable Markov chain analysis on the long-term feature subset to extract the steady-state distribution of the long-term preference layer and establish an implicit preference migration matrix; performing seasonal and holiday interest pattern analysis on the periodic feature subset to establish a periodic interest layer; constructing a micro-trend amplifier in the short-term feature subset, the micro-trend amplifier being constructed based on a preset nearest neighbor time window, calculating the amplification factor under the constraint of sudden behavior intensity in the preset nearest neighbor time window, and establishing a micro-trend amplification layer after bias correction of the amplification factor; and fusing the long-term preference layer, periodic interest layer, and micro-trend amplification layer with a multi-layer residual coupling network to establish a preference state machine.

[0077] Furthermore, the aforementioned tourism preference modeling system based on user social behavior mining is also used to: establish a psychological profile of the target user based on the time-series behavioral feature vector, wherein the psychological profile is constructed through risk-taking tendency features, independence features, and social features; and after constructing a modulating factor for tourism preferences using the psychological profile, generate compensation for the target user's tourism preferences based on the modulating factor.

[0078] Furthermore, the aforementioned travel preference modeling system based on user social behavior mining is also used for: performing context-aware recognition on the target user; when the target user's context awareness triggers an event-driven mechanism, generating local driving instructions; the context-aware recognition includes geographical location change recognition and social topic concentration change recognition; and adaptively executing local parameter updates of the preference state machine according to the local driving instructions to complete the generation of the target user's travel preferences.

[0079] Furthermore, the aforementioned tourism preference modeling system based on user social behavior mining is also used for: performing synchronous analysis of the preference evolution of multiple friends based on the friend relationship chain and interaction records, and establishing a dynamic embedding of group preferences; and using the dynamic embedding of group preferences to perform target user interest diffusion network compensation management.

[0080] Furthermore, the aforementioned tourism preference modeling system based on user social behavior mining is also used to: establish a self-updating mechanism, which is used to perform new data collection of target users at a preset period and use the new data collection results to perform incremental update management of the preference state machine.

[0081] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The tourism preference modeling method and specific example based on user social behavior mining in the foregoing embodiment one are also applicable to the tourism preference modeling system based on user social behavior mining in this embodiment. Through the foregoing detailed description of the tourism preference modeling method based on user social behavior mining, those skilled in the art can clearly understand the tourism preference modeling system based on user social behavior mining in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0082] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. 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 this application. Therefore, this application 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 disclosed herein.

[0083] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for modeling travel preferences based on user social behavior mining, characterized in that, The method includes: The system collects cross-domain social behavior data from target users, including interaction data on tourism topics, interaction data on non-tourism topics, friend relationship chains and interaction records, and location check-in data. Multimodal feature extraction is performed on the cross-domain social behavior data to establish text sentiment topic vectors, image scene recognition vectors, temporal behavior feature vectors, and friend influence weight matrices. Joint embedding analysis is then performed to establish an initial embedding vector set. Construct a preference state machine comprising a long-term preference layer, a periodic interest layer, and a micro-trend amplification layer using the initial embedding vector set; Construct a target user interest diffusion network, which updates the interest propagation path based on the friend influence weight matrix, and uses the target user interest diffusion network to perform relation-driven reinforcement of the state transition probability of the preference state machine. The target user's travel preferences are generated based on the enhanced preference state machine and visualized through a preference radar chart.

2. The method for modeling travel preferences based on user social behavior mining as described in claim 1, characterized in that, The step of extracting multimodal features from the cross-domain social behavior data and establishing text sentiment topic vectors, image scene recognition vectors, temporal behavior feature vectors, and friend influence weight matrices includes: Semantic segmentation and sentiment polarity recognition are performed on the tourism topic interaction data and non-tourism topic interaction data to establish a mapped multidimensional topic distribution vector and sentiment intensity vector. Extract published image data from tourism topic interaction data and non-tourism topic interaction data, perform scene recognition, object detection and human composition analysis on the published image data, extract geographical landmark tags and travel mode tags, use the extracted tags to perform cross-image multi-image aggregation analysis, and establish image scene recognition vectors; The location check-in data is encoded using a time series encoding method with a multi-scale sliding window to extract periodic travel patterns, travel time preferences, and intensity of sudden behaviors. Fourier time-frequency transform is then used to perform periodic and non-periodic segmentation to establish a time-series behavioral feature vector. A graph structure analysis was performed on the friend relationship chain to calculate the frequency of friend interactions, topic overlap rate, and number of trips together. A friend influence weight matrix was established based on the calculation results.

3. The tourism preference modeling method based on user social behavior mining as described in claim 2, characterized in that, The joint embedding analysis, which establishes an initial embedding vector set, includes: Perform semantic similarity attention analysis on multidimensional topic distribution vectors to establish a first-level attention factor; A dual-channel attention mechanism for positive and negative emotions is constructed on the emotional intensity vector. The weight coefficients for positive and negative emotions are calculated separately to construct an emotional polarity response matrix. A second-level attention factor is established based on the emotional polarity response matrix. Cross-channel cross-calculation is performed using the first-level attention factor and the second-level attention factor to establish a topic-emotion joint modulation vector; The recent sentiment fluctuation index of the target user is obtained. After establishing a dynamic adjustment factor using the recent sentiment fluctuation index, multi-layer residual fusion is performed on the topic-emotion joint modulation vector based on the dynamic adjustment factor to establish the modulated text sentiment topic vector. After concatenating the modulated text sentiment topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix, deep embedding mapping is performed to establish the initial embedding vector set.

