Entrepreneurship guidance interaction method for smart campus

Abstract

The invention discloses an entrepreneurship guidance interaction method for a smart campus, relates to the technical field of smart campuses, and aims at entrepreneurship guidance of the smart campus through the following contents: 1, multi-source data fusion and preprocessing, and labeling integration of academic, community, competition and other information; 2, mining explicit and implicit features through clustering and deep learning, and constructing a multi-dimensional portrait; 3, updating portraits and weights in real time by an online learning mechanism; 4, reinforcement learning driving personalized recommendation, and dynamic optimization pushing through positive and negative feedback and a trial and error process; 5, multi-edge interaction and cooperative enhancement of teachers and students are carried out, and teacher portraits are brought into a group feedback closed loop, so that accurate matching and joint guidance are promoted; through deep feature extraction, continuous iterative updating and multi-role cooperation, the ability of capturing potential entrepreneurship demands of students is effectively improved, reasonable resource allocation is enhanced, the teaching practice effect is guaranteed, and a solid support is provided for refined and dynamic services of entrepreneurship education in smart schools.

Description

Technical Field

[0001] The present invention relates to the technical field of smart campuses, and in particular to an entrepreneurial guidance interaction method for smart campuses. Background Art

[0002] As universities increasingly prioritize innovation and entrepreneurship education, students often participate in a variety of academic and practical activities during their time on campus, including club activities, competitions, incubation programs, and cross-departmental collaborations. These activities often generate large amounts of diverse data, encompassing everything from academic performance and course attendance to club positions, competition awards, library checkout lists, and even interactions in online forums. Efficiently integrating this vast and diverse data source and creating accurate student profiles would make it possible to timely recommend appropriate tutoring courses, incubation resources, and mentors to students at different stages of their entrepreneurial journey, supporting them throughout the entire process from interest discovery to practical implementation. However, existing campus information systems are often fragmented across different departments or platforms and lack a unified data interface that can vertically link students' multi-dimensional behaviors, leading to frequent information silos. Similarly, the professional backgrounds and coaching expertise of teachers (or mentors) are scattered across multiple systems, making it difficult to quickly retrieve and match information with student needs at critical moments. It can be seen from this that if personalized entrepreneurial education is to be carried out on a large scale, it is necessary to build deeply interconnected data pipelines and dynamic portrait models in a smart campus environment, break down departmental data barriers, and continuously iterate portrait accuracy based on actual interactive feedback, providing more appropriate guidance and cooperation channels for both students and teachers.

[0003] The core technical problem that the present invention needs to solve is: how to build an adaptive and evolving user portrait model and achieve accurate matching under the premise that multi-source data is highly dispersed and student needs change rapidly, otherwise there will be the risk of data distortion and resource mismatch.

[0004] Specifically, a student's innovative potential and entrepreneurial interest are not solely reflected in their academic performance and major affiliation, but are also largely reflected in information such as club activities, interdisciplinary competitions, and presentations at online seminars. If data collection and processing are not comprehensive, an objective portrayal of students' comprehensive abilities and potential interests will be lacking. Furthermore, student needs often evolve with academic stages and personal cognition. Without a dynamic update mechanism, the preference labels or ability scores in the profiles will gradually become disconnected from their actual status, resulting in mismatched recommendations. The cold start problem is also unavoidable: new students or potential entrepreneurs with no previous activity history may not receive appropriate resources or timely mentoring due to a lack of historical behavioral data.

[0005] What is more serious is that if the platform only relies on one-way information flow or overly static label classification, students will often receive resources that do not match their actual abilities or interests, and teachers will find it difficult to accurately find student groups with relevant needs or cooperation potential. This will lead to waste of resources, missed opportunities and even a decrease in platform stickiness, making it difficult for the smart campus entrepreneurial environment to form an effective closed loop.

[0006] To this end, the present invention provides an entrepreneurial guidance interaction method for smart campus. Summary of the Invention

[0007] (1) Technical problems solved

[0008] In response to the shortcomings of the existing technology, the present invention provides an interactive method for entrepreneurial guidance in smart campuses, which is achieved through the following: first, multi-source data fusion and preprocessing, labeling and integration of academic, club, competition and other information; second, clustering and deep learning to mine explicit and implicit features and construct multi-dimensional portraits; third, online learning mechanisms to update portraits and weights in real time; fourth, reinforcement learning drives personalized recommendations, and dynamic optimization and push through positive and negative feedback and trial and error processes; fifth, multi-lateral interaction and collaborative reinforcement between teachers and students, incorporating teacher portraits into the group feedback loop to promote accurate matching and joint guidance. Through deep feature extraction, continuous iterative updates and multi-role collaboration, the ability to capture students' potential entrepreneurial needs is effectively improved, the rational allocation of resources is strengthened, and the effectiveness of teaching practice is guaranteed, providing solid support for the refined and dynamic services of entrepreneurial education in smart campuses; thereby solving the technical problems recorded in the background technology.

[0009] (2) Technical solution

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0011] The interactive method for entrepreneurial guidance in smart campuses includes: when detecting that student multidimensional data (such as club records, competition performance, and book borrowing) is fragmented, the original information is cleaned, deduplicated, and uniformly labeled, and key information is mapped to a unique identifier using structured fields to generate a multi-source preprocessed dataset;

[0012] After obtaining the cleaned and labeled multi-dimensional features, clustering algorithms and deep learning models are used to perform parallel feature extraction on student behavior sequences and text information, and cluster group portraits are generated through NLP semantic analysis;

[0013] If a student's profile or behavior is significantly different from the original cluster, online learning and local iteration are performed on the newly generated click, browse, and feedback events based on streaming data processing, and the corresponding profile weights and cluster affiliations are updated, forming an adaptive closed loop of profile-recommendation-feedback-re-profile.

[0014] After completing the instant update of the user profile, the reinforcement learning model is called to conduct trial and error iteration based on the recommended actions and user feedback. The Q value and policy parameters are modified based on the exponential amplification or attenuation mechanism of positive and negative feedback, and the optimal push solution is output to continuously enhance the efficiency of personalized resource matching.

[0015] If it is identified that teacher portraits and multilateral collaboration need to be introduced, an identifier will be established for the teacher side and feature extraction and group clustering will be performed in parallel with the student portraits. The multilateral collaborative extended data set will be updated collaboratively through teacher-student similarity and group feedback, thereby achieving multi-party discussion, mutual evaluation and two-way matching.

[0016] Preferably, various data sources related to student entrepreneurship on and off campus are identified, and distributed crawling and database interface docking are used to automatically crawl the above multi-source data, and each piece of data is indexed and mapped with a unique ID; for each collected record, key information fields are extracted and basic metadata is marked;

[0017] Preferably, the obtained preliminary meta-information tags are aggregated into the same student dimension according to ID, and the credibility of each record is calculated. i ), the specific formula is as follows:

[0018]

[0019] Where: rec i It represents the set of records of the same student in all sources about the same data entity; S is the number of comparable data sources; Indicates the record value corresponding to the student at source s; It represents the reference value obtained after preliminary aggregation of the same field of student i;

[0020] DIST(x,y) represents the difference measure function between any two record segments x and y, α s is the credibility weight of source s, β is the scaling factor of the difference;

[0021] If the reliability of some records Υ(rec i ) is far below the given credibility threshold Υ min , conduct manual review or automatic elimination;

[0022] After ensuring data consistency, we generate refined labels for each field and append them to the corresponding student records to form a multi-source preprocessed dataset.

