Intelligent research journey recommendation method and system based on collaborative filtering and LSTM fusion

Abstract

The invention provides an intelligent research journey recommendation method and system based on collaborative filtering and LSTM fusion, and the method comprises the steps: taking a multi-modal behavior time sequence as the input of a pre-constructed research behavior space-time attention network, and obtaining a time sequence preference implicit vector; calculating the matching degree of the time sequence preference implicit vector and the dynamic constraint factor, and correcting the time sequence preference implicit vector by using the constraint factor feature vector based on a calculation result to obtain a scene adaptive time sequence preference vector; extracting a time sequence behavior feature vector of the current research journey group, and performing weighted fusion on the time sequence behavior feature vector and the scene adaptive time sequence preference vector to obtain a final time sequence preference vector; the final time sequence preference vector is used for outputting a personalized research and study journey recommendation list. According to the multi-dimensional intelligent recommendation method, the group preference, the personal time sequence behavior, the research education attribute and the real-time resource state can be fused, so that the recommendation accuracy and the user experience are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method and system for intelligent recommendation of study tour itineraries based on the fusion of collaborative filtering and LSTM. Background Technology

[0002] Study tour recommendation is a special recommendation scenario that combines educational attributes with personalized needs. It involves matching users' grade level, subject preferences, educational goals with the educational qualifications, theme suitability, and safety conditions of study tour resources.

[0003] Among related technologies, the recommendation technology for study tours mainly suffers from the following problems: Traditional collaborative filtering methods rely on user-trip interaction data, and the recommendation effect drops significantly when the study tour platform faces the cold start problem. Furthermore, it does not fully integrate the educational characteristics of study tours, resulting in a disconnect between the recommendation results and educational goals. Although the LSTM-based temporal behavior modeling method can capture the dynamic behavior sequence of users, it ignores the reference value of group preferences and lacks the ability to model the multidimensional educational attributes of study tours, making it difficult to achieve recommendations in educational study tour scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent recommendation method and system for study tour itineraries based on the fusion of collaborative filtering and LSTM, so as to at least solve the technical problems of inaccurate recommendation results and inability to meet user needs in existing recommendation methods.

[0005] To achieve the above objectives, the present invention provides the following technical solution.

[0006] According to one embodiment of this application, a method for intelligent recommendation of study tour itineraries based on the fusion of collaborative filtering and LSTM is provided, including the following steps:

[0007] Collect and construct a multimodal behavioral time series that integrates explicit and implicit user behavioral data;

[0008] Multimodal behavioral time-series sequences are used as input to a pre-constructed spatiotemporal attention network for study tour behavior to obtain a temporal preference latent vector. The matching degree between the temporal preference latent vector and dynamic constraint factors is calculated. Based on the calculation results, the temporal preference latent vector is corrected using the feature vector of the constraint factors to obtain a scene-adaptive temporal preference vector. The temporal behavioral feature vector of the current study tour group is extracted and weighted and fused with the scene-adaptive temporal preference vector to obtain the final temporal preference vector. The fusion weights are dynamically adjusted according to the richness of the individual user's behavioral sequence.

[0009] The final time-series preference vector, along with the group preference vector mined based on collaborative filtering, the user's study tour feature vector, and the real-time resource status vector, are used to generate a comprehensive user preference vector through an attention mechanism. The recommendation score is then calculated based on the comprehensive vector and the feature vectors of candidate study tour itineraries. Combined with real-time resource availability filtering and itinerary combination optimization, a personalized study tour itinerary recommendation list is output.

[0010] Furthermore, the construction of multimodal behavioral time series includes the following steps:

[0011] Structured encoding is performed on explicit behavioral data to generate explicit behavioral feature sub-vectors;

[0012] The implicit behavior data is quantized and semantically encoded in multiple dimensions to obtain implicit behavior feature sub-vectors;

[0013] The explicit and implicit behavioral feature vectors at the same timestamp are fused to generate a single-time period behavioral feature vector. Then, all single-time period behavioral feature vectors are sorted according to the order of their timestamps to obtain a multimodal behavioral time series.

[0014] Furthermore, the pre-constructed spatiotemporal attention network for research-based learning behaviors includes:

[0015] The time attention branch is used to divide time windows according to the study tour cycle, and calculate the time attention weight based on the interval between each time window and the current time and its relevance to the user's current study tour education needs, and output the time attention feature vector.

[0016] The spatial attention branch is used to extract the research space features corresponding to the behaviors in the input sequence, calculate the spatial attention weights based on the spatial features, and output the spatial attention feature vector.

[0017] The LSTM layer is used to introduce the research and study education feature vector as a correction term in the LSTM gated computation, and outputs the hidden state vector representing the temporal evolution characteristics of the user's research and study behavior.

[0018] The fusion layer is used to perform element-wise weighted fusion of the temporal attention feature vector, spatial attention feature vector, and hidden state vector to obtain the fused feature vector.

[0019] Furthermore, the dynamic constraint factors include user status factors, education policy factors, and resource dynamic factors; the step of calculating the matching degree between the time-series preference latent vector and the dynamic constraint factors includes: vectorizing the user status factors, education policy factors, and resource dynamic factors respectively to generate corresponding feature vectors; calculating the subdivided matching degree between the time-series preference latent vector and the feature vector of each factor respectively; and weighting and fusing the subdivided matching degrees based on preset weights to obtain the overall matching degree.

