A system for evaluating students' comprehensive quality

Abstract

The present application discloses a student comprehensive quality evaluation system, which relates to the field of data analysis, including: a collection module, which collects students' data including static and dynamic data; static data includes students' basic information and educational background; dynamic data includes students' behavioral data and social data; a feature module, which extracts basic information features and educational background features of static data based on the collected data; extracts behavioral features and social features of dynamic data; behavioral features include behavioral frequency features and behavioral cycle features; social features include thematic social content features and emotional social content features; a model module, which constructs and trains a comprehensive quality evaluation model based on a neural network based on the extracted features, and calculates the comprehensive quality score of students. In view of the problem of low accuracy of student comprehensive quality evaluation in the prior art, the present application provides a student comprehensive quality evaluation system, which improves the evaluation accuracy through static and dynamic multi-dimensional evaluation.

Description

Technical Field

[0001] The present application relates to the technical field of data analysis, and in particular to a system for evaluating the comprehensive quality of students. Background Art

[0002] At present, the comprehensive quality evaluation of college students mainly adopts a combination of qualitative and quantitative methods, through student self-evaluation, teacher evaluation, academic performance and other means. However, these methods often rely on subjective judgment and limited examination dimensions, and the accuracy and comprehensiveness of the evaluation results need to be improved. At the same time, with the development of information technology, students generate a large amount of behavioral data and social data during their school years, which contain important information about the development of students' quality. However, the existing evaluation methods lack effective collection, analysis and utilization of these dynamic data, resulting in insufficient information dimensions for evaluation.

[0003] In addition, the comprehensive quality evaluation of students involves a variety of heterogeneous data sources, such as basic information of students, academic performance, daily behavior, social activities, etc. These data have great differences in format, scale, and semantics. How to effectively integrate and merge them to form a comprehensive and accurate student portrait is another challenge faced by the comprehensive quality evaluation of students.

[0004] In the related art, for example, Chinese patent document CN117350596A provides a method, device, equipment and storage medium for evaluating the comprehensive quality of students, which can be applied to the field of big data or finance. Obtain student characteristic data, which includes at least static data and dynamic data, and the dynamic data includes at least behavioral data. Perform statistical analysis on the behavioral data to obtain behavioral stability and social index. According to the static data, behavioral stability, social index and the calculation rules of each level of indicators, the index values ​​of each indicator are calculated in order from low to high, and the weight of the highest-level indicator is determined. Based on the weight and index value of the highest-level indicator, the comprehensive quality evaluation score is calculated. However, the feature data source of this scheme is relatively single, so the evaluation accuracy of this scheme needs to be further improved. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In view of the problem of low accuracy in the evaluation of students' comprehensive qualities in the prior art, the present application provides a system for evaluating students' comprehensive qualities, which improves the evaluation accuracy through static and dynamic multi-dimensional evaluation.

[0007] 2. Technical solution

[0008] The purpose of this application is achieved through the following technical solutions.

[0009] The embodiment of this specification provides a student comprehensive quality evaluation system, including: a collection module, which collects students' data including static and dynamic data; the static data includes students' basic information and academic background; the dynamic data includes students' behavioral data and social data; a feature module, which extracts basic information features and academic background features of static data based on the collected data; extracts behavioral features and social features of dynamic data; behavioral features include behavioral frequency features and behavioral cycle features; social features include thematic social content features and emotional social content features; a model module, which constructs and trains a comprehensive quality evaluation model based on a neural network based on the extracted features, and calculates the student's comprehensive quality score.

[0010] Among them, basic information features refer to feature vectors extracted from students' static data that reflect their basic attributes, usually including students' demographic characteristics (such as gender, age) and identity characteristics (such as student number, major), etc. In this application, basic information features are extracted from students' basic information data collected by the acquisition module. For example, the student's gender can be mapped to a binary feature (male is 1, female is 0), the student's age can be normalized to a real-valued feature (with a value range of 0 to, and then the student number and major can be converted into multiple binary features using one-hot encoding. Finally, the basic information features can be represented as a fixed-length 0-1 vector.

[0011] Among them, the educational background feature refers to the feature vector extracted from the student's static data that reflects his or her educational background and academic level, usually including the student's education stage, school category, grade ranking, etc. In this application, the educational background feature is extracted from the student's educational background data collected by the acquisition module. For example, the student's education stage (high school, undergraduate, master's degree, etc.) can be converted into an integer feature using ordinal coding, and the school category (985, 211, ordinary undergraduate, etc.) can be converted into multiple binary features using unique hot coding, and then the student's grade ranking can be normalized into a real-valued feature (with a value range of 0 to. Finally, the educational background feature can be represented as a mixed-type vector of a fixed length.

[0012] Among them, behavioral features refer to feature vectors extracted from students' dynamic data that reflect their daily behavior patterns, usually including various behavior categories, frequencies, and periodicities of students on and off campus. In this application, behavioral features include two sub-features: behavioral frequency features and behavioral cycle features, which are extracted from student behavior data collected by the acquisition module. For example, the number of times a student enters and exits the library in a month can be counted as the frequency feature of borrowing behavior, and the number of times a student goes to the cafeteria for meals per week can be counted as the frequency feature of dining behavior; the time distribution of students' computer behavior in a day can be analyzed, and its periodicity (such as frequency differences in the morning, afternoon, and evening) can be extracted as the periodic characteristics of computer behavior. Ultimately, different types of behavioral frequency features and periodic features constitute different dimensions of behavioral features.

[0013] Among them, social features refer to feature vectors extracted from students' dynamic data that reflect their social behavior and social psychology, usually including the subject matter, interaction intensity, emotional polarity, etc. of students' interactive content in social networks. In this application, social features include two sub-features: thematic social content features and emotional social content features, which are extracted from the student social data collected by the acquisition module respectively. For example, a topic model (such as LDA) can be used to perform topic mining on the students' social text content, and the topic distribution representing their social interests can be obtained as thematic social content features; the emotional dictionary can be used to match the students' social text content, and the proportion and polarity of the emotional words therein can be counted to obtain the positivity / negativity representing their social emotions as emotional social content features. Finally, the thematic social content features and emotional social content features constitute different dimensions of social features respectively.

[0014] Furthermore, the collection module includes: a static data collection unit, which obtains students' basic information and educational background from the student information management system; the basic information includes name, gender and age; the educational background includes major and grades; a behavioral data collection unit, which obtains students' behavioral data from the campus card usage record system; the behavioral data includes attendance records; a social data collection unit, which obtains students' public social data from social networks through a web crawler method; the social data includes the content of students' posts on social networking platforms.

[0015] Among them, social data refers to the general term for various types of data related to people's social behaviors and social relationships generated on social network platforms. In a broad sense, social data covers multiple aspects such as personal information, social relationship chains, social interaction behaviors, and social content of social network users; but in specific scenarios, social data usually refers specifically to social content data publicly released by users on social networks in the form of text, images, videos, etc., reflecting subjective information such as users' interests, opinions, and emotions. In this student comprehensive quality evaluation system, social data specifically refers to the content of posts publicly released by students on social network platforms, which are collected by the social data collection unit through web crawler methods.

[0016] Specifically: Data source: The social data collection unit first determines the main social networking platforms where students are active, such as Weibo, WeChat, QQ Space, Zhihu, etc., as the target source for data collection. Data acquisition: For each target platform, the social data collection unit develops a customized web crawler program to automatically crawl students' publicly published posts on a regular basis. The crawler needs to deal with technical difficulties such as the anti-crawling mechanism, paging loading, asynchronous rendering, etc. of each platform in order to obtain complete and real-time social post data. Data filtering: The crawled social post data may be mixed with the content of other users' posts forwarded and commented on by students. It is necessary to filter according to the publisher ID of the post, and only retain the original post data published by the students themselves. Data cleaning: Due to the colloquialism and irregularity of students' posts, the crawled social post data often contains a lot of noise, such as emoticons, Internet buzzwords, typos, etc., and needs to go through a series of text preprocessing steps (such as removing stop words, word form normalization, error correction, etc.) to clean it and improve data quality. Data storage: The cleaned social post data is generally stored in JSON format, and each piece of data contains multiple fields such as post ID, publishing time, text content, image URL, etc. To facilitate subsequent feature extraction and analysis and mining, document-based databases such as MongoDB are usually used for storage. Privacy protection: Since social data may involve students' personal privacy, during the entire process of collection, storage, and use, on the one hand, it is necessary to comply with relevant data protection regulations, such as desensitization, encrypted storage, access permission control, etc., to maximize the protection of students' privacy security; on the other hand, the collection of user personal information requires the user's consent before it can be carried out.

[0017] Furthermore, the collection module also includes: an identification unit, which performs a SHA-256 hash operation on the student's student ID number to generate a 256-bit hash value as the student's unique identifier; the identification unit generates a unique data identifier for static data, behavioral data and social data through the MD5 hash algorithm based on the obtained unique identifier; an association unit, based on the MongoDB database, uses the student's unique identifier as the id field of the document, stores the static data, behavioral data, social data and corresponding data identifiers as nested objects of the document, and generates student-centered data records in JSON format.

