Micro-learning data refining and enhancing method for improving evolution rule generator

Abstract

The invention relates to the technical field of data analysis, in particular to a micro-learning data refining and enhancing method for an improved evolution rule generator, which comprises the following steps: constructing the improved evolution rule generator; an improved evolution rule generator is adopted, and association rules are extracted from the micro-learning data; and refining and enhancing the micro-learning data according to the association rule. Wherein the improved evolution rule generator comprises an external cycle and an internal cycle; the external loop runs a genetic algorithm used for association rule generation, and the variation process of the genetic algorithm is driven by the internal loop to optimize individuals; and the artificial bee colony algorithm is circularly operated inside to add and integrate the micro-learning data into reproduction of the association rule. According to the scheme, the embedded EA is adopted to discover, generate and construct the hidden association rules, so that the problem that a large number of association rules cannot be found due to scarcity and scarcity of micro-open learning data sources is solved, and the requirements of data refinement and enhancement are met.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a micro-learning data refining and enhancement method for an improved evolutionary rule generator. Background Art

[0002] In the ever-changing electronic society, the way people access and consume learning resources is very different from the traditional way. In the emerging open learning, the time frame for people to access open educational resources (OER) on the move is smaller, or in other words, the scattered time of learning is fragmented. Many previous works have shown that knowledge can be effectively learned through micro-learning; micro-learning, which can also be understood as fragmented learning, is different from systematic course learning. It is not restricted by time and space, content, or teaching. You can learn anytime and anywhere, and learn whenever you want.

[0003] As the rise of open educational resources has gained widespread support throughout higher education and adult education, these emerging learning paradigms have attracted the attention of many researchers from educational, social and computational perspectives. The use of open educational resources in micro-learning mode is consistent with the characteristics of a modern electronic society, in which mobile and pervasive computing dominate; as an emerging educational phenomenon, micro-learning is more user-centered and requires a different learning schedule, different from campus learning, and even different from standard distance learning and mobile learning;

[0004] Meanwhile, the combination of artificial intelligence (AI) and online / mobile learning has been widely studied to promote the recommendation and adaptation of personalized learning resources, however, mature products are rarely found in the education field, especially the provision of micro-open educational resources. This situation can be attributed to the fact that AI is inherently data-hungry, while the availability, readiness and quality of data in education, especially micro-open learning, are relatively low.

[0005] To ensure a balance between accuracy and interpretability, the refinement and enhancement of data highly rely on association rules that co-occur with learning activities, but due to reasons such as data shortage, the current demand for building recommendation and adaptation models to drive personalized micro-open educational resource delivery cannot be met by adopting well-discovered and manually designed off-the-shelf rules.

[0006] Therefore, there is an urgent need for a micro-learning data refinement and enhancement method that improves the evolutionary rule generator, using nested EA to discover, generate and construct hidden association rules, to solve the problem of being unable to discover a large number of association rules caused by the scarcity and rarity of micro-open learning data sources, so as to meet the needs of data refinement and enhancement. Summary of the invention

[0007] The present invention aims to provide a micro-learning data refinement and enhancement method with an improved evolutionary rule generator, which adopts nested EA to discover, generate and construct hidden association rules, and solves the problem of being unable to discover a large number of association rules caused by the scarcity and rarity of micro-open learning data sources, so as to meet the needs of data refinement and enhancement.

[0008] The present invention provides the following basic scheme: a micro-learning data refinement and enhancement method for improving the evolutionary rule generator, comprising:

[0009] Build and improve the evolutionary rule generator;

[0010] An improved evolutionary rule generator is used to extract association rules from micro-learning data;

[0011] According to association rules, micro-learning data is refined and enhanced.

[0012] The improved evolutionary rule generator includes: an outer loop and an inner loop;

[0013] The outer loop runs a genetic algorithm for association rule generation, and the mutation process of the genetic algorithm is driven by the inner loop to optimize individuals; the inner loop runs an artificial bee colony algorithm to integrate micro-learning data augmentation into the reproduction of association rules.