[0020] Furthermore, the step of correcting the temporal preference latent vector based on the calculation results using the constraint factor eigenvector includes:

[0021] The correction level is determined based on the comparison between the overall matching degree and the adaptive threshold.

[0022] A multi-factor hierarchical correction model is constructed to differentiate the latent vector of time series preferences based on the correction level;

[0023] The residual vector is weighted and compensated to generate a corrected temporal preference latent vector. The corrected temporal preference latent vector is then optimized by study tour feature constraints. The vector after study tour feature constraint optimization and normalization is determined as a scene adaptive temporal preference vector that includes user temporal behavior features and dynamic constraint features of the current study tour scenario.

[0024] Furthermore, in the multi-factor hierarchical correction model:

[0025] When the comparison result indicates a severe mismatch, a full reconstruction correction is applied, including:

[0026] The basic preference vector is regenerated based on the eigenvector of the dynamic constraint factor and then fused with the original temporal preference latent vector to obtain the corrected temporal preference latent vector. , represented as:

[0027]

[0028] In the formula, Indicates the reconstructed weights; This represents the regenerated basic preference vector. This represents the original time-series preference latent vector;

[0029] When the comparison result indicates a moderate mismatch, incremental compensation correction is used, including:

[0030] The residual vector between the dynamic constraint factor and the original temporal preference latent vector is calculated, and the residual vector is weighted and compensated based on the adaptive correction coefficient to generate the corrected temporal preference latent vector. , represented as:

[0031]

[0032]

[0033] In the formula, k represents the category of dynamic factors, u corresponds to user status factors, p corresponds to education policy factors, and r corresponds to resource dynamic factors; This represents the original time-series preference latent vector; This represents the adaptive correction coefficient corresponding to the k-th type of dynamic constraint factor; Represents the residual vector of the k-th type of dynamic constraint factor; This represents the eigenvector of the k-th type of dynamic constraint factor. The projection function represents the projection of the latent vector of temporal preferences. Project onto the feature space of the k-th type of constraint factor;

[0034] No correction is made when the comparison result is a match.

[0035] Furthermore, the step of extracting the temporal behavioral feature vector of the current study tour group and weightedly fusing it with the scene-adaptive temporal preference vector to obtain the final temporal preference vector includes:

[0036] Determine the target study tour group to which the user belongs based on the user's study tour characteristics;

[0037] Extract the group temporal behavior feature vector of the target study group;

[0038] Construct a multi-dimensional richness assessment model based on research-based learning behavior types to obtain individual behavior richness factors. , represented as:

[0039]

[0040] In the formula, B represents the set of study tour behavior types. Indicates the weight of behavior type. This represents the sequence length of a user's behavior under a certain type. Indicates the effective threshold of the study tour behavior sequence. This represents the maximum length of the user behavior sequence within the target group G;

[0041] Calculate the research and study feature vector of the current user With the central eigenvector of the target group G The cosine similarity is used, and semantic matching degree correction based on research and study needs is introduced to obtain the individual and group feature similarity factor, expressed as:

[0042]

[0043] In the formula, Represents cosine similarity. A semantic embedding vector representing the user's current study and research needs. This represents a vector of typical study tour needs of the target group;

[0044] The dynamic fusion weight is obtained by weighting the richness factor and the similarity factor. The scene adaptive temporal preference vector and the group temporal behavior feature vector are then weighted and fused using the dynamic fusion weight to obtain the final temporal preference vector.

[0045] Furthermore, in the step of generating the user's comprehensive preference vector through the attention mechanism, a hierarchical masked attention network for the research and study scenario is adopted, specifically including:

[0046] Based on the characteristics and attributes of the study tour, the multi-source vector is divided into a target feature layer and a constraint feature layer. The target feature layer includes the final time-series preference vector and the user study tour education feature vector, while the constraint feature layer includes the group preference vector and the real-time resource status vector.

[0047] Based on the user's research and study target vector, a binary mask matrix is ​​generated. The target layer vector is multiplied element by element with the mask matrix to filter out features that are not related to the research and study target, and the intra-layer attention weight of the target layer is calculated.

[0048] A constraint priority mask is introduced, and the intra-layer attention weights of the constraint layer are calculated after masking the constraint layer vector.

[0049] By fusing the target layer vector and the constraint layer vector through two intra-layer attention weights, the target fusion vector and the constraint fusion vector are obtained respectively.

[0050] Dynamically determine inter-layer attention weights based on the constraint strength of user-driven learning scenarios. , represented as:

[0051]

[0052] In the formula, Indicates the degree of resource scarcity. Indicates group similarity. Represents the target fusion vector;

[0053] A comprehensive preference vector is generated through hierarchical fusion. , represented as:

[0054]

[0055] In the formula, Represents the target fusion vector. Represents the constraint fusion vector. This represents the interlayer attention weight.