[0023] Preferably, the label vector ∑(u j); convert the plain text label into a numerical feature dimension, set the number of clusters to K; introduce the similarity function Ψ(u j ,Ω k ) to represent the student u and the kth cluster center vector Ω k The similarity between them, similarity function Ψ(u j ,Ω k )The larger the value, the more likely the student u j More inclined to belong to cluster k; specifically as follows:

[0024] Ψ(u j ,Ω k )=Γ×exp{-Δ·D(∑(u j ),Ω k )}

[0025] Where: ∑(u j ) is the eigenvector of student u, Ω k is the kth cluster center vector; D(·,·): used to measure ∑(u j ) and the cluster center vector Ω k The distance function; Γ is the basic similarity coefficient; Δ is the distance scaling coefficient;

[0026] For each student j , according to the maximum similarity function Ψ(u j ,Ω k ) rule to determine its belonging cluster k * :

[0027]

[0028] After completing the cluster assignment, the cluster center vector Ω is assigned according to the assignment result. k Update; iterate until the cluster allocation result converges or reaches the preset maximum number of iterations, until the user group portrait is formed; cluster clusters and their cluster center vectors Ω k The feature summary information is output as a cluster group portrait;

[0029] Preferably, a sequence of operations or events related to the time axis is extracted from a multi-source preprocessed data set and converted into a fixed-length or variable-length sequence; a deep representation learning is performed on the behavior sequence to obtain the student's implicit state vector θ(u j );Extract text semantic vector Ξ(u j ); Combine the obtained cluster information with the implicit state vector Θ(u j ) and text semantic vector Ξ(u j ) to fuse and obtain the user portrait vector Po(u j ):

[0030] Po(u j )=[∑(u j )||Θ(u j )||Ξ(u j )]

[0031] Where: || represents the vector concatenation operation, Σ(u j ) is the label vector, Θ(u j ) is the implicit state vector output by the sequence model, Ξ(u j ) is the text semantic vector;

[0032] The user portrait vector Po(u j ) Summarize and generate a multi-dimensional portrait dataset;

[0033] Preferably, the streaming data processing module is deployed in the platform interaction layer to collect student u j In the entrepreneurial data on the platform, whenever a new key behavioral event is captured, a behavioral event data is generated;

[0034] To prevent frequent or invalid updates, trigger thresholds and rules are set. If the trigger conditions are met, real-time scheduling of online learning and portrait correction is performed on the user portrait vector Po(u j ) performs iterative operations; before the trigger rule is established, each behavior event data is temporarily stored in a fast cache area, merged and packaged into a real-time incremental stream;

[0035] Preferably, when the update is triggered, the deep feature extraction model is called to extract the student u j The newly added behavior sequence and text data are locally iterated, and the online gradient update strategy is adopted for the sequence model:

[0036] The newly added behavior segments are regarded as mini-batches, and the implicit state vector Θ(u j ), perform incremental updates. For text parsing models, a small amount of iterative training can be performed for new keywords or text fragments to correct the text semantic vector Ξ(u j ) in vector representation;

[0037] When obtaining the new implicit state vector Θ(u j ) or text semantic vector Ξ(u j ), and the original user portrait vector Po(u j ) to fuse and generate a new user portrait vector Po new (u j ), the new user portrait vector Po new (u j ) is written into a new data object to correct the data set in real time to represent the latest portrait status of the student.

[0038] If the new user portrait vector Po new (u j ) has deviated significantly from the original corresponding cluster in the feature space, and local clustering fine-tuning is performed on the student or the cluster to which it belongs;

[0039] Preferably, the state S(t) can be represented by the user portrait vector Po(u j ) is combined with the recent user behavior context; the action A(t) corresponds to the set of entrepreneurial guidance or resource push provided to students. In each cycle, one or more recommended actions A(t) are selected and a recommended action record is generated;

[0040] When students j Feedback is provided on the platform for the recommendations of this cycle, corresponding feedback information records are generated, and the immediate reward r(t) is automatically calculated. The aforementioned state, action, and reward information are encapsulated into the reinforcement strategy environment information;

[0041] The state S(t) in the enhanced policy environment information refers to the user portrait vector Po(u j )'s latest image information.

[0042] Preferably, a policy function π(S) is learned in reinforcement learning, which inputs the current state S(t) and outputs the selected action A(t). To facilitate online updating, a Q function approximation based on a deep network can be introduced:

[0043] Q(S(t),A(t))≈Φ(W,S(t),A(t))

[0044] Where: Q(·,·) represents the state-action value function; W represents the learnable weights in the deep network; Φ(·) represents the result of the network forward propagation, which outputs the value estimate of the action in a specific state;

[0045] In the recommendation strategy, the exploration mechanism is introduced:

[0046] If the exploration probability ∈ is triggered at a certain stage, a possible recommended action A(t) is randomly selected to discover potential new resources or new areas; if the exploration is not triggered, The optimal action estimated using the current Q function.

[0047] Whenever a new immediate reward r(t) is obtained, the Q function is corrected online and gradient descent is performed according to the following formula to update the learnable weight W, so that the strategy π(S) continuously approaches the optimal solution;

[0048] Preferably, reinforcement learning updates the Q function after obtaining the immediate reward r(t), and synchronously transmits the user's explicit feedback on the recommended content back to the online learning;

[0049] When a resource is highly recognized by users, the user portrait vector Po(u j ) within the user group; if negative reviews appear multiple times, the corresponding weight is reduced, thus forming a multi-round loop of profiling-recommendation-feedback-re-profiling;

[0050] If some students u j For recommendations that consistently deviate from the mainstream behavior of the same group during multiple rounds of interaction, we use the clustering online fine-tuning logic to shift them to clusters that better match the group profile.

[0051] The recommended actions, immediate rewards, Q-function update information, etc. recorded in all reinforcement learning stages are organized and output as a reinforcement learning recommendation dataset.

[0052] Preferably, a tutor data table is added to the platform database to extract the tutor's guidance characteristics;

[0053] After processing missing values ​​or conflicting information during cleaning and normalization, all available fields are integrated into the new tutor basic dataset object, and the tutor data is clustered and feature mined; a new tutor portrait group is introduced. Characterize the typical tutor group image, extract the vector of the teacher's achievements, papers and teaching evaluation text, and form the tutor's text feature Ξ(t p ), and then combine the quantitative labels to construct the mentor portrait vector To(t p ); clustering information of teachers and tutor portrait vector To(t p ) are integrated into a new data object, the teacher-student collaboration dataset;

[0054] Preferably, when having a mentor portrait vector To(t p ) and the student portrait vector Po(u j ) and define the matching degree Ψ ts (j,p) to evaluate teacher-student matching:

[0055]

[0056] In the formula: Po(u j ) is the portrait vector of the jth student; To(t p ) is the portrait vector of the p-th tutor; Γ ts is the teacher-student similarity benchmark coefficient; Δ ts is the distance attenuation coefficient, used to control Po(u j ) and To(t p ), the influence of the difference on the similarity; M ts is a learnable symmetric positive definite matrix; α is the power exponent;

[0057] Matching degree Ψ tsThe larger the value of (j,p), the more compatible the student j is with the teacher p in terms of their profile dimensions. The teacher can be recommended as the student's entrepreneurial mentor or invited to join the project team.