[0014] Furthermore, the genetic algorithm for generating association rules includes:

[0015] S101. Generate frequent itemsets and rare itemsets using the entire micro-learning dataset and minimum support; initialize iteration conditions;

[0016] S102, coding individuals using the Michigan method;

[0017] S103, using the set probability to select individuals from the entire population as ancestors;

[0018] S104, iteratively performing crossover and mutation on the ancestors to generate offspring;

[0019] S105, determine whether the iteration termination condition is met, if so, execute S1010; if not, execute S106;

[0020] S106. For each offspring O i ∈O, randomly select two individuals and compare their fitness;

[0021] S107. Use individuals with higher fitness to produce offspring through crossover and mutation;

[0022] S108, training and evaluating offspring using a comprehensive fitness function;

[0023] S109, putting the offspring into the population, replacing the individual with the lowest fitness as the ancestor, updating the number of iterations, and executing S104;

[0024] S1010: Output the generated association rules.

[0025] Furthermore, the comprehensive fitness function is:

[0026]

[0027] Where Sup(O i ) for the offspring O i The support degree of Conf(O i ) for the offspring O i Confidence, Int(O i ) for the offspring O i Interest level, Com(O i ) for the offspring O i understandability;

[0028] w1, w2, w3, and w4 are the weights corresponding to the four measurements;

[0029] For θ i Normalize it so that its value is in the range of [0,1].

[0030] Furthermore, n association rules are expressed in the form of A→B, where A and B are discontinuous item sets;

[0031] Both A and B belong to the entire data source I;

[0032] Support and confidence are used to measure the strength of association rules, which are expressed as:

[0033]

[0034]

[0035] The interest level is the lift, expressed as:

[0036]

[0037] The measure of comprehensibility is expressed as:

[0038]

[0039] Furthermore, the artificial bee colony algorithm includes:

[0040] S401, scanning all offspring O generated during the crossover and mutation process of the genetic algorithm i ; Initialize iteration conditions;

[0041] S402, for each offspring O i , check O i Whether it includes: MRI, MFI or MZG;

[0042] S403, if O i Including MRI, the rare item set and rare association rule discovery method are used to generate rare association rules, and its closure C i ;

[0043] S404: If Oi contains a negative item set, the negative association rule discovery method is used to generate the negative association rule to generate the enhanced item set A. i , and evaluate C according to the comprehensive fitness function i or A i Adaptability;

[0044] S405. If the fitness meets the preset fitness, C i or A i Added as a starter solution;

[0045] S406, flip C i or A i The itemsets in , generate optimized offspring;

[0046] S407, replace the current offspring with the optimized offspring; update NO i NO i e ;

[0047] S408, determine whether the iteration termination condition is met, if yes, execute S4016; if no, execute S409;

[0048] S409, selecting a new solution according to probability;

[0049] S4010, Flip C i or A i The itemsets in , generate optimized offspring;

[0050] S4011, evaluate the fitness of the optimized offspring;

[0051] S4012, replacing the old offspring with the optimized offspring;

[0052] S4013, if the fitness of the updated offspring is not improved compared with the old offspring, then new offspring are generated according to the preset model; i.e., the wandering bee stage;

[0053] S4014, evaluate the fitness of new offspring;

[0054] S4015, replace the old offspring with the new offspring, update the number of iterations, and execute S408; #Update NO i ;

[0055] S4016. Treat the new offspring as offspring produced by mutation.

[0056] Further, the following formula is adopted in S406 to flip C i or A i Itemsets in :

[0057]

[0058] The S4010 adopts the following flip C i or A i The itemsets in include:

[0059]

[0060] Furthermore, in S409, the probability generated by the following formula is used to select a new solution:

[0061]

[0062] where i∈{1,2,…,N}, N is the size of the population.

[0063] Further, the preset model is:

[0064]

[0065] in and are the upper and lower limits of the neighborhood search, and rand(0,1) is a random value between 0 and 1.

[0066] Furthermore, the rare item set and rare association rule discovery method includes:

[0067] Scan the entire microlearning dataset;

[0068] According to the support, the itemsets are divided into non-existing itemsets, frequent itemsets, and rare itemsets;

[0069] When the sets of frequent itemsets and rare itemsets are not empty, repeat the following steps:

[0070] Generate item sets;

[0071] Merge all rare itemsets to form R;

[0072] For each item set R j :

[0073] ScanRj all subsets of ;

[0074] If R j All subsets of are frequent itemsets, then R j Set as MRI;

[0075] For each MRI:

[0076] Find the closure of MRI, denoted by

[0077] Generate rare association rules

[0078] Furthermore, the negative association rule discovery method includes:

[0079] Frequent itemsets and rare itemsets generated using the entire micro-learning dataset and minimum support;

[0080] For each frequent itemset F i :

[0081] For F i For every pair of X and Y in and

[0082] If Confidence(X→Y)≥Minimum Confidence&&Lift(X→Y)≥1, then output rule X→Y and record it as the positive association rule PAR; where Confidence is the confidence, Minimum Confidence is the minimum confidence, and Lift is the lift;

[0083] On the contrary, if Output rules Record as negative association rule;

[0084] like Output rules Record as negative association rule;

[0085] like Output rules Record as negative association rule;

[0086] For each rare itemset R i :

[0087] For R i For every pair of X and Y in R, X∩Y=R i and At the same time, Support(X)>MinimumSupport, Support(Y)>MinimumSupport: where is the support;

[0088] If Confidence(X→Y)≥Minimum Confidence&&Lift(X→Y)≥1, then output rule X→Y and record it as a positive association rule;

[0089] On the contrary, if Output rules Record as negative association rule;

[0090] like Output rules Record as negative association rule;

[0091] like Output rules Record as negative association rule;

[0092] Output positive association rules and negative association rules.

[0093] Beneficial effects of this scheme: Taking into account the current scarcity of data and rules representing open learning activities, this scheme constructs an evolutionary rule generator consisting of an outer loop and an inner loop; the outer loop runs a genetic algorithm (GA) to generate association rules, which can be effectively generated from a small number of available data sources in micro-open learning scenarios; while the inner loop optimizes the generated candidate recommendation alternative resources by considering rare and negative association rules (NARs), etc.; these optimization rules are further applied to refine and expand the data representing learners' open learning behavior into a low-dimensional, descriptive and interpretable form.

[0094] As a new and eye-catching learning scenario, the construction of micro-open learning recommendation mechanism relies heavily on the association rules hidden behind learner behaviors in open learning activities. Therefore, a new method of discovering, generating and constructing hidden association rules is used using nested EA (purification algorithm); usually using a small amount of data, EA has been applied to optimize the discovery of a limited number of rules and the construction of existing rules, so as to generate more derivatives through the evolutionary process, thereby greatly supplementing the data and reducing the impact caused by the scarcity and rarity of data sources. It is helpful for the current research and application of artificial intelligence in educational data.

[0095] In summary, this scheme uses nested EA to discover, generate and construct hidden association rules to solve the problem of being unable to discover a large number of association rules caused by the scarcity and rarity of micro-open learning data sources, so as to meet the needs of data refinement and enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 A schematic flow chart of an embodiment of a micro-learning data refining and enhancement method for improving an evolutionary rule generator according to the present invention;

[0097] Figure 2 It is a logistic regression prediction view in an embodiment of a micro-learning data refinement and enhancement method for improving an evolutionary rule generator of the present invention;

[0098] Figure 3 This is a graph showing the accuracy of linear discriminant analysis of predicted visual field in an embodiment of a micro-learning data refinement and enhancement method for improving an evolutionary rule generator of the present invention;

[0099] Figure 4 This is a graph showing the accuracy of logistic regression for exploration prediction in an embodiment of a micro-learning data refinement and enhancement method for improving an evolutionary rule generator of the present invention;

[0100] Figure 5 This is a graph showing the accuracy of linear discriminant analysis for predictive exploration in an embodiment of a micro-learning data refinement and enhancement method for an improved evolutionary rule generator of the present invention. DETAILED DESCRIPTION

[0101] The following is further described in detail through specific implementation methods:

[0102] Description: GA: genetic algorithm;

[0103] NAR: Rare and Negative Association Rule;

[0104] OER: Open Educational Resources;

[0105] EA: Evolutionary Algorithm;

[0106] PLE: Personalized Learning Environment;

[0107] ARM: Micro-learning association rule mining;

[0108] MLaaS: Micro-learning as a service;

[0109] PLE: Personalized Learning Environment;

[0110] SaaS: Smart Micro Open Learning Software as a Service;

[0111] EDM: Educational Data Mining;

[0112] LA: Learning Analytics;

[0113] OERs: non-microlearning;

[0114] RAR: association rules;

[0115] MFI: Maximum frequent itemset;

[0116] MRI: Minimal rare itemset;

[0117] MZG: Minimal Zero Generator.