[0056] Furthermore, it also includes steps to enhance the target layer vector and constraint layer vector, wherein:

[0057] The steps for enhancing the target layer vector include:

[0058] Calculate the similarity between the user's research and study target vector and the target vector, and then apply a weighted enhancement to the target vector, as shown below:

[0059]

[0060]

[0061] In the formula, This represents the final temporal preference vector after enhancement processing. Represents the original final time series preference vector. Represents the cosine similarity function. Represents the user's research and study target vector; This represents the enhanced user study tour feature vector. This represents the feature vector of the original user's study tour education;

[0062] The steps for enhancing the constraint layer vectors include:

[0063] The constraint vector is enhanced based on resource scarcity and group similarity, and is expressed as follows:

[0064]

[0065]

[0066] In the formula, This represents the collaborative filtering group preference vector enhanced by scenario constraints. This represents the original collaborative filtering group preference vector. This indicates the similarity in characteristics between a user and their study tour group. This represents the real-time resource state vector after scene constraint enhancement. This represents the original real-time resource state vector. This indicates the scarcity of resources for study tours.

[0067] According to one embodiment of the present invention, an intelligent recommendation system for study tour itineraries based on the fusion of collaborative filtering and LSTM is provided, comprising the following modules:

[0068] The sequence construction module is used to collect and construct multimodal behavioral time-series sequences that integrate explicit and implicit user behavioral data.

[0069] The sequence processing module takes the multimodal behavioral temporal sequence as input to a pre-constructed spatiotemporal attention network for study tour behavior to obtain a temporal preference latent vector; calculates the matching degree between the temporal preference latent vector and dynamic constraint factors; and corrects the temporal preference latent vector based on the constraint factor feature vector using the calculation results to obtain a scene-adaptive temporal preference vector; extracts the temporal behavioral feature vector of the current study tour group and performs a weighted fusion with the scene-adaptive temporal preference vector to obtain the final temporal preference vector; wherein, the fusion weight is dynamically adjusted according to the richness of the individual user's behavioral sequence.

[0070] The recommendation module is used to generate a comprehensive user preference vector by combining the final time-series preference vector with the group preference vector mined based on collaborative filtering, the user's study tour feature vector, and the real-time resource status vector through an attention mechanism. The module then calculates a recommendation score based on the comprehensive vector and the feature vectors of candidate study tour itineraries. Finally, it combines real-time resource availability filtering and itinerary combination optimization to output a personalized study tour itinerary recommendation list.

[0071] Compared with existing technologies, the beneficial effects of the intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM in this invention are as follows: This invention collects and constructs a multimodal behavioral time-series sequence that integrates explicit and implicit user behavioral data; uses the multimodal behavioral time-series sequence as input to a pre-constructed spatiotemporal attention network for study tour behavior to obtain a temporal preference latent vector; calculates the matching degree between the temporal preference latent vector and dynamic constraint factors, and corrects the temporal preference latent vector based on the constraint factor feature vector based on the calculation results to obtain a scene-adaptive temporal preference vector; extracts the temporal behavioral feature vector of the current study tour group, and performs weighted fusion with the scene-adaptive temporal preference vector to obtain the final temporal preference vector; wherein, the fusion weight is dynamically adjusted according to the richness of the individual user's behavioral sequence; the final temporal preference vector, together with the group preference vector mined based on collaborative filtering, the user's study tour education feature vector, and the real-time resource status vector, generates a user comprehensive preference vector through an attention mechanism, and calculates the recommendation score based on the comprehensive vector and the feature vector of the candidate study tour itineraries; combined with real-time resource availability screening and itinerary combination optimization, a personalized study tour itinerary recommendation list is output. The recommendation method of this invention can integrate a multi-dimensional intelligent recommendation approach that combines group preferences, individual temporal behavior, research and education attributes, and real-time resource status to improve recommendation accuracy and user experience. Attached Figure Description

[0072] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0073] In the attached diagram:

[0074] Figure 1This is a flowchart illustrating the implementation of the intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM in this invention.

[0075] Figure 2 This is a sub-process of the intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM in this invention;

[0076] Figure 3 This is a structural block diagram of the spatiotemporal attention network for research and learning behavior in this invention;

[0077] Figure 4 This is another sub-process of the intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM in this invention;

[0078] Figure 5 This is a structural block diagram of the intelligent recommendation system for study tour itineraries based on the fusion of collaborative filtering and LSTM, as proposed in this invention. Detailed Implementation

[0079] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0080] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0081] Figure 1 This is a flowchart illustrating the implementation of the intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM in this invention.

[0082] Please refer to Figure 1 According to an embodiment of the present invention, a method for intelligent recommendation of study tour itineraries based on the fusion of collaborative filtering and LSTM is provided, comprising the following steps:

[0083] Step S1: Collect and construct a multimodal behavioral time series that integrates explicit and implicit user behavioral data;

[0084] Specifically, in step S1, by collecting users' full-dimensional behavioral data in the study tour scenario, and through structured encoding, semantic quantization and temporal integration, a multimodal behavioral temporal sequence is constructed to represent users' study tour preferences, providing high-quality input data for the subsequent spatiotemporal attention network to extract temporal preference latent vectors;

[0085] Explicit behavioral data refers to user-initiated actions with clear intentions, while implicit behavioral data represents user-generated actions that indirectly reflect preferences.

