CN-122019877-A - Intelligent study journey recommendation method and system based on collaborative filtering and LSTM fusion

CN122019877ACN 122019877 ACN122019877 ACN 122019877ACN-122019877-A

Abstract

The invention provides a learning journey intelligent recommendation method and system based on collaborative filtering and LSTM fusion, wherein the method comprises the steps of taking a multi-mode behavior time sequence as the input of a pre-constructed learning behavior time-space attention network to obtain a time sequence preference hidden vector; the matching degree of the time sequence preference hidden vector and the dynamic constraint factor is calculated, the time sequence preference hidden vector is corrected by utilizing the constraint factor feature vector based on the calculation result to obtain a scene self-adaptive time sequence preference vector, the time sequence behavior feature vector of the current study journey group is extracted, the time sequence behavior feature vector and the scene self-adaptive time sequence preference vector are subjected to weighted fusion to obtain a final time sequence preference vector, and the final time sequence preference vector is used for outputting a personalized study journey recommendation list. The multi-dimensional intelligent recommendation method can integrate group preference, personal time sequence behavior, research education attribute and real-time resource state, so as to improve recommendation accuracy and user experience.

Inventors

TIAN LONG
Du Xianghe
XIE JUANJUAN
MAO CHANGSHENG
NI XIANGDONG
WU YAPING
CUI LU

Assignees

时代数媒科技股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (10)