[0118] The embodiment is basically as shown in the attached Figure 1 A micro-learning data refinement and enhancement method for improving an evolutionary rule generator is shown, comprising:

[0119] Build and improve the evolutionary rule generator;

[0120] The improved evolutionary rule generator is used to extract association rules (RAR) from micro-learning data. The association rules include positive association rules (PAR) and negative association rules (NAR).

[0121] According to association rules, micro-learning data is refined and enhanced.

[0122] The improved evolutionary rule generator includes: an outer loop and an inner loop;

[0123] The outer loop runs the genetic algorithm for association rule generation, and the mutation process of the genetic algorithm is driven by the inner loop to optimize individuals;

[0124] The inner loop runs the artificial bee colony algorithm to integrate micro-learning data augmentation into the reproduction of association rules to obtain better adaptability;

[0125] In the outer loop, the genetic algorithm is used to generate the rules of interest. After the encoding process is completed, the genetic algorithm is applied for multi-objective processing to extract the association rules of interest from the frequent itemsets and rare itemsets.

[0126] The above genetic algorithm for generating association rules includes:

[0127] S101. Generate frequent itemsets and rare itemsets using the entire micro-learning dataset and minimum support; initialize iteration conditions;

[0128] S102, coding individuals using the Michigan method;

[0129] S103, using the set probability to select individuals from the entire population as ancestors;

[0130] S104, iteratively performing crossover and mutation on the ancestors to generate offspring;

[0131] S105, determine whether the iteration termination condition is met, if so, execute S1010; if not, execute S106;

[0132] S106. For each offspring O i ∈O, randomly select two individuals and compare their fitness;

[0133] S107. Use individuals with higher fitness to produce offspring through crossover and mutation;

[0134] S108, training and evaluating offspring using a comprehensive fitness function;

[0135] S109, putting the offspring into the population, replacing the individual with the lowest fitness as the ancestor, updating the number of iterations, and executing S104;

[0136] S1010, outputting the generated association rules;

[0137] The fitness is calculated according to the set comprehensive fitness function;

[0138] Comprehensive fitness function:

[0139]

[0140] Where Sup(O i ) for the offspring O i The support degree of Conf(O i ) for the offspring O i Confidence, Int(O i ) for the offspring O i Interest level, Com(O i ) for the offspring O i understandability;

[0141] w1, w2, w3, and w4 are weights corresponding to the four measurements (i.e., support, confidence, interest, and comprehensibility);

[0142] For θ i Normalize it so that its value is in the range of [0,1];

[0143] In this scheme, association rules and confidence support framework, n association rules are expressed in the form of A→B, where A and B are discontinuous item sets, i.e., they have nothing in common;

[0144] An itemset is defined as: a set of learning activities performed in an open learning scenario, where an itemset is defined as a learning activity consisting of a number of open educational resources accessed within it;

[0145] Given that item sets A and B are disjoint, let their union set T represent the set of all micro-open learning activities, i.e., transactions in data mining;

[0146] In addition, they all belong to the entire data source I; the intuitive meaning of the association rule between the antecedent A and the subsequent B is: if A occurs, then B may also occur; according to the typical IF-THEN structure, the strength of the association rule is measured by support and confidence, which are expressed as:

[0147]

[0148]

[0149] If the support of any item set is lower than the preset minimum support, but greater than or equal to the minimum frequency support threshold, the item is defined as a rare item (Das, 2015).

[0150] Interest is promotion:

[0151]

[0152] If the confidence value of an association rule is greater than the support value, then the association rule is of interest;

[0153] The measure of understandability determines the importance of the rule according to the number of itemsets presented in the antecedent and subsequent domains, which can be expressed as:

[0154]

[0155] The above artificial bee colony algorithm includes:

[0156] S401, scanning all offspring O generated during the crossover and mutation process of the genetic algorithm i ; Initialize iteration conditions;

[0157] S402, for each offspring O i , check O i Whether it includes: MRI, MFI or MZG;

[0158] S403, if O i Including MRI, the rare item set and rare association rule discovery method are used to generate rare association rules, and its closure C i ;

[0159] S404. If Oi contains a negative item set (i.e., the learning activities missing in T), the negative association rule discovery method is used to generate the negative association rules to produce the enhanced item set A. i , and evaluate C according to the comprehensive fitness function i or A iAdaptability;