[0018] Among them, SHA-256 is a secure hash algorithm (Secure HashAlgorithm), which belongs to the SHA-2 cryptographic hash function family. It can map messages of any length to a fixed-length hash value of 256 bits (32 bytes), and has an avalanche effect, that is, a slight change in the input will cause a drastic change in the output. SHA-256 is currently recognized as one of the more secure and reliable hash algorithms. In this application, the identification unit uses the SHA-256 algorithm to perform a hash operation on the student number of each student, and generates a 256-bit fixed-length hash value as the unique identifier of the student. Since the student number is unique, performing a SHA-256 hash process on it will not introduce conflicts. At the same time, the SHA-256 algorithm is irreversible, and the original student number cannot be derived from the hash value, which can effectively protect the privacy of students. In addition, using a fixed-length hash value as a student identifier facilitates subsequent data association and query operations.

[0019] Among them, MD5 is another widely used cryptographic hash function that can map messages of arbitrary length to a fixed-length hash value of 128 bits (16 bytes). Compared with SHA-256, MD5 has poor security and has been proven to have conflict vulnerabilities, but it is still used in general data verification, query indexing and other occasions. In this application, the identification unit uses the MD5 algorithm to generate a 128-bit unique identifier for each student's static data, behavioral data, and social data. The data identifier plays the role of an index key, which facilitates subsequent rapid query and association of different types of data. The reason for using MD5 instead of SHA-256 is that the data identifier is only used inside MongoDB, and the identifier space is large enough and the probability of conflict is low, so the MD5 algorithm with faster calculation speed is selected.

[0020] Among them, MongoDB is a NoSQL database based on distributed file storage. It uses the JSON-like BSON format to store data, supports dynamic patterns, nested objects, arrays and other complex data types, has the characteristics of high scalability, high availability, high performance, etc., and is suitable for storing unstructured and semi-structured large data sets. In this application, the association unit is based on the MongoDB database, creates a JSON format data document for each student, stores its static data, behavioral data, and social data as nested objects in the document, and establishes a student-centered association data view. Compared with traditional relational databases, MongoDB is more suitable for storing student data, a complex data structure with multiple layers of nesting and diverse data types. At the same time, MongoDB supports a rich query language and indexing method, and can flexibly perform data query and aggregation analysis based on multiple dimensions such as student identification, data identification, and data fields.

[0021] Among them, in MongoDB, each document contains a special field called "_id", which is used as the unique identifier of the document. The id field can be an Object Id type automatically generated by MongoDB, or it can be any type specified by the application (such as string, integer, etc.), but it must be unique within a collection. In this application, the association unit uses the student ID (hash value generated by SHA-256) as the id field value of the MongoDB document to ensure that each student has only one corresponding data document in the database to avoid duplication and redundancy. When you need to query a student's data, you only need to retrieve the id field based on its student ID to quickly locate the unique data document, supporting efficient data access.

[0022] Among them, in MongoDB, the document adopts JSON structure, which supports nested sub-objects and sub-arrays. Through nested objects, related data can be organized together to form a hierarchical data model. The advantage of nested objects is that table connections in relational databases are avoided, and the cohesion and query efficiency of data are improved. In this application, the association unit makes full use of the nested object characteristics of MongoDB, and stores students' static data, behavioral data, social data and corresponding data identifiers as nested sub-objects in student documents. This data organization method intuitively reflects the data relationship centered on students, and the complete information of students can be obtained through a single query without the need for multi-table association. At the same time, nested data objects can be further nested as needed to build a more complex and fine-grained data model.

[0023] Among them, JSON (Java Script Object Notation) is a lightweight data exchange format that uses a text format independent of programming languages ​​to store and represent data. Data in JSON format is organized in the form of key-value pairs, supports multiple data types such as numbers, strings, Boolean values, arrays, objects, etc., and can flexibly represent structured, semi-structured and unstructured data. In this application, the MongoDB document generated by the association unit is essentially a data record in JSON format. Each student data record corresponds to a JSON object, and the fields of the object include "_id" (student ID), "static" (static data object), "behavior" (behavior data object), "social" (social data object), etc. The value of each field can be a basic type, an array or a nested JSON sub-object. Using JSON format to store student data has the following advantages: it is compatible with MongoDB's BSON storage format and can be written directly to the database; it is convenient for data transmission and parsing, and mainstream programming languages ​​support JSON serialization and deserialization operations; it is convenient for data visualization and manual inspection, and JSON itself is a highly readable plain text format. Therefore, JSON has become the de facto standard data format in the field of big data processing.

[0024] Further, the acquisition module also includes: a data format conversion unit, which uses the Protocol Buffers serialization protocol to convert the associated JSON format data records into binary format; a data verification unit, which uses the CRC32 verification algorithm to perform integrity verification on the converted binary format data records and generate a verification code. Among them, Protocol Buffers (abbreviated as Protobuf) is a platform-independent, language-independent, and scalable data serialization protocol developed by Google. It defines structured data as a .proto file, and then uses the Protobuf compiler to generate a data access class in the target language, thereby realizing the serialization and deserialization operations of the data. Compared with text formats such as JSON and XML, the binary data serialized by Protobuf is more compact, occupies less space, and has a faster parsing speed. In this application, the data format conversion unit uses the Protobuf protocol to convert the JSON format student data records in MongoDB into corresponding binary formats. The specific steps are as follows: define a .proto file for student data records, specify the name, data type, identification number, etc. of each field; use the Protobuf compiler to generate a data access class in the target language (such as Python, Java, etc.) based on the .proto file; map the student data records in JSON format to an instance of the data access class, and call the corresponding serialization method to generate Protobuf data in binary format.

[0025] Among them, CRC32 (CyclicRedundancyCheck3) is an error detection algorithm based on cyclic redundancy check, which is mainly used for integrity verification during data transmission and storage. The CRC32 algorithm generates a 32-bit check code by performing polynomial division on the data, which is used to verify whether the data has errors or tampering during transmission or storage. If the data changes, the recalculated CRC32 check code will be inconsistent with the original check code, thereby detecting an error. In this application, the data verification unit uses the CRC32 algorithm to perform integrity verification on the binary student data records serialized by Protobuf.

[0026] The specific steps are as follows: After the data format conversion is completed, the CRC32 algorithm is immediately executed on the Protobuf binary data to calculate the 32-bit check code; the CRC32 check code is used as metadata and stored or transmitted together with the Protobuf binary data; when reading or receiving data, the CRC32 algorithm is re-executed on the Protobuf binary data to calculate a new check code; compare the new and old check codes to see if they are consistent. If they are inconsistent, it means that there is an error in the data and it needs to be retransmitted or restored. More specifically, the purpose of introducing the CRC32 check is to: ensure the integrity of the data during storage and transmission, and to detect and correct data errors in a timely manner; by comparing the check code, quickly determine whether the data has changed, and avoid unnecessary data analysis and processing; the CRC32 algorithm is simple to implement and has high computational efficiency, which is suitable for real-time verification scenarios of massive data. The CRC32 check is used in combination with Protobuf encoding to form an effective data fault tolerance mechanism.

[0027] Among them, the check code is the calculation result of the data verification algorithm (such as CRC3, which is generally a binary code or hexadecimal string of fixed length. The function of the check code is to characterize the integrity of the original data and can be used for equivalence judgment and error detection. For a given data content, the check code calculated again should be consistent with the original check code; if the calculation result is inconsistent, it means that the data content has changed. In this application, the CRC32 check code plays the role of "fingerprint" of the student data record. On the one hand, when writing data, the data verification unit calculates the CRC32 value of cada student data records in real time, and uses it as a check code and stores it persistently together with the serialized Protobuf binary data to ensure that each data has an independent reference system. On the other hand, when reading data, the data verification unit recalculates the CRC32 value of each student data record and compares it with the original check code queried. Once an inconsistency is found, it is marked as dirty data in time or the data recovery process is automatically initiated to maximize the accuracy of data processing. In addition, the CRC32 check code can also be used in data caching and incremental update scenarios. The check code can be used to quickly determine whether the data version has changed, thereby optimizing data access and synchronization strategies.

[0028] Furthermore, the feature module includes: a static feature extraction unit, which uses a regular expression matching method to extract students' basic information features and educational background features; the basic information features include gender features and age features; the educational background features include professional category features and grade features; the extracted basic information features and educational background features are combined into a static feature vector; a behavioral feature extraction unit, which uses time series analysis and frequent pattern mining methods to extract students' behavioral features, and the behavioral features include behavioral frequency features and behavioral periodic features.

[0029] Among them, regular expression (Regular Expression, referred to as Regex) is a formal language for matching string patterns. It describes the structure and content of the target string by defining a series of special characters and metacharacters. Regular expression matching refers to the process of using regular expressions to search, extract, replace, and other operations on strings. Common regular expression engines include PCRE, Python's re module, etc. In this application, the static feature extraction unit uses a regular expression matching method to extract structured features from students' basic information and academic background data. The specific steps are as follows: analyze the text format of student data, summarize the writing mode of each field, such as gender is usually expressed as "male / female", grades are usually expressed as "excellent / good / pass", etc.; for each target field, write a corresponding regular expression pattern for matching and extracting the corresponding feature value; use a regular expression engine to scan the student data one by one, identify key fields such as gender, age, major category, and grade, and convert the extraction results into a structured feature representation.

