A data processing method for improving a customized travel recommendation algorithm model

Abstract

The application discloses a data processing method for improving a customized tourism recommendation algorithm model, relates to the technical field of big data processing, and effectively solves three core problems of a tourism recommendation system: for data sparsity, multi-source data fusion is realized through time window segmentation interval 1h / 24h / 7d time series data set construction and cross-modal contrast learning; for personalized missing, an attention mechanism is used to calculate user interest weight Ayh and three types of interest decay models are established; for poor dynamic adaptability, an asynchronous data flow framework and an environment perception mechanism with a response less than or equal to a fixed time length are constructed, double matching degree thresholds are set to realize real-time adjustment, and finally, a closed-loop optimization system is formed through reinforcement learning, so that the recommendation accuracy and user experience are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, specifically to a data processing method for an improved customized travel recommendation algorithm model. Background Technology

[0002] Travel recommendation algorithms originated from collaborative filtering technology in the 1990s. With the development of the internet and big data, they have gradually integrated content recommendation, knowledge-based recommendation, and hybrid recommendation technologies. In recent years, advanced technologies such as deep learning, reinforcement learning, and graph neural networks have been introduced, further improving the personalization and accuracy of recommendations.

[0003] However, current travel recommendation systems generally suffer from the following problems:

[0004] 1. Data sparsity: The mismatch between the dimensions of behavioral data and the feature data of scenic spots leads to insufficient recommendation accuracy;

[0005] 2. Lack of personalization: Traditional collaborative filtering algorithms do not fully consider users' real-time preferences and scene differences;

[0006] 3. Poor dynamic adaptability: It lacks a response mechanism to real-time factors such as holidays and weather. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an improved data processing method for customized travel recommendation algorithm models, which solves the technical deficiencies mentioned in the background section.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a data processing method for an improved customized travel recommendation algorithm model, applied to a user interaction system of a mobile device, comprising the following steps:

[0009] S1. Collect user group behavior-related data and travel preference-related data, and construct a user preference data set; through the location module of mobile devices and external data interfaces, collect environmental information including user location, timestamp, weather, and social media activity in the user's area, and establish an environmental information data set;

[0010] The environmental information dataset is segmented into time windows, and the user preference dataset of the user group is synchronized and time-aligned to construct a time-series dataset.

[0011] S2. Based on the time-series dataset, construct a time-series user profile; extract environmental context features that affect the accuracy of recommendations based on the environmental information dataset, and use a multilayer perceptron neural network combined with an attention mechanism to model the weight influence of users on different recommendation factors, while dynamically adjusting the recommendation strategy.

[0012] S3. Establish an interest classification mechanism to divide user interests into short-term interests, long-term interests, and context-triggered interests; then adjust the priority of recommended content and optimize the recommendation order; finally, optimize the weight of user interests based on historical user feedback.

[0013] S4. Prioritize matching short-term interest recommendation lists, then optimize the recommendation order by combining long-term interests and context-triggered interests; next, build and combine a real-time behavior analysis model to monitor user interaction, and trigger an intelligent interaction adjustment mechanism when the recommended content does not meet the user's needs.

[0014] Preferably, S2 specifically includes:

[0015] The time-series dataset includes: representing the temporal characteristics of user group behavior as dimension d1; representing the environmental context characteristics of user group as dimension d2; representing the fused feature dimension after timestamp alignment as F=d1+d2; then extracting timestamp information from the time-series dataset and combining it with behavior-related data to divide it into time windows;

[0016] Based on the time-series dataset, user preference data sets are extracted, and user behavior patterns are extracted based on time series modeling methods. Cluster analysis is performed on the behavioral characteristics of users in different time windows to form time-series user feature vectors. Combined with environmental information such as user geographic location, social media activity, and weather, a time-series user profile model is constructed.

[0017] We use an attention-based feature fusion model to analyze the correlation between user behavior and contextual features, and generate user interest preference vectors in different scenarios. Finally, we integrate the temporal features of user behavior, the results of context-aware modeling, and the interest preference vectors into a user profile feature vector, where each time window corresponds to a comprehensive representation that includes user behavior features, contextual features, and interest preference vectors.

[0018] Preferably, S2 further includes:

[0019] An environmental information dataset is extracted, and an information gain-based feature selection algorithm is used to perform feature engineering on environmental information, including timestamps, weather conditions, and social media activity in the user's location, to extract context-related features. These features are then correlated with the user preference dataset to generate context-aware feature vectors.

[0020] User profile feature vectors and context-aware feature vectors are input into an MLP model to extract high-dimensional feature representations. Then, an attention mechanism is used to calculate the user interest weight Ayh for recommendation factors, including attraction type, activity type, and time preference. The formula for calculating the user interest weight Ayh is as follows:

[0021] ;

[0022] In the formula, F is the fusion feature dimension, which is the fusion representation of user profile features and context-aware features, and depends on the dimensions of user profile features and context features;

[0023] W f is the weight matrix of the MLP model, used to map the input features F to the hidden layers; its value is an h×(d1+d2) matrix, where h is the dimension of the hidden layers;

[0024] bf is the bias term of the MLP model, used to adjust the output of the hidden layer; its value is an h×1 vector.

[0025] W a The weight matrix for the attention mechanism is used to calculate the user's attention score for each recommendation factor. Its value is a k×h matrix, where k is the number of recommendation factors.

[0026] The Softmax function is used to normalize attention scores into a probability distribution; its output ranges from (0, 1], and the sum of all output values ​​is 1.

[0027] This is the hyperbolic tangent activation function, used to introduce a nonlinear transformation, with an output range of (-1, 1).