[0086] Please refer to Figure 2 In step S1, the construction of the multimodal behavioral time series includes the following steps:

[0087] Step S11: Perform structured encoding on the explicit behavior data to generate explicit behavior feature sub-vectors;

[0088] In step S11, the present invention assigns a unique identifier code to each dominant behavior, represented by one-hot encoding; then, the key attributes of each dominant behavior are extracted and quantized into numerical features, and the behavior type encoding vector and the attribute quantization vector are concatenated in sequence to form a feature sub-vector of a single dominant behavior;

[0089] Step S12: Perform multi-dimensional quantization and semantic encoding on the implicit behavior data to obtain implicit behavior feature sub-vectors;

[0090] In step S12, for implicit behavioral data with no explicit intent but implicit preferences, a combination of quantitative statistics and semantic mapping is used to encode unstructured implicit behaviors into implicit behavioral feature vectors with the same dimension as explicit behavioral feature vectors.

[0091] Step S13: Merge the explicit behavior feature vectors and implicit behavior feature vectors under the same timestamp to generate a single time period behavior feature vector, and sort all the single time period behavior feature vectors according to the order of the timestamps to obtain a multimodal behavior time series sequence.

[0092] In step S13 of the present invention, explicit behavioral feature sub-vectors and implicit behavioral feature sub-vectors under the same time dimension are fused by timestamp alignment and time sequence sorting to form a complete multimodal behavioral time sequence. The length of the sequence is the number of active periods of the user on the platform, and each element in the sequence corresponds to the comprehensive user behavior feature of a time period.

[0093] Please continue to refer to Figure 1 The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM provided by this invention also includes:

[0094] Step S2: Use the multimodal behavioral temporal sequence as input to the pre-constructed study tour behavior spatiotemporal attention network to obtain the temporal preference latent vector; calculate the matching degree between the temporal preference latent vector and the dynamic constraint factor, and correct the temporal preference latent vector based on the calculation result using the constraint factor feature vector to obtain the scene-adaptive temporal preference vector; extract the temporal behavioral feature vector of the current study tour group, and perform weighted fusion with the scene-adaptive temporal preference vector to obtain the final temporal preference vector; wherein, the fusion weight is dynamically adjusted according to the richness of the individual user's behavior sequence;

[0095] Step S2 in this embodiment of the invention is used to obtain the final temporal preference vector that characterizes the user's personalized needs and scenario adaptability;

[0096] Among them, the pre-constructed spatiotemporal attention network for study tour behavior in this invention is used to capture the temporal evolution pattern, spatial preference characteristics, and temporal dynamic characteristics of users' study tour behavior from multimodal behavioral time-series sequences;

[0097] Specifically, such as Figure 3 As shown, in this embodiment of the invention, the pre-constructed spatiotemporal attention network for study tour behavior includes a time attention branch, which is used to divide time windows according to the study tour cycle, and calculate time attention weights and output time attention feature vectors based on the interval between each time window and the current time and its correlation with the user's current study tour education needs.

[0098] In this embodiment, common study tour cycles include 1-day tours, 2-day tours, and 3-day tours. The multimodal behavior time sequence is divided into time windows according to the study tour cycle as the basic unit. If the user's behavior spans multiple natural days, multiple consecutive time windows are generated. For example, the user's browsing, collection, and search behaviors within 3 days correspond to 3 time windows respectively.

[0099] In the calculation of attention weights, the attention weight of each time window is calculated based on two dimensions: time decay and demand relevance. Then, the behavioral feature vector of each time window is multiplied element by element with the corresponding time attention weight, and the weighted feature vectors of all windows are summed to obtain the time attention feature vector.

[0100] Furthermore, the pre-constructed spatiotemporal attention network for study tour behavior in this invention also includes a spatial attention branch, which is used to extract the study tour spatial features corresponding to the behavior in the input sequence, calculate the spatial attention weight based on the spatial features, and output the spatial attention feature vector.

[0101] Specifically, for spatial feature extraction, this invention extracts the corresponding research and study space features from each behavioral data in the multimodal behavioral time series, including the region, spatial type, and spatial level. For the calculation of spatial attention weights, spatial attention weights are calculated based on the user interaction frequency and preference matching degree of spatial features. Then, the spatial feature vector is weighted and fused with the corresponding spatial attention weights to obtain the spatial attention feature vector.

[0102] Furthermore, the pre-constructed spatiotemporal attention network for study tour behavior in this invention also includes an LSTM layer, which is used to introduce the study tour education feature vector as a correction term in the LSTM gated computation and output a hidden state vector that represents the temporal evolution characteristics of the user's study tour behavior.

[0103] In the LSTM layer of this invention, a user-led study tour feature vector is introduced as a correction term in the calculation of the input gate, forget gate, and output gate of the LSTM. This correction term is constructed based on the traveler's educational stage, study tour type preference, and training objective suitability. Then, through the gating mechanism of the LSTM layer and the time-series iterative calculation, a hidden state vector representing the time-series evolution characteristics of the user's study tour behavior is output.

[0104] The pre-constructed spatiotemporal attention network for research and study behavior in this invention also includes a fusion layer, which is used to perform element-wise weighted fusion of the temporal attention feature vector, the spatial attention feature vector and the hidden state vector to obtain a fused feature vector.

[0105] Furthermore, the dynamic constraint factors in this invention include user status factors, education policy factors, and resource dynamic factors;

[0106] The steps of this invention for calculating the matching degree between the temporal preference latent vector and the dynamic constraint factor include:

[0107] The user status factor, education policy factor, and resource dynamic factor are vectorized to generate corresponding feature vectors.