[0058] Add online seminar rooms and discussion groups, incorporate all discussion and sharing records into new behavioral event data, and use clustering information and the matching degree Ψ obtained in the current step TS , automatically assigning better teacher-student groups or project teams;

[0059] If it is detected that the needs of some students in the group are not met, the online update strategy will be triggered and the platform administrator will be prompted.

[0060] Preferably, if it is detected that the teacher-student discussion group or the student mutual evaluation group has a highly positive evaluation or a negative evaluation of a resource or tutor, the weights of the portraits of all students and tutors are fine-tuned synchronously, and the following group feedback factor ω is defined: g :

[0061] ω g =κ g exp(μ gpos ·f gpos -μ gneg ·f gneg )

[0062] Where: κ g is the group feedback benchmark coefficient, which is greater than 0; f gpos Positive consensus within the group; f gneg is the negative consensus within the group; μ gpos 、μ gneg are the amplification weights of positive and negative feedback respectively;

[0063] When the group feedback factor ω g If the score is higher than expected, the label of the corresponding field of fit between the student and the tutor will be strengthened, or the Q value in the reinforcement learning strategy will be adjusted accordingly.

[0064] After detecting significant group feedback, a portrait of the tutor group can be created. Or mentor portrait vector To(t p ) also performs online updates, and changes the mentor portrait vector To(t p ) and the corresponding student portrait vector Po(u j ) are written into the multilateral collaborative extension dataset.

[0065] (3) Beneficial effects

[0066] The present invention provides an interactive method for entrepreneurship guidance in smart campuses, which has the following beneficial effects:

[0067] Through distributed crawling and ETL technology, multi-dimensional information inside and outside the campus (such as club positions, competition awards, and book borrowing) is uniformly calibrated under the student's unique identifier, which can reduce the image distortion caused by information islands. It adopts automatic correction of outliers and missing values ​​and standardized processing of different types of data (text, numbers), paving the way for data quality in subsequent steps.

[0068] Combining cluster analysis and deep learning feature extraction provides a multi-dimensional depiction of student profiles, combining group commonalities with individual differences. By clustering cleaned data, subgroups of students with similar entrepreneurial needs or capabilities can be quickly identified. Leveraging time series models like RNN or Transformer, and NLP parsing text content, the vast amount of semantic information generated by students during platform interactions (such as business plans and project questions) can be refined into key high-order vector features. Single-dimensional label profiles are expanded into three-dimensional representations that integrate explicit behaviors, latent preferences, and textual sentiment analysis, enabling educational administrators to better understand students' true needs at the teaching level.

[0069] With the help of streaming data processing modules and online learning algorithms (online gradient descent, online random forest, etc.), continuous correction of student portraits can be achieved: when new user behaviors or extreme feedback (such as strong rejection of a resource) are monitored, local iteration of the model is triggered, allowing the portrait to maintain high fidelity over time.

[0070] Based on the positive and negative feedback (clicks, ignores, favorites, negative reviews, etc.) given by students to each recommended content, a value assessment is conducted, and optimization is iteratively sought. Positive and negative feedback are amplified or reduced using power functions and exponential decay respectively. During gradient updates, it accurately distinguishes between situations of high recognition or serious dissatisfaction with the platform, thereby calibrating the focus of resource push more quickly.

[0071] By incorporating faculty into the profiling system and combining the clustering and deep feature analysis methods from Step 2, we construct a mentor profile that mirrors the student profile, encompassing dimensions like professional experience, teaching style, and project preferences. This similarity measurement enables precise two-way matching: helping students find mentors who best meet their needs while enabling mentors to focus on students whose strengths align with their own, truly fostering a mutually beneficial relationship between mentor and student.

[0072] The group feedback factor is used to assess the recognition of a mentor or resource at the group level, and this is weighted and integrated with individual student feedback. When disagreements arise within the group, they are not blindly averaged. Instead, subgroups are divided based on dispersion or minority opinions, reflecting inclusiveness and adaptation to diverse needs.

[0073] It not only solves common problems such as cold start, distorted portraits, and missing mentor roles, but also comprehensively improves the efficiency and depth of interaction in entrepreneurial guidance through two-way portraits, real-time updates, and multilateral collaboration. It has obvious promoting value for the innovative development of smart campuses and teaching reforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 The figure is a flow chart of the interactive method for entrepreneurship guidance in smart campus of the present invention. DETAILED DESCRIPTION

[0075] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0076] See also Figure 1 The present invention provides an interactive method for entrepreneurial guidance in a smart campus, comprising:

[0077] Step 1: When it is detected that student multidimensional data (society records, competition performance, book borrowing, etc.) is fragmented, the original information is cleaned, deduplicated, and uniformly labeled. The key information is mapped to a unique identifier using structured fields to generate a multi-source preprocessed dataset.

[0078] The step 1 includes the following:

[0079] Step 101: Multi-source data collection and preliminary indexing

[0080] Identify various data sources related to student entrepreneurship on and off campus, including but not limited to academic system databases, club activity and competition records, entrepreneurial project application and progress logs, library loan systems, and platform browsing and interaction logs. Utilize distributed crawling and database interface integration (e.g., distributed data processing frameworks based on Spark or Flink) to automatically crawl and pull these multi-source data, and index and map each piece of data with a unique ID.

[0081] For each collected record, extract key information fields (such as student number, major, community, position, project name, etc.), and perform basic metadata tagging (such as timestamp, source system identifier, data type, etc.) to form preliminary metadata tags;

[0082] When used, a unified ID enables cross-system data association, providing a traceable index foundation for subsequent deep feature extraction and profiling. Leveraging distributed crawling and integration, the system significantly increases the breadth and efficiency of data collection, enabling timely coverage of student behaviors and records across diverse scenarios. Preliminary metadata tagging across multiple source systems significantly enhances the operability of subsequent data cleaning and integration.

[0083] Step 102: Data consistency check and label preprocessing

[0084] Aggregate the obtained preliminary meta-information tags into the same student dimension according to ID, and calculate the credibility of each record based on the following data consistency scoring formula: i ), in order to distinguish the importance of different sources and fields, a configurable weight coefficient is introduced. The specific formula is as follows:

[0085]

[0086] Where: rec i It represents the set of records of the same student in all sources about the same data entity; S is the number of comparable data sources; Indicates the record value corresponding to the student at source s; represents a reference value obtained after preliminary aggregation of the same field of student i (e.g., a reference text or value selected based on the highest frequency of occurrence or the most credible source);

[0087] DIST(x,y) represents the difference measure function between any two record segments x and y. Different comparison methods can be defined here based on the data type, such as string edit distance or embedding vector cosine distance.