[0028] Preferably, S2 further includes:

[0029] Real-time collection and analysis of multimodal data, including user historical data, text reviews, click behavior, and social interactions, are performed. A user interest weight threshold A is assigned to each recommendation factor, and this threshold is compared with the user interest weight Ayh to evaluate the user's level of interest in various recommendation factors, including attraction type, activity type, and time preference. Specific evaluation content is as follows:

[0030] When the user interest weight Ayh i ≥ User interest weight threshold A i This indicates that the user's interest in the current recommendation factor is "qualified".

[0031] When the user interest weight Ayh i <User interest weight threshold A i This indicates that the user's interest in the current recommendation factor is "unqualified".

[0032] Among them, the user interest weight Ayh i Let Ayh represent the user interest weight of the i-th recommendation factor, and let A be the user interest weight threshold. i Let A represent the user interest weight threshold of the i-th recommendation factor; i is the category index of the recommendation factor.

[0033] Finally, based on the comparative evaluation results, the priority of the recommended content is dynamically adjusted:

[0034] For recommendation factors that meet the "qualification" criteria for interest, increase the recommendation priority;

[0035] For recommendation factors that are "unqualified" in terms of interest, reduce their recommendation priority or exclude them.

[0036] Preferably, S3 specifically includes:

[0037] Read users' historical behavior data and extract user interest features, including clicks, browsing, favorites, and reservations, from the historical behavior data and time-series user profiles; then classify user interests into three categories: short-term interests, long-term interests, and context-triggered interests.

[0038] For short-term interests, identification and modeling are performed based on behavioral data, including the user's clicks and browsing history over the past week.

[0039] For long-term interests, identification and modeling are performed based on behavioral data, including the user's continuous attention records over the past three months.

[0040] For context-triggered interests, identification and modeling are performed based on contextual feature data, including weather, holidays, and geographical location. A decision tree-based rule engine is used, which includes classifying user interests and assigning initial weights to each category. The classified user interests and their corresponding initial weights are then output.

[0041] Preferably, S3 further includes:

[0042] An interest decay model is established for each user interest type, specifically including:

[0043] For short-term interest, an exponential decay function is used. Modeling is performed, where I0 represents the initial interest intensity, λ represents the decay rate, and t represents the time interval;

[0044] For long-term interests, a linear decay model is used for modeling.

[0045] For context-triggered interest, an attention-based context-aware model is used to capture the correlation between user behavior and contextual features;

[0046] Next, based on behavioral data and time-series user profiles, the interest decay coefficient Bsj for each type of interest is calculated;

[0047] For short-term interests, the initial interest intensity I0 and decay rate λ are calculated based on short-term user behavior data.

[0048] For long-term interests, the initial interest intensity I0 and decay rate λ are calculated based on long-term user behavior data.

[0049] For context-triggered interest, the dynamic decay rate λ is calculated by combining contextual feature data; the specific calculation formula for the interest decay coefficient Bsj is as follows:

[0050] ;

[0051] In the formula, for short-term interest, the larger the decay rate λ, the faster the interest decays;

[0052] For long-term interests, the smaller the decay rate λ, the slower the interest decays.

[0053] For context-triggered interest, the decay rate λ is adjusted according to the timeliness of the context features;

[0054] t represents the time interval, which is the length of time from the moment the interest arises to the current moment; e represents the base of the natural logarithm.

[0055] The range of the interest decay coefficient Bsj is obtained through experimental data fitting and dynamic adjustment of context. The effective range of the interest decay coefficient Bsj is set to 0 < Bsj ≤ 1. When the interest decay coefficient Bsj is closer to 1, it indicates a greater degree of interest decay. When the interest decay coefficient Bsj is closer to 0, it indicates a smaller degree of interest decay.

[0056] Preferably, S3 further includes:

[0057] When interest has waned, we use deep reinforcement learning algorithms to build interest matching rules and adjust the priority and order of recommended content based on user feedback. This includes building interest matching rules, designing reward functions, and optimizing interest weights.

[0058] Secondly, a reward function was designed to optimize the weight of user interest based on user satisfaction metrics.

[0059] Finally, the optimized user interest weights are fed back to the recommendation system to generate the next round of recommendation lists, and user interaction behavior is monitored in real time. When the recommended content does not meet the user's needs, an intelligent interaction adjustment mechanism is triggered to further optimize the user interest weights.

[0060] Preferably, S4 specifically includes:

[0061] Based on the time-series user profile, short-term interest features are extracted and a short-term interest recommendation list is constructed. The short-term interest features include statistical information on the user's recent clicks, browsing, collection and booking records, as well as the trend of the user's interest preferences in different time intervals.

[0062] Then, the user's long-term interest characteristics are modeled, including calculating the long-term preference distribution based on the user's historical behavior data, and optimizing the recommended content in combination with the user's stable preference characteristics;

[0063] Secondly, extract users' contextual triggering interest features, including information such as geographical location, weather and social activity in the current environmental information data set, calculate users' contextual interest weights, and dynamically adjust the ranking strategy of recommended content based on contextual interest weights;

[0064] Finally, during the recommendation process, following a hierarchical optimization strategy, the recommendation list is first matched with short-term interest recommendations, and the recommendation order is optimized based on the weight calculation results of long-term interests and context-triggered interests. Specifically, the distribution of users' long-term interest preferences is calculated, and the recommended content is optimized by combining stable preference features. Secondly, context-triggered interest features, including geographical location, weather, and social activity, are extracted, and context-triggered interest weights are calculated. Finally, the recommendation ranking strategy is dynamically adjusted based on the comprehensive weight of long-term interests and context-triggered interests.