[0030] Among them, a time series is a sequence of data points arranged in chronological order, reflecting the state or value of a thing or phenomenon at different time points. Time series analysis is a branch of statistics, which mainly studies the problem of feature extraction, pattern recognition, trend prediction, anomaly detection, etc. of time series data. Common time series analysis methods include moving average, autoregression, ARIMA, etc. In this application, the behavior feature extraction unit adopts the method of time series analysis to extract the time features reflecting the characteristics of the students' behavior from their original behavior data (such as campus card swiping records). The specific steps are as follows: sort the students' behavior data according to timestamps and convert them into a time series format; perform preprocessing operations such as sampling, interpolation, and smoothing on the behavior time series to eliminate the influence of noise and missing values; use moving average, difference and other techniques to transform the behavior time series, and extract the statistical features such as mean, variance, and trend of the behavior; use visualization tools such as periodograms and autocorrelation diagrams to perform periodic analysis on the behavior time series, and identify the periodic features such as daily cycle, weekly cycle, and monthly cycle of student behavior; combine the extracted statistical features and periodic features into a structured behavior feature vector.

[0031] Among them, frequent patterns refer to meaningful sub-patterns that appear repeatedly in large-scale data sets. Frequent pattern mining is a classic problem in the field of data mining, which aims to discover frequent co-occurring feature combinations from a large amount of data and reveal the intrinsic connection between things. Common frequent pattern mining algorithms include Apriori, FP-growth, Eclat, etc. In this application, the behavior feature extraction unit adopts the method of frequent pattern mining to discover recurring behavior combination patterns from the student's behavior sequence data. The specific steps are as follows: convert the student's original behavior data into a transaction form, each transaction represents a time window (such as a day, a week) in which a behavior occurs, and the items in the transaction represent the behavior type that occurs in the time window; configure the minimum support threshold, use the frequent pattern mining algorithm (such as Apriori) to scan the behavior transaction data, and generate frequent behavior item sets; calculate the support of each frequent behavior item set as the frequency feature of the behavior combination pattern; extract the association rules and time series patterns of the behavior according to the length and composition of the frequent behavior item set as the structural features of the behavior combination pattern; integrate the frequency features and structural features of the behavior combination pattern into the behavior feature vector.

[0032] Furthermore, the behavior feature extraction unit performs the following steps: converting the preprocessed data into a time series format, where each time series represents a student's behavior record, and each element of the time series contains the timestamp and behavior type of the behavior; using a sliding window method, the time series is segmented at different time granularities to obtain multiple subsequences; wherein the time granularity includes days, weeks, and months; the size and step size of the sliding window are set according to the time granularity; performing frequent pattern mining on each subsequence to obtain frequent item sets and sequential patterns of student behavior; wherein the frequent item set represents a combination of behavior types that appear simultaneously with a frequency greater than a threshold; the sequential pattern represents a continuous The sequence of behavioral types that appear; the Apriori algorithm is used for frequent pattern mining; a counting-based method is used to count the number of behaviors of each student in a sliding window of different time granularities, and the average daily and weekly behavioral frequencies are calculated as behavioral frequency features; a method based on periodic analysis is used to perform frequency domain analysis on the behavioral time series of each student using Fourier transform, and the behavioral distribution with a period of one week and one month is extracted as the behavioral periodic feature; wherein, the Fourier transform converts the time series from the time domain to the frequency domain, and determines the periodic law of the behavior by analyzing the main frequency components in the frequency domain; the extracted behavioral frequency features and behavioral periodic features are organized into behavioral feature vectors.

[0033] Among them, frequent item sets refer to the set of items in the transaction data set whose frequency exceeds the minimum support threshold. The support of an item set is defined as the ratio of the number of transactions containing the item set to the total number of transactions. Frequent item sets reflect the co-occurrence relationship between different items in the data set and are the basic output result of frequent pattern mining. In this application, by performing frequent pattern mining on student behavior subsequences, frequent co-occurrence relationships between different behavior types, namely frequent item sets, can be found. For example, if the item set {attending class, self-study} is frequent, it means that students often study while attending class. Frequent item sets reveal a combination pattern of students' behavioral preferences, embody the correlation between different behavior types, and are of reference value for understanding students' comprehensive behavioral habits.

[0034] Among them, the sequential pattern is a frequent pattern containing temporal information, which represents a sequence of item sets that appear frequently in a specific order. Compared with frequent item sets, the sequential pattern not only considers the co-occurrence relationship within the item set, but also considers the order between different item sets, so it is possible to discover the temporal regularity in the data set. In this application, by mining frequent patterns on student behavior subsequences, the sequential dependency between different behavior types, namely the sequential pattern, can be discovered. For example, if the sequence <{class}, {eating}, {self-study}> is frequent, it means that a typical behavior path for students is to attend class first, then eat, and finally study. The sequential pattern reveals the temporal order of student behavior and reflects the rationality of students' schedules, which is of great value for optimizing student management and providing personalized services.

[0035] Among them, Apriori is a classic frequent item set mining algorithm. Based on the a priori principle, it generates candidate item sets iteratively and calculates their support, and finally obtains all frequent item sets. The core of the Apriori algorithm is that "any non-empty subset of a frequent item set must also be frequent", which is the so-called Apriori property. According to this, k-item candidate item sets can be generated through known k-1 frequent item sets, greatly reducing the search space. In this application, the Apriori algorithm is used to mine frequent patterns on student behavior subsequences. The specific steps are as follows: map each behavior subsequence to a transaction, and the behavior type is mapped to the item in the transaction; scan the transaction data set, calculate the support of each behavior type, and generate a 1-frequent item set; generate k-item candidate item sets from k-1 frequent item sets, and ensure that the candidate item sets meet the Apriori property through connection and pruning operations; scan the transaction data set, calculate the support of each candidate item set, and filter out k-frequent item sets; repeat until no higher-order candidate item sets can be generated; output all frequent item sets as results. The Apriori algorithm is easy to implement and is suitable for frequent pattern mining of small and medium-sized data sets. The use of the Apriori algorithm in student behavior analysis can achieve a balance between accuracy and efficiency.

[0036] Among them, sliding window is a local processing method for time series data. Given a time window of fixed length, the time series is divided into several overlapping or non-overlapping subsequences by continuously sliding the starting position of the window. The sliding window reduces the computational complexity while ensuring data coverage by controlling the window size and sliding step. In this application, the sliding window method is used to segment the student behavior time series. Since different time granularities correspond to different behavioral patterns, multiple sliding windows are used to sample the time series from three granularities of day, week, and month. Take the weekly granularity as an example: set the window size to 7 days, the sliding step to 1 day, and obtain several behavioral subsequences with a length of one week by moving the starting position of the window. These subsequences not only contain complete weekly behaviors, but also avoid truncation of behaviors through overlapping sliding methods. Through sliding window sampling, behavioral patterns at different time scales can be captured, the length of the time series can be reduced, and the subsequent computational overhead can be reduced.

[0037] Among them, Fourier transform is a time-frequency domain conversion method that can convert time series from time domain representation to frequency domain representation. In the time domain, the time series is represented as a sequence of values ​​at different times; in the frequency domain, the time series is represented as amplitude and phase components of different frequencies. By analyzing the frequency components with larger amplitudes in the frequency domain, the periodic characteristics of the time series can be determined. In this application, a periodic analysis method based on Fourier transform is used to extract the periodic characteristics of student behavior on a weekly and monthly scale. The specific steps are as follows: interpolation, smoothing and other preprocessing of the original behavior time series of students are performed to obtain time sampling at fixed intervals; fast Fourier transform (FFT) is performed on the time series to obtain the amplitude spectrum and phase spectrum of different frequency components; the main frequency components in the amplitude spectrum are analyzed. If there are significant weekly frequency (such as 1 / 7 day) or monthly frequency (such as 1 / 30 day) components, it is considered that the student's behavior has weekly or monthly periodicity; for the frequency components with significant periodicity, the amplitude value is used to represent the intensity of the behavior under the period, and the phase value is used to represent the time offset of the behavior under the period; the intensity and time offset of each period are used as the characteristics of the behavior period. Fourier transform realizes the periodicity mining of student behavior through time-frequency domain conversion, revealing the regularity of student behavior at different time scales.

[0038] Further, the feature module also includes: a social feature extraction unit, which uses a natural language processing method to extract the social content features of students. Among them, natural language processing (Natural Language Processing, NLP) is an important branch in the field of artificial intelligence, which aims to realize the understanding, generation and application of human natural language by computers. NLP integrates the theories and technologies of multiple disciplines such as linguistics, computer science, mathematics, etc., and mainly studies how to convert human language into symbolic representations that can be processed by computers, and performs intelligent processing such as semantic understanding, information extraction, and machine translation. Common NLP tasks include word segmentation, part-of-speech tagging, named entity recognition, syntactic analysis, semantic role labeling, text classification, sentiment analysis, text summarization, etc. NLP is the basis of intelligent applications such as human-computer interaction, knowledge graphs, and search engines. In this application, the social feature extraction unit uses NLP methods to perform semantic analysis and feature extraction on students' social media text data (such as Moments, Weibo, QQ Space, etc.).

[0039] Furthermore, the social feature extraction unit includes: the social feature extraction unit includes a topic extraction subunit, a sentiment analysis subunit and a feature organization subunit; the topic extraction subunit adopts a latent Dirichlet allocation algorithm, by representing the post content as a document-topic probability distribution and a topic-term probability distribution, and estimating and judging the probability distribution, the post content is divided into several topics; wherein the document-topic probability distribution represents the probability that each post belongs to each topic, and the topic-term probability distribution represents the probability that each topic contains each term; the probability distribution of each student discussing different topics is calculated as a topic social content feature; the sentiment analysis subunit adopts a dictionary-based method, by constructing a sentiment dictionary and matching the sentiment words in the post content, the post content is divided into positive, negative and neutral sentiment categories; the proportion of posts of each student in different sentiment categories is calculated as the sentiment social content feature; the feature organization subunit sequentially splices the obtained topic social content features and sentiment social content features into a vector by using a feature splicing method, and normalizes the spliced ​​vector by using a minimum-maximum normalization method to obtain a social feature vector.