[0088] α s is the credibility weight of source s, which needs to be configured during the initialization of this step and has a value between 0 and 1; β is the scaling factor of the difference, which has a value greater than 0. The greater the difference between the record and the reference value, the closer the exponential term is to 0;

[0089] γ(rec i ) value is higher, indicating that the record is more consistent with the reference value and has higher credibility;

[0090] If the reliability of some records γ(rec i ) is far below the given credibility threshold Υ min , then it is judged that the record may be an input error or an extreme outlier in the student dimension. Such abnormal records can be further manually reviewed or automatically eliminated to ensure data reliability during subsequent labeling processing;

[0091] After ensuring data consistency, detailed labels are generated for each field, such as academic feature labels, community role labels, entrepreneurial project labels, and individual behavior labels. This process uses structured rules or text feature extraction to map entries from different data sources to a unified label system and attach these labels to the corresponding student records. All records that have passed consistency verification and completed labeling constitute a new fused dataset, recorded as a multi-source preprocessed dataset, which then outputs data objects that can be read in subsequent steps:

[0092] When using, the reliability Υ (rec i ) can automatically measure and filter out the most representative or credible records when there is inconsistency in multi-source data, thereby effectively reducing the interference of redundant and erroneous data on the accuracy of subsequent portraits; multi-dimensional labeling is completed in this stage, providing a unified and clear feature entry for subsequent portrait modeling, and avoiding chain contamination of low-quality data in the future by eliminating or reviewing abnormal records.

[0093] Step 2: After obtaining the cleaned and labeled multi-dimensional features, clustering algorithms and deep learning models are used to perform parallel feature extraction on student behavior sequences and text information, and cluster group portraits are generated through NLP semantic analysis;

[0094] The second step includes the following:

[0095] Step 201: Multi-dimensional clustering and user group profiling

[0096] Read the label vector and related numerical features (such as professional information, project experience, activity participation, etc.) of each student from the multi-source preprocessed dataset. The label vector is recorded as ∑(u j ), where u j Represents the vector expression of the integration of all discrete or continuous features corresponding to the j-th student;

[0097] A mapping table and internal coding rules are used to convert plain text labels (such as community positions, competition awards, etc.) into numerical feature dimensions that can be processed by clustering algorithms.

[0098] Select K-Means++ or hierarchical clustering algorithms as the basic framework, and set the number of clusters to K (which can be preliminarily set by business needs or strategies such as silhouette coefficient); introduce the similarity function ψ(u j ,Ω k ) to represent the student u and the kth cluster center vector Ω k The similarity between them, similarity function Ψ(u j ,Ω k )The larger the value, the more likely the student u j More likely to belong to cluster k, as follows:

[0099] Ψ(u j ,Ω k )=Γ×exp{-Δ·D(∑(u j ),Ω k )}

[0100] Where: ∑(u j ) is the eigenvector of student u, Ω k is the kth cluster center vector;

[0101] D(·,·): used to measure ∑(u j ) and the cluster center vector Ω k The distance function (which can be Euclidean distance, cosine distance, etc., depending on the type of feature dimension);

[0102] Γ is the basic similarity coefficient, which is used to control the scaling of the overall similarity metric and is greater than 0; Δ is the distance scaling coefficient, which is used to adjust the sensitivity of distance to similarity attenuation and is greater than 0;

[0103] During initialization, several cluster center vectors Ω can be randomly selected or selected based on multiple iterations. k As the initial center;

[0104] For each student j , according to the maximum similarity function Ψ(u j ,Ω k ) rule to determine its belonging cluster k * :

[0105]

[0106] After completing the cluster assignment, the cluster center vector Ω is assigned according to the assignment result. k Update (for example, in K-Means, take all ∑(u j ), or distance-based merging strategy in hierarchical clustering);

[0107] Iterate until the cluster assignment results converge or the preset maximum number of iterations is reached, until each cluster represents a group of students with similar entrepreneurial needs or ability characteristics, forming a user group portrait;

[0108] Cluster ID and its cluster center vector Ω k Together with the characteristic summary information (such as the average distribution of the cluster in terms of community participation and project focus areas), it is output as a cluster group portrait; at the same time, for each student u jRecord the cluster ID to which it belongs, so that it can be quickly indexed and referenced in the dynamic update or reinforcement learning recommendation phase (for example, inference can be made based on the portraits of similar groups of students during cold start);

[0109] When used, by defining the similarity function Ψ(u j ,Ω k ) This exponential similarity measure is more flexible than the traditional simple distance comparison method. It can perform attenuation control for different distance measures and improve the reliability of outliers and feature noise. The clustering results can not only divide users into significant categories, but also extract common characteristics of the group (i.e., user group portraits), providing a direct reference for rapid cluster-level migration in dynamic updates.

[0110] Step 202: Deep feature extraction and soft skills characterization

[0111] To capture the temporal behavior characteristics of students in platform or campus activities (e.g., weekly frequency of checking project resources, monthly trend in number of competitions participated in, etc.), we extract timeline-related operation or event sequences from multi-source preprocessed datasets and convert them into fixed-length or variable-length sequences.

[0112] Use RNN (such as LSTM) or Transformer-based sequence model to perform deep representation learning on the behavior sequence and obtain the student's implicit state vector θ(u j ); Implicit state vector Θ(u j ) can be considered as a code of the evolution trajectory of students' entrepreneurial interests, which can be used to help determine their possible future interest shifts or skill improvement directions;

[0113] For text information such as business plans and entrepreneurial problem descriptions submitted by students, a custom Embedding model (e.g., secondary fine-tuning based on BERT) is used to extract semantic vectors, and the keyword preferences and sentiment tendencies reflected in the text are mapped to a high-dimensional vector space to form a text semantic vector Ξ(u j ); by comparing with labeled information, we can further identify the soft skills presented by students in their language expression, such as leadership and stress resistance: in the internal feature system, these soft skill dimensions are recorded and quantified in specific dimensions to form a richer portrait description of the students.

[0114] The obtained cluster information (especially cluster group portrait) is combined with the implicit state vector Θ(u j ) and text semantic vector Ξ(u j ) to fuse and obtain the user portrait vector Po(u j ):

[0115] Po(u j )=[∑(u j)||Θ(u j )||Ξ(u j )]

[0116] Where: || represents vector concatenation operation, ∑(u j ) is the label vector, Θ(u j ) is the implicit state vector output by the sequence model, Ξ(u j ) is the text semantic vector;

[0117] Po(u j ) is the final output, representing student u j The global portrait description at the multi-dimensional level (label features, temporal behavior, text semantics, etc.) is converted into the user portrait vector Po(u j ) Aggregate and generate a new data object, which is recorded as a multidimensional portrait dataset;

[0118] When used, the clustering results are combined with the time series features and text semantic vectors of deep learning, overcoming the limitations of relying solely on static labels or single feature dimensions, and achieving multi-angle characterization of user portraits; the user portrait vector Po(u j ) takes into account both the commonality at the group level (from cluster information) and the behavioral evolution at the individual level (from the implicit state vector Θ(u j )) and language style (from text semantic vector Ξ(u j )) This lays a high-dimensional and semantically explanatory data foundation for subsequent dynamic updates and personalized recommendations;

[0119] Through deep models such as RNN or Transformer, it is possible to represent, learn, and predict the potential evolution paths of students' interests or skills, providing the platform with more forward-looking entrepreneurial guidance strategies at different time points; it defines a bidirectional deep representation learning structure for sequential behavior and text semantics, and brings the soft skills and interest evolution that are difficult to quantify in traditional user portrait systems into a computable scope, filling the gaps in static portraits.