[0065] Preferably, S4 further includes:

[0066] A user interest matching model is constructed, taking into input user historical behavior data, time-series user profiles, and contextual feature data. The model is trained using a reinforcement learning framework, including deep learning rules. Then, the recommended content matching degree (Cnr) is calculated using the user interest matching model, with the specific calculation formula as follows:

[0067] ;

[0068] Where n represents the total number of interest categories, j represents the index of the interest category, i.e. the different classification numbers of user interests; wj is the weight coefficient of the j-th interest category, Sj is the similarity between the recommended content and the j-th interest category, and f(t) is the timeliness correction factor.

[0069] Preferably, S4 further includes:

[0070] A first matching threshold C1 and a second matching threshold C2 are preset, and compared with the matching degree Cnr of the recommended content to evaluate the real-time matching degree between the recommended content and the user's interests, and to respond in real time to changes in user needs; and the first matching threshold C1 is greater than the second matching threshold C2. The specific evaluation content is as follows:

[0071] When the matching degree of the recommended content Cnr is greater than or equal to the first matching degree evaluation threshold C1, it indicates that the recommended content is highly matched with the user's interests. The system makes recommendations directly without adjusting the recommendation strategy and monitors user interaction feedback data.

[0072] When the second matching threshold C2 ≤ recommended content matching degree Cnr < first matching threshold C1: it indicates that the recommended content does not match the user's interests well, but is not lower than the minimum acceptable second matching threshold C2. In this case, the recommended content is optimized, and the specific adjustment methods include:

[0073] (1) Increase the weight of short-term interests and adjust the order of recommended content;

[0074] (2) Incorporate environmental factors to increase the proportion of personalized recommendations;

[0075] When the matching degree of the recommended content Cnr is less than the second matching degree evaluation threshold C2, it indicates that the matching degree of the recommended content with the user's interests is not up to standard and is lower than the minimum acceptable range of the second matching degree evaluation threshold C2. At this time, the intelligent interaction adjustment mechanism is triggered to immediately replace the recommended content and recalculate the matching degree of the recommended content Cnr.

[0076] After the recommended content is pushed, the system continuously collects user feedback data, including clicks, dwell time, favorites, and reservations, and calculates feedback scores; it uses reinforcement learning models to adjust the recommendation strategy so that the recommendation system can adapt to changes in user interests.

[0077] This invention provides a data processing method for an improved customized travel recommendation algorithm model. It has the following beneficial effects:

[0078] (1) This improved customized tourism recommendation algorithm model data processing method addresses the data sparsity problem. This technical solution achieves deep integration of multi-dimensional data by constructing a time-series dataset and a cross-domain knowledge graph; it uses sliding window statistics and time-series clustering methods to extract local change features and global patterns of user behavior; it uses a heterogeneous data co-representation mapping network to uniformly represent structured and unstructured data; it enhances the correlation of information in different modalities through cross-modal contrastive learning; and finally, it integrates user profile feature vectors with context-aware feature vectors, which significantly improves data utilization efficiency and recommendation accuracy.

[0079] (2) This improved customized tourism recommendation algorithm model's data processing method addresses the problem of lack of personalization. The scheme designs a user interest weight calculation model based on an attention mechanism; extracts high-dimensional feature representations through an MLP model; and employs an attention mechanism-based weight matrix W. a Calculate the user's interest weight Ayh for each recommendation factor; combine it with a preset threshold A i Dynamic evaluation is conducted; three types of interest decay models are established; and personalized optimization of the recommendation strategy is achieved through reinforcement learning-driven interest matching rules.

[0080] (3) This improved customized tourism recommendation algorithm model data processing method addresses the problem of poor dynamic adaptability. The solution constructs a multi-layer asynchronous feedback data stream processing framework; adopts time window segmentation technology to capture real-time behavior changes; dynamically adjusts recommendation parameters through environmental perception interest regulation mechanism; sets dual matching degree thresholds to achieve intelligent interactive adjustment; and finally forms a complete adaptive system that includes real-time behavior analysis, dynamic weight adjustment and closed-loop feedback optimization. Attached Figure Description

[0081] Figure 1 This is a flowchart illustrating the steps of a data processing method for an improved customized tourism recommendation algorithm model according to the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] Example 1

[0084] Please see Figure 1 This invention provides a data processing method for an improved customized travel recommendation algorithm model, applied to a user interaction system of a mobile device, comprising the following steps:

[0085] S1. Collect user group behavior-related data and travel preference-related data, and construct a user preference data set; through the location module of mobile devices and external data interfaces, collect environmental information including user location, timestamp, weather, and social media activity in the user's area, and establish an environmental information data set;

[0086] The environmental information dataset is segmented into time windows, and the user preference dataset of the user group is synchronized and time-aligned to construct a time-series dataset.

[0087] S2. Based on the time-series dataset, construct a time-series user profile; extract environmental context features that affect the accuracy of recommendations based on the environmental information dataset, and use a multilayer perceptron neural network combined with an attention mechanism to model the weight influence of users on different recommendation factors, while dynamically adjusting the recommendation strategy.

[0088] S3. Establish an interest classification mechanism to divide user interests into short-term interests, long-term interests, and context-triggered interests; then adjust the priority of recommended content and optimize the recommendation order; finally, optimize the weight of user interests based on historical user feedback.

[0089] S4. Prioritize matching short-term interest recommendation lists, then optimize the recommendation order by combining long-term interests and context-triggered interests; next, build and combine a real-time behavior analysis model to monitor user interaction, and trigger an intelligent interaction adjustment mechanism when the recommended content does not meet the user's needs.

[0090] In this embodiment, S1 specifically includes:

[0091] The extracted user behavior and travel preference-related data, including user clicks, browsing, favorites and booking records, are preprocessed to construct a user preference dataset; at the same time, environmental information is collected through the positioning module of mobile devices and external data interfaces, and is preprocessed to establish an environmental information dataset.