[0160] S405. If the fitness meets the preset fitness, C i or A i Add as initial solution (i.e. food source);

[0161] S406, flip C i or A i The item sets in generate optimized offspring; that is, the hiring bee stage

[0162] Specifically, use the following formula to flip C i or A i Itemsets in :

[0163]

[0164] S407, replace the current offspring with the optimized offspring; that is, update NO i NO i e ; NO i The new O i ;

[0165] S408, determine whether the iteration termination condition is met, if yes, execute S4016; if no, execute S409;

[0166] S409, selecting a new solution based on probability; i.e., the scout bee stage;

[0167] Specifically, the new solution is selected using the probability generated by the following formula:

[0168]

[0169] Where i∈{1,2,…,N}, N is the size of the population; for the onlooker, the possibility of selecting a source of information from the candidate data is determined by the probability generated by the above formula; where the population is the set of all solutions;

[0170] S4010, Flip C i or A i The itemsets in , generate optimized offspring;

[0171] Specifically, use the following formula to flip C i or A i The itemsets in include:

[0172]

[0173] S4011, evaluate the fitness of the optimized offspring;

[0174] S4012, replacing the old offspring with the optimized offspring;

[0175] S4013, if the fitness of the updated offspring is not improved compared with the old offspring, then new offspring are generated according to the preset model; i.e., the wandering bee stage;

[0176] The preset models are:

[0177]

[0178] in and are the upper and lower limits of the neighborhood search, respectively, and rand(0,1) is a random value between 0 and 1;

[0179] S4014, evaluate the fitness of new offspring;

[0180] S4015, replace the old offspring with the new offspring, update the number of iterations, and execute S408; #Update NO i ;

[0181] S4016. Treat the new offspring as offspring produced by mutation;

[0182] The above rare item sets and rare association rule discovery methods are:

[0183] Scan the entire microlearning dataset;

[0184] According to the support, the itemsets are divided into non-existing itemsets, frequent itemsets, and rare itemsets;

[0185] When the sets of frequent itemsets and rare itemsets are not empty, repeat the following steps:

[0186] Generate item sets;

[0187] Merge all rare itemsets to form R;

[0188] For each item set R j :

[0189] ScanR j all subsets of ;

[0190] If R j All subsets of are frequent itemsets, then R j Set as MRI;

[0191] For each MRI:

[0192] Find the closure of MRI, denoted by

[0193] Generate rare association rules

[0194] The above negative association rule discovery method is:

[0195] Frequent itemsets and rare itemsets generated using the entire micro-learning dataset and minimum support;

[0196] For each frequent itemset F i :

[0197] For F i For every pair of X and Y in and

[0198] If Confidence(X→Y)≥Minimum Confidence&&Lift(X→Y)≥1, then output rule X→Y and record it as the positive association rule PAR; where Confidence is the confidence, Minimum Confidence is the minimum confidence, and Lift is the lift;

[0199] On the contrary, if Output rules Record as negative association rule;

[0200] like Output rules Record as negative association rule;

[0201] like Output rules Record as negative association rule;

[0202] For each rare itemset R i :

[0203] For R i For every pair of X and Y in R, X∩Y=R i and At the same time, Support(X)>MinimumSupport, Support(Y)>MinimumSupport: where is the support;

[0204] If Confidence(X→Y)≥Minimum Confidence&&Lift(X→Y)≥1, then output rule X→Y and record it as a positive association rule;

[0205] On the contrary, if Output rules Record as negative association rule;

[0206] like Output rules Record as negative association rule;

[0207] like Output rules Record as negative association rule;

[0208] Output positive association rules and negative association rules.

[0209] The above algorithms are responsible for the generation of CAR by generating knowledge representation in the form of A → B. This has a similar working principle, where A is the connection of the predicted value and B is the predicted class instead of the rule.

[0210] Feature generation is different from data refinement in a single genetic algorithm and data enhancement in an artificial bee colony algorithm. Using a nested evolutionary rule generator can generate better quality data, thereby further pushing the boundaries of universality, especially micro-learning.