[0120] Step 3: If a student's profile or behavior is significantly different from the original cluster, online learning and local iteration are performed on the newly generated click, browse, and feedback events based on streaming data processing. The corresponding profile weight and cluster affiliation are updated, forming an adaptive closed loop of profile-recommendation-feedback-re-profile.

[0121] The step three includes the following:

[0122] Step 301: Streaming data capture and image update triggering

[0123] By deploying a streaming data processing module (based on Flink, KafkaStreams or other real-time streaming systems) at the platform interaction layer, students’ uj The platform collects entrepreneurial data such as visit tracks, click records, search keywords, and the progress of potential entrepreneurial projects. Whenever a new key behavioral event is captured (such as browsing a marketing page multiple times in a row or submitting a new business plan draft), a behavioral event data is generated.

[0124] To prevent frequent or invalid updates, set trigger thresholds and rules, for example:

[0125] If L consecutive behavioral events all point to the same new preference topic (such as investment and financing learning materials), then student u j A significant increase in interest in the topic triggers an update;

[0126] If abnormal behavior is detected that does not match the cluster group profile described above (e.g., students belonging to the marketing operations group frequently visit the technology development zone), an update can be triggered to re-evaluate their profile;

[0127] If the triggering conditions are met, real-time scheduling of online learning and portrait correction is performed on the user portrait vector Po(u j ) performs iterative operations;

[0128] Before triggering a rule, each behavioral event data is temporarily stored in a fast cache, merged, and packaged into a dedicated real-time incremental stream for unified call upon triggering. Only when an update is triggered is the relevant event incorporated into the calculation input required for the next step, thus reducing the pressure on the online learning algorithm and ensuring efficient execution.

[0129] During use, the streaming data processing module tracks changes in student behavior in a fine-grained manner. This allows for rapid capture of shifts in interest or the emergence of new needs in campus entrepreneurship scenarios. It also sets thresholds for determining whether to update the student profile, effectively reducing frequent and worthless updates caused by noise or sporadic operations, ensuring stability in subsequent online learning.

[0130] Traditional educational information systems often rely on batch or scheduled updates. Combining streaming monitoring with threshold triggering better meets the ever-changing real-time needs of students' entrepreneurial needs. Incorporating cluster group profiling into the judgment criteria can distinguish between abnormal behavior deviations and normal behavior continuations at the update trigger level, avoiding a one-size-fits-all threshold strategy.

[0131] Step 302: Online learning and image correction

[0132] When the update is triggered, the online version of the deep feature extraction model (such as RNN / Transforme or NLP parsing model) is called to perform the task on the student u j The newly added behavior sequence and text data are locally iterated;

[0133] For sequence models (such as LSTM), an online gradient update strategy is adopted:

[0134] The newly added behavior segment is regarded as a mini-batch, and the hidden state vector Θ(u j ) is incrementally updated according to the loss function. For text parsing models (such as custom Embedding fine-tuning), a small number of iterative training can be carried out for new keywords or text segments to correct the vector representation in the text semantic vector Ξ(u j );

[0135] After obtaining the new hidden state vector Θ(u j ) or text semantic vector Ξ(u j ), it is fused with the original user portrait vector Po(u j ), and the following adaptive weighted update formula can be defined to generate a new user portrait vector Po new (u j ):

[0136] Po new (u j ) = (1 - k)Po old (u j ) + kΦ[Θ [[ID=3P]] new (u j ), Ξ new (u j )]

[0137] In the formula: Po old (u j ) refers to the old user portrait vector saved in the multi-dimensional portrait dataset last time; Θ new (u j ), Ξ new (u j ) respectively represent the latest hidden state vector and text semantic vector obtained after online learning;

[0138] Φ(·) is a feature fusion operation (such as direct splicing or non-linear transformation); κ is the portrait correction rate (0 < k ≤ 1), which determines the fusion ratio of new information and the old portrait;

[0139] After the update is completed, the new user portrait vector Po new (u j ) is written into the new data object to real-time correct the dataset to represent the latest portrait state of the student. ]

[0140] If the new user portrait vector Po new (u j ) has deviated significantly from the original corresponding clustering cluster in the feature space (the similarity function Ψ(u j , Ω k) measure), perform local clustering fine-tuning on the student or the cluster he belongs to:

[0141] Recalculate student u j and the cluster center vector Ω k The similarity of the old and new clusters is calculated and the decision of whether to migrate to a new cluster is made based on the maximum similarity principle. The cluster centers of the migrated old and new clusters are updated to maintain the availability of the cluster group portrait.

[0142] If the student u j If the profile characteristics of a student are significantly different from any existing clusters, you can set a threshold based on business needs and create a new emerging interest cluster to cover students in the transition period or the early cold start period.

[0143] When used, through online gradient update or local iterative training, it is possible to capture the student’s latest interest change or skill improvement in a timely manner, so that the implicit state vector Θ(u j ) and text semantic vector Ξ(u j ) Maintain dynamic accuracy; Po new (u j ) integrates old portraits with new feedback, taking into account the continuity of historical behaviors and the rapid response to new behaviors, avoiding the shock caused by a one-time rewrite; online fine-tuning of group allocation can make the clustering structure formed in the second step remain reasonable when facing the evolution of the student group, and continuously adapt to the dynamic changes of the entire campus entrepreneurial ecosystem.

[0144] Step 4: After completing the instant update of the user profile, the reinforcement learning model is used to perform trial and error iteration based on the recommended actions and user feedback. The Q value and policy parameters are modified based on the exponential amplification or attenuation mechanism of positive and negative feedback, and the optimal push solution is output to continuously enhance the efficiency of personalized resource matching.

[0145] The step 4 includes the following contents:

[0146] Step 401: Reinforcement Learning Environment Definition and State Construction

[0147] The state S(t) can be represented by the user portrait vector Po(u j ) is combined with the recent user behavior context, where t represents the time step or recommendation cycle; the action A(t) corresponds to the entrepreneurial guidance or resource push set provided to students, such as recommending an entrepreneurial course or inviting them to participate in an incubation project, etc.

[0148] In each cycle, one or more recommended actions A(t) are selected and a record of recommended actions is generated for subsequent evaluation;

[0149] When students jFeedback is provided on the platform for recommendations in this cycle (such as clicks, favorites, ignores, negative reviews, etc.), corresponding feedback information records are generated, and the immediate reward r(t) is automatically calculated. The following exponential reward function is designed:

[0150]

[0151] Where: Ω is the positive return benchmark coefficient, which can be set by the platform administrator to control the overall scale of the reward value;

[0152] f cos (t) is the frequency of statistical positive interactions (such as clicks, collections, and positive comments); f neg (t) is the frequency of statistical negative interactions (such as ignoring, bad reviews, and quick exits); ζ is the exponential amplification coefficient of positive feedback, which is greater than 0;

[0153] γ is the exponential attenuation coefficient for negative feedback, and its value is greater than 0;

[0154] Encapsulate the aforementioned state, action, and reward information into enhanced strategy environment information;

[0155] The state S(t) in the enhanced policy environment information refers to the user portrait vector Po(u j ) to ensure that the reinforcement learning algorithm makes decisions based on the most timely user characteristics;

[0156] When used, explicitly set the student portrait vector Po(u j ) and merged with its recent behavioral context into a state, which can more comprehensively reflect the user's preferences and needs at the current moment. The encapsulated reinforcement policy environment information environment object ensures that the previous data object and feedback record can be seamlessly connected during the next algorithm execution, achieving unified management and call.