[0092] Furthermore, user behavior and travel preference-related data, as well as environmental information, are segmented according to preset time intervals. First, the user behavior and travel preference-related data, as well as environmental information, are timestamped and segmented according to preset time intervals, including hourly, daily, or weekly intervals. Next, the data within each time interval is aggregated and features are extracted, including calculating the frequency of user behavior, preference changes, and statistical values ​​of environmental features within each time interval. Finally, the segmented data is stored as a time-series dataset.

[0093] Example 2

[0094] S2 specifically includes:

[0095] The time-series dataset includes: representing the temporal characteristics of user group behavior as dimension d1; representing the environmental context characteristics of user group as dimension d2; representing the fused feature dimension after timestamp alignment as F=d1+d2; then extracting timestamp information from the time-series dataset and combining it with behavior-related data to divide it into time windows;

[0096] Based on the time-series dataset, user preference data sets are extracted, and user behavior patterns are extracted based on time series modeling methods. Cluster analysis is performed on the behavioral characteristics of users in different time windows to form time-series user feature vectors. Combined with environmental information such as user geographic location, social media activity, and weather, a time-series user profile model is constructed.

[0097] We use an attention-based feature fusion model to analyze the correlation between user behavior and contextual features, and generate user interest preference vectors in different scenarios. Finally, we integrate the temporal features of user behavior, the results of context-aware modeling, and the interest preference vectors into a user profile feature vector, where each time window corresponds to a comprehensive representation that includes user behavior features, contextual features, and interest preference vectors.

[0098] S2 also includes:

[0099] An environmental information dataset is extracted, and an information gain-based feature selection algorithm is used to perform feature engineering on environmental information, including timestamps, weather conditions, and social media activity in the user's location, to extract context-related features. These features are then correlated with the user preference dataset to generate context-aware feature vectors.

[0100] User profile feature vectors and context-aware feature vectors are input into an MLP model to extract high-dimensional feature representations. Then, an attention mechanism is used to calculate the user interest weight Ayh for recommendation factors, including attraction type, activity type, and time preference. The formula for calculating the user interest weight Ayh is as follows:

[0101] ;

[0102] In the formula, F is the fusion feature dimension, which is the fusion representation of user profile features and context-aware features, and depends on the dimensions of user profile features and context features;

[0103] W f is the weight matrix of the MLP model, used to map the input features F to the hidden layers; its value is an h×(d1+d2) matrix, where h is the dimension of the hidden layers;

[0104] bf is the bias term of the MLP model, used to adjust the output of the hidden layer; its value is an h×1 vector.

[0105] W a The weight matrix for the attention mechanism is used to calculate the user's attention score for each recommendation factor. Its value is a k×h matrix, where k is the number of recommendation factors.

[0106] The Softmax function is used to normalize attention scores into a probability distribution; its output ranges from (0, 1], and the sum of all output values ​​is 1.

[0107] This is the hyperbolic tangent activation function, used to introduce a nonlinear transformation, with an output range of (-1, 1).

[0108] S2 also includes:

[0109] Real-time collection and analysis of multimodal data, including user historical data, text reviews, click behavior, and social interactions, are performed. A user interest weight threshold A is assigned to each recommendation factor, and this threshold is compared with the user interest weight Ayh to evaluate the user's level of interest in various recommendation factors, including attraction type, activity type, and time preference. Specific evaluation content is as follows:

[0110] When the user interest weight Ayh i≥ User interest weight threshold A i This indicates that the user's interest in the current recommendation factor is "qualified".

[0111] When the user interest weight Ayh i <User interest weight threshold A i This indicates that the user's interest in the current recommendation factor is "unqualified".

[0112] Among them, the user interest weight Ayh i Let Ayh represent the user interest weight of the i-th recommendation factor, and let A be the user interest weight threshold. i Let A represent the user interest weight threshold of the i-th recommendation factor; i is the category index of the recommendation factor.

[0113] Finally, based on the comparative evaluation results, the priority of the recommended content is dynamically adjusted:

[0114] For recommendation factors that meet the "qualification" criteria for interest, increase the recommendation priority;

[0115] For recommendation factors that are "unqualified" in terms of interest, reduce their recommendation priority or exclude them.

[0116] In this embodiment, the intelligence level of the tourism recommendation system is significantly improved through time-series dataset processing and multi-dimensional feature analysis; local features of user behavior are accurately captured based on time window segmentation technology (preset interval 1h / 24h / 7d) and sliding window statistics (mean / variance / maximum / minimum); global behavior patterns are extracted through time-series cluster analysis (k-means cluster number n=5).

[0117] An attention mechanism feature fusion model (weight matrix W) is constructed by combining environmental context features (dimension d2). a (Dimension k×h); Generate high-dimensional feature representations (input dimension d1+d2) using an MLP model (hidden layer dimension h=64);

[0118] The user interest weight Ayh is calculated using the formula (including the weight matrix W). f Quantify recommendation factor preferences using dimensions h×(d1+d2) and bias term bf (size h×1); set a dynamic threshold Ai to assess interest eligibility (when Aiyh) i Priority is increased when Ai is greater than or equal to 7; finally, behavioral temporal features, context-aware vectors, and interest preference vectors are integrated to form a comprehensive user profile. The sliding window size for behavioral temporal features is w=7, the information gain threshold for context-aware vectors is IG>0.3, and the activation output range of the interest preference vector tanh is (-1, 1]. This scheme achieves a complete closed-loop optimization from raw data to personalized recommendations, enabling the system to capture interest changes in real time, dynamically adjust recommendation strategies, and accurately predict user needs. The calculation formula is: In the formula, x represents the input value, that is, the value that needs to be nonlinearly transformed;

[0119] Example 3

[0120] S3 specifically includes:

[0121] Read users' historical behavior data and extract user interest features, including clicks, browsing, favorites, and reservations, from the historical behavior data and time-series user profiles; then classify user interests into three categories: short-term interests, long-term interests, and context-triggered interests.

