A college student academic activity recommendation method based on a dynamic knowledge graph

Abstract

The application provides a college student academic activity recommendation method based on a dynamic knowledge graph, and belongs to the technical field of artificial intelligence and knowledge reasoning. Through joint modeling of a static student portrait and a dynamic knowledge graph, a semantic injection mechanism, a relationship modulation time evolution unit, a time position coding based on a semester cycle, and sequence-level contrast learning are introduced to capture the structural dependency and time continuity of student interest. A student's historical participation sequence is modeled through multi-view sequence encoding and attention weighted aggregation based on a Transformer, and the individualized interest of the student is embodied. Finally, scoring is performed through a multilayer perceptron and a Sigmoid function, thereby realizing intelligent matching and personalized recommendation of students and academic competitions, scientific research projects, or campus activities. The application scheme can realize more accurate personalized project recommendation, and improve the accuracy and interpretability of student academic activity recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and knowledge reasoning technology, and particularly relates to a college student academic activity recommendation method based on a dynamic knowledge graph. BACKGROUND

[0002] At present, colleges and universities usually rely on teacher experience or static rules for student academic activity recommendation, such as simply matching related competition activities according to the student's major or grades, however, this approach has the following shortcomings:

[0003] Lack of dynamics, the existing method mainly uses static student portrait data, which cannot reflect the dynamic characteristics of the change of student interest over time;

[0004] Limited feature expression capability, traditional feature splicing or statistical feature extraction methods are difficult to capture the complex semantic association between students and between students and projects;

[0005] Insufficient personalization of recommendation results, existing recommendation methods and systems ignore the potential interest transfer law embodied in the behavior sequence of students participating in activities and competitions at different times, resulting in low recommendation accuracy and poor student participation enthusiasm.

[0006] In recent years, knowledge graphs and graph neural networks have been widely used in recommendation systems, however, most methods only focus on static graph modeling, fail to effectively integrate the dynamic knowledge evolution of the time dimension, and lack a modeling mechanism for student interest sequences in the education scenario. SUMMARY

[0007] To solve the above problems, the present application proposes a college student academic activity recommendation method based on a dynamic knowledge graph, which aims to capture the structural dependency and time continuity of student interest by jointly modeling static student portraits and dynamic knowledge graphs, introducing a semantic injection mechanism, a relationship modulation time evolution unit, a time position coding based on a semester cycle, and sequence-level contrastive learning, and thereby realizing intelligent matching and personalized recommendation of students and academic competitions, research projects or campus activities.

[0008] A college student academic activity recommendation method based on a dynamic knowledge graph, comprising the following steps:

[0009] Step S1, static information learning, a static knowledge graph is constructed from student static basic information, neighbor node information is aggregated using a structure learning based on a relational graph convolution network, and a static portrait embedding of each student is learned;

[0010] Step S2, dynamic information learning, by learning the behavior and participation records of students over time, a dynamic knowledge graph containing four-tuples of student-activity-relation-time is constructed, and through the cascade of distributed embedding learning, entity embedding and relation embedding of the dynamic knowledge graph are obtained, forming academic activity embedding containing dynamic information;

[0011] Step S3, time information encoding, through continuous position encoding, semester period encoding, event time difference encoding and time encoding fusion, the time information encoding of academic activities is obtained to adapt to the strong correlation between academic activities and time;

[0012] Step S4, multi-view sequence encoding, the static portrait embedding and comprehensive entity embedding output by static information learning and dynamic information learning are input into the Transformer encoder to obtain the academic activity sequence of each student;

[0013] Step S5, attention weighted aggregation, the comprehensive entity embedding of other academic activities that the student has not participated in obtained from dynamic information learning is attention weighted aggregated with the academic activity sequence to obtain a student interest embedding vector;

[0014] Step S6, matching prediction, the student interest embedding vector is concatenated with the candidate item embedding vector, input into the multi-layer perception MLP and Sigmoid function for scoring, and the matching probability of the student to the candidate item is output, and the system ranks according to the score to recommend academic activities for the student.

[0015] Further, the step S1 specifically comprises:

[0016] Input student basic information knowledge graph , Including entity nodes and their semantic relations , , ,..., Semantic relations , , ,..., Composing a semantic relation set R;

[0017] A relation graph convolution network R-GCN is used to aggregate the neighborhood information of entity nodes to realize semantic enhancement:

[0018]

[0019] Wherein, represents the embedding of the node of the layer, represents the embedding of the node of the Embedded, Indicates the first Layer nodes neighboring nodes Embedded, node The first layer of embeddings is obtained through random initialization. For the first The learnable parameter matrix of the layer relation r, The normalization constant is This represents the learnable parameter matrix of a node itself. It is a non-linear activation function. Represents nodes The set of nodes that represent relation r, i.e., the set of neighboring nodes;

[0020] After multiple aggregations and updates, the static portrait of the student node is embedded in S.

[0021] Further, step S2 specifically includes:

[0022] Step S21, semantic injection, introduces a large model semantic injection mechanism, embedding and mapping the sentence vectors of student attribute descriptions and project text introductions to the entity embedding space, thereby aligning semantic knowledge with entity embedding;

[0023] Step S22, relation modulation gating: The relation embedding from the previous time step and the relation embedding generated by semantic encoding using the text encoding model in the current time step are used as input. The update gate and reset gate of the relation modulation gating GRU are modulated through the relation vector to obtain the comprehensive relation embedding of the current time step.

[0024] Step S23, structure learning, embedding the semantically enhanced initial entities. The integrated relationship of the modulated and gated output of the relationship at this time step is embedded into the input of the relationship graph convolutional network, allowing each entity node to aggregate neighbor information to obtain the relationship. ;

[0025] Step S24, entity gating: The semantically enhanced initial entity embedding is gating with the entity embedding that has undergone structure learning at the current time step to obtain the comprehensive entity embedding at the current time step. After cascading through multiple time steps, the comprehensive entity embedding of any entity node is finally obtained.

[0026] Further, step S21 specifically includes:

[0027] Extract sentence vector embeddings from student attribute descriptions and project introduction texts using a pre-trained large language model. The entity embeddings are then mapped to the entity space via linear projection, ultimately yielding semantically enhanced initial entity embeddings. and initial relation embedding :

[0028]

[0029]

[0030] wherein, , respectively represent the initial entity embedding and the initial relation embedding obtained by initializing the encoding, , is a fusion weight, , is a learnable linear projection matrix, , is an entity and relation sentence embedding obtained by encoding the attribute description or the item introduction text by a pre-trained large language model.