[0157] Step 402: Strategy Learning and Dynamic Trial and Error

[0158] In reinforcement learning, a policy function π(S) is learned, which inputs the current state S(t) and outputs the selected action A(t). To facilitate online updates, a Q function approximation based on a deep network can be introduced:

[0159] Q(S(t),A(t))≈Φ(W,S(t),A(t))

[0160] Where: Q(·,·) represents the state-action value function; W represents the learnable weights in the deep network; Φ(·) represents the result of the network forward propagation, which outputs the value estimate of the action in a specific state;

[0161] In practical implementation, a Deep Q-Network (DQN) or an Actor-Critic-based network structure can be used, depending on the scale of the campus entrepreneurship application. In the recommendation strategy, to avoid missing out on potentially high-value actions due to over-utilization of the current optimal strategy, it is necessary to introduce exploration mechanisms such as ∈-greedy or UCB (upper confidence bound):

[0162] If the exploration probability ∈ is triggered at a certain stage, a possible recommended action A(t) is randomly selected to discover potential new resources or new areas; if the exploration is not triggered, The optimal action estimated using the current Q function.

[0163] Whenever a new instant reward r(t) is obtained, the Q function is corrected online according to the following formula (taking DQN as an example), that is, the target value y(t) and the estimated value The error between:

[0164]

[0165] Where: η is the discount factor used to balance immediate benefits with future benefits; A′ is the set of actions that can be taken at the next moment;

[0166] Use root mean square error, Huber loss, or other advanced loss functions to perform gradient descent on Φ(W,·) and update the learnable weight W, so that the strategy π(S) continuously approaches the optimal solution.

[0167] W represents the learnable weight matrix or tensor in the deep network, which is used to connect the neurons in the input layer, hidden layer or output layer. Kaiming or Xavier strategy is usually used to make it conform to normal or uniform distribution to keep the gradient stable during initialization. During the training process, the loss function is optimized by combining stochastic gradient descent (SGD) or Adam. Perform backpropagation update: After obtaining user / tutor interaction data from the training samples in each iteration, calculate the target function (such as recommendation error or reinforcement learning Q value difference) Press again The method is iteratively corrected to ensure that the network can adaptively learn deep patterns and output updated user portrait embeddings, value functions or policy decision results under inputs such as multi-source portrait features, temporal behaviors or teacher-student feedback.

[0168] When using it, with the help of strategies such as Q function approximation and ∈-greedy, continuous trial and error is carried out in a dynamic environment, which not only retains the use of existing high-value actions, but also continuously explores potential new strategies, and combines the immediate reward r(t) with the user portrait vector Po(u j) combined with deep Q networks can help deep Q networks or actor-critic models to deeply understand user personality traits and achieve truly personalized entrepreneurial resource recommendations. Online iterative updates can optimize strategies in real time after each round of user feedback, providing more accurate action choices for the next recommendation cycle. The introduction of reinforcement learning and deep Q networks can adapt to diverse resource types and students' long-term growth trajectories, and discover and introduce more cross-disciplinary or cross-field opportunities through exploration mechanisms.

[0169] Step 403: Joint fine-tuning of profile and strategy

[0170] After receiving the immediate reward r(t), reinforcement learning can not only update the Q function, but also synchronously transmit the user's explicit feedback on the recommended content (such as how long the user stays after clicking, whether they continue to view related information, etc.) back to the online learning;

[0171] When a resource is highly recognized by users, the platform can increase the user portrait vector Po(u j ) within the field; if negative comments appear multiple times, the corresponding weight will be reduced, thus forming a multi-round loop of profiling-recommendation-feedback-re-profiling; if some students u j For recommendations that consistently deviate from the mainstream behavior of the same group during multiple rounds of interaction, we call the clustering online fine-tuning logic to transfer them to a more matching group portrait cluster. By observing the student's action selection and reward trends at the reinforcement learning level, we can more accurately judge u j A genuine interest in certain entrepreneurial skills or fields;

[0172] The recommended actions, immediate rewards, Q-function update information, etc. recorded in all reinforcement learning stages are organized and output as a reinforcement learning recommendation dataset, which can be regarded as an extension of the real-time correction dataset in the interactive feedback dimension, containing richer user-system interaction sequences and policy parameter history.

[0173] When in use, the positive and negative feedback signals of reinforcement learning are mapped back to user portraits and group portraits, which not only allows the system strategy to gradually learn which recommendations are suitable for a certain type of student, but also helps the portrait system to continuously improve the representation of students' real abilities and needs; the reinforcement learning recommendation dataset systematically stores the core elements of the recommendation process (actions, rewards, strategy parameters), making it easier for management to conduct large-scale teaching strategy evaluation, entrepreneurship support effectiveness analysis and other subsequent work.

[0174] Step 5: If it is identified that teacher profiling and multilateral collaboration are needed, an identifier is created for the teacher side and feature extraction and group clustering are performed in parallel with the student profiling. The multilateral collaborative extended dataset is updated collaboratively through teacher-student similarity and group feedback, thereby achieving multi-party discussion, mutual evaluation, and two-way matching.

[0175] The step five includes the following:

[0176] Step 501: Teacher multi-source data expansion and portrait modeling

[0177] To include teachers or entrepreneurial mentors (hereinafter referred to as t p Identify the p-th tutor), add a tutor data table to the platform database, and extract tutor characteristics such as professional background, practical project experience, tutoring style, and teaching achievements;

[0178] Assign a unique identifier to each tutor, handle missing values ​​or conflicting information during cleaning and normalization, and integrate all available fields into a new tutor basic dataset object. Similar to the modeling of student group portraits, clustering and feature mining can be performed on tutor data.

[0179] If the cluster center vector Ω has been defined in the second step k Or similarity measure Ψ(u j ,Ω k ), introduce new teacher portrait groups Characterize typical tutor profiles (such as technical tutors, market strategy tutors, management and coaching tutors, etc.); for text-based teacher achievements, papers, and teaching evaluations, the same NLP model as the student text analysis (such as BERT) can be used to extract vectors to form the tutor's text features Ξ(t p ), and then combined with quantitative labels (such as the number of tutored students, common tutoring areas, etc.), the tutor portrait vector To(t p );

[0180] The teacher clustering information and the tutor portrait vector To(t p ) is integrated into a new data object, the teacher-student collaborative dataset, to provide basic portrait support for subsequent teacher-student multilateral interactions. In the newly added teacher portrait group In addition, the original k The parsing structure ensures clear feature separation when multiple user types coexist;

[0181] When in use, by integrating and profiling the tutors' multi-source data, the platform can accurately portray the tutors' strengths, styles and experiences in entrepreneurship education or coaching, laying a data foundation for subsequent teacher-student collaboration and precise matching; when the teacher-student collaboration data set and student portrait data coexist, it can provide the platform with a more complete portrait of both teachers and students, presenting a panoramic view of the smart campus entrepreneurial ecosystem.