[0040] Among them, Latent Dirichlet Allocation (LDA) is an unsupervised text topic mining algorithm, which is based on probabilistic graphical models and Bayesian inference, and can divide documents in a text corpus into several topics. The LDA algorithm assumes that each document is a mixture of multiple topics in different proportions, and each topic is composed of multiple terms with different probabilities. By estimating the two probability distribution parameters of document-topic and topic-term, LDA can reveal the latent semantic structure of the text corpus. LDA is a representative algorithm in the field of topic models and is widely used in scenarios such as text classification and recommendation systems. In this application, the topic extraction subunit uses the LDA algorithm to divide the content of students' social posts into topics. The specific steps are as follows: perform preprocessing such as word segmentation and stop word removal on social posts to obtain a normalized word sequence; treat each post as a "document" and all posts constitute a text corpus; set the number of topics K, randomly initialize the document-topic probability distribution α and the topic-term probability distribution β; use approximate inference algorithms such as Gibbs sampling to iteratively update the topic distribution of each post and the term distribution of each topic until the parameters converge; for each student, count the topic distribution of the posts he posted to obtain the discussion probability of the student on each topic; select the top N topics with the largest probability as the main social topics of the student, and their probability values ​​as the social content features of the topic. Through LDA topic division, we can not only discover the collective interest distribution of the student group, but also analyze the participation of different students in different topics and explore the personalized social preferences of students.

[0041] Among them, the dictionary-based method is a simple and effective text sentiment analysis method. It builds a sentiment word dictionary, matches the sentiment words appearing in the text, and assigns corresponding sentiment weights according to the part of speech and degree of the sentiment words, and finally calculates the sentiment value of the text. Commonly used Chinese sentiment dictionaries include Hownet and others. Compared with the sentiment analysis method based on machine learning, the dictionary-based method has the advantages of no need for labeled data, strong domain adaptability, and good interpretability. In this application, the sentiment analysis subunit uses a dictionary-based method to classify the sentiment of students' social post content. The specific steps are as follows: construct a Chinese sentiment word dictionary, including sentiment words with positive and negative polarities, and assign corresponding sentiment weights (such as 1 / -; segment the content of each post and match the sentiment words in the sentiment dictionary; calculate the sentiment score for each sentiment word according to its sentiment weight; perform weighted summation of the sentiment scores of all sentiment words in the post to obtain the sentiment value of the post; divide the posts into three categories: positive, negative, and neutral according to the positive and negative and size of the sentiment values; for each student, count the sentiment classification ratio of the posts he / she posts as the feature of emotional social content. By matching the sentiment dictionary and accumulating the sentiment scores, the emotional tendency of students' posts can be judged more accurately, and then the emotional state and expression habits of students on social media can be analyzed.

[0042] Among them, feature splicing is a commonly used multi-dimensional feature fusion method, which connects feature vectors of different sources and different types end to end to obtain a longer combined feature vector. Splicing can achieve information aggregation and complementarity at the feature level and improve the representation ability of the machine learning model. However, feature splicing may also introduce redundancy and noise between features, so it is necessary to perform appropriate dimensionality reduction and normalization on the spliced ​​feature vector. In this application, the feature organization subunit adopts the feature splicing method to integrate the theme social content features and the emotional social content features into a unified social feature vector. The specific steps are as follows: vectorize the theme social content features, each theme corresponds to a dimension, and the discussion probability of the theme is used as the value of the dimension; vectorize the emotional social content features, each of the three emotional categories of positive, negative, and neutral corresponds to a dimension, and the proportion of the emotional category is used as the value of the dimension; splice the theme feature vector and the emotional feature vector in order to obtain a longer social feature vector; due to the large difference in the numerical range of different dimensions, it is necessary to normalize the spliced ​​social feature vector. Through feature concatenation, the topic dimension and the sentiment dimension are organically combined to comprehensively depict the social media portrait of students. The concatenated social feature vector can be used for downstream tasks such as student clustering and anomaly detection.

[0043] Among them, minimum-maximum standardization is a commonly used data normalization method. By linearly mapping the value of the feature to the interval [0, 1], the differences in the feature dimension and numerical range are eliminated, making the features from different sources comparable. The calculation formula of minimum-maximum standardization is: x'=(x-min) / (max-min), where x is the original feature value, x' is the normalized feature value, min and max are the minimum and maximum values ​​of the feature of this dimension, respectively. In this application, the feature organization subunit uses the minimum-maximum standardization method to normalize the spliced ​​social feature vector. The specific steps are as follows: for each dimension of the social feature vector, calculate the minimum and maximum values ​​of the feature values ​​of all students under this dimension; for each dimension feature value of each student, use the minimum-maximum standardization formula to convert it.

[0044] Furthermore, the feature module also includes: a feature fusion unit, which fuses the static feature vector, the behavioral feature vector and the social feature vector to obtain a comprehensive feature vector of the student.

[0045] Furthermore, the feature fusion unit includes: a static feature encoder, a behavioral feature encoder, a social feature encoder, a feature mapping layer and a feature splicing layer; the static feature encoder uses an autoencoder network to encode the static feature vector, and extracts the high-level semantic representation of the static feature through unsupervised learning to obtain a static feature view; the autoencoder is composed of a multi-layer feedforward neural network, including an input layer, a hidden layer and an output layer; the input layer receives the static feature vector, the hidden layer encodes the input feature to obtain a compressed representation of the static feature, and the output layer decodes the compressed representation to obtain a reconstructed static feature vector; the behavioral feature encoder uses an autoencoder network to encode the behavioral feature vector, and extracts the high-level semantic representation of the behavioral feature through unsupervised learning. The behavioral feature view is obtained; the social feature encoder adopts an autoencoder network to encode the social feature vector, and extracts the high-level semantic representation of the social features through unsupervised learning to obtain the social feature view; the feature mapping layer adopts a feedforward neural network to map the static feature view, the behavioral feature view and the social feature view to a shared semantic space, and learns the common representation of different views by minimizing the differences between different views; the feature concatenation layer adopts a vector concatenation method to concatenate the static feature view, the behavioral feature view and the social feature view output by the feature mapping layer in a shared semantic space to obtain the student's comprehensive feature vector; each element in the comprehensive feature vector represents a high-level semantic feature, which integrates the feature information of the students in different dimensions.

[0046] Among them, the autoencoder is an unsupervised neural network model, which consists of two parts, an encoder and a decoder, and its goal is to learn the compressed representation of the input data. The encoder maps the high-dimensional input to a low-dimensional latent space, and the decoder reconstructs the latent space representation into the original input. By minimizing the reconstruction error, the autoencoder can automatically extract high-level features of the input data. Variants such as stacked autoencoders, sparse autoencoders, and denoising autoencoders have expanded the representation ability and generalization performance of the autoencoder. Autoencoders are commonly used for tasks such as data dimensionality reduction, feature learning, and anomaly detection. In this application, the static feature encoder, behavioral feature encoder, and social feature encoder respectively use autoencoder networks to encode the corresponding feature vectors. Take the static feature encoder as an example: the input layer receives static feature vectors, such as the student's gender, age, family background, etc.; the hidden layer maps the input features to the latent space in a nonlinear way through a multi-layer feedforward neural network to extract the abstract representation of the static features; the output layer decodes the latent space representation into the original static feature vector; by minimizing the reconstruction error, such as mean square error or cross entropy, the parameters of the encoder and decoder are optimized so that the decoded feature vector is as similar as possible to the original feature vector; the decoder is removed, the trained encoder is retained, and the static feature vector is mapped to the latent space to obtain a static feature view. The processing process of the behavioral feature encoder and the social feature encoder is similar. The high-level features extracted by the autoencoder filter out the redundant information of the original features, making them more concise and robust.

[0047] Among them, unsupervised learning is a large paradigm of machine learning, which discovers the intrinsic structure and regularity of data from unlabeled data. Unlike supervised learning, which relies on manually annotated training data, unsupervised learning only uses the statistical characteristics of the data to perform clustering, density estimation, dimensionality reduction, etc. on the data through similarity measurement between samples. Common unsupervised learning algorithms include K-means clustering, hierarchical clustering, principal component analysis, independent component analysis, autoencoder, etc. Unsupervised learning is widely used in fields such as anomaly detection, recommendation systems, and semantic representation. In this application, the autoencoder network automatically extracts high-level semantic representations of static features, behavioral features, and social features through unsupervised learning without manual annotation. Specifically, the autoencoder network takes the feature vector as input, and learns the compressed representation of the feature vector through parameter optimization of the encoder and decoder; the optimization goal is to minimize the reconstruction error, that is, the difference between the decoded feature vector and the original feature vector, and this process does not require the label information of the feature vector; through unsupervised feature extraction, the autoencoder network explores the intrinsic pattern of the feature data and obtains the denoised and abstract feature representation, that is, the static feature view, the behavioral feature view, and the social feature view. Unsupervised learning reduces the dependence on labeled data and improves the automation of feature fusion.