[0122] For short-term interests, identification and modeling are performed based on behavioral data, including the user's clicks and browsing history over the past week.

[0123] For long-term interests, identification and modeling are performed based on behavioral data, including the user's continuous attention records over the past three months.

[0124] For context-triggered interests, identification and modeling are performed based on contextual feature data, including weather, holidays, and geographical location. A decision tree-based rule engine is used, which includes classifying user interests and assigning initial weights to each category. The classified user interests and their corresponding initial weights are then output.

[0125] S3 also includes:

[0126] An interest decay model is established for each user interest type, specifically including:

[0127] For short-term interest, an exponential decay function is used. Modeling is performed, where I0 represents the initial interest intensity, λ represents the decay rate, and t represents the time interval;

[0128] For long-term interests, a linear decay model is used for modeling.

[0129] For context-triggered interest, an attention-based context-aware model is used to capture the correlation between user behavior and contextual features;

[0130] Next, based on behavioral data and time-series user profiles, the interest decay coefficient Bsj for each type of interest is calculated;

[0131] For short-term interests, the initial interest intensity I0 and decay rate λ are calculated based on short-term user behavior data.

[0132] For long-term interests, the initial interest intensity I0 and decay rate λ are calculated based on long-term user behavior data.

[0133] For context-triggered interest, the dynamic decay rate λ is calculated by combining contextual feature data; the specific calculation formula for the interest decay coefficient Bsj is as follows:

[0134] ;

[0135] In the formula, for short-term interest, the larger the decay rate λ, the faster the interest decays;

[0136] For long-term interests, the smaller the decay rate λ, the slower the interest decays.

[0137] For context-triggered interest, the decay rate λ is adjusted according to the timeliness of the context features;

[0138] t represents the time interval, which is the length of time from the moment the interest arises to the current moment; e represents the base of the natural logarithm.

[0139] The range of the interest decay coefficient Bsj is obtained through experimental data fitting and dynamic adjustment of context. The effective range of the interest decay coefficient Bsj is set to 0 < Bsj ≤ 1. When the interest decay coefficient Bsj is closer to 1, it indicates a greater degree of interest decay. When the interest decay coefficient Bsj is closer to 0, it indicates a smaller degree of interest decay.

[0140] S3 also includes:

[0141] When interest has waned, we use deep reinforcement learning algorithms to build interest matching rules and adjust the priority and order of recommended content based on user feedback. This includes building interest matching rules, designing reward functions, and optimizing interest weights.

[0142] Secondly, a reward function was designed to optimize the weight of user interest based on user satisfaction metrics.

[0143] Finally, the optimized user interest weights are fed back to the recommendation system to generate the next round of recommendation lists, and user interaction behavior is monitored in real time. When the recommended content does not meet the user's needs, an intelligent interaction adjustment mechanism is triggered to further optimize the user interest weights.

[0144] In this embodiment, by reading users' historical behavior data and time-series user profiles, user interest features are extracted and categorized into three types: short-term interests, long-term interests, and context-triggered interests, thereby achieving precise interest modeling. Short-term interests are identified and modeled using users' click and browsing behavior data from the past week, reflecting real-time changes in user interests. Long-term interests are based on users' continuous attention records over the past few months, depicting stable preference trends. Context-triggered interests combine contextual feature data such as weather, holidays, and geographical location to capture dynamic changes in user interests influenced by environmental factors. A rule engine is used to classify user interests and assign initial weights to optimize the rationality of interest weights. Furthermore, an interest decay model is established for each type of interest. Short-term interests are modeled using an exponential decay function, with the initial interest intensity I0 and decay rate λ used to calculate the decay trend of interest over time. Long-term interests are modeled using a linear decay model to ensure the smooth decay of users' long-term attention points. Context-triggered interests employ a context-aware model based on an attention mechanism. The dynamic decay rate λ is calculated by combining contextual feature data to reflect the impact of environmental changes on user interests. An interest decay coefficient Bsj is calculated based on behavioral data and time-series user profiles to ensure the recommendation system accurately perceives changes in interest. The value range of the interest decay coefficient Bsj is 0 < Bsj ≤ 1; when Bsj = 1, it indicates no interest decay, and when Bsj is close to 0, it indicates that interest has decayed. When user interest has decayed, a deep reinforcement learning algorithm is used to construct interest matching rules, and the priority and order of recommended content are dynamically adjusted based on user feedback to ensure the adaptability and accuracy of the recommended content. Furthermore, a reward function is designed to optimize user interest weights based on user satisfaction indicators, enabling the system to adapt to changes in user needs. The optimized user interest weights are fed back to the recommendation system to generate the next round of recommendation lists and to monitor user interaction behavior in real time. When recommended content does not meet user needs, an intelligent interaction adjustment mechanism is triggered to further optimize user interest weights, improving the long-term intelligence level of the recommendation system and user satisfaction.

[0145] Example 4

[0146] S4 specifically includes:

[0147] Based on the time-series user profile, short-term interest features are extracted and a short-term interest recommendation list is constructed. The short-term interest features include statistical information on the user's recent clicks, browsing, collection and booking records, as well as the trend of the user's interest preferences in different time intervals.