[0031] Further, the step S22 specifically comprises:

[0032] At time step , the semantic enhanced initial relation embedding is inputted together with the semantic encoded input at the current time step to generate the relation embedding using the text encoding model, and the comprehensive relation embedding at time step is obtained by modulating the update gate and the reset gate of the relation vector modulated gated recurrent unit (GRU) with the relation embedding,

[0033] For any time step t, the semantic enhanced initial relation embedding is inputted together with the semantic encoded input at the current time step to capture the time evolution difference of different relations, and the comprehensive relation embedding at the current time step is obtained by encoding the semantic encoded input using the text encoding model, :

[0034]

[0035]

[0036]

[0037] wherein, represents a relation modulated gate vector, represents a relation modulated candidate relation embedding, represents the comprehensive relation embedding at the previous time step, is the relation embedding at the current time step based on the text encoding model, and the symbol represents an element-wise product between vectors, , , These are the learnable linear transformation matrices corresponding to the update gates. For bias terms, Represents the hyperbolic tangent function. , , These are the learnable linear transformation matrices corresponding to the reset gate. This is a bias term.

[0038] Further, step S24 specifically includes:

[0039] At time step Embed the semantically enhanced initial entity The entity embedding obtained through structure learning at the current time step Input entity gating, output any entity node at the current time. Integrated entity embedding For any time step t, The current moment's integrated entity embedding for:

[0040]

[0041] in, It is a time-controlled weight matrix. It is a bias term. It was the previous step Moment-based integrated entity embedding, The entity embedding at this moment, after structure learning, is cascaded through multiple time steps to finally obtain the comprehensive entity embedding of any entity node. .

[0042] Furthermore, step S3 specifically includes:

[0043] Step S31, Continuous Position Encoding: The timestamp is encoded using a continuous position encoding method based on sine and cosine functions. For the i-th academic activity in the student activity sequence, its position code is defined as:

[0044]

[0045]

[0046]

[0047] in, This represents the consecutive position encoding embedding of the i-th academic activity. d represents the hidden layer dimension. denotes the value of the i-th academic activity in the 2k+1 dimension of the position encoding embedding, denotes the value of the i-th academic activity in the 2k+1 dimension of the position encoding embedding,

[0048] Step S32, semester period encoding: the timestamp of the i-th academic activity is mapped to a multi-level periodic time feature:

[0049]

[0050] wherein, denotes the timestamp of the i-th academic activity, denotes the semester number, denotes the week number within the semester, denotes the semester phase;

[0051] For any periodic time feature its period encoding is defined as:

[0052]

[0053]

[0054]

[0055] wherein, denotes the semester period encoding embedding of the i-th academic activity, denotes the value of the i-th academic activity in the 2k dimension of the semester period encoding embedding, denotes the value of the i-th academic activity in the 2k+1 dimension of the semester period encoding embedding, denotes the length of the corresponding period;

[0056] Step S33, inter-event time difference encoding: the time difference between the i-th academic activity and its previous academic activity is defined as:

[0057]

[0058] The time difference is expanded using a multi-scale sine function:

[0059]

[0060] wherein, is the inter-event time difference encoding embedding of the i-th academic activity at M time scales, M is the number of time scales, denotes the inter-event time difference encoding embedding of the i-th academic activity at the m-th time scale, the inter-event time difference encoding at the m-th time scale is defined as:

[0061]

[0062]

[0063]

[0064] wherein, denotes the value of the 2kth dimension of the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, denotes the value of the 2k+1th dimension of the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, is the normalization factor corresponding to the mth time scale;

[0065] Step S34, time encoding fusion: the final time encoding of the ith academic activity is denoted as:

[0066] ;

[0067] wherein, denotes the time encoding fusion embedding of the ith academic activity.

[0068] Further, the step S4 specifically comprises:

[0069] Let the comprehensive entity embedding of each academic activity entity fuse the time encoding information of a specific student :

[0070] ;

[0071] wherein, is the academic activity entity embedding, i takes values from 1 to n, and then the academic activity sequence of each student in the time dimension is taken as the input sequence , the first element S of the sequence is the student static portrait embedding from static information learning;

[0072] For each input sequence , multi-view enhancement is performed, and the enhanced view sequence and the original sequence are respectively encoded by Transformer to obtain and , then the InfoNCE loss is used to close the embedding distance of the same sequence under the original view and the enhanced view, and to close the distance of the embedding of other sequences, and the InfoNCE loss is:

[0073]

[0074] wherein, denotes the InfoNCE loss, denotes the interest embedding of the i-th negative sample after being processed by the Transformer encoder, denotes the enhanced view interest embedding of the same original input sequence after being processed by the Transformer encoder, denotes the i-th negative sample after being randomly sampled from all academic activities, denotes the interest embedding of the i-th negative sample after being processed by the Transformer encoder, is a temperature coefficient, denotes a similarity function;

[0075] Finally, the academic activity entity embedding is fused with the student personalized interest, and the academic activity sequence of each student is obtained after being processed by the Transformer encoder .

[0076] Further, the step S5 specifically includes:

[0077] defining a comprehensive entity embedding of other academic activities that the student has not participated in embedding the candidate item, calculating the attention weight between the candidate item and each activity in the input sequence:

[0078]

[0079] wherein n denotes the number of historical academic activities that the student has participated in, denotes the relevance score function of the candidate item and the i-th historical academic activity;

[0080] based on the attention weight the historical academic activity vector is weighted and summed to obtain a student interest embedding vector of the same type as :

[0081]

[0082] wherein, is the historical academic activity vector, and the student interest embedding vector denotes the dynamic interest representation of the student in the context of the candidate item .

[0083] Further, the step S6 specifically includes:

[0084] concatenating the student interest embedding vector and the candidate item embedding :

[0085] ​​

[0086] concatenate the vectors Input the multi-layer perception:

[0087]

[0088] wherein, is a Sigmoid function, and output represents the matching probability of the student and the candidate item, , is a learnable parameter matrix, , is a bias term;

[0089] The system ranks according to the score and recommends the top N items with the highest scores to the student.