[0048] Among them, high-level semantic representation refers to the abstract and conceptual representation of data, which is more concise and robust than the original data, and can characterize the intrinsic structure and attributes of the data. High-level semantic representation maps the original data to a semantically rich space through feature transformation and combination, in which the intrinsic patterns and laws of the data are more obvious. Through multi-layer nonlinear transformation, the deep learning model can automatically learn the hierarchical semantic representation of data for downstream tasks such as classification, clustering, and retrieval. Word vectors, image features, knowledge graphs, etc. are all examples of high-level semantic representation. In this application, the static feature encoder, the behavioral feature encoder, and the social feature encoder learn the high-level semantic representation of feature data, namely, the static feature view, the behavioral feature view, and the social feature view, respectively. Take the static feature encoder as an example: through a multi-layer feedforward neural network, the static features are mapped layer by layer to a more abstract latent space; each dimension of the latent space corresponds to a high-level semantic feature, such as the student's family background features, learning ability features, etc. These high-level features characterize the concepts and internal connections of the student's static attributes; the high-level semantic representation consists of a few key features, which removes the redundancy and noise in the original features and is more concise and robust. High-level semantic representation reveals the intrinsic characteristics of students and provides more effective features for subsequent feature fusion.

[0049] In this application, the static feature view, behavioral feature view, and social feature view are generated by the static feature encoder, behavioral feature encoder, and social feature encoder, respectively, representing the high-level features of students in different dimensions: the static feature view is composed of the basic attribute characteristics of the students, such as gender, age, family background, etc., reflecting the inherent attributes and external environment of the students; the behavioral feature view is composed of the daily behavioral characteristics of the students, such as academic activities, campus activities, consumption behaviors, etc., reflecting the behavioral patterns and living habits of the students during their time at school; the social feature view is composed of the social text characteristics of the students, such as social themes, emotional tendencies, etc.

[0050] 3. Beneficial effects

[0051] Compared with the prior art, the advantages of this application are:

[0052] This application collects static and dynamic data of students and extracts rich features from them, including basic information features, educational background features, behavior frequency features, behavior cycle features, theme social content features, and emotional social content features. By portraying the characteristics of students from all aspects and angles, it overcomes the limitation of the single evaluation dimension of existing methods, achieves accurate portrait of students, and provides a reliable basis for comprehensive quality evaluation.

[0053] This application not only utilizes static data in the student management system, but also integrates dynamic data such as campus behavior data and social network data. Through time series analysis, frequent pattern mining, topic model, sentiment analysis and other data mining and natural language processing technologies, it deeply mines high-level semantic information such as behavior patterns, interest preferences, and emotional states contained in dynamic data, expanding the information dimension of evaluation.

[0054] This application innovatively designs a feature fusion unit, uses an autoencoder network to learn high-level representations of static, behavioral, and social features, and achieves fusion of multi-view features in a shared semantic space through feature mapping and splicing. This method can extract the essence of student features from multiple views, reveal the inherent connections between different features, and form more comprehensive, in-depth, and accurate comprehensive student features.

[0055] This application uses the integrated features after fusion to construct and train a multi-layer neural network evaluation model, overcoming the shortcomings of traditional evaluation methods that rely on manually designed features and scoring rules. The neural network model can automatically learn the complex associations and high-level abstractions between features, has strong representation and generalization capabilities, and can significantly improve the evaluation and prediction level of students' comprehensive qualities. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0057] Figure 1 is an exemplary module diagram of a student comprehensive quality evaluation system according to some embodiments of this specification;

[0058] Figure 2 is a schematic diagram of a collection module according to some embodiments of this specification;

[0059] Figure 3 is a schematic diagram of a feature module according to some embodiments of this specification;

[0060] Figure 4 is an exemplary flow chart of obtaining a behavior feature vector according to some embodiments of this specification;

[0061] Figure 5 is an exemplary flow chart of obtaining a behavior social vector according to some embodiments of the present specification. DETAILED DESCRIPTION

[0062] The method and system provided in the embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0063] Figure 1 is an exemplary module diagram of a student comprehensive quality evaluation system according to some embodiments of this specification, such as Figure 1 As shown in Figure 1, the student comprehensive quality evaluation system includes data collection module, data preprocessing module, feature extraction module, feature fusion module and evaluation prediction module. In the data collection stage, the data collection module connects with the student management system, campus behavior database and social network platform to obtain multi-source heterogeneous data of students.

[0064] Figure 2 This is a schematic diagram of the collection module shown in some embodiments of this specification. The data collection module obtains students' basic information data and academic background data from the student management system through the API interface, including static attribute data such as student name, gender, age, home address, enrollment time, department and major. At the same time, the data collection module obtains students' daily behavior record data from the campus behavior database through regular data synchronization, including dynamic data such as the time, location, type, and frequency of students' class attendance, book borrowing, physical exercise, campus consumption and other behaviors. In addition, the data collection module obtains students' social text data from the social networking platform through data crawling and OAuth authorization, including text content data such as posts, comments, and private messages posted by students.

[0065] Specifically, a school has a student information management system that stores the basic information and educational background of each student. The static data collection unit connects to the database of the system and executes SQL query statements to obtain the name, gender, age, major and grade data of the specified student. For example, for the student with the student number "20230001", the query results are: Name: Zhang San, Gender: Male, Age: 20, Major: Computer Science and Technology, Grade: [{"Course": "Data Structure", "Grade": 90}, {"Course": "Java Programming", "Grade": 85}]. Behavioral data collection unit: The school's campus card usage record system records the attendance information of each student. The behavioral data collection unit obtains the attendance record of the specified student in the specified time period by interacting with the API interface of the system. For example, for the student with the student number "20230001", the attendance record for the past week is: [{"Date":

[0066] "2023-04-17", "status": "OK"}, {"date": "2023-04-18", "status":

[0067] "Normal"}, {"Date":"2023-04-19","Status":"Absent"}, {"Date":"2023-04-20","Status":"Normal"}, {"Date":"2023-04-21","Status":"Normal"}]. Social Data Collection Unit: A student registered an account on a social networking platform and published some posts publicly. The social data collection unit uses web crawler technology to regularly crawl the public content of the platform and filter out the posts belonging to the student based on the student's social account ID. For example, for the student with student number "20230001", the posting content in the last month is: [{"Title":"College life is so happy","Content":"Today I participated in the school's spring outing and met many new friends. College life is so colorful!","Time":"2023-04-15"}, {"Title":"Learning data structure is so difficult","Content":"The data structure course is too abstract and there are too many homeworks. I feel like I am about to fail. Please give me some advice!","Time":"2023-04-10"}].

[0068] Identification unit: In order to protect student privacy, the identification unit first performs a SHA-256 hash operation on the student's ID number and converts the student ID number into a 256-bit fixed-length hash value as the student's unique identifier. For example, the hash identifier of the student ID "20230001" is: "f5a5a608b1c4d342d2d1f19b51d398a5e5fef3c3ac4c4d5c6c1d2b0e59f7f9c". Next, the identification unit uses the MD5 hash algorithm to generate a 128-bit unique identifier for each static data, behavioral data, and social data based on the student ID and data type. For example, the identifier of a piece of social data of the above student is: "d2a2d3a0c56f8c7c4f5d5e7f8c1d2b0e".

[0069] Association unit: The association unit is based on the MongoDB database and stores all the student data in a JSON format document. The id field of the document is the unique identifier of the student, and other fields include static data objects, behavioral data arrays, social data arrays, and corresponding data identifiers. For example, the data document of the above student is:

[0070] {"_id": "f5a5a608b1c4d342d2d1f19b51d398a5e5fef3c3ac4c4d5c6c1d2b0e59f7f9c", "static": {"name": "Zhang San", "gender": "male", "age": 20, "major":

[0071] "Computer Science and Technology", "scores": [{"course":"Data Structure","score":90},

[0072] {"course":"Java Programming","score":85}],"data_id":

[0073] "a0b1c2d3e4f5a6b7c8d9e0f1a2b3c4d5"}, "behavior": [{ "date": "2023-04-17", "status": "OK", "data_id": "f1e2d3c4b 5a6f7e8d9c0b1a2d3c4e5f6"}, {"date": "2023-04-18", "status": "OK", "data_id": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6 ”},{“date”:“2023-04-19”,”status”:“false”,”data_id”:“e1f2d3c4b5a6e7f8d9c0b1a2d3c4e5f6”},{“date”:“2023-04-20”,”status”:“normal”,”data_id:c1d2e3f4a5b6c7d8e9f0a1b2c3d4e5f6”},{“date”:“2023-04-21”,”status”:“normal”,”data_id:

[0074] b1c2d3e4f5a6b7c8d9e0f1a2b3c4d5e6"}],"social":[{"title":"College life is so happy","content":"I participated in the school's spring outing today and made many new friends. College life is so colorful!","time":"2023-04-15","data_id":"d2a2d3a0c56f8c7c4f5d5e7f8c1d2b0e"},{"title":"Learning data structure is so difficult","content":"The data structure course is too abstract and there are too many homeworks. I feel like I'm going to fail. Please give me some advice!","time":"2023-04-10","data_id":"c3b4a5d6e7f8c9d0a1b2c3d4e5f6a7b8"}]}.

[0075] Data format conversion unit: In order to improve the efficiency of data transmission and storage, the data format conversion unit uses the Protocol Buffers serialization protocol developed by Google to convert the above JSON format data documents into binary format. The converted data size can be reduced by more than 50%. Data verification unit: In order to ensure the integrity and consistency of the data, the data verification unit performs CRC32 verification on the converted binary data and generates a 32-bit verification code. The verification code is stored together with the binary data for subsequent data verification and error correction.