[0148] Then, the user's long-term interest characteristics are modeled, including calculating the long-term preference distribution based on the user's historical behavior data, and optimizing the recommended content in combination with the user's stable preference characteristics;

[0149] Secondly, extract users' contextual triggering interest features, including information such as geographical location, weather and social activity in the current environmental information data set, calculate users' contextual interest weights, and dynamically adjust the ranking strategy of recommended content based on contextual interest weights;

[0150] Finally, during the recommendation process, following a hierarchical optimization strategy, the recommendation list is first matched with short-term interest recommendations, and the recommendation order is optimized based on the weight calculation results of long-term interests and context-triggered interests. Specifically, the distribution of users' long-term interest preferences is calculated, and the recommended content is optimized by combining stable preference features. Secondly, context-triggered interest features, including geographical location, weather, and social activity, are extracted, and context-triggered interest weights are calculated. Finally, the recommendation ranking strategy is dynamically adjusted based on the comprehensive weight of long-term interests and context-triggered interests.

[0151] S4 also includes:

[0152] A user interest matching model is constructed, taking into input user historical behavior data, time-series user profiles, and contextual feature data. The model is trained using a reinforcement learning framework, including deep learning rules. Then, the recommended content matching degree (Cnr) is calculated using the user interest matching model, with the specific calculation formula as follows:

[0153] ;

[0154] Where n represents the total number of interest categories, j represents the index of the interest category, i.e. the different classification numbers of user interests; wj is the weight coefficient of the j-th interest category, Sj is the similarity between the recommended content and the j-th interest category, and f(t) is the timeliness correction factor.

[0155] S4 also includes:

[0156] A first matching threshold C1 and a second matching threshold C2 are preset, and compared with the matching degree Cnr of the recommended content to evaluate the real-time matching degree between the recommended content and the user's interests, and to respond in real time to changes in user needs; and the first matching threshold C1 is greater than the second matching threshold C2. The specific evaluation content is as follows:

[0157] When the matching degree of the recommended content Cnr is greater than or equal to the first matching degree evaluation threshold C1, it indicates that the recommended content is highly matched with the user's interests. The system makes recommendations directly without adjusting the recommendation strategy and monitors user interaction feedback data.

[0158] When the second matching threshold C2 ≤ recommended content matching degree Cnr < first matching threshold C1: it indicates that the recommended content does not match the user's interests well, but is not lower than the minimum acceptable second matching threshold C2. In this case, the recommended content is optimized, and the specific adjustment methods include:

[0159] (1) Increase the weight of short-term interests and adjust the order of recommended content;

[0160] (2) Incorporate environmental factors to increase the proportion of personalized recommendations;

[0161] When the matching degree of the recommended content Cnr is less than the second matching degree evaluation threshold C2, it indicates that the matching degree of the recommended content with the user's interests is not up to standard and is lower than the minimum acceptable range of the second matching degree evaluation threshold C2. At this time, the intelligent interaction adjustment mechanism is triggered to immediately replace the recommended content and recalculate the matching degree of the recommended content Cnr.

[0162] After the recommended content is pushed, the system continuously collects user feedback data, including clicks, dwell time, favorites, and reservations, and calculates feedback scores; it uses reinforcement learning models to adjust the recommendation strategy so that the recommendation system can adapt to changes in user interests.

[0163] In this embodiment, this step constructs a user interest matching model to comprehensively evaluate short-term, long-term, and context-triggered interest features of users, thereby optimizing the calculation of the recommended content matching coefficient Cnr and ensuring the accuracy of the recommended content. Specifically, short-term interest features are extracted using statistical information from users' recent clicks, browsing, favorites, and booking records, combined with trends in interest preferences across different time intervals, reflecting recent fluctuations in user interests. Long-term interest features are calculated based on users' historical behavior data to determine long-term preference distribution, and combined with users' stable preference features to optimize recommended content, ensuring the continued relevance of recommendations. Context-triggered interest features calculate users' contextual interest weights based on information such as geographical location, weather, and social activity in the current environmental information dataset, and dynamically adjust the ranking strategy of recommended content based on these contextual interest weights, enhancing the real-time adaptability of recommendations.

[0164] Subsequently, a user interest matching model was trained using a reinforcement learning framework, and a recommended content matching degree evaluation mechanism was used for real-time optimization. The first matching degree evaluation threshold C1 and the second matching degree evaluation threshold C2 were used to evaluate the rationality of the recommended content matching degree coefficient Cnr. C1 was set as a higher matching standard, and C2 was set as the lowest acceptable range.

[0165] When the matching coefficient of the recommended content Cnr is greater than or equal to the first matching evaluation threshold C1, the system directly recommends and monitors user interaction feedback data to improve the system's learning ability; when the second matching evaluation threshold C2 is less than or equal to the matching coefficient of the recommended content Cnr and less than the first matching evaluation threshold C1, the system increases the weight of short-term interests, recalculates the cosine similarity between the recommended content and the user's interest vector, generates a new recommendation list in descending order of the updated scores, adjusts the order of recommended content, or increases the proportion of personalized recommended content in combination with environmental factors to optimize the recommended content.

[0166] When the matching coefficient Cnr of the recommended content is less than the second matching evaluation threshold C2, the intelligent interaction adjustment mechanism is triggered. The recommended content is immediately replaced and the matching coefficient Cnr is recalculated to ensure the adaptability and accuracy of the recommended content. Finally, based on user feedback data, including clicks, dwell time, collection and booking status, a feedback score is calculated, and a reinforcement learning model is used to dynamically adjust the recommendation strategy, enabling the system to adapt to changes in user interests and improve the personalization capability and long-term effectiveness of the recommendation system.