[0090] The beneficial technical effects of the present application are:

[0091] Deep integration of static semantics and dynamic knowledge graph is realized, and representation consistency and interpretability are enhanced: the present application can capture the long-term attributes and dynamic behavior characteristics of students at the same time, the mechanism makes the embedding have richer semantic structure, which helps to improve the matching ability between student background information and item content, thereby improving the accuracy and interpretability of academic planning;

[0092] The relationship modulation gate is proposed, and the evolution of the relationship is explicitly controlled to change the entity over time: traditional dynamic graph embedding methods only update the entity over time, but it is difficult to reflect the modulation effect of the evolution of different relationships on the behavior of the entity. The present application uses a relationship gating unit, which makes the relationship vector directly affect the update process of the gating mechanism, can capture the differences in the evolution of different relationship types such as competition activities and research activities on the state of the student, and thus improve the modeling ability of complex educational behavior graphs;

[0093] Multi-scale time coding of semester cycle and time difference is introduced to improve the ability to capture the periodic and dense behavior rules in the educational scene: the present application proposes a joint modeling mechanism of three types of time signals, standard position coding, semester cycle coding and event time difference, so that the sequence model can accurately describe the real behavior rhythm of the student in the early and late semesters, examination weeks and activity-intensive periods, and effectively improve the adaptation ability of the model to time-sensitive tasks such as item selection and activity participation timing;

[0094] Propose a static student image guided individualized sequence modeling framework to enhance the individual difference sensitivity of interest modeling: by embedding the student as the first element of the sequence into multi-view sequence modeling, the model continuously references the individual characteristics of the student in the self-attention calculation, thereby realizing the joint modeling of student-project bilateral semantics, which can significantly improve the ability of the model to model the interests of students with different backgrounds, abilities and preferences, and realize more accurate education recommendation;

[0095] Improve the robustness and discriminability of interest representation through sequence-level contrastive learning: the invention constructs multi-view enhancement on student behavior sequences, and uses sequence-level contrastive loss to narrow the representation of different views of the same student and distinguish the representation of different student sequences, effectively reducing the representation deviation caused by noise activities, behavior sparsity, etc. This mechanism improves the robustness and generalization ability of interest representation, making academic planning perform better in cold start and low-frequency participating students;

[0096] Improve the accuracy, stability and explainability of the recommendation result, and realize personalized academic activity recommendation: the invention combines structure learning, time continuity modeling, time coding and semantic injection to realize dynamic, structured and semantic comprehensive modeling of student interest. The recommendation result not only has high accuracy, but also can be output in an interpretable manner through embedding explanation and attention weight, meeting the transparency requirements of the education context;

[0097] Good scalability, can be extended to research topic selection, course planning and career development, etc. Multiple scenarios: the architecture of the invention is suitable for any form of education behavior and activity recommendation, and can be smoothly extended to research direction selection, learning path planning, campus resource matching, etc. It has wide practical application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0098] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0099] Figure 1 is a process schematic diagram of a college student academic activity recommendation method based on a dynamic knowledge graph provided by an embodiment of the present application;

[0100] Figure 2 is a structural schematic diagram of an embedding learning layer provided by an embodiment of the present application;

[0101] Figure 3 is a structural schematic diagram of a semantic injection module provided by an embodiment of the present application;

[0102] Figure 4 is a structural schematic diagram of a relationship gate unit provided by an embodiment of the present application;

[0103] Figure 5 is a structural schematic diagram of an entity gate unit provided by an embodiment of the present application;

[0104] Figure 6 is a structural schematic diagram of a student interest extraction layer provided by an embodiment of the present application;

[0105] Figure 7 is a structural schematic diagram of a matching prediction layer provided by an embodiment of the present application. DETAILED DESCRIPTION

[0106] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0107] A college student academic activity recommendation method based on a dynamic knowledge graph, as shown in Figure 1 , the method comprises the following steps:

[0108] Step S1, static information learning, to construct a static knowledge graph with student static basic information (including gender, college, major, etc.), to aggregate neighbor node information by using a structure learning based on a relationship graph convolution network, and to learn a static portrait embedding of each student.

[0109] Specifically, an input student portrait knowledge graph , includes entity nodes (such as students, affiliated colleges, affiliated majors, and attributes) and their relationships , , ,..., such as "student-college is-computer college", "student-major is-computer science", and semantic relationships , , ,..., constitute a semantic relationship set R.

[0110] Each entity node aggregates neighbor node information into the node by using a relationship graph convolution network (R-GCN), to achieve semantic enhancement:

[0111]

[0112] wherein, representing the embedding of the layer node , representing the embedding of the layer node , representing the embedding of the neighbor nodes of the layer node , the embedding of the node The embedding of the first layer is initialized randomly, is the learnable parameter matrix of the layer relationship r, is a normalization constant, is the learnable parameter matrix of the node itself, is a nonlinear activation function, representing the set of nodes that exhibit the relationship r, i.e., the neighbor node set.

[0113] After multiple aggregation updates, the static portrait embedding S of the student node is output.

[0114] Step S2, dynamic information learning, by learning the behavior and participation records of students over time, a dynamic knowledge graph containing four-tuple form of student- academic activity- relationship- time is constructed, representing the student- academic activity- relationship knowledge graph at time tn. Distributed embedding learning takes entity embedding, relationship embedding, and randomly initialized relationship embedding as input. Through the cascade of distributed embedding learning, distributed embedding learning learns the entity embedding and relationship embedding of the knowledge graph at a certain time, and passes it to the next time distributed embedding learning for learning. Finally, the academic activity embedding containing dynamic information is obtained.

[0115] Step S21, semantic injection, introduce a large model semantic injection mechanism, map the sentence vector embedding of student attribute description and project text introduction to the entity embedding space, realize the alignment of semantic knowledge and entity embedding. Specifically:

[0116] Use the pre-trained large language model to extract the sentence vector embedding of the student attribute description and the project introduction text , and map it to the entity space through linear projection, finally get the semantic enhanced initial entity embedding and the initial relationship embedding :

[0117]

[0118]

[0119] wherein,​ respectively represent initial entity embeddings and initial relation embeddings obtained by initializing the encoding, is a fusion weight, is a learnable linear projection matrix, is an entity and relation sentence embedding obtained by encoding the attribute description or item introduction text by a pre-trained large language model.

[0120] Step S22, relation modulation gating, at time step the semantic-enhanced initial relation embedding and the input at the current time step the relation embedding generated by semantic encoding using the text encoding model are jointly used as input to modulate the update gate and the reset gate of the relation vector- modulated GRU, to obtain the comprehensive relation embedding at time step , so that the relation change can explicitly affect the entity evolution process, and this mechanism improves the time sensitivity and semantic expression ability of the model in the multi-relation dynamic graph. For any time step t, the semantic-enhanced initial relation embedding and the input at the current time step after semantic encoding capture the differences in the time evolution of different relations, is the relation embedding obtained by semantic encoding using the text encoding model, and the comprehensive relation embedding at the current time step is output :

[0121]

[0122]

[0123]

[0124] wherein, represents a relation-modulated gating vector for determining the time update amplitude of the relation embedding, represents a relation-modulated candidate relation embedding, indicating the candidate relation state generated after fusing the history and new semantics at the current time step, represents the comprehensive relation embedding at the previous time step, used to represent the accumulated historical relation semantic information up to the current time, is the relation embedding at the current time step based on the text encoding model, and the symbol represents an element-wise product between vectors, used to explicitly model the fine-grained interaction between the historical relation state and the current relation input, ,​​​​​​ are the learnable linear transformation matrices corresponding to the update gate, where Specifically designed for extracting interactive items The dynamic modulation information of relationships contained within it For bias terms, Represents the hyperbolic tangent function. , , These are the learnable linear transformation matrices corresponding to the reset gate. This is a bias term.