[0076] Preferably, the present application also includes a data preprocessing module. First, the data preprocessing module deduplicates student data from different sources. Since the basic information of students may be stored repeatedly in different submodules of the student management system, and daily behavior records may also generate duplicate data due to the data synchronization mechanism. The data preprocessing module identifies and merges data through key fields such as student ID, eliminates redundant data, and ensures that the information of each student is unique. Secondly, the data preprocessing module fills in data items with missing values. For example, the home address information of some students may not be registered, and some behavior records of some students may be lost due to system failure. For missing values, the data preprocessing module uses an appropriate filling strategy for estimation, such as using the mean, median, mode and other statistical values ​​of the field to fill in numerical missing values, using the most popular category label to fill in categorical missing values, or using a machine learning algorithm to predict missing values ​​based on other features.

[0077] Thirdly, the data preprocessing module removes invalid data from the original data. Invalid data includes null values, outliers, inconsistent data formats, etc. For example, the student's date of birth should not be later than the date of enrollment, and the student's behavior timestamp should not exceed the normal time range. The data preprocessing module sets reasonable threshold conditions, marks and deletes data items that do not meet the conditions, and ensures the validity and consistency of the data.

[0078] Finally, the data preprocessing module converts the cleaned multi-source heterogeneous data into a unified representation to facilitate the processing of the feature extraction module. For structured data (such as basic information and educational background), the data preprocessing module organizes it into a relational data table, where each row represents a student and each column represents an attribute field. For behavioral sequence data (such as daily behavior records), the data preprocessing module converts it into a time series format, and each student's behavior record is sorted by timestamp to form a behavior sequence. For text data (such as social network posts), the data preprocessing module performs natural language processing operations such as word segmentation, stop word removal, and part-of-speech tagging to convert the text into a bag-of-words vector or word embedding vector representation.

[0079] Figure 3It is a schematic diagram of the feature module shown in some embodiments of this specification. In the feature extraction stage, the feature extraction module performs multi-dimensional feature engineering on the preprocessed student data to construct static features, behavioral features and social features respectively. The static feature extraction unit receives the structured student data transmitted by the data preprocessing module, including the basic information data table and the academic background data table. Assume that the basic information data table contains N students and M attribute fields, and the academic background data table contains N students and P attribute fields.

[0080] For the basic information data table, the static feature extraction unit first identifies the categorical attribute fields and the numerical attribute fields. For example, gender, home address, etc. are categorical attributes, and age is a numerical attribute. For categorical attributes, the static feature extraction unit uses one-hot encoding or serial number encoding to perform feature conversion. One-hot encoding maps each categorical value to a binary vector, where the vector length is equal to the number of categorical values, the corresponding position is 1, and the rest are 0. For example, the gender attribute "male" is encoded as [1, 0], and the gender attribute "female" is encoded as [0, 1]. Serial number encoding maps each categorical value to a unique integer, such as the address attribute "XXX" is encoded as 1, "XXX" is encoded as 2, and so on. For numerical attributes, the static feature extraction unit uses the minimum-maximum normalization method to map it to the [0, 1] interval, and the formula is: Where x is the original value, min and max are the minimum and maximum values ​​of the attribute among all students. Normalization eliminates the scale differences between different numerical attributes.

[0081] For the academic background data table, the static feature extraction unit focuses on extracting the department and major information to which the student belongs. In order to convert the textual professional name into a semantic feature vector, the static feature extraction unit uses the Word2Vec word embedding method. First, the static feature extraction unit constructs a corpus of professional names and uses the professional names of all students as training samples. Then, the Skip-gram or CBOW model is used to learn the word vector representation of the professional name, mapping each professional name to a real-valued vector space of a fixed length (such as 100 dimensions). The professional word vector obtained by training can characterize the semantic similarity between majors. For example, the word vectors of "computer science" and "software engineering" are close in space.

[0082] Finally, the static feature extraction unit concatenates the one-hot encoding, serial number encoding, and normalized value of each attribute field in the basic information data table, as well as the professional word vector extracted from the education background data table, into a vector with a length of L (L = k1 + k2 + ... + k M +100), where k iIndicates the number of values ​​of the i-th classification attribute. The static feature vector comprehensively describes the inherent attribute information of the students and is the basic representation of the student portrait. The static feature vector output by the static feature extraction unit can be expressed as: [0, 1, 0, …, 0.8, 0.2, 0, …, 0.15, 0.32, …, 0.78, 0.64, …], where the first k1 elements correspond to the one-hot encoding of the gender attribute, the next k2 elements correspond to the serial number encoding of the address attribute, the next element corresponds to the normalized value of the age attribute, and the last 100 elements correspond to the word vector representation of the major name.

[0083] Figure 4 This is an exemplary flow chart for obtaining a behavior feature vector according to some embodiments of this specification. The behavior feature extraction unit receives the behavior sequence data transmitted by the data preprocessing module. There are N students, each student has T timestamped behavior records, and each record contains information such as behavior type, behavior duration, and behavior occurrence time. First, the behavior feature extraction unit performs time series analysis on the behavior sequence data of each student. Specifically, the behavior feature extraction unit counts the occurrence frequency, average duration, time interval distribution, and other indicators of each student in different behavior types. For example, for the behavior type of "borrowing books", the behavior feature extraction unit counts the number of times student A borrows books in a month (occurrence frequency), the average duration of each borrowing (average duration), the mean and variance of the borrowing time interval (time interval distribution), etc. These statistical indicators reflect the activity level and regularity of students in different behavior types, forming a behavior frequency feature.

[0084] Secondly, the behavior feature extraction unit uses frequent pattern mining algorithms, such as Apriori and FP-growth, to mine the frequent co-occurring behaviors and behavior sequence patterns in the student behavior records. For example, the Apriori algorithm can be used to find that the two behaviors of "attending class" and "homework" often appear at the same time in the behavior record of student A, forming the frequent co-occurring behavior "attending class=>homework". The FP-growth algorithm can be used to find that the three behaviors of "borrowing books", "reading", and "discussing" often appear in sequence in the behavior record of student B, forming the frequent behavior sequence pattern "borrowing books=>reading=>discussing". These frequent patterns reflect the correlation and sequence between students' behaviors, forming the characteristics of behavioral association rules.

[0085] Thirdly, the behavior feature extraction unit uses periodic detection algorithms, such as Fourier transform and autocorrelation analysis, to identify the periodic laws of student behavior. For example, the "exercise" behavior sequence of student C is subjected to Fourier transform, and it is found that its spectrum graph has a significant peak near 7 days a week, indicating that the student has regular exercise behavior every week. The "social" behavior sequence of student D is subjected to autocorrelation analysis, and it is found that its autocorrelation coefficient has a significant positive value at an interval of 30 days, indicating that the student has regular social behavior every month. These periodic laws reflect the time characteristics of student behavior and form behavioral cycle characteristics. Finally, the behavior feature extraction unit combines the above-extracted behavior frequency features, behavior association rule features, behavior cycle features, etc. to form a behavior feature vector. Assuming that a total of M behavioral statistical indicators, K frequently co-occurring behaviors, L frequent behavior sequences, and P behavior cycles are extracted, the length of the behavior feature vector is M+K+L+P. The behavior feature vector comprehensively depicts the daily behavior pattern information of students and is an important part of student portraits. The behavior feature vector output by the behavior feature extraction unit can be expressed as: [0.8, 0.2, 0.5, …, 1, 0, 0, …, 0, 1, 0, …, 7, 30, …], where the first M elements correspond to the values ​​of M behavior statistical indicators, the next K elements correspond to the occurrence or not of K frequent co-occurring behaviors (1 indicates occurrence, 0 indicates non-occurrence), the next L elements correspond to the occurrence or not of L frequent behavior sequences, and the last P elements correspond to the period lengths of P behavior periods.

[0086] Figure 5It is an exemplary flow chart for obtaining a behavior social vector according to some embodiments of this specification. The social feature extraction unit receives the social text data from the data preprocessing module. There are N students, each of whom has M social media posts or comments. Each text undergoes preprocessing steps such as word segmentation and stop word removal. First, the social feature extraction unit performs topic analysis on the social text data of each student. Specifically, the social feature extraction unit splices all the student's texts into a document, and uses topic model methods such as LDA (Latent Dirichlet Allocation) or LSA (Latent Semantic Analysis) to map the document to a low-dimensional topic space. The topic model assumes that each document is generated by a mixture of multiple topics, and each topic is a probability distribution on the vocabulary. Two important results can be obtained through the topic model: one is the document-topic distribution, that is, the probability that each document belongs to each topic; the other is the topic-vocabulary distribution, that is, the probability that each topic contains each vocabulary. Based on the document-topic distribution, the social feature extraction unit can identify the topics most discussed by students, such as "learning", "life", "entertainment", etc. Based on the topic-vocabulary distribution, the social feature extraction unit can find out the keywords under each topic. For example, the keywords under the topic of "learning" may include "homework", "exam", "grade", etc. Topics and keywords together characterize the semantic features of students' social content and form the thematic social content features.

[0087] Secondly, the social feature extraction unit performs sentiment analysis on the social text data of each student. Specifically, the social feature extraction unit uses a dictionary method or a machine learning method to determine the sentiment tendency of each text. The dictionary method is based on the sentiment dictionary, matches the sentiment words that appear in the text, and calculates the sentiment score of the text based on factors such as the part of speech and degree adverbs of the sentiment words. A score greater than 0 indicates positive sentiment, less than 0 indicates negative sentiment, and equal to 0 indicates neutral sentiment. The machine learning method trains a sentiment classification model based on annotated sentiment text corpus, and then uses the trained model to predict sentiment for new texts. Regardless of the method used, the social feature extraction unit can obtain the sentiment polarity (positive, neutral, negative) and sentiment intensity (such as positive +1, negative -2, etc.) of the student text. Sentiment polarity and sentiment intensity characterize the sentiment characteristics of the student's social content and form the sentiment social content characteristics.