[0108] The subdivided matching degree between the temporal preference latent vector and the feature vector of each factor is calculated separately. Based on the preset weights, the subdivided matching degrees are weighted and fused to obtain the overall matching degree.

[0109] Specifically, for calculating the subdivided matching degree, this invention calculates the temporal preference latent vector separately. The cosine similarity with the feature vectors of the three factors is used as the subdivision matching degree. For the overall matching degree, the subdivision matching degree is weighted and fused based on preset weights to obtain the overall matching degree. For example, the preset weights include: user status is 0.4, education policy is 0.3, and resource dynamics is 0.3. The overall matching degree of this invention ranges from 0 to 1. The closer the value is to 1, the better the temporal preference latent vector matches the dynamic constraint factor.

[0110] Furthermore, in one implementation of the present invention, such as Figure 4 As shown, in step S2:

[0111] The step of correcting the latent vector of temporal preference using the constraint factor feature vector based on the calculation results includes:

[0112] Step S21: Based on the comparison results between the overall matching degree and the adaptive threshold, classify the correction level;

[0113] The correction levels of this invention include matching, moderate mismatch, and severe mismatch;

[0114] Step S22: Construct a multi-factor hierarchical correction model and perform differentiated correction on the latent vector of time series preferences based on the correction level;

[0115] Specifically, first preset two adaptive thresholds. and The threshold of the former is greater than the threshold of the latter.

[0116] For the overall matching degree M:

[0117] when At this point, the matching level indicates that the temporal preference latent vector and the dynamic constraint factor are highly compatible and require no correction.

[0118] when When the mismatch level is moderate, it indicates that some factors are biased, and incremental compensation correction is required.

[0119] when When the mismatch level is reached, it indicates a severe mismatch, meaning that the factors have serious biases and require full reconstruction and correction.

[0120] Specifically, in one implementation of the present invention, in the multi-factor hierarchical correction model:

[0121] When the comparison result indicates a severe mismatch, a full reconstruction correction is applied, including:

[0122] The basic preference vector is regenerated based on the eigenvector of the dynamic constraint factor and then fused with the original temporal preference latent vector to obtain the corrected temporal preference latent vector. , represented as:

[0123]

[0124] In the formula, Indicates the reconstructed weights; This represents the regenerated basic preference vector. This represents the original time-series preference latent vector;

[0125] When the comparison result indicates a moderate mismatch, incremental compensation correction is used, including:

[0126] The residual vector between the dynamic constraint factor and the original temporal preference latent vector is calculated, and the residual vector is weighted and compensated based on the adaptive correction coefficient to generate the corrected temporal preference latent vector. , represented as:

[0127]

[0128]

[0129] In the formula, k represents the category of dynamic factors, u corresponds to user status factors, p corresponds to education policy factors, and r corresponds to resource dynamic factors; This represents the original time-series preference latent vector; This represents the adaptive correction coefficient corresponding to the k-th type of dynamic constraint factor; Represents the residual vector of the k-th type of dynamic constraint factor; This represents the eigenvector of the k-th type of dynamic constraint factor. The projection function represents the projection of the latent vector of temporal preferences. Project onto the feature space of the k-th type of constraint factor;

[0130] No correction is made when the comparison result is a match.

[0131] Step S23: Perform weighted compensation on the residual vector to generate a corrected temporal preference latent vector. Perform research and study feature constraint optimization on the corrected temporal preference latent vector. Determine the vector after research and study feature constraint optimization and normalization as a scene adaptive temporal preference vector containing user temporal behavior features and dynamic constraint features of the current research and study scenario.

[0132] In the study tour feature constraint optimization of the present invention, the modified temporal preference latent vector and the study tour feature vector (such as safety level, course suitability, and instructor qualification) are constrained and projected to ensure that the modified vector meets the requirements of study tour education; the vector after constraint projection is the scene adaptive temporal preference vector, which simultaneously contains the user's temporal behavior features and the dynamic constraint features of the current study tour scene.

[0133] Furthermore, the step of extracting the temporal behavioral feature vector of the current study tour group and weightedly fusing it with the scene-adaptive temporal preference vector to obtain the final temporal preference vector includes:

[0134] The target study tour group to which a user belongs is determined based on the user's study tour characteristics; among them, the K-means clustering algorithm is used to divide the target study tour group, and the clustering characteristics include: school level, study tour type, theme preference, and trip length preference;

[0135] Extract the group temporal behavior feature vector of the target study tour group; wherein, from the database, extract the multimodal behavior temporal sequence of all users in the target study tour group, extract the group temporal preference latent vector of each user through the study tour behavior spatiotemporal attention network, and then take the mean of all vectors to obtain the group temporal behavior feature vector of the target group;

[0136] Construct a multi-dimensional richness assessment model based on research-based learning behavior types to obtain individual behavior richness factors. In which the present invention The value range is [0,1], and the closer it is to 1, the richer the individual behavior and the higher the credibility of the individual preference; in this embodiment, the individual behavior richness factor is expressed as:

[0137]

[0138] In the formula, B represents the set of study tour behavior types. Indicates the weight of behavior type. This represents the sequence length of a user's behavior under a certain type. Indicates the effective threshold of the study tour behavior sequence. This represents the maximum length of the user behavior sequence within the target group G;