[0125] Step S23, structure learning, integrates structural information into entity embedding, at time step Embed the semantically enhanced initial entity Embedded relationship between the modulated gated output and the time step relationship The input is fed into a graph convolutional network, allowing each entity node to aggregate neighbor information to obtain... The same operation is performed at each time step.

[0126] Step S24, entity gating, in order to embed and fuse historical information into entities, at the time step Embed the semantically enhanced initial entity The entity embedding obtained through structure learning at the current time step Input an improved entity-gated recurrent unit to capture the continuous characteristics of entity changes over time, and output any entity node at the current time. Integrated entity embedding For any time step t, The current moment's integrated entity embedding for:

[0127]

[0128] in, It is a time-controlled weight matrix. It is a bias term. It was the previous step Moment-based integrated entity embedding, The entity embedding at this moment, after structure learning, is cascaded through multiple time steps to finally obtain the comprehensive entity embedding of any entity node. The system obtained entity embeddings at multiple time steps. With relational distributed embedding, input is provided for subsequent interest extraction, ultimately enabling comprehensive entity embedding of each entity node. All contain dynamic information.

[0129] Step S3, time information encoding, in the academic activity recommendation scenario, student behavior not only has obvious time sequence, but also is strongly influenced by periodic time factors such as semester rhythm, in order to fully depict the time characteristics, the application models the time information from three levels of sequential position, semester period and time interval between events, to adapt to the strong correlation between academic activities and time.

[0130] Step S31, continuous position encoding: in order to describe the relative sequence of student behavior, the application encodes the timestamp by using a continuous position encoding method based on sine and cosine functions. For the i-th academic activity in the academic activity sequence, the position encoding is defined as:

[0131]

[0132]

[0133]

[0134] wherein, represents the continuous position encoding embedding of the i-th academic activity, d represents the dimension of the hidden layer, represents the value of the i-th academic activity in the 2k dimension of the position encoding embedding, represents the value of the i-th academic activity in the 2k+1 dimension of the position encoding embedding. This encoding method can effectively distinguish the relative position of different behaviors in the sequence, so that the model remains sensitive to the behavior sequence.

[0135] Step S32, semester period encoding: considering that academic behavior is highly related to the structure of the semester (such as semester, week number and teaching stage), the timestamp is mapped to periodic time features such as semester, week number and stage, and periodic feature modeling is realized by sine and cosine basis expansion, specifically:

[0136] The application maps the timestamp of academic activity i to multi-level periodic time features:

[0137]

[0138] wherein, represents the timestamp of academic activity i, represents the semester number, represents the week number in the semester, represents the semester stage (such as teaching period, examination week or holiday).

[0139] For any periodic time feature , the period encoding is defined as:

[0140]

[0141]

[0142]

[0143] wherein, denotes the period cycle encoding embedding of the i-th academic activity, denotes the value of the 2k-th dimension of the period cycle encoding embedding of the i-th academic activity, denotes the value of the 2k+1-th dimension of the period cycle encoding embedding of the i-th academic activity, denotes the length of the corresponding period (e.g. the number of weeks in a semester), which design enables the model to capture the recurring learning behavior patterns and phase changes within a semester.

[0144] Step S33, inter-event time difference encoding: based on the time interval between two adjacent participating behaviors, the time difference is calculated and discretized. The present application realizes the modeling of dynamic characteristics such as behavior density and activity interval length by using multi-scale sine function expansion (such as short time difference, weekly difference, monthly difference) on the time difference. Specifically:

[0145] To characterize the time interval characteristics between adjacent academic behaviors, the time difference between the i-th academic activity and its previous academic activity is defined as:

[0146]

[0147] To model the behavior rhythm changes at different time scales at the same time, the present application uses multi-scale sine function expansion on the time difference:

[0148]

[0149] wherein, is the inter-event time difference encoding embedding of the i-th academic activity under M time scales, M is the number of time scales, and the time scale can control short-term or long-term activity changes, denotes the inter-event time difference encoding embedding of the i-th academic activity under the m-th time scale, and the inter-event time difference encoding under the m-th time scale is defined as:

[0150]

[0151]

[0152]

[0153] wherein, denotes the value of the 2k-th dimension of the inter-event time difference encoding embedding of the i-th academic activity under the m-th time scale, represents the value of the 2k+1 dimension of the event-to-event time difference encoding of the ith academic activity in the mth time scale, is the normalization factor corresponding to the mth time scale, which can correspond to short time scale (such as hour level), medium time scale (such as week level) and long time scale (such as month level) respectively. This encoding method can effectively characterize the behavior density change, learning activity interval length and dynamic change characteristics of academic rhythm.

[0154] Step S34, time encoding fusion: integrating the above-mentioned various time information, the final time encoding of the ith academic activity is represented as:

[0155] ;

[0156] wherein, represents the time encoding fusion embedding of the ith academic activity.

[0157] Step S4, multi-view sequence encoding: let the comprehensive entity embedding of each academic activity entity fuse the time encoding information of a specific student :

[0158] ;

[0159] i is valued from 1 to n, and then the academic activity sequence of each student in the time dimension is taken as an input sequence , the first element S of the sequence is the student static embedding vector from static information learning, and the subsequent elements are academic activity entity embeddings interacting with the student from dynamic information learning, so that the subsequent multi-head attention always refers to the individual differences of the student when calculating the inter-project dependency.

[0160] To improve the robustness of the interest representation, the present application constructs a multi-view enhancement (such as random activity sampling, semantic enhancement, etc.) for each input sequence , and encodes the enhanced view sequence and the original sequence through Transformer to obtain and , then uses InfoNCE loss to narrow the embedding distance of the same sequence under the original view and the enhanced view, and to distance the embedding distance of other sequences, thereby improving the robustness and discriminability of the interest representation under multi-view disturbance, and the InfoNCE loss is:

[0161]

[0162] wherein, represents the InfoNCE loss, This represents the interest embedding obtained after the original input sequence (unenhanced view) is processed by the Transformer encoder. This represents the enhanced view interest embedding obtained by the Transformer encoder after applying multi-view enhancement to the same original input sequence. This represents the result of randomly sampling all academic activities. One negative sample, Indicates the first Interest embeddings obtained from negative samples through a Transformer encoder It is the temperature coefficient. This represents a similarity function (such as cosine similarity).

[0163] Transformer encoding employs a multi-head self-attention mechanism, which calculates the dependency weights of each item in the input sequence using this mechanism.

[0164]

[0165] Where Q, K, and V are the query, key, and value vectors of the input sequence, respectively, obtained by linearly mapping the input sequence to a learnable weight matrix. It is a normalization function. This is the scaling factor.

[0166] Ultimately, this allows academic activities to be embedded in reality. By incorporating students' individual interests and processing them through a Transformer encoder, a sequence of academic activities for each student is obtained. .