[0167] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data processing method for an improved customized travel recommendation algorithm model, applied to a user interaction system of a mobile device, characterized in that: Includes the following steps: S1. Collect user group behavior-related data and travel preference-related data, and construct a user preference data set; through the location module of mobile devices and external data interfaces, collect environmental information including user location, timestamp, weather, and social media activity in the user's area, and establish an environmental information data set; The environmental information dataset is segmented into time windows, and the user preference dataset of the user group is synchronized and time-aligned to construct a time-series dataset. S2. The time-series dataset includes: representing the time-series characteristics of user group behavior as dimension d1; representing the environmental context characteristics of user group as dimension d2; representing the fused feature dimension after timestamp alignment as F=d1+d2; then extracting the timestamp information from the time-series dataset and combining it with behavior-related data to divide it into time windows; A time-series user profile is constructed based on a time-series dataset. This involves extracting user preference data sets and extracting user behavior patterns using time series modeling methods. Cluster analysis is performed on user behavior characteristics within different time windows to form a time-series user feature vector. Finally, a time-series user profile model is constructed by combining user geographic location, social media activity, and weather information. We use an attention-based feature fusion model to analyze the correlation between user behavior and contextual features, and generate user interest preference vectors in different scenarios. Finally, we integrate user behavior temporal features, context-aware modeling results and interest preference vectors into user profile feature vectors, where each time window corresponds to a comprehensive representation containing user behavior features, contextual features and interest preference vectors. Based on the environmental information dataset, we extract environmental context features that affect the accuracy of recommendations, and use a multilayer perceptual neural network combined with an attention mechanism to model the weight of different recommendation factors for users, while dynamically adjusting the recommendation strategy. S3. Establish an interest classification mechanism to divide user interests into short-term interests, long-term interests, and context-triggered interests; then adjust the priority of recommended content and optimize the recommendation order; finally, optimize the weight of user interests based on historical user feedback. S4. Prioritize matching short-term interest recommendation lists, then optimize the recommendation order by combining long-term interests and context-triggered interests; next, build and combine a real-time behavior analysis model to monitor user interaction, and trigger an intelligent interaction adjustment mechanism when the recommended content does not meet the user's needs.

2. The data processing method for an improved customized travel recommendation algorithm model according to claim 1, characterized in that: S2 also includes: An environmental information dataset is extracted, and an information gain-based feature selection algorithm is used to perform feature engineering on environmental information, including timestamps, weather conditions, and social media activity in the user's location, to extract context-related features. These features are then correlated with the user preference dataset to generate context-aware feature vectors. User profile feature vectors and context-aware feature vectors are input into an MLP model to extract high-dimensional feature representations. Then, an attention mechanism is used to calculate the user interest weight Ayh for recommendation factors, including attraction type, activity type, and time preference. The formula for calculating the user interest weight Ayh is as follows: ； In the formula, F is the fusion feature dimension, which is the fusion representation of user profile features and context-aware features, and depends on the dimensions of user profile features and context features; W f is the weight matrix of the MLP model, used to map the input features F to the hidden layers; its value is an h×(d1+d2) matrix, where h is the dimension of the hidden layers; bf is the bias term of the MLP model, used to adjust the output of the hidden layer; its value is an h×1 vector. W a The weight matrix for the attention mechanism is used to calculate the user's attention score for each recommendation factor. Its value is a k×h matrix, where k is the number of recommendation factors. The Softmax function is used to normalize attention scores into a probability distribution; its output ranges from (0, 1], and the sum of all output values ​​is 1. This is the hyperbolic tangent activation function, used to introduce a nonlinear transformation, with an output range of (-1, 1).

3. The data processing method for an improved customized tourism recommendation algorithm model according to claim 1, characterized in that: S2 also includes: Real-time collection and analysis of multimodal data, including user historical data, text reviews, click behavior, and social interactions, are performed. A user interest weight threshold A is assigned to each recommendation factor, and this threshold is compared with the user interest weight Ayh to evaluate the user's level of interest in various recommendation factors, including attraction type, activity type, and time preference. Specific evaluation content is as follows: When the user interest weight Ayh i ≥ User interest weight threshold A i This indicates that the user's interest in the current recommendation factor is "qualified". When the user interest weight Ayh i <User interest weight threshold A i This indicates that the user's interest in the current recommendation factor is "unqualified". Among them, the user interest weight Ayh i Let Ayh represent the user interest weight of the i-th recommendation factor, and let A be the user interest weight threshold. i Let A represent the user interest weight threshold of the i-th recommendation factor; i is the category index of the recommendation factor. Finally, based on the comparative evaluation results, the priority of the recommended content is dynamically adjusted: For recommendation factors that meet the "qualification" criteria for interest, increase the recommendation priority; For recommendation factors that are "unqualified" in terms of interest, reduce their recommendation priority or exclude them.

4. The data processing method for an improved customized tourism recommendation algorithm model according to claim 1, characterized in that: S3 specifically includes: Read users' historical behavior data and extract user interest features, including clicks, browsing, favorites, and reservations, from the historical behavior data and time-series user profiles; then classify user interests into three categories: short-term interests, long-term interests, and context-triggered interests. For short-term interests, identification and modeling are performed based on behavioral data, including the user's clicks and browsing history over the past week. For long-term interests, identification and modeling are performed based on behavioral data, including the user's continuous attention records over the past three months. For context-triggered interests, identification and modeling are performed based on contextual feature data, including weather, holidays, and geographical location. A decision tree-based rule engine is used, which includes classifying user interests and assigning initial weights to each category. The classified user interests and their corresponding initial weights are then output.