[0182] Traditional education systems often focus solely on student-side profiling, ignoring the multidimensional capabilities of teachers or mentors. This step builds on the previous unified profiling approach by providing teachers with the same rich label and semantic vector representations as students, achieving a two-way profiling approach.

[0183] Step 502: Multilateral collaboration and enhanced interaction between teachers and students

[0184] In the case of having a mentor portrait vector To(t p ) and the student portrait vector Po(u j ) and define the matching degree Ψ ts (j,p) to evaluate teacher-student matching:

[0185]

[0186] In the formula: Po(u j ) is the portrait vector of the jth student; To(t p ) is the portrait vector of the p-th teacher (or entrepreneurial mentor); Γ ts is the teacher-student similarity benchmark coefficient, which is used for overall scaling;

[0187] Δ ts is the distance attenuation coefficient, used to control Po(u j ) and To(t p ), the influence of differences on similarity;

[0188] M ts It is a learnable symmetric positive definite matrix that applies differentiated weights or correlations to different image dimensions;

[0189] α is the power exponent, which is greater than 0 and controls the nonlinear amplification or reduction of the distance metric. When α>0, when α=1, it degenerates into ordinary exponential distance decay; when α>1, it makes a steeper decay for large distances;

[0190] ψ ts The larger the value of (j,p), the more compatible the student j is with the teacher p in terms of their profile dimensions. The teacher can be recommended as the student's entrepreneurial mentor or invited to join the project team.

[0191] New online seminar rooms and discussion groups allow students and mentors to interact in real time on specific projects and topics. In addition to teacher-student interaction, students can also recommend or evaluate projects and learning resources to each other, such as sharing financing cases they have discovered. More user-generated content is accumulated through ratings and comments, and the weight of the portrait is updated in combination with the online learning mechanism. All discussion and sharing records are incorporated into the new behavioral event data. With the help of the clustering information constructed in the previous article (including the student group Ω k With teacher portrait group ), and the matching degree Ψ obtained in the current step TS, it can automatically assign better teacher-student groups or project teams to avoid the waste of resources caused by manual random allocation; if it is detected that some students’ needs are not met in the group (such as encountering a bottleneck in the marketing direction but there is no marketing tutor involved), it will trigger the online update strategy and prompt the platform administrator to try to invite more suitable tutors to join the group.

[0192] When used, the two-way matching of teachers and students and the online seminar function can form a multilateral collaborative mechanism at the platform level, so that entrepreneurial guidance is no longer limited to automatic push, but combines the active choices of mentors and students to enhance the depth and accuracy of interaction; the mutual evaluation and sharing of resources among students introduce more user-generated information, which not only enriches the content ecology of the platform, but also provides more detailed social data support for subsequent portraits and recommendation algorithms.

[0193] Step 503: Multilateral feedback and advanced profiling iteration

[0194] Combined with the recommended action records and feedback information records generated by the reinforcement learning process in the fourth step, a group comprehensive feedback mechanism is introduced in this step: if it is detected that the teacher-student discussion group or the student mutual evaluation group has a highly positive evaluation (or conversely a negative evaluation) of a certain resource or tutor, the portrait weights of all relevant users (students and tutors) are fine-tuned synchronously. The following group feedback factor ω can be defined: g :

[0195] ω g =κ g exp(μ gpos ·f gpos -μ gneg ·f gneg )

[0196] Where: κ g is the group feedback benchmark coefficient, which is greater than 0; f gpos Positive consensus within the group (such as collective likes and the number of positive comments); f gneg Negative consensus within the group (such as number of disapproval or low ratings);

[0197] μ gpos 、μ gneg are the amplification weights of positive and negative feedback, respectively, and both values ​​are greater than 0;

[0198] When the group feedback factor ω g When it is higher than expected, that is, the group as a whole has a high degree of support for a certain resource or mentor, the corresponding student and mentor portrait in the relevant field of the fit label can be strengthened, or the Q value in the reinforcement learning strategy can be corrected accordingly; after detecting significant group feedback, the mentor group portrait can be Or mentor portrait vector To(t p) also performs online updates; for example, if a tutor is particularly good at market analysis according to the feedback from most students, then improve the tutor portrait vector Po(u j ) in order to increase the recommendation priority of the tutor in this field in the subsequent tutor-student matching mechanism;

[0199] The modified mentor portrait vector To(t p ) and the corresponding student portrait vector Po(u j ) are written into the multilateral collaborative expansion data set, ultimately presenting a more dynamic and malleable multilateral portrait evolution at the platform level;

[0200] The multilateral collaborative expansion dataset and the previous reinforcement learning recommendation dataset can be combined into a multi-dimensional data asset for the platform, encompassing both individual-level reinforcement learning feedback (student-system) and multilateral group-level feedback (teacher-student interaction), forming an extended cycle of profiling-multilateral collaboration-feedback-re-profiling.

[0201] The inclusion of group-level feedback allows the platform to more quickly identify universal consensus or overall pain points, avoiding relying solely on individual feedback from a single student to adjust the system. The synchronous online update of teacher portraits ensures that the matching degree between tutor expertise and student needs continues to improve, and realizes the dynamic linkage of student portraits, tutor portraits and group interactions. On the basis of reinforcement learning, the dimension of comprehensive group feedback is further introduced, so as not to focus only on the clicks or pop-ups of individual behaviors, but also take into account the consensus or disputes within the group. This is especially critical for scenarios characterized by group interaction, such as university entrepreneurship classrooms and team projects.

[0202] In the present application, all parameters (including vector features, distance metrics, similarity calculations, and feedback counts) are normalized or standardized to achieve dimensional consistency. For example, features with physical units (time, score, number of times, etc.) are first converted into dimensionless or unified benchmarks before being substituted into the formula, thereby ensuring that exponential functions and matrix operations are not restricted by units, so that all core formulas can operate reasonably at the numerical level, and will not lead to insufficient disclosure or technical impracticality due to dimensional inconsistency.

[0203] See also Figure 1 The present invention provides an entrepreneurial guidance interactive system for smart campus, including:

[0204] The multi-source data fusion module, when detecting that student multi-dimensional data (society records, competition performance, book borrowing, etc.) is fragmented, cleans, deduplicates, and uniformly labels the original information, and maps key information to unique identifiers using structured fields to generate a multi-source pre-processed dataset;

[0205] The multi-dimensional portrait modeling module obtains cleaned and labeled multi-dimensional features, uses clustering algorithms and deep learning models to perform parallel feature extraction on student behavior sequences and text information, and generates cluster group portraits through NLP semantic analysis;

[0206] The dynamic update module detects significant differences between student profiles or behaviors and the original clusters. It then performs online learning and local iteration on newly generated clicks, browsing, and feedback events based on streaming data processing, and updates the corresponding profile weights and clustering affiliations, forming an adaptive closed loop of profile-recommendation-feedback-re-profile.

[0207] The personalized recommendation module, after completing the instant update of the user profile, calls the reinforcement learning model to conduct trial and error iteration based on the recommended actions and user feedback. It also modifies the Q value and policy parameters based on the exponential amplification or attenuation mechanism of positive and negative feedback, and outputs the optimal push solution to continuously enhance the efficiency of personalized resource matching.