[0167] Step S5, attention-weighted aggregation, embeds the comprehensive entity of other academic activities that the student has not yet participated in, obtained from dynamic information learning. and Perform attention-weighted aggregation to predict what students are interested in. Vector of similar academic activities Specifically:

[0168] First, define the integrated entity embedding of other academic activities that students have not yet participated in. For candidate item embedding, calculate the attention weights between the candidate item and each activity in the input sequence:

[0169]

[0170] Where n represents the number of history academic activities the student has participated in. Indicates candidate item embedding The relevance score function of the ith historical academic activity, which can adopt a dot product form or a learnable bilinear function definition.

[0171] Then, based on the attention weight The historical academic activity vectors are weighted and summed to obtain the student interest embedding vector of the same type as the student interest embedding vector:

[0172]

[0173] wherein, is the historical academic activity vector, which represents the dynamic interest representation of the student in the candidate item context, which can integrate the historical behavior sequence and the semantic relevance of the candidate item.

[0174] The final output student interest embedding vector will be input into the matching prediction layer to calculate the matching score of the student and the candidate item, which integrates the long-term preference and short-term behavior characteristics of the student, and can be directly used in the prediction stage.

[0175] Step S6, matching prediction, concatenating the student interest embedding vector and the candidate item embedding vector, inputting a multi-layer perception MLP and a Sigmoid function for scoring, outputting the matching probability of the student to the candidate item, and the system ranks according to the score to recommend academic activities for the student. Specifically,

[0176] The student interest embedding vector is concatenated with the candidate item embedding :

[0177]

[0178] The concatenated vector is input into a multi-layer perception:

[0179]

[0180] wherein, is a Sigmoid function, and the output represents the matching probability of the student and the candidate item, , is a learnable parameter matrix, , is a bias term.

[0181] The system recommends the top N items with the highest scores to the student according to the scoring ranking.

[0182] The application also provides a college student academic activity recommendation system based on a dynamic knowledge graph, which comprises an embedding learning layer, a student interest extraction layer and a matching prediction layer. The embedding learning layer is used to learn distributed embedding representations of students, projects and related entities from static and dynamic knowledge graphs, serving as a basis for subsequent interest modeling and recommendation. The student interest extraction layer is realized by a multi-view sequence encoding module based on a Transformer and an attention weighted aggregation module. The entity embeddings of students, competitions and activities output by the embedding learning layer are input into the multi-view sequence encoding module based on the Transformer, which models the historical participation sequence of the students. The goal of the student interest extraction layer is to capture the evolution of user interest and obtain a more accurate representation of user interest. The multi-view sequence encoding module uses a Transformer encoder to capture the time features in the user interest sequence, obtaining a user behavior sequence with time features. The attention weighted aggregation module aggregates the user interest with time relationships into the final user interest representation using attention calculation. The matching prediction layer is used to match and predict the academic activities of students.

[0183] As shown in Figure 2 , the embedding learning layer comprises a static information learning module, a dynamic information learning module and a time information encoding module. The static information learning is realized by a structure learning module. The dynamic information learning module comprises a semantic injection module and a distributed embedding learning unit, which comprises a structure learning module, a relationship gating unit and an entity gating unit.

[0184] Static information learning module: a static knowledge graph is constructed based on the static basic information of students (including gender, college, major, etc.), and a structure learning based on a relational graph convolution network is used to aggregate neighbor node information, so as to learn the static portrait embedding of each student, which specifically comprises:

[0185] Input student portrait knowledge graph , which comprises entity nodes (such as students, affiliated colleges, affiliated majors, and attributes) and their relationships , , ,..., such as "student-college is-computer college", "student-major is-computer science", and semantic relationships , , ,..., constitute a semantic relationship set R.

[0186] A relational graph convolution network (R-GCN) is used to aggregate neighbor node information into each entity node, achieving semantic enhancement:

[0187]

[0188] where, denotes the embedding of the layer node , denotes the embedding of the layer node , denotes the embedding of the layer node 's neighbor nodes , The embedding of the first layer is obtained by random initialization, is the learnable parameter matrix of the layer relation r, is a normalization constant, denotes the learnable parameter matrix of the node itself, is a nonlinear activation function, denotes the set of nodes that exhibit the relation r with the node , i.e., the set of neighbor nodes.

[0189] After multiple aggregation updates, the static portrait embedding S of the student node is output.

[0190] Semantic injection module: by inputting the attribute description of the student and academic activity, project introduction text, using a pre-trained large language model (LLM) to extract the sentence embedding corresponding to the text, and through linear projection, mapping to the initial embedding (E / R) space of the entity or relation itself of the original graph. As shown in Figure 3 , specifically includes:

[0191] Using a pre-trained large language model to extract the sentence vector embedding of the student attribute description and project introduction text , and through linear projection mapping to the entity space, finally obtaining the semantic enhanced initial entity embedding and initial relation embedding :

[0192]

[0193]

[0194] where, , denotes the initial entity embedding and initial relation embedding obtained by initialization encoding, , is the fusion weight, , is a learnable linear projection matrix, , , This refers to the embedding of entity and relation sentences obtained by encoding attribute descriptions or project introduction texts using a pre-trained large language model.

[0195] Relationship gating unit: such as Figure 4 As shown, the relation gating unit mainly includes three operations: update, reset, and candidate relation. The update gate controls the fusion ratio of historical information from the previous time step with the current input information; the closer the update gate is to 1, the more historical information is retained. Compared to traditional gated recurrent units, the update gate in the relation gating unit is no longer a uniform rule, but rather each relation adjusts the forgetting and updating ratio based on its own historical embedding and current input. The reset gate controls the contribution of historical information in calculating candidate hidden states; the candidate states, combined with the update gate, determine the final output. Specifically, this includes:

[0196] At time step Embed the semantically enhanced initial relation With this time step Relation embeddings generated using semantic encoding models based on text encoding. Using the relation vector as input, the update and reset gates of the relation modulation gated GRU are modulated to obtain the time step. Embedded comprehensive relationship of time This mechanism allows changes in relationships to explicitly influence entity evolution, enhancing the model's temporal sensitivity and semantic expressiveness in multi-relational dynamic graphs. For any time step t, The semantically encoded input at this time step Capturing the temporal evolution differences of different relationships, To output the comprehensive relation embedding at the current time step, which is obtained by semantic encoding using a text encoding model. :

[0197]

[0198]

[0199]

[0200] in, It is the gate vector for relation modulation, used to determine the temporal update magnitude of relation embedding. It is a candidate relation embedding of relation modulation, representing the candidate relation state generated at the current time step after fusing history and new semantics. This represents the comprehensive relation embedding from the previous time step, used to characterize the accumulated historical relation semantic information up to the current moment. For the current time step, the relation embedding is based on a text encoding model, and the symbol is... denotes element-wise multiplication between vectors, which is used to explicitly model the fine-grained interaction between the historical relational state and the current relational input, , , are the learnable linear transformation matrices corresponding to the update gate, respectively, where is specially designed to extract the relational dynamic modulation information contained in the interaction term is the bias term, denotes the hyperbolic tangent function, , , , are the learnable linear transformation matrices corresponding to the reset gate, respectively, is the bias term.