4. The method for modeling travel preferences based on user social behavior mining as described in claim 3, characterized in that, The execution of deep embedding mapping includes: Principal component compression is performed on the modulated text sentiment topic vector, image scene recognition vector, temporal behavior feature vector, and friend influence weight matrix. After removing low-contribution dimensions, a first embedding subspace is established. A multi-head graph convolutional network is constructed within the first embedding subspace. The multi-head graph convolutional network is used to perform joint propagation of the friend relationship chain connection matrix and spliced ​​features to establish a second embedding subspace containing topology awareness. The adaptive gating-based dual-stream Transformer fuses the global attention stream and the local time series stream in the second embedding subspace and performs deep embedding mapping.

5. The method for modeling travel preferences based on user social behavior mining as described in claim 1, characterized in that, The construction of a preference state machine comprising a long-term preference layer, a periodic interest layer, and a micro-trend amplification layer using the initial embedding vector set includes: The initial embedded vector set is sliced ​​into multi-scale time windows to establish long-term feature subsets, periodic feature subsets, and short-term feature subsets, respectively. Perform latent variable Markov chain analysis on the long-term feature subset to extract the steady-state distribution of the long-term preference layer and establish the latent preference transition matrix; Seasonal and holiday interest pattern analysis is performed on the aforementioned periodic feature subset to establish a periodic interest layer; A micro-trend amplifier is constructed in the short-term feature subset. The micro-trend amplifier is constructed based on a preset nearest neighbor time window. The amplification factor is calculated under the constraint of sudden behavior intensity in the preset nearest neighbor time window. After bias correction of the amplification factor, a micro-trend amplification layer is established. The long-term preference layer, the periodic interest layer, and the micro-trend amplification layer are fused into a multi-layer residual coupling network to establish a preference state machine.

6. The method for modeling travel preferences based on user social behavior mining as described in claim 1, characterized in that, The step of generating target user travel preferences based on the enhanced preference state machine includes: A psychological profile of the target user is established based on the time-series behavioral feature vector, and the psychological profile is constructed through risk-taking tendency features, independence features, and social features; After constructing a modulating factor for travel preferences using the aforementioned psychological profile, compensation for the target user's travel preferences is generated based on the modulating factor.

7. The method for modeling travel preferences based on user social behavior mining as described in claim 1, characterized in that, The step of generating target user travel preferences based on the enhanced preference state machine also includes: Context-aware recognition is performed on the target user. When the target user's context awareness triggers the event-driven mechanism, a local driving instruction is generated. The context-aware recognition includes geolocation change recognition and social topic concentration change recognition. The local parameters of the preference state machine are updated adaptively according to the local driving instructions to complete the generation of the target user's travel preferences.

8. The method for modeling travel preferences based on user social behavior mining as described in claim 1, characterized in that, The construction of the target user interest diffusion network includes: Based on the aforementioned friend relationship chain and interaction records, a synchronous analysis of the preference evolution of multiple friends is performed to establish a dynamic embedding of group preferences. The aforementioned dynamic embedding of group preferences is used for target user interest diffusion network compensation management.

9. The method for modeling travel preferences based on user social behavior mining as described in claim 1, characterized in that, The step of generating target user travel preferences based on the enhanced preference state machine also includes: A self-updating mechanism is established, which is used to collect new data of the target user at a preset period and use the results of the new data collection to manage the incremental update of the preference state machine.

10. A tourism preference modeling system based on user social behavior mining, characterized in that, The steps for implementing the travel preference modeling method based on user social behavior mining as described in any one of claims 1 to 9 include: The cross-domain social behavior data collection module is used to collect cross-domain social behavior data of target users. The cross-domain social behavior data includes tourism topic interaction data, non-tourism topic interaction data, friend relationship chain and interaction records, and location check-in data. The initial embedding vector set establishment module is used to extract multimodal features from the cross-domain social behavior data, establish text sentiment topic vectors, image scene recognition vectors, temporal behavior feature vectors and friend influence weight matrices, and perform joint embedding analysis to establish an initial embedding vector set. A preference state machine construction module is used to construct a preference state machine including a long-term preference layer, a periodic interest layer, and a micro-trend amplification layer using the initial embedding vector set; A relationship-driven reinforcement execution module is used to construct a target user interest diffusion network. The target user interest diffusion network updates the interest propagation path based on the friend influence weight matrix and uses the relationship-driven reinforcement of the state transition probability of the preference state machine to execute the target user interest diffusion network. The visualization module is used to generate target user travel preferences based on the enhanced preference state machine and visualize them through a preference radar chart.

Images