1. The intelligent research journey recommending method based on collaborative filtering and LSTM fusion is characterized by comprising the following steps of: Acquiring and constructing a multi-mode behavior time sequence integrating dominant behavior data and recessive behavior data of a user; The method comprises the steps of taking a multi-modal behavior time sequence as input of a pre-built learning behavior space-time attention network to obtain a time sequence preference hidden vector, calculating the matching degree of the time sequence preference hidden vector and a dynamic constraint factor, correcting the time sequence preference hidden vector by utilizing the constraint factor feature vector based on a calculation result to obtain a scene self-adaptive time sequence preference vector, extracting the time sequence behavior feature vector of a current learning journey group, and carrying out weighted fusion with the scene self-adaptive time sequence preference vector to obtain a final time sequence preference vector, wherein the fusion weight is dynamically adjusted according to the richness of a user individual behavior sequence; generating a user comprehensive preference vector by an attention mechanism by combining the final time sequence preference vector with a group preference vector, a user research education feature vector and a real-time resource state vector which are mined based on collaborative filtering; And calculating a recommendation score based on the comprehensive vector and the feature vector of the candidate study journey, and outputting a personalized study journey recommendation list by combining real-time resource availability screening and journey combination optimization.
2. The collaborative filtering and LSTM fusion-based intelligent research trip recommendation method according to claim 1, wherein the construction of the multi-modal behavioral time series sequence comprises the following steps: carrying out structural coding on the dominant behavior data to generate dominant behavior feature sub-vectors; Carrying out multidimensional quantization and semantic coding on the implicit behavior data to obtain implicit behavior feature sub-vectors; The dominant behavior feature sub-vector and the recessive behavior feature sub-vector under the same time stamp are fused to generate a single-period behavior feature vector, and all the single-period behavior feature vectors are ordered according to the sequence of the time stamp to obtain the multi-mode behavior time sequence.
3. The collaborative filtering and LSTM fusion-based intelligent research itinerary recommendation method of claim 2, wherein the pre-constructed spatiotemporal attention network for research activities comprises: the time attention branch is used for dividing a time window according to the study period, calculating time attention weight according to the interval of the current time and the correlation between the current time and the current study education requirement of a user, and outputting a time attention feature vector; the spatial attention branch is used for extracting the scientific spatial features corresponding to the behaviors in the input sequence, calculating the spatial attention weight based on the spatial features and outputting a spatial attention feature vector; The LSTM layer is used for introducing a research education feature vector into LSTM gating calculation to serve as a correction term and outputting a hidden state vector representing time sequence evolution features of the research behaviors of the user; And the fusion layer is used for carrying out element-by-element weighted fusion on the time attention feature vector, the space attention feature vector and the hidden state vector to obtain a fusion feature vector.
4. The method for intelligent recommendation of a research trip based on collaborative filtering and LSTM fusion of claim 3, wherein the dynamic constraint factors include user state factors, education policy factors and resource dynamic factors, and the step of calculating the matching degree of the timing preference hidden vector and the dynamic constraint factors comprises the following steps: the user state factors, the education policy factors and the resource dynamic factors are respectively vectorized to generate corresponding feature vectors, the subdivision matching degree between the time sequence preference hidden vectors and the factor feature vectors is respectively calculated, and all subdivision matching degrees are weighted and fused based on preset weights to obtain the overall matching degree.
5. The intelligent recommendation method for learning journey based on collaborative filtering and LSTM fusion according to claim 4, wherein the step of correcting the time preference hidden vector by using the constraint factor feature vector based on the calculation result comprises: Dividing correction grades according to the comparison result of the overall matching degree and the self-adaptive threshold value; Constructing a multi-factor hierarchical correction model, and performing differential correction on the time preference hidden vector based on the correction level; And carrying out weighted compensation on the residual vector, generating a corrected time sequence preference hidden vector, carrying out research feature constraint optimization on the corrected time sequence preference hidden vector, and determining the vector subjected to the research feature constraint optimization as a scene self-adaptive time sequence preference vector containing the time sequence behavior feature of the user and the dynamic constraint feature of the current research scene.
6. The collaborative filtering and LSTM fusion-based intelligent research trip recommendation method according to claim 5, wherein, in a multi-factor hierarchical correction model: When the comparison result is that the weight is not matched, adopting full reconstruction type correction, wherein the method comprises the following steps: Based on the dynamic constraint factor feature vector, regenerating a basic preference vector, and fusing the basic preference vector with an original time sequence preference hidden vector to obtain a corrected time sequence preference hidden vector Expressed as: In the formula, Representing the reconstruction weights; Representing the regenerated base preference vector, Representing an original timing preference hidden vector; when the comparison result is that the intermediate degree is not matched, incremental compensation type correction is adopted, and the method comprises the following steps: Calculating a residual vector of the dynamic constraint factor and the original time sequence preference hidden vector, and performing weighted compensation on the residual vector based on the self-adaptive correction coefficient to generate a corrected time sequence preference hidden vector Expressed as: Wherein k represents the category of dynamic factors, u corresponds to user state factors, p corresponds to education policy factors, and r corresponds to resource dynamic factors; representing an original timing preference hidden vector; representing the self-adaptive correction coefficient corresponding to the k-th type dynamic constraint factor; A residual vector representing a kth class of dynamic constraint factors; feature vectors representing the k-th class of dynamic constraint factors, Representing projection functions representing hidden vectors of timing preference Projecting to a feature space of a kth class constraint factor; When the comparison result is a match, no correction is performed.