[0208] If the collaborative enhancement module recognizes the need to introduce teacher portraits and multilateral collaboration, it will establish an identifier for the teacher side and perform feature extraction and group clustering parallel to the student portraits. It will also collaboratively update the multilateral collaborative expansion data set through teacher-student similarity and group feedback, thereby achieving multi-party discussion, mutual evaluation and two-way matching.

[0209] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0210] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0211] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0212] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0213] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. An interactive method for entrepreneurial guidance in a smart campus, characterized by: include, When it is detected that the student multidimensional data is in a fragmented state, the original information is preprocessed and the key information is mapped to a unique identifier using structured fields to generate a multi-source preprocessed dataset; After obtaining cleaned and labeled multi-dimensional features, clustering algorithms and deep learning models are used to perform parallel feature extraction on student behavior sequences and text information, and cluster group portraits are generated through NLP semantic analysis; If a student's profile or behavior is detected to be significantly different from the original cluster, online learning and local iteration are performed on the newly generated click, browse, and feedback events based on streaming data processing, and the corresponding profile weights and cluster affiliations are updated, forming an adaptive closed loop of profile-recommendation-feedback-re-profile. After completing the instant update of the user profile, the reinforcement learning model is called to conduct trial and error iteration based on the recommended actions and user feedback. The Q value and policy parameters are modified based on the exponential amplification or attenuation mechanism of positive and negative feedback to output the optimal push solution. If it is identified that teacher portraits and multilateral collaboration need to be introduced, an identifier is established for the teacher side and feature extraction and group clustering are performed in parallel with student portraits. The multilateral collaborative extended dataset is updated collaboratively through teacher-student similarity and group feedback.

2. The interactive method for entrepreneurial guidance in smart campus according to claim 1, characterized in that: Identify various data sources related to student entrepreneurship on and off campus, automatically capture them using distributed crawling and database interface docking, index and map each piece of data with a unique ID, extract key information fields from each collected record, and tag basic metadata.

3. The interactive method for entrepreneurial guidance in smart campus according to claim 2, characterized in that: Aggregate the obtained preliminary meta-information tags into the same student dimension according to ID, and calculate the credibility of each record; If the credibility of some records is far below the given credibility threshold, the abnormal or conflicting entries will be eliminated, and refined labels will be generated for each field and attached to the corresponding student records to form a multi-source preprocessing dataset.

4. The interactive method for entrepreneurial guidance in smart campus according to claim 3, characterized in that: After reading each student's label vector and related numerical features, a label vector is generated. For each student, the cluster to which they belong is determined according to the rule of maximizing the similarity function. After the cluster assignment is completed, the cluster center vector is updated according to the assignment result. Iterate until the cluster assignment results converge or the preset maximum number of iterations is reached, until a user group portrait is formed, and the cluster clusters and their cluster center vectors and feature summary information are output as cluster group portraits.

5. The interactive method for entrepreneurial guidance in smart campus according to claim 4, characterized in that: Extract the behavior sequence for deep representation learning to obtain the student's implicit state vector in the time dimension; The text semantic vector is extracted from the submitted text information, the obtained cluster information is fused with the latent state vector and the text semantic vector to obtain the user portrait vector, and the user portrait vector is summarized to generate a multidimensional portrait dataset.

6. The interactive method for entrepreneurial guidance in smart campus according to claim 1, characterized in that: Collect students' entrepreneurial data on the platform and generate behavioral event data whenever a new key behavioral event is received; Set trigger thresholds and rules: If the trigger conditions are met, real-time scheduling of online learning and profile correction will iteratively calculate the user profile vector. Before the trigger rules are established, the data of each behavioral event will be temporarily stored in a fast cache area, merged and packaged into a real-time incremental stream.

7. The interactive method for entrepreneurial guidance in smart campus according to claim 1, characterized in that: When the update is triggered, the deep feature extraction model is called to perform local iteration on the student's newly added behavior sequence and text data. The online gradient update strategy is used for the sequence model: After the correction generates a new implicit state vector or text semantic vector, it is fused with the original user portrait vector to generate a new user portrait vector. The new user portrait vector is added to the new data object real-time correction data set to represent the student's latest portrait status; If the new user portrait vector has deviated significantly from the original corresponding cluster in the feature space, local clustering fine-tuning will be performed on the student or the cluster to which it belongs.

8. The interactive method for entrepreneurial guidance in smart campus according to claim 7, characterized in that: The state is formed by combining the user portrait vector and the recent user behavior context. The corresponding action is the entrepreneurial guidance or resource push set provided to students. One or more recommended actions are selected in each cycle and a recommended action record is generated. When students give feedback on the recommendations of this cycle, corresponding feedback information records are generated and immediate rewards are automatically calculated, encapsulating the state, action, and reward information into reinforcement strategy environment information.

9. The interactive method for entrepreneurial guidance in smart campus according to claim 8, characterized in that: In reinforcement learning, a policy function is learned, which outputs the selected action after inputting the current state. This can introduce a Q function approximation based on a deep network and introduce an exploration mechanism into the recommendation strategy: If the exploration probability is triggered at a certain stage, a possible recommended action is randomly selected to discover potential new resources or new areas; if exploration is not triggered, the optimal action estimated by the current Q function is used, and after obtaining a new immediate reward, the learning weight gradient is updated to make the policy function continuously approach the optimal solution.

10. The interactive method for entrepreneurial guidance in smart campus according to claim 9, characterized in that: Reinforcement learning updates the Q function after receiving immediate rewards, synchronously transmitting users' explicit feedback on recommended content back to online learning. When a resource is highly recognized by users, the weight of the user's preference for that field in the user portrait vector is increased; If negative reviews appear multiple times, the corresponding weight will be reduced, thus forming a multi-round closed loop of profiling-recommendation-feedback-re-profiling; If some students' recommendations deviate from the mainstream behavior of the same group over multiple rounds of interaction, the clustering online fine-tuning logic is invoked to transfer them to a cluster that better matches the group portrait; Summarize the recommended actions, immediate rewards, and Q-function update information recorded in the reinforcement learning stage, organize them, and output them as a reinforcement learning recommendation dataset.

11. The interactive method for entrepreneurial guidance in smart campus according to claim 10, characterized in that: Add a tutor data table to the platform database, extract the tutor's guidance characteristics, and integrate all available fields into a new tutor basic dataset object after preprocessing; A new teacher portrait group is introduced to represent the typical tutor group portrait. Vectors are extracted from text-based teacher achievements, papers and teaching evaluations to form the tutor's text features. Then, the tutor portrait vector is constructed by combining quantitative labels. The teacher clustering information and tutor portrait vector are integrated into a new data object, the teacher-student collaborative dataset.

12. The interactive method for entrepreneurial guidance in smart campus according to claim 11, characterized in that: In multi-person group scenarios between tutors and students, interaction events are recorded and group collaboration data is generated. The group feedback factor is calculated based on the accumulation of positive and negative evaluations. The weights of the tutor and student portraits are modified by combining individual feedback and group consensus, and then updated online. If the same discussion group has serious differences in positive and negative opinions on a recommended resource or project, the group will be split into multiple subgroups, or individual correction rates will be set to fine-tune labels, and the new portraits after multi-party feedback will be synchronized to the multilateral collaborative expansion dataset.

Images