[0201] The structure learning module fuses the structure information into the entity embedding, and at time step , the semantic-enhanced initial entity embedding and the comprehensive relational embedding output by the relational modulation gate at the current time step are input into the relational graph convolution network, so that each entity node aggregates neighbor information to obtain , and the same operation is performed at each time step.

[0202] The entity gating unit: as shown in Figure 5 , the dynamic balance between the historical entity embedding and the current entity embedding is realized through the gating mechanism, so that the model can flexibly capture the feature evolution of the entity over time while maintaining the time continuity. When the length of the historical knowledge graph sequence is longer and the number of stacked layers of the R-GCN is larger, problems such as gradient disappearance may occur. Therefore, for the entity distributed embedding learning with time characteristics, the entity gating unit is used to slow down these problems to enhance the maintenance ability of the graph structure information. Specifically, it includes:

[0203] At time step , the semantic-enhanced initial entity embedding and the entity embedding obtained by structure learning at the current time step are input into the improved entity gating recurrent unit to capture the continuity feature of the entity over time, and the comprehensive entity embedding of any entity node at the current time is output. For any time step t, , the comprehensive entity embedding at the current time is:

[0204]

[0205] wherein, is a time-controlled weight matrix, is a bias term, is the integrated entity embedding of the previous time step is the integrated entity embedding of the current time step is the entity embedding of the current time step learned by structure learning, and the integrated entity embedding of any entity node is obtained through concatenation of multiple time steps The system obtains entity embedding and relationship distributed embedding at multiple time steps, provides input for subsequent interest extraction, and finally makes the integrated entity embedding of each entity node contain dynamic information.

[0206] Time information encoding module: in the academic activity recommendation scene, student behavior not only has obvious time sequence, but is also strongly influenced by periodic time factors such as semester rhythm. In order to fully depict this time characteristic, the time information encoding module models the time information from three aspects of sequential position, semester period and event time interval, to adapt to the strong correlation between academic activities and time, including:

[0207] Continuous position encoding: in order to describe the relative sequence of student behavior, the invention adopts continuous position encoding based on sine and cosine functions to encode the timestamp. For the i-th academic activity in the student activity sequence, the position encoding is defined as:

[0208]

[0209]

[0210]

[0211] wherein, represents the continuous position encoding embedding of the i-th academic activity, d represents the hidden layer dimension, represents the value of the i-th academic activity in the 2k dimension of the position encoding embedding, represents the value of the i-th academic activity in the 2k+1 dimension of the position encoding embedding. This encoding method can effectively distinguish the relative position of different behaviors in the sequence, so that the model remains sensitive to the behavior sequence.

[0212] Semester period encoding: considering that academic behavior is highly related to the structure of the semester (such as semester, week, and teaching stage), the timestamp is mapped to periodic time features such as semester, week, and stage, and periodic feature modeling is achieved through sine and cosine basis expansion, specifically:

[0213] The invention maps the timestamp of academic activity i to multi-level periodic time features:

[0214]

[0215] where, denotes the timestamp of academic activity i, denotes the semester number, denotes the week number within the semester, denotes the phase of the semester (e.g., teaching period, exam week, or holiday).

[0216] For any periodic time feature its period encoding is defined as:

[0217]

[0218]

[0219]

[0220] where, denotes the semester period encoding embedding of the i-th academic activity, denotes the value of the i-th academic activity in the 2k-th dimension of the semester period encoding embedding, denotes the value of the i-th academic activity in the 2k+1-th dimension of the semester period encoding embedding, denotes the length of the corresponding period (e.g., the number of weeks in a semester), which allows the model to capture recurring learning behavior patterns and phase changes within a semester.

[0221] Inter-event time difference encoding: based on the time interval between two adjacent participation behaviors, the time difference is calculated and discretized. The invention uses multi-scale sine function expansion (such as short time difference, weekly difference, monthly difference) on the time difference to model the behavior density, activity interval length, etc. Dynamic characteristics. Specifically:

[0222] To characterize the time interval between adjacent academic behaviors, the time difference between the i-th academic activity and its previous academic activity is defined as:

[0223]

[0224] To model the behavior rhythm changes at different time scales, the invention uses multi-scale sine function expansion on the time difference:

[0225]

[0226] where, is the inter-event time difference encoding embedding of the i-th academic activity at M time scales, M is the number of time scales, and the time scale can control short-term or long-term activity changes, denotes the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, and the inter-event time difference encoding at the mth time scale is defined as:

[0227]

[0228]

[0229]

[0230] wherein, denotes the value of the 2kth dimension of the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, denotes the value of the 2k+1th dimension of the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, is a normalization factor corresponding to the mth time scale, which can correspond to a short time scale (such as a hourly level), a medium time scale (such as a weekly level), and a long time scale (such as a monthly level), respectively. The encoding mode can effectively depict the behavior density variation, the interval length of learning activities, and the dynamic variation characteristics of academic rhythm.

[0231] Time encoding fusion: integrating the above-mentioned various time information, the final time encoding of the ith academic activity is represented as:

[0232] ;

[0233] wherein, denotes the time encoding fusion embedding of the ith academic activity.

[0234] As shown in Figure 6 , the student interest extraction layer includes a multi-view sequence encoding module and an attention weighted aggregation module, specifically:

[0235] Multi-view sequence encoding module: let the comprehensive entity embedding of each academic activity entity fuse the time encoding information of a specific student :

[0236] ;

[0237] The value of i is 1 to n, and then the academic activity sequence of each student in the time dimension is taken as an input sequence , the first element S of the sequence is a student static embedding vector from static information learning, and the subsequent elements are academic activity entity embeddings from dynamic information learning and interacting with the student, so that the subsequent multi-head attention always refers to the individual differences of the student when calculating the inter-project dependency.

[0238] To improve the robustness of the interest representation, the present application performs a multi-head self-attention operation on each input sequence Construct multi-view enhancement (such as random activity sampling, semantic enhancement, etc.), and encode the enhanced view sequence and the original sequence through the Transformer respectively and Then, the InfoNCE loss is used to narrow the embedding distance of the same sequence under the original view and the enhanced view, and to widen the distance between the embedding of other sequences, so as to improve the robustness and discriminability of the interest representation under multi-view disturbance. The InfoNCE loss is:

[0239]

[0240] wherein, represents the InfoNCE loss, represents the interest embedding obtained by the Transformer encoder after the original input sequence (unenhanced view), represents the enhanced view interest embedding obtained by the Transformer encoder after applying multi-view enhancement to the same original input sequence, represents the th negative sample obtained by randomly sampling all academic activities, represents the interest embedding of the th negative sample obtained by the Transformer encoder, is the temperature coefficient, represents the similarity function (such as cosine similarity).