[0088] Specifically, take a piece of social text data of student A as an example: "Today's exam is too difficult, I don't want to study anymore. But I still have to work hard and can't give up!" If the dictionary method is used for sentiment analysis, the social feature extraction unit first loads the predefined sentiment dictionary. The sentiment dictionary contains various sentiment words and their corresponding sentiment scores, such as "happy" +2, "sad" -2, "angry" -3, etc. Then, the social feature extraction unit performs word segmentation and part-of-speech tagging on the text to obtain "Today's exam is too difficult, I don't want to study anymore. But I still have to work hard and can't give up! " Then, the social feature extraction unit traverses each word, and if the word appears in the sentiment dictionary, its corresponding sentiment score is added to the sentiment score of the text. If a negative word such as "not" or "no" appears before the word, the sentiment score is multiplied by -1. If an adverb of degree such as "too" or "very" appears before a sentiment word, the sentiment score is multiplied by 1.5. Finally, the social feature extraction unit obtains a sentiment score of -2.5 for the text, which is less than 0. Therefore, the text is judged to have negative sentiment, and the sentiment intensity is |-2.5|=2.5.

[0089] If the machine learning method is used for sentiment analysis, the social feature extraction unit first needs to prepare annotated sentiment text corpus. Each text in the corpus is annotated with its corresponding sentiment polarity, such as "positive", "negative", and "neutral". Then, the social feature extraction unit extracts features from the corpus. Common features include bag-of-words model, TF-IDF, Word2Vec, etc. Taking the bag-of-words model as an example, the social feature extraction unit combines all the texts in the corpus into a large document, counts the number of times each word appears in it, and forms a word frequency vector. Then, the social feature extraction unit selects a suitable machine learning algorithm, such as naive Bayes, support vector machine, logistic regression, etc., and uses the extracted feature vector and the corresponding sentiment label to train the sentiment classification model. After the training is completed, the social feature extraction unit performs the same feature extraction on the text of student A to obtain the word frequency vector of the text, and then inputs it into the trained sentiment classification model for prediction. Finally, the probability that the text belongs to negative sentiment is 0.8, the probability of positive sentiment is 0.1, and the probability of neutral sentiment is 0.1. Therefore, it is judged that the text belongs to negative sentiment, and the sentiment intensity is 0.8.

[0090] Finally, the social feature extraction unit combines the topic social content features and the emotional social content features into a social feature vector. Assuming that T topics are identified through the topic model, and each topic has K keywords, the topic social content features can be represented as a T×K-dimensional matrix. Assuming that the emotional social content features include 1 emotional polarity and 1 emotional intensity, the emotional social content features can be represented as a 2-dimensional vector. The topic social content feature matrix is ​​flattened into a T×K-dimensional vector, and then the 2-dimensional emotional social content feature vector is spliced, and finally a T×K+2-dimensional social feature vector is obtained. The social feature vector comprehensively depicts the social network interaction information of students and is an important supplement to student portraits.

[0091] Specifically, the social feature extraction unit identified three main topics in the social text data of student A through the LDA topic model. Each topic contains 5 keywords. At the same time, the sentiment polarity of the student's social text is obtained through the sentiment dictionary method as negative (-, and the sentiment intensity is 2.5. Then the theme social content features and emotional social content features of student A are: Theme social content features (3×5 dimensional matrix): Topic 1: [study, 0.3], [exam, 0.2], [homework, 0.15], [grade, 0.1], [pressure, 0.05], Topic 2: [life, 0.25], [classmate, 0.2], [Dormitory, 0.15], [Canteen, 0.1], [Society, 0.05], Topic 3: [Entertainment, 0.2], [Movie, 0.15], [Music, 0.1], [Game, 0.1], [Sports, 0.05]. Emotional social content features (2-dimensional vector): [Polarity, -1], [Intensity, 2.5]. Next, the social feature extraction unit needs to flatten the theme social content feature matrix into a vector. Specifically, the social feature extraction unit concatenates the keyword probability values ​​of each theme in the order of theme 1 to theme 3 to form a 3×5=15-dimensional vector: Theme social content features ( 15-dimensional vector): [0.3, 0.2, 0.15, 0.1, 0.05, 0.25, 0.2, 0.15, 0.1, 0.05, 0.2, 0.15, 0.1, 0.1, 0.05]. Finally, the social feature extraction unit concatenates the tiled thematic social content feature vector and the emotional social content feature vector to obtain the final social feature vector with a vector length of 15+2=17 dimensions: Social feature vector (17 dimensions): [0.3, 0.2, 0.15, 0.1, 0.05, 0.25, 0.2, 0.15, 0.1, 0.05, 0.2, 0.15, 0 .1, 0.1, 0.05, -1, 2.5]. This 17-dimensional social feature vector comprehensively describes the content characteristics of student A's social media, including the main topics he is concerned about, the keyword weights under the topics, the overall emotional tendency and emotional intensity, and other information. From this vector, we can see that student A's social content mainly involves three major themes: study, life, and entertainment. Among them, the study topic is the main one, and the most mentioned keywords are "study", "exam", "life", etc. The overall emotional tendency is negative, and the emotional expression is relatively strong. This information is of great value for understanding student A's social behavior and psychological state.

[0092] The three submodules of static feature extraction, behavioral feature extraction, and social feature extraction are independent of each other and can be executed in parallel to improve the efficiency of feature engineering. The three types of feature vectors extracted respectively depict the characteristics of students' inherent attributes, daily behavior patterns, social interaction preferences, and other different perspectives, reflecting the diversity and richness of student portraits. The feature extraction module transmits the three types of feature vectors to the feature fusion module, which further learns the association and mapping between features to form a compact and comprehensive student representation vector. Multi-dimensional feature extraction and fusion can describe the comprehensive quality of students from multiple levels, both globally and locally, and provide more accurate and comprehensive feature representation for subsequent quality evaluation.

[0093] Feature fusion module, after feature extraction, each student's static features, behavioral features and social features are represented by an N-dimensional, M-dimensional and K-dimensional vector respectively. Now it is necessary to fuse the features of these three views into a unified and compact representation. The feature fusion module adopts a multi-view representation learning method based on autoencoders. Autoencoders are an unsupervised neural network model consisting of an encoder and a decoder. The encoder maps the input data to a low-dimensional latent space, and the decoder reconstructs the latent space representation into the original input. By minimizing the reconstruction error, the autoencoder can learn a high-level, abstract feature representation of the input data. Specifically, the feature fusion module designs an independent autoencoder network for the features of each view. Taking static features as an example, the corresponding autoencoder architecture is as follows: input layer (N-dimensional) -> hidden layer 1 (N / 2-dimensional) -> hidden layer 2 (N / 4-dimensional) -> encoding layer (D-dimensional) -> hidden layer 3 (N / 4-dimensional) -> hidden layer 4 (N / 2-dimensional) -> output layer (N-dimensional).

[0094] Among them, the dimension D of the encoding layer is much smaller than the original feature dimension N, which represents the compressed representation of static features in the latent space. Similarly, behavioral features and social features also learn D-dimensional latent space representations through their respective autoencoder networks. During the training phase, the feature fusion module uses unlabeled student feature data to minimize the reconstruction error of each autoencoder through the back-propagation algorithm, and iteratively updates the network parameters so that the encoding layer can restore the original features as accurately as possible. At the same time, in order to achieve the fusion of features from different views in a shared semantic space, the feature fusion module adds a similarity constraint term on top of the encoding layers of the three autoencoders to encourage the encoding representations of different views to be close to each other. Assume that the static, behavioral, and social feature encodings of the i-th student are s respectively. i , b i 、c i , then the similarity constraint can be defined as: Where ||*|| represents the Euclidean distance of the vector. Minimizing the similarity constraint can make the encoding representations of different views semantically aligned, reflecting the idea of ​​multi-view learning.

[0095] In the test phase, the feature fusion module uses the trained encoder network to map the three view features of each student to the D-dimensional latent space and obtains its corresponding encoding representation s i , b i 、c i Then, the feature fusion module uses a simple splicing method to merge the three encoding vectors into a 3D comprehensive feature vector: i =[s i ; b i ;c i ], the fused feature vector f i It contains information on students' static attributes, daily behaviors, social interactions, etc., and can comprehensively portray students' comprehensive qualities and development characteristics.

[0096] Specifically, a student's static features (such as grade and grades) are 50-dimensional, behavioral features (such as borrowing books and using computers) are 100-dimensional, and social features (such as topics and emotions) are 80-dimensional. After learning the autoencoder network, the features of these three views are compressed into a 20-dimensional latent space. The fused student comprehensive features are a 60-dimensional vector, with the first 20 dimensions representing static feature encoding, the middle 20 dimensions representing behavioral feature encoding, and the last 20 dimensions representing social feature encoding. Although the fusion feature dimension is greatly reduced, the information it contains is richer and more abstract, showing the high-level semantics of the student's status. This comprehensive feature can be further input into downstream tasks, such as student clustering and anomaly detection, to give full play to the advantages of multi-view fusion.