7. The intelligent research trip recommendation method based on collaborative filtering and LSTM fusion according to any one of claims 2 to 6, wherein the step of extracting the time sequence behavior feature vector of the current research trip group and performing weighted fusion with the scene adaptive time sequence preference vector to obtain a final time sequence preference vector; Comprising the following steps: determining a target study group to which the user belongs based on the study characteristics of the user; extracting group time sequence behavior feature vectors of a target study group; Constructing a multidimensional richness assessment model based on the research behavior type to obtain individual behavior richness factors Expressed as: Wherein B represents a set of behavioral types of research, The behavior type weight is represented as a weight, Representing the length of the sequence of the user under a certain class of behavior, Representing the effective threshold of the sequence of the scientific behaviour, Representing the maximum length of the sequence of user behaviors within the target group G; computing a current user's research feature vector Center feature vector with target group G And introducing semantic matching degree correction of research requirements to obtain individual and group feature similarity factors, wherein the factors are expressed as: In the formula, Representing the degree of cosine similarity, A semantic embedded vector representing the current research needs of the user, Representative research demand vectors representing target populations; Representing the current user's research feature vector, A central feature vector representing the target group G; obtaining dynamic fusion weights through the weighted products of the richness factors and the similarity factors, and carrying out weighted fusion on the scene self-adaptive time sequence preference vectors and the group time sequence behavior feature vectors by utilizing the dynamic fusion weights to obtain final time sequence preference vectors.
8. The collaborative filtering and LSTM fusion-based research itinerary intelligent recommendation method according to claim 7, wherein in the step of generating a user comprehensive preference vector through an attention mechanism, a research scene hierarchical mask attention network implementation is adopted, specifically comprising: Dividing the multi-source vector into a target feature layer and a constraint feature layer according to the research feature attribute, wherein the target feature layer comprises a final time sequence preference vector and a user research education feature vector, and the constraint feature layer comprises a group preference vector and a real-time resource state vector; Generating a binary mask matrix based on a user research target vector, multiplying the target layer vector by the mask matrix element by element, filtering non-research target related features, and calculating the in-layer attention weight of the target layer; introducing constraint priority masks, and calculating in-layer attention weights of constraint layers after masking constraint layer vectors; respectively fusing the target layer vector and the constraint layer vector through the attention weights in the two layers to respectively obtain a target fusion vector and a constraint fusion vector; dynamic determination of inter-layer attention weights based on constraint intensity of user research scene Expressed as: In the formula, The resource tension is indicated as such, The degree of similarity of the population is represented, Representing a target fusion vector; Representing a user study target vector; representing cosine similarity; Generating comprehensive preference vectors through hierarchical fusion Expressed as: In the formula, The target fusion vector is represented as such, The constraint fusion vector is represented as a function of the constraint, Indicating the inter-layer attention weight.
9. The collaborative filtering and LSTM fusion-based intelligent research trip recommendation method of claim 8, further comprising the step of enhancing target layer vectors and constraint layer vectors, wherein: the step of enhancing the target layer vector includes: Calculating the similarity between the user research target vector and the target vector, and weighting and enhancing the target vector, wherein the weighted and enhanced target vector is expressed as follows: In the formula, Representing the final timing preference vector after the enhancement process, Representing the original final timing preference vector, Representing the degree of cosine similarity, Representing a user study target vector; Representing the enhanced user's educational character vector, Representing an original user's educational education feature vector; the step of enhancing the constraint layer vector comprises the following steps: And enhancing the constraint vector according to the resource tension and the group similarity, wherein the constraint vector is expressed as follows: In the formula, Represents collaborative filtering group preference vectors enhanced by scene constraints, Representing the original collaborative filtering population preference vector, Representing the feature similarity of the user to the belonging study population, Representing the real-time resource status vector enhanced by the scene constraint, Representing the original real-time resource state vector, Representing the strength of the research resource.
10. Recommendation system for implementing a collaborative filtering and LSTM fusion based intelligent recommendation method for a research trip according to any one of claims 1 to 9, characterized in that it comprises the following modules: The sequence construction module is used for acquiring and constructing a multi-mode behavior time sequence integrating the dominant behavior data and the recessive behavior data of the user; The sequence processing module is used for taking the multi-mode behavior time sequence as the input of the pre-built study behavior space-time attention network to obtain a time sequence preference hidden vector, calculating the matching degree of the time sequence preference hidden vector and a dynamic constraint factor, correcting the time sequence preference hidden vector by utilizing the constraint factor feature vector based on a calculation result to obtain a scene self-adaptive time sequence preference vector, extracting the time sequence behavior feature vector of the current study travel group, and carrying out weighted fusion with the scene self-adaptive time sequence preference vector to obtain a final time sequence preference vector, wherein the fusion weight is dynamically adjusted according to the richness of the user individual behavior sequence; The recommendation module is used for generating a user comprehensive preference vector through an attention mechanism by combining the final time sequence preference vector with a group preference vector, a user study education feature vector and a real-time resource state vector which are mined based on collaborative filtering, calculating a recommendation score based on the comprehensive vector and the feature vector of the candidate study journey, and outputting a personalized study journey recommendation list by combining real-time resource availability screening and journey combination optimization.