5. The data processing method for an improved customized tourism recommendation algorithm model according to claim 1, characterized in that: S3 also includes: An interest decay model is established for each user interest type, specifically including: For short-term interest, an exponential decay function is used. Modeling is performed, where I0 represents the initial interest intensity, λ represents the decay rate, and t represents the time interval; For long-term interests, a linear decay model is used for modeling. For context-triggered interest, an attention-based context-aware model is used to capture the correlation between user behavior and contextual features; Next, based on behavioral data and time-series user profiles, the interest decay coefficient Bsj for each type of interest is calculated; For short-term interests, the initial interest intensity I0 and decay rate λ are calculated based on short-term user behavior data. For long-term interests, the initial interest intensity I0 and decay rate λ are calculated based on long-term user behavior data. For context-triggered interest, the dynamic decay rate λ is calculated by combining contextual feature data; the specific calculation formula for the interest decay coefficient Bsj is as follows: ； In the formula, for short-term interest, the larger the decay rate λ, the faster the interest decays; For long-term interests, the smaller the decay rate λ, the slower the interest decays. For context-triggered interest, the decay rate λ is adjusted according to the timeliness of the context features; t represents the time interval, which is the length of time from the moment the interest arises to the current moment; e represents the base of the natural logarithm. The range of the interest decay coefficient Bsj is obtained through experimental data fitting and dynamic adjustment of context. The effective range of the interest decay coefficient Bsj is set to 0 < Bsj ≤ 1. When the interest decay coefficient Bsj is closer to 1, it indicates a greater degree of interest decay. When the interest decay coefficient Bsj is closer to 0, it indicates a smaller degree of interest decay.

6. The data processing method for an improved customized tourism recommendation algorithm model according to claim 1, characterized in that: S3 also includes: When interest has waned, we use deep reinforcement learning algorithms to build interest matching rules and adjust the priority and order of recommended content based on user feedback. This includes building interest matching rules, designing reward functions, and optimizing interest weights. Secondly, a reward function was designed to optimize the weight of user interest based on user satisfaction metrics. Finally, the optimized user interest weights are fed back to the recommendation system to generate the next round of recommendation lists, and user interaction behavior is monitored in real time. When the recommended content does not meet the user's needs, an intelligent interaction adjustment mechanism is triggered to further optimize the user interest weights.

7. The data processing method for an improved customized tourism recommendation algorithm model according to claim 1, characterized in that: S4 specifically includes: Based on the time-series user profile, short-term interest features are extracted and a short-term interest recommendation list is constructed. The short-term interest features include statistical information on the user's recent clicks, browsing, collection and booking records, as well as the trend of the user's interest preferences in different time intervals. Then, the user's long-term interest characteristics are modeled, including calculating the long-term preference distribution based on the user's historical behavior data, and optimizing the recommended content in combination with the user's stable preference characteristics; Secondly, extract users' contextual triggering interest features, including information such as geographical location, weather and social activity in the current environmental information data set, calculate users' contextual interest weights, and dynamically adjust the ranking strategy of recommended content based on contextual interest weights; Finally, during the recommendation process, following a hierarchical optimization strategy, the recommendation list is first matched with short-term interest recommendations, and the recommendation order is optimized based on the weight calculation results of long-term interests and context-triggered interests. Specifically, the distribution of users' long-term interest preferences is calculated, and the recommended content is optimized by combining stable preference features. Secondly, context-triggered interest features, including geographical location, weather, and social activity, are extracted, and context-triggered interest weights are calculated. Finally, the recommendation ranking strategy is dynamically adjusted based on the comprehensive weight of long-term interests and context-triggered interests.

8. The data processing method for an improved customized tourism recommendation algorithm model according to claim 1, characterized in that: S4 also includes: A user interest matching model is constructed, taking into input user historical behavior data, time-series user profiles, and contextual feature data. The model is trained using a reinforcement learning framework, including deep learning rules. Then, the recommended content matching degree (Cnr) is calculated using the user interest matching model, with the specific calculation formula as follows: ； Where n represents the total number of interest categories, j represents the index of the interest category, i.e. the different classification numbers of user interests; wj is the weight coefficient of the j-th interest category, Sj is the similarity between the recommended content and the j-th interest category, and f(t) is the timeliness correction factor.

9. The data processing method for an improved customized tourism recommendation algorithm model according to claim 1, characterized in that: S4 also includes: A first matching threshold C1 and a second matching threshold C2 are preset, and compared with the matching degree Cnr of the recommended content to evaluate the real-time matching degree between the recommended content and the user's interests, and to respond in real time to changes in user needs; and the first matching threshold C1 is greater than the second matching threshold C2. The specific evaluation content is as follows: When the matching degree of the recommended content Cnr is greater than or equal to the first matching degree evaluation threshold C1, it indicates that the recommended content is highly matched with the user's interests. The system makes recommendations directly without adjusting the recommendation strategy and monitors user interaction feedback data. When the second matching threshold C2 ≤ recommended content matching degree Cnr < first matching threshold C1: it indicates that the recommended content does not match the user's interests well, but is not lower than the minimum acceptable second matching threshold C2. In this case, the recommended content is optimized, and the specific adjustment methods include: Increase the weighting of short-term interests and adjust the order of recommended content; Incorporate environmental factors to increase the proportion of personalized recommendations; When the matching degree of the recommended content Cnr is less than the second matching degree evaluation threshold C2, it indicates that the matching degree of the recommended content with the user's interests is not up to standard and is lower than the minimum acceptable range of the second matching degree evaluation threshold C2. At this time, the intelligent interaction adjustment mechanism is triggered to immediately replace the recommended content and recalculate the matching degree of the recommended content Cnr. After the recommended content is pushed, the system continuously collects user feedback data, including clicks, dwell time, favorites, and reservations, and calculates feedback scores; it uses reinforcement learning models to adjust the recommendation strategy so that the recommendation system can adapt to changes in user interests.

Images