[0097] Evaluation prediction module, the student comprehensive feature vector output by the feature fusion module is D-dimensional, and each student corresponds to a comprehensive quality score y, which represents the overall level of the student's performance in school, with a value range of [0, 100]. Now it is necessary to predict the comprehensive quality score of the student based on the comprehensive features of the student. The evaluation prediction module adopts a three-layer feedforward neural network model with the following structure: input layer (D dimension) -> hidden layer 1 (H1 dimension) -> hidden layer 2 (H2 dimension) -> output layer (1 dimension), where the input layer receives the D-dimensional student comprehensive feature vector, the dimensions H1 and H2 of hidden layer 1 and hidden layer 2 are hyperparameters and need to be adjusted according to the actual data, and the output layer has only one node, representing the predicted comprehensive quality score. In the training stage, the evaluation prediction module uses the labeled student data (comprehensive features + comprehensive quality score) to learn the network parameters. Specifically, assuming that there are N labeled student samples, the training goal of the model is to minimize the mean square error (MSE) between the predicted score and the true score: Among them, y i is the real comprehensive quality score of the i-th student, is the model based on the comprehensive characteristics x of the i-th student i The predicted score is calculated as: Among them, f represents the feedforward neural network model, and θ represents all the parameters of the model (including the weight matrix and bias vector of each layer).

[0098] In order to minimize the loss function L, the evaluation prediction module uses the gradient descent algorithm to continuously update the model parameters θ. Specifically, in each training iteration, the model first calculates the predicted scores and loss values ​​of a small batch of student samples, and then calculates the gradient of the loss function for each parameter through the back-propagation algorithm, and finally updates the parameters according to the direction of the gradient and the learning rate. Repeat this process until the loss function converges or reaches the preset number of iterations. In the testing phase, the evaluation prediction module uses the trained model parameters to conduct a comprehensive quality evaluation of new student samples (only comprehensive features, no score labels). Specifically, for a new student feature vector x, the model inputs it into the neural network, and after layer-by-layer nonlinear transformation and parameter calculation, a predicted comprehensive quality score is finally obtained at the output layer. This score reflects the student's overall performance in all aspects during his or her time at school and can be used as a quantitative indicator of the student's development potential.

[0099] Specifically, the input of the evaluation prediction module is a 200-dimensional student comprehensive feature vector. The two hidden layers of the model are set to 100 dimensions and 50 dimensions respectively, and the output layer is 1 dimension. In the training phase, the model uses 1,000 labeled student samples for parameter learning. Each sample includes 200-dimensional comprehensive features and 1 comprehensive quality score (such as 85.5 points). After multiple rounds of iterative optimization, the mean square error of the model on the training set dropped to 2.5, indicating that the model can better fit the comprehensive quality level of students. In the testing phase, the model predicted 100 new student samples, and the comprehensive quality scores obtained showed a high correlation with the teacher's subjective evaluation, proving the effectiveness of the model.

Claims

1. A student comprehensive quality evaluation system, comprising: The collection module collects students’ static and dynamic data; Static data includes students’ basic information and academic background; dynamic data includes students’ behavioral data and social data; The feature module extracts the basic information features and educational background features of static data based on the collected data; and extracts the behavioral features and social features of dynamic data; Behavioral characteristics include behavioral frequency characteristics and behavioral cycle characteristics; social characteristics include thematic social content characteristics and emotional social content characteristics; The model module constructs and trains a comprehensive quality evaluation model based on a neural network according to the extracted features, and calculates the comprehensive quality score of the students.

2. The student comprehensive quality evaluation system according to claim 1 is characterized by: Acquisition module, including: The static data collection unit obtains the basic information and educational background of students from the student information management system; the basic information includes name, gender and age; the educational background includes major and grades; The behavior data collection unit obtains students' behavior data from the campus card usage record system; the behavior data includes attendance records; The social data collection unit obtains students' public social data from social networks through a web crawler method; the social data includes the content posted by students on the social network platform.

3. The student comprehensive quality evaluation system according to claim 2 is characterized by: The acquisition module also includes: The identification unit performs a SHA-256 hash operation on the student's ID number to generate a 256-bit hash value as the student's unique identification. Based on the obtained unique identification, the identification unit generates a unique data identification for static data, behavioral data, and social data through the MD5 hash algorithm. The association unit, based on the MongoDB database, uses the student's unique identifier as the id field of the document, stores static data, behavioral data, social data and corresponding data identifiers as nested objects of the document, and generates student-centered data records in JSON format.

4. The student comprehensive quality evaluation system according to claim 3 is characterized by: The acquisition module also includes: The data format conversion unit uses the Protocol Buffers serialization protocol to convert the associated JSON format data records into binary format; The data verification unit uses the CRC32 verification algorithm to perform integrity verification on the converted binary format data records and generate a verification code.

5. The student comprehensive quality evaluation system according to any one of claims 1 to 4, characterized in that: Feature modules, including: The static feature extraction unit uses a regular expression matching method to extract the basic information features and educational background features of the students; the basic information features include gender features and age features; the educational background features include major category features and grade features; the extracted basic information features and educational background features are combined into a static feature vector; The behavior feature extraction unit uses time series analysis and frequent pattern mining methods to extract students' behavior features, which include behavior frequency features and behavior period features.

6. The student comprehensive quality evaluation system according to claim 5 is characterized by: The behavior feature extraction unit performs the following steps: Convert the data into a time series format. Each time series represents a student's behavior record. Each element of the time series contains the timestamp and behavior type of the behavior. The sliding window method is used to segment the time series at different time granularities to obtain multiple subsequences. The time granularity includes day, week and month. The size and step length of the sliding window are set according to the time granularity. Frequent pattern mining is performed on each subsequence to obtain frequent item sets and sequential patterns of student behaviors; frequent item sets represent combinations of behavior types that appear simultaneously with a frequency greater than a threshold; sequential patterns represent sequences of behavior types that appear continuously with a frequency greater than a threshold; frequent pattern mining uses the Apriori algorithm; According to the acquired frequent item sets and sequential patterns, a counting-based method is used to count the number of behaviors of each student in sliding windows of different time granularities, and the average daily and weekly behavior frequencies are calculated as behavior frequency features; According to the acquired frequent item sets and sequential patterns, a method based on periodic analysis was adopted, using Fourier transform to perform frequency domain analysis on the behavioral time series of each student, and extract the behavioral distribution with a period of one week and one month as the behavioral periodic characteristics; Among them, Fourier transform converts the time series from time domain to frequency domain, and judges the periodic law of behavior by analyzing the main frequency components in the frequency domain; The extracted behavior frequency features and behavior period features are organized into behavior feature vectors.

7. The student comprehensive quality evaluation system according to claim 6 is characterized by: Feature modules also include: The social feature extraction unit uses natural language processing methods to extract students' social content features.

8. The student comprehensive quality evaluation system according to claim 7 is characterized by: Social feature extraction unit, including: The social feature extraction unit includes a topic extraction subunit, a sentiment analysis subunit, and a feature organization subunit; The topic extraction subunit uses the latent Dirichlet allocation algorithm to divide the post content into several topics by representing the post content as a document-topic probability distribution and a topic-term probability distribution, and estimating and judging the probability distribution. The document-topic probability distribution represents the probability that each post belongs to each topic, and the topic-term probability distribution represents the probability that each topic contains each term. The probability distribution of each student discussing different topics is calculated as the feature of the topic social content. The sentiment analysis subunit uses a dictionary-based approach to classify the post content into positive, negative, and neutral sentiment categories by constructing a sentiment dictionary and matching sentiment words in the post content; the proportion of posts with different sentiment categories for each student is calculated as the sentiment social content feature; The feature organization subunit sequentially splices the obtained thematic social content features and emotional social content features into a vector using a feature splicing method, and normalizes the spliced ​​vector using a minimum-maximum normalization method to obtain a social feature vector.

9. The student comprehensive quality evaluation system according to claim 8 is characterized by: Feature modules also include: The feature fusion unit fuses the static feature vector, the behavioral feature vector and the social feature vector to obtain the comprehensive feature vector of the student.

10. The student comprehensive quality evaluation system according to claim 9 is characterized by: Feature fusion unit, including: static feature encoder, behavior feature encoder, social feature encoder, feature mapping layer and feature concatenation layer; The static feature encoder uses an autoencoder network to encode the static feature vector, and extracts the high-level semantic representation of the static feature through unsupervised learning to obtain the static feature view; the autoencoder is composed of a multi-layer feedforward neural network, including an input layer, a hidden layer, and an output layer; the input layer receives the static feature vector, the hidden layer encodes the input feature to obtain a compressed representation of the static feature, and the output layer decodes the compressed representation to obtain a reconstructed static feature vector; The behavior feature encoder uses an autoencoder network to encode the behavior feature vector and extracts the high-level semantic representation of the behavior feature through unsupervised learning to obtain the behavior feature view. The social feature encoder uses an autoencoder network to encode the social feature vector and extracts the high-level semantic representation of the social features through unsupervised learning to obtain the social feature view. The feature mapping layer uses a feedforward neural network to map static feature views, behavioral feature views, and social feature views into a shared semantic space, and learns the common representation of different views by minimizing the differences between them. The feature splicing layer uses a vector splicing method to splice the static feature view, behavioral feature view, and social feature view output by the feature mapping layer in a shared semantic space to obtain the student's comprehensive feature vector; each element in the comprehensive feature vector represents a high-level semantic feature, which integrates the student's feature information in different dimensions.

Images