[0139] Calculate the research and study feature vector of the current user With the central eigenvector of the target group G The cosine similarity is used, and semantic matching degree correction based on research and study needs is introduced to obtain the individual and group feature similarity factor, expressed as:

[0140]

[0141] In the formula, Represents cosine similarity. A semantic embedding vector representing the user's current study and research needs. This represents a vector of typical study tour needs of the target group;

[0142] Furthermore, this invention obtains a dynamic fusion weight by weighted product of richness factor and similarity factor. Using this dynamic fusion weight, the scene adaptive temporal preference vector and the group temporal behavior feature vector are weighted and fused to obtain the final temporal preference vector; wherein, for the dynamic fusion weight... , and weight Normalize to the 0-1 interval, For the weights of the scene adaptive temporal preference vector, 1− The weights of the group's temporal behavioral feature vector;

[0143] This invention uses dynamic fusion weights to perform weighted fusion of two vectors, as shown below:

[0144]

[0145] in, Represents the final time series preference vector. This represents a scene-adaptive temporal preference vector. This represents the feature vector of group temporal behavior.

[0146] Please continue to refer to Figure 1 The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM of the present invention further includes the following steps:

[0147] Step S3: Combine the final time-series preference vector with the group preference vector mined based on collaborative filtering, the user's study tour feature vector, and the real-time resource status vector to generate a comprehensive user preference vector through an attention mechanism. Calculate the recommendation score based on the comprehensive vector and the feature vectors of candidate study tour itineraries. Combine real-time resource availability filtering and itinerary combination optimization to output a personalized study tour itinerary recommendation list.

[0148] Specifically, in step S3 of this invention, the step of generating a comprehensive user preference vector through an attention mechanism is implemented using a hierarchical masked attention network for study tour scenarios, specifically including:

[0149] Based on the characteristics and attributes of the study tour, the multi-source vector is divided into a target feature layer and a constraint feature layer. The target feature layer includes the final time-series preference vector and the user study tour education feature vector, while the constraint feature layer includes the group preference vector and the real-time resource status vector.

[0150] Based on the user's research and study target vector, a binary mask matrix is ​​generated. The target layer vector is multiplied element by element with the mask matrix to filter out features that are not related to the research and study target, and the intra-layer attention weight of the target layer is calculated.

[0151] A constraint priority mask is introduced, and the intra-layer attention weights of the constraint layer are calculated after masking the constraint layer vector. The target layer vector and the constraint layer vector are fused by the two intra-layer attention weights respectively to obtain the target fusion vector and the constraint fusion vector.

[0152] In this invention, the inter-layer attention weights are dynamically determined based on the constraint strength of the user's research and learning scenario. , represented as:

[0153]

[0154] In the formula, Indicates the degree of resource scarcity. Indicates group similarity. Represents the target fusion vector;

[0155] Furthermore, this invention generates a comprehensive preference vector through hierarchical fusion. , represented as:

[0156]

[0157] In the formula, Represents the target fusion vector. Represents the constraint fusion vector. This represents the interlayer attention weight.

[0158] Furthermore, it also includes steps to enhance the target layer vector and constraint layer vector.

[0159] In one implementation of the present invention, the step of enhancing the target layer vector includes: calculating the similarity between the user's research target vector and the target vector, and performing weighted enhancement on the target vector, as expressed as:

[0160]

[0161]

[0162] In the formula, This represents the final temporal preference vector after enhancement processing. Represents the original final time series preference vector. Represents the cosine similarity function. Represents the user's research and study target vector; This represents the enhanced user study tour feature vector. This represents the feature vector of the original user's study tour education;

[0163] In one implementation of the present invention, the step of enhancing the constraint layer vector includes: enhancing the constraint vector based on resource scarcity and group similarity, as expressed as:

[0164]

[0165]

[0166] In the formula, This represents the collaborative filtering group preference vector enhanced by scenario constraints. This represents the original collaborative filtering group preference vector. This indicates the similarity in characteristics between a user and their study tour group. This represents the real-time resource state vector after scene constraint enhancement. This represents the original real-time resource state vector. This indicates the scarcity of resources for study tours.

[0167] This invention integrates a multi-dimensional intelligent recommendation method that combines group preferences, individual temporal behavior, research and education attributes, and real-time resource status to improve recommendation accuracy and user experience.

[0168] like Figure 5 As shown, according to an embodiment of the present invention, an intelligent recommendation system for study tour itineraries based on the fusion of collaborative filtering and LSTM is provided, comprising the following modules:

[0169] Sequence construction module 41 is used to collect and construct multimodal behavioral time series sequences that integrate explicit and implicit user behavioral data;

[0170] The sequence processing module 42 is used to take the multimodal behavioral temporal sequence as input to the pre-constructed study tour behavior spatiotemporal attention network to obtain the temporal preference latent vector; calculate the matching degree between the temporal preference latent vector and the dynamic constraint factor; based on the calculation result, use the constraint factor feature vector to correct the temporal preference latent vector to obtain the scene-adaptive temporal preference vector; extract the temporal behavioral feature vector of the current study tour group, and perform weighted fusion with the scene-adaptive temporal preference vector to obtain the final temporal preference vector; wherein, the fusion weight is dynamically adjusted according to the richness of the individual user's behavior sequence;

[0171] The recommendation module 43 is used to generate a comprehensive user preference vector by combining the final time-series preference vector with the group preference vector mined based on collaborative filtering, the user's study tour feature vector, and the real-time resource status vector through an attention mechanism. The recommendation score is calculated based on the comprehensive vector and the feature vector of the candidate study tour itinerary. Combined with real-time resource availability filtering and itinerary combination optimization, a personalized study tour itinerary recommendation list is output.