[0241] The Transformer encoder is stacked by multiple encoding layers, and each encoding layer is composed of two sub-modules, i.e. multi-head self-attention and feed-forward neural network for feed-forward propagation. After each sub-module, residual addition and normalization are introduced to alleviate the gradient vanishing problem and improve the stability of model training.

[0242] The Transformer encoding adopts a multi-head self-attention mechanism to calculate the dependency weight of each item in the input sequence:

[0243]

[0244] wherein Q, K, V are the query, key and value vectors of the input sequence, which are obtained by linear mapping of the input sequence and the learnable weight matrix, is a normalization function, is a scaling factor.

[0245] Finally, the academic activity entity embedding is fused with the personalized interest of students, and the academic activity sequence of each student is obtained by processing through the Transformer encoder .

[0246] Attention-weighted aggregation module: comprehensive entity embedding of other academic activities that the student has not participated in obtained from dynamic information learning with Attention-weighted aggregation is performed to predict the student's interest in academic activities of the same type as vector , specifically:

[0247] First, define the comprehensive entity embedding of other academic activities that the student has not participated in as a candidate item embedding, calculate the attention weight between the candidate item and each activity in the input sequence:

[0248]

[0249] where n represents the number of historical academic activities that the student has participated in, represents the candidate item embedding and the relevance score function of the i-th historical academic activity, the relevance score function can be defined in the form of dot product or learnable bilinear function.

[0250] Then, based on the attention weight weighted sum of the historical academic activity vector, the student interest embedding vector of the same type as is obtained:

[0251]

[0252] where, is the historical academic activity vector, which represents the dynamic interest representation of the student in the candidate item context, which can integrate the historical behavior sequence and the semantic relevance of the candidate item.

[0253] The final output student interest embedding vector will be input into the matching prediction layer to calculate the matching score of the student and the candidate item, which integrates the student's long-term preference and short-term behavior characteristics, and can be directly used in the prediction stage.

[0254] As shown in Figure 7 , the matching prediction layer includes a multi-layer perception MLP and a Sigmoid function, which concatenates the student interest embedding vector and the candidate item embedding vector, inputs them into the multi-layer perception MLP and the Sigmoid function for scoring, and outputs the matching probability of the student to the candidate item. The system ranks according to the score to recommend academic activities for the student. Specifically:

[0255] The student interest embedding vector with candidate item embedding Concatenate:

[0256]

[0257] Input the concatenated vector into a multi-layer perception:

[0258]

[0259] where, is a sigmoid function, and the output represents the matching probability of the student and the candidate item, , is a learnable parameter matrix, , is a bias term.

[0260] The system ranks according to the score and recommends the top N items with the highest scores to the student.

[0261] The present embodiment verifies and implements the proposed college student academic activity recommendation method based on a dynamic knowledge graph in the following manner:

[0262] Input and data structure: static data comes from the school student management system (gender, college, major, etc.), and dynamic data comes from the activity management system (competition registration record, activity participation record, etc.).

[0263] Model training: use the cross-entropy loss function:

[0264]

[0265] Use the Adam optimizer, set the learning rate to 0.001, and train the model in the GPU environment, support batch data input, represents the real participating academic activity, represents the student interest academic activity predicted by the model.

[0266] Inference stage: for each student, input the latest behavior sequence, output the interest vector through the student interest extraction layer, match with all candidate item embeddings, and generate a recommendation list.

[0267] Compared with the existing recommendation method based on static rules or collaborative filtering, the technical scheme of the present invention has the following effects:

[0268] Improved recommendation accuracy: by introducing dynamic knowledge graph and time series modeling, the evolution trend of student interest is accurately captured, and the recommendation accuracy is improved by about 12% to 18%;

[0269] Modeling efficiency is improved: the modular embedded learning structure reduces the training time by about 25% under the same computing resources;

[0270] Model interpretability is enhanced: through student interest extraction, it is suitable for academic planning in the education background, and the traceable recommendation basis can realize visual interpretation.

[0271] Strong adaptability: it can be seamlessly expanded to scientific research topics, course recommendation and other scenarios, and has good migration and universality.