Description

Intelligent study journey recommendation method and system based on collaborative filtering and LSTM fusion Technical Field The invention belongs to the technical field of artificial intelligence, and particularly relates to an intelligent research journey recommendation method and system based on collaborative filtering and LSTM fusion. Background The study journey recommendation is a special recommendation scene with combination of educational attributes and personalized requirements, and is characterized by matching multidimensional characteristics such as school segments, discipline preference, educational objective and educational qualification of study resources, theme suitability and safety condition of users. In the related technology, the study journey recommendation technology mainly has the following problems that the traditional collaborative filtering method relies on user-journey interaction data, recommendation effect is obviously reduced when a study platform faces a cold start problem, education features of the study are not fully fused, recommendation results are disjointed with education targets, the LSTM-based time sequence behavior modeling method can capture dynamic behavior sequences of users, but neglects reference values of group preference, and has insufficient modeling capability on multidimensional education attributes of the study journey, so that recommendation in an educational study scene is difficult to realize. Disclosure of Invention The embodiment of the invention aims to provide an intelligent study journey recommending method and system based on collaborative filtering and LSTM fusion, so as to at least solve the technical problem that the recommending result is inaccurate and the user requirement cannot be met in the existing recommending method. In order to achieve the above purpose, the present invention provides the following technical solutions. According to one embodiment of the application, an intelligent research journey recommendation method based on collaborative filtering and LSTM fusion is provided, and the method comprises the following steps: Acquiring and constructing a multi-mode behavior time sequence integrating dominant behavior data and recessive behavior data of a user; The method comprises the steps of taking a multi-modal behavior time sequence as input of a pre-built learning behavior space-time attention network to obtain a time sequence preference hidden vector, calculating the matching degree of the time sequence preference hidden vector and a dynamic constraint factor, correcting the time sequence preference hidden vector by utilizing the constraint factor feature vector based on a calculation result to obtain a scene self-adaptive time sequence preference vector, extracting the time sequence behavior feature vector of a current learning journey group, and carrying out weighted fusion with the scene self-adaptive time sequence preference vector to obtain a final time sequence preference vector, wherein the fusion weight is dynamically adjusted according to the richness of a user individual behavior sequence; And generating a user comprehensive preference vector by a concentration mechanism by combining the final time sequence preference vector with a group preference vector, a user study education feature vector and a real-time resource state vector which are mined based on collaborative filtering, calculating a recommendation score based on the comprehensive vector and a feature vector of a candidate study trip, and outputting a personalized study trip recommendation list by combining real-time resource availability screening and trip combination optimization. Further, the construction of the multi-mode behavior time sequence comprises the following steps: carrying out structural coding on the dominant behavior data to generate dominant behavior feature sub-vectors; carrying out multidimensional quantization and semantic coding on the implicit behavior data to obtain implicit behavior feature sub-vectors; The dominant behavior feature sub-vector and the recessive behavior feature sub-vector under the same time stamp are fused to generate a single-period behavior feature vector, and all the single-period behavior feature vectors are ordered according to the sequence of the time stamp to obtain the multi-mode behavior time sequence. Further, the pre-constructed learning behavior time-space attention network comprises: the time attention branch is used for dividing time windows according to the study period, calculating time attention weights based on the intervals between each time window and the current time and the correlation between each time window and the current study education requirement of a user, and outputting time attention feature vectors; The spatial attention branch is used for extracting the scientific spatial features corresponding to the behaviors in the input sequence, calculating the spatial attention weight based on the s