[0172] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0173] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0174] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0175] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0176] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0177] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0178] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0179] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for intelligent recommendation of study tour itineraries based on the fusion of collaborative filtering and LSTM, characterized in that, Includes the following steps: Collect and construct a multimodal behavioral time series that integrates explicit and implicit user behavioral data; Using the multimodal behavioral temporal sequence as input to a pre-constructed spatiotemporal attention network for research and study behavior, we obtain the temporal preference latent vector; The matching degree between the temporal preference latent vector and the dynamic constraint factor is calculated. Based on the calculation result, the temporal preference latent vector is corrected using the constraint factor feature vector to obtain the scene-adaptive temporal preference vector. The temporal behavior feature vector of the current study tour group is extracted and weighted and fused with the scene-adaptive temporal preference vector to obtain the final temporal preference vector. The fusion weight is dynamically adjusted according to the richness of the individual user's behavior sequence. The final time-series preference vector, together with the group preference vector mined based on collaborative filtering, the user's research and education feature vector, and the real-time resource status vector, are used to generate a comprehensive user preference vector through an attention mechanism; The recommendation score is calculated based on the comprehensive vector and the feature vector of the candidate study tour itinerary. Combined with real-time resource availability filtering and itinerary combination optimization, a personalized study tour itinerary recommendation list is output.

2. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in claim 1, characterized in that, The construction of multimodal behavioral time series includes the following steps: Structured encoding is performed on explicit behavioral data to generate explicit behavioral feature sub-vectors; Implicit behavioral data is quantized and semantically encoded in multiple dimensions to obtain implicit behavioral feature sub-vectors; The explicit and implicit behavioral feature vectors at the same timestamp are fused to generate a single-time period behavioral feature vector. Then, all single-time period behavioral feature vectors are sorted according to the order of their timestamps to obtain a multimodal behavioral time series.

3. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in claim 2, characterized in that, The pre-constructed spatiotemporal attention network for study tour behavior includes: The time attention branch is used to divide the time window according to the study tour cycle, calculate the time attention weight with the interval of the current time and its correlation with the user's current study tour education needs, and output the time attention feature vector. The spatial attention branch is used to extract the research space features corresponding to the behaviors in the input sequence, calculate the spatial attention weights based on the spatial features, and output the spatial attention feature vector. The LSTM layer is used to introduce the research and study education feature vector as a correction term in the LSTM gated computation, and outputs the hidden state vector representing the temporal evolution characteristics of the user's research and study behavior. The fusion layer is used to perform element-wise weighted fusion of the temporal attention feature vector, spatial attention feature vector, and hidden state vector to obtain the fused feature vector.

4. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in claim 3, characterized in that, Dynamic constraint factors include user status factors, education policy factors, and resource dynamic factors; The steps for calculating the matching degree between the temporal preference latent vector and the dynamic constraint factor include: User status factors, education policy factors, and resource dynamic factors are vectorized to generate corresponding feature vectors. The subdivision matching degree between the time-series preference latent vector and the feature vector of each factor is calculated separately. Based on the preset weights, the subdivision matching degrees are weighted and fused to obtain the overall matching degree.

5. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in claim 4, characterized in that, The steps for correcting the latent vector of temporal preferences based on the calculation results using the constraint factor eigenvector include: The correction level is determined based on the comparison between the overall matching degree and the adaptive threshold. A multi-factor hierarchical correction model is constructed to differentiate the latent vector of time series preferences based on the correction level; The residual vector is weighted and compensated to generate a corrected temporal preference latent vector. The corrected temporal preference latent vector is then optimized by study tour feature constraints. The vector optimized by study tour feature constraints is determined as a scene-adaptive temporal preference vector that includes user temporal behavior features and dynamic constraint features of the current study tour scenario.

6. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in claim 5, characterized in that, In the multi-factor hierarchical correction model: When the comparison result indicates a severe mismatch, a full reconstruction correction is applied, including: The basic preference vector is regenerated based on the eigenvector of the dynamic constraint factor and then fused with the original temporal preference latent vector to obtain the corrected temporal preference latent vector. , represented as: In the formula, Indicates the reconstructed weights; This represents the regenerated basic preference vector. This represents the original time-series preference latent vector; When the comparison result indicates a moderate mismatch, incremental compensation correction is used, including: The residual vector between the dynamic constraint factor and the original temporal preference latent vector is calculated, and the residual vector is weighted and compensated based on the adaptive correction coefficient to generate the corrected temporal preference latent vector. , represented as: In the formula, k represents the category of dynamic factors, u corresponds to user status factors, p corresponds to education policy factors, and r corresponds to resource dynamic factors; This represents the original time-series preference latent vector; This represents the adaptive correction coefficient corresponding to the k-th type of dynamic constraint factor; Represents the residual vector of the k-th type of dynamic constraint factor; This represents the eigenvector of the k-th type of dynamic constraint factor. The projection function represents the projection of the latent vector of temporal preferences. Project onto the feature space of the k-th type of constraint factor; No correction is made when the comparison result is a match.

7. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM according to any one of claims 2 to 6, characterized in that, The steps include: extracting the temporal behavioral feature vector of the current study tour group and weighting and fusing it with the scene-adaptive temporal preference vector to obtain the final temporal preference vector; include: Determine the target study tour group to which the user belongs based on the user's study tour characteristics; Extract the temporal behavioral feature vector of the target study group; Construct a multi-dimensional richness assessment model based on research-based learning behavior types to obtain individual behavior richness factors. , represented as: In the formula, B represents the set of study tour behavior types. Indicates the weight of behavior type. This represents the sequence length of a user's behavior under a certain type. Indicates the effective threshold of the study tour behavior sequence. This represents the maximum length of the user behavior sequence within the target group G; Calculate the research and study feature vector of the current user With the central eigenvector of the target group G The cosine similarity is used, and semantic matching degree correction based on research and study needs is introduced to obtain the individual and group feature similarity factor, expressed as: In the formula, Represents cosine similarity. A semantic embedding vector representing the user's current study and research needs. This represents a vector of typical study tour needs of the target group; This represents the research and study feature vector of the current user. Represents the central eigenvector of the target group G; The dynamic fusion weight is obtained by weighting the richness factor and the similarity factor. The scene adaptive temporal preference vector and the group temporal behavior feature vector are then weighted and fused using the dynamic fusion weight to obtain the final temporal preference vector.

8. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in claim 7, characterized in that, In the step of generating a comprehensive user preference vector through an attention mechanism, a hierarchical masked attention network for research and study scenarios is used, specifically including: Based on the characteristics and attributes of the study tour, the multi-source vector is divided into a target feature layer and a constraint feature layer. The target feature layer includes the final time-series preference vector and the user study tour education feature vector, while the constraint feature layer includes the group preference vector and the real-time resource status vector. Based on the user's research and study target vector, a binary mask matrix is ​​generated. The target layer vector is multiplied element by element with the mask matrix to filter out features that are not related to the research and study target, and the intra-layer attention weight of the target layer is calculated. A constraint priority mask is introduced, and the intra-layer attention weights of the constraint layer are calculated after masking the constraint layer vector. By fusing the target layer vector and the constraint layer vector through two intra-layer attention weights, the target fusion vector and the constraint fusion vector are obtained respectively. Dynamically determine inter-layer attention weights based on the constraint strength of user-driven learning scenarios. , represented as: In the formula, Indicates the degree of resource scarcity. Indicates group similarity. Represents the target fusion vector; Represents the user's research and study target vector; Indicates cosine similarity; A comprehensive preference vector is generated through hierarchical fusion. , represented as: In the formula, Represents the target fusion vector. Represents the constraint fusion vector. This represents the interlayer attention weight.

9. The intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in claim 8, characterized in that, It also includes steps for enhancing the target layer vector and constraint layer vector, wherein: The steps for enhancing the target layer vector include: Calculate the similarity between the user's research and study target vector and the target vector, and then apply a weighted enhancement to the target vector, as shown below: In the formula, This represents the final temporal preference vector after enhancement processing. Represents the original final time series preference vector. Represents cosine similarity. Represents the user's research and study target vector; This represents the enhanced user study tour feature vector. This represents the feature vector of the original user's study tour education; The steps for enhancing the constraint layer vectors include: The constraint vector is enhanced based on resource scarcity and group similarity, and is expressed as follows: In the formula, This represents the collaborative filtering group preference vector enhanced by scenario constraints. This represents the original collaborative filtering group preference vector. This indicates the similarity in characteristics between a user and their study tour group. This represents the real-time resource state vector after scene constraint enhancement. This represents the original real-time resource state vector. This indicates the scarcity of resources for study tours.

10. A recommendation system for implementing the intelligent recommendation method for study tour itineraries based on the fusion of collaborative filtering and LSTM as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The sequence construction module is used to collect and construct multimodal behavioral time-series sequences that integrate explicit and implicit user behavioral data. The sequence processing module is used to take the multimodal behavior time sequence as input to the pre-constructed study behavior spatiotemporal attention network to obtain the temporal preference latent vector; The matching degree between the temporal preference latent vector and the dynamic constraint factor is calculated. Based on the calculation result, the temporal preference latent vector is corrected using the constraint factor feature vector to obtain the scene-adaptive temporal preference vector. The temporal behavior feature vector of the current study tour group is extracted and weighted and fused with the scene-adaptive temporal preference vector to obtain the final temporal preference vector. The fusion weight is dynamically adjusted according to the richness of the individual user's behavior sequence. The recommendation module is used to generate a comprehensive user preference vector by combining the final time-series preference vector with the group preference vector mined based on collaborative filtering, the user's study tour feature vector, and the real-time resource status vector through an attention mechanism. The module then calculates a recommendation score based on the comprehensive vector and the feature vectors of candidate study tour itineraries. Finally, it combines real-time resource availability filtering and itinerary combination optimization to output a personalized study tour itinerary recommendation list.

Images