[0272] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1.A method for recommending college students' academic activities based on a dynamic knowledge graph, characterized in that, The method comprises: Step S1, static information learning, constructing a static knowledge graph with student static basic information, aggregating neighbor node information by structure learning based on a relational graph convolution network, and learning to obtain a static portrait embedding of each student; Step S2, dynamic information learning, constructing a dynamic knowledge graph containing student-activity-relation-time quadruples by learning the behavior and participation records of students over time, obtaining entity embedding and relation embedding of the dynamic knowledge graph through cascading of distributed embedding learning, and forming learning activity embedding containing dynamic information; Step S3, time information encoding, obtaining time information encoding of learning activities through continuous position encoding, semester cycle encoding, event time difference encoding and time encoding fusion to adapt to the strong correlation between learning activities and time; Step S4, multi-view sequence encoding, inputting the static portrait embedding and comprehensive entity embedding output by the static information learning and the dynamic information learning into a Transformer encoder to obtain a learning activity sequence of each student; Step S5, attention weighted aggregation, performing attention weighted aggregation on the comprehensive entity embedding of other learning activities that the student has not participated in and the learning activity sequence to obtain a student interest embedding vector; Step S6, matching prediction, splicing the student interest embedding vector and the candidate item embedding vector, inputting them into a multi-layer perception MLP and a Sigmoid function for scoring, and outputting a matching probability of the student to the candidate item, and the system ranks according to the score to recommend learning activities for the student. 2.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 1, characterized in that, The step S1 further comprises: Input student basic information knowledge graph , Including entity nodes and semantic relationships thereof , , ,..., Semantic relationships , , ,..., Form a semantic relationship set R; The neighbor information of the entity node is aggregated by using a relational graph convolution network R-GCN to realize semantic enhancement: wherein, denotes the embedding of the layer node , denotes the embedding of the layer node , denotes the embedding of the layer node 's neighbor nodes , The embedding of the first layer is obtained by random initialization, is a learnable parameter matrix for the layer relation r, is a normalization constant, denotes a learnable parameter matrix for the node itself, is a non-linear activation function, denotes the set of nodes that exhibit a relation r with the node , i.e. the set of neighbor nodes; After multiple aggregation updates, the static portrait embedding S of the student node is output. 3.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 2, characterized in that, The step S2 further comprises: Step S21, semantic injection, introducing a large model semantic injection mechanism to map sentence vector embedding of student attribute description and item text introduction to entity embedding space to realize alignment of semantic knowledge and entity embedding; Step S22, relation modulation gating, taking the relation embedding of the previous time step and the relation embedding generated by semantic coding of the text coding model at the current time step as inputs, modulating the update gate and the reset gate of the relation vector through the relation vector, and obtaining the comprehensive relation embedding at the current time step; Step S23, structure learning, semantic-enhanced initial entity embedding and the integrated relational embedding input of the relationship graph convolution network modulated by the relationship of the current time step, so that each entity node aggregates neighbor information to obtain ; Step S24, entity gating, performing entity gating on the initial entity embedding enhanced by semantics and the entity embedding learned by structure learning at the current time step to obtain the comprehensive entity embedding at the current time, and finally obtaining the comprehensive entity embedding of any entity node through cascading of multiple time steps. 4.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 3, characterized in that, The step S21 further comprises: Using a pre-trained large language model to extract student attribute descriptions, sentence vector embeddings of project introduction texts , and finally obtain semantic enhanced initial entity embeddings by linear projection mapping and initial relationship embeddings : wherein, , respectively represent initial entity embeddings and initial relation embeddings obtained by initializing the encoding, , is a fusion weight, , is a learnable linear projection matrix, , are entity and relation sentence embeddings obtained by encoding the attribute description or item introduction text by a pre-trained large language model. 5.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 4, characterized in that, The step S22 further comprises: At the time step The initial relation embedding enhanced with semantics With the current time step The relation embedding generated by semantic coding using the text coding model As input, the update gate and the reset gate of the relation vector modulated gated GRU are modulated to obtain the time step The comprehensive relation embedding at the moment ; For any time step t, let the input at the current time step be semantically encoded as capture the temporal evolution differences of different relations, the relation embedding obtained by using the text encoding model for semantic encoding, output the integrated relation embedding at the current time step : wherein, denotes the gating vector for relation modulation, denotes the candidate relation embedding for relation modulation, denotes the comprehensive relation embedding of the previous time step, is the relation embedding for the current time step encoded based on the text encoding model, symbol denotes the element-wise product between vectors, , , are the learnable linear transformation matrices corresponding to the update gate, is the bias term, denotes the hyperbolic tangent function, , , are the learnable linear transformation matrices corresponding to the reset gate, is the bias term. 6.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 5, characterized in that, The step S24 further comprises: At time step The initial entity embedding enhanced by semantics The entity embedding learned by structure learning at the current time step Input entity gating, output the comprehensive entity embedding of any one entity node at the current time For any time step t, The comprehensive entity embedding at the current time Is:​ wherein, is a time-controlled weight matrix, is a bias term, is the integrated entity embedding at the previous time step is the entity embedding learned by structure at the current time step, is the integrated entity embedding at the current time step . 7.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 1, characterized in that, The step S3 further comprises: Step S31, continuous position encoding: a continuous position encoding method based on a sine and cosine function is used to encode the timestamp, and for the i-th learning activity in the learning activity sequence, the position encoding is defined as: wherein, represents a continuous positional encoding embedding for the i-th academic activity, , d represents a hidden layer dimension, represents a value of the i-th academic activity in the 2k-th dimension of the positional encoding embedding, represents a value of the i-th academic activity in the 2k+1-th dimension of the positional encoding embedding; Step S32, Semester Cycle Encoding: Encode the timestamp of academic activity i Mapped to multi-level periodic time features: wherein, denotes a timestamp of the academic activity i, denotes a semester number, denotes a week number within the semester, denotes a semester phase; For any periodic temporal feature whose period encoding is defined as: wherein, denotes the semester period encoding embedding of the i-th academic activity, denotes the value of the 2k-th dimension of the semester period encoding embedding of the i-th academic activity, denotes the value of the 2k+1-th dimension of the semester period encoding embedding of the i-th academic activity, denotes the length of the corresponding period; Step S33, event time difference encoding: the time difference between the i-th learning activity and the previous learning activity is defined as: The time difference is expanded by a multi-scale cosine function: wherein, is the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, M being the number of time scales, denotes the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, the inter-event time difference encoding at the mth time scale being defined as: wherein, represents the value of the 2kth dimension of the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, represents the value of the 2k+1th dimension of the inter-event time difference encoding embedding of the ith academic activity at the mth time scale, is a normalization factor corresponding to the mth time scale; Step S34, time code fusion: the final time code of the i-th learning activity is expressed as: ； wherein, denotes the time-encoding fusion embedding of the i-th academic activity. 8.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 1, characterized in that, The step S4 further includes: Let the comprehensive entity embedding of each academic activity entity Fusing time encoding information specific to a student : ； wherein, For academic activity entity embedding, i takes values from 1 to n, and then the academic activity sequence of each student in the time dimension is taken as the input sequence The first element S of the sequence is the student static portrait embedding from static information learning; For each input sequence perform multi-view augmentation, and encode the augmented view sequence and the original sequence respectively by Transformer to get and Then, the InfoNCE loss is used to close the embedding distance of the same sequence under the original view and the augmented view, and to close the distance of the embedding of other sequences, and the InfoNCE loss is: in, Indicates InfoNCE loss, This represents the interest embedding obtained after the original input sequence has been processed by the Transformer encoder. This represents the enhanced view interest embedding obtained by the Transformer encoder after applying multi-view enhancement to the same original input sequence. This represents the result of randomly sampling all academic activities. One negative sample, Indicates the first Interest embeddings obtained from negative samples through a Transformer encoder It is the temperature coefficient. Represents the similarity function; Embedding the academic activity entity finally Fusing the personalized interests of students, processed by the Transformer encoder, to obtain the academic activity sequence of each student . 9.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 1, characterized in that, The step S5 further includes: embed the comprehensive entity that defines the other academic activity the student has not yet participated in To embed the candidate item, compute the attention weights between the candidate item and each activity in the input sequence: wherein n represents the number of historical academic activities that the student has participated in, representing a candidate item a relevance score function for the i-th historical academic activity; Based on the attention weights The historical academic activity vectors are weighted summed to obtain a student interest embedding vector of the same type as The historical academic activity vectors are weighted summed to obtain a student interest embedding vector of the same type as wherein, is a historical academic activity vector, student interest embedding vector represents a student's dynamic interest representation in a candidate item context. 10.The college student academic activity recommendation method based on dynamic knowledge graph according to claim 1, characterized in that, The step S6 further includes: Embedding student interest vectors Embedding with candidate items Concatenating: concatenating vectors inputting a multilayer perceptron: wherein, is a Sigmoid function, and the output denotes the matching probability of a student with a candidate item, , is a learnable parameter matrix, , is a bias term; The system recommends the top N items with the highest scores to the student according to the scoring ranking.

Images