[0211] Feature generation is activated by representing the profile of the learner with a vector; first, a CAR-based method is applied to identify the learner's learning style and filter out the association rules related to it, and then the method is applied to produce a set of refined behavioral features to evaluate or predict complementary features with each learner's values; the working principle is that each vector produces an empty feature set at the beginning of the iteration, so the rules that best suit the primary learning style (whether generated or original) are checked, verified and adopted; these rules fill the candidate features into the vector as the first and support the acceptance of a prominent feature as the first feature after comparison; by referring to these rules again, the feature is connected into a complete feature-value pair, and its current value is retained or replaced with a new predicted value.

[0212] The exponential growth in computational complexity is too high to try all possible eigenvalues; therefore, this feature generation method is heuristic and greedy, and if a new eigenvalue pair is added to the vector representing a learner's profile, it will not be modified or removed through backtracking. Similarly, the feature generation process keeps adding the second eigenvalue pair and subsequent eigenvalue pairs to the vector until a termination condition is reached.

[0213] AI-based educational computing builds and optimizes the intelligent delivery mechanism of adaptive micro-open educational resources by iteratively consuming refined and enhanced data and supporting evidence-based decision-making processes.

[0214] In this embodiment, micro-learning data in a structured dataset is used;

[0215] The candidate solutions of the association rules are encoded into a representation that can be understood by a computer; taking the learning activity dataset (micro-learning data) of Luna et al. as an example, part of the description of the learning activity is shown in Table 1;

[0216] The association rules extracted from the dataset are described as follows:

[0217] IFTime_on_forum=Yes AND Number_read=High

[0218] THEN Final_grade = pass.

[0219] Its encoding method is shown in Table 2:

[0220] Table 1: Example description of the learning activity dataset

[0221]

[0222] Table 2: Examples of encoded association rules

[0223]

[0224] Table 3: Learning activity data - positive and negative items

[0225]

[0226] In the inner loop of EA (i.e., ABC algorithm operation), the data is augmented by combining existing item sets with non-current learning activities to generate candidate item sets; NAR is used to enhance learning activities, and the data is presented through the sample learning activity sets containing six types of transactions (i.e., learning activities) in Table 3. "Original learning activities" refer to the learning activities performed by each learner and are recorded as resources accessed by each learner in the database. The labels used for learning activities include: A for forum posting, B for video playback, C for teaching resource browsing, D for general content browsing, E for URL clicks to subpages, and F for quiz attempts. In addition, "enhanced learning activities" illustrate the current items and their corresponding negative items (i.e., lack of learning activities).

[0227] According to the definition in the “Feature Generation” section, the CAR-based approach is applied to identify the learner’s learning style before adding feature value pairs to the vector representing the learner’s personal features. Taking the work of Truong and Jeyakumar and Sreenath as an example, Felder-Silverman’s learning style mainly exists in the literature of computer-assisted learning, including visual, verbal, active, reflective, sensitive, intuitive, etc. Before starting learning, no quantitative survey of learners’ learning styles was conducted, and concepts were borrowed from the work to determine learners’ learning styles by collecting “hints” in learning activities. For example, a learner’s high access to exercises can be attributed to the hint of active learning style; if learners stay longer on abstract materials such as theory, while learners are considered to be more sensitive by consuming more practical content, learners tend to be intuitive. Therefore, the hints of Felder-Silverman’s learning style are summarized. By presetting thresholds to distinguish whether students have low, medium, and high incidence of behaviors during learning, a digital hint formula for learning styles is designed. Thereafter, an example of CAR mining for learning style identification for adaptive micro-open learning is as follows:

[0228] IF "Intuitive" = [2.42, 4.77]

[0229] THEN"Verbal"value=[3.23,3.47]

[0230] Table 4: Example results of EA-based rule generator operation

[0231]

[0232] In simple terms, this rule refers to the fact that if the value of a student's intuitive learning style is within the mentioned score range, then he / she tends to be a verbal learner. It is worth noting that only the antecedent can be a combination of multiple attribute values. Since the teaching philosophy adopted in this program does not allow for the overlap of learning styles, it should be a unique class. For example, the CAR below indicates that if a learner's verbal and visual styles are assessed within a given range, then the learner is an intuitive learner who may prefer to consume abstract and raw information and be theory-oriented.

[0233] Results are evaluated in a semantic manner, and the rules discovered and newly generated from the selected dataset conforming to the standard IF-THEN structure are evaluated. In the MITx dataset, many transactions carry an extremely high number of “VideoPlayed”, i.e., 197757, which is the total number of videos available in MITx. These noisy transactions are filtered out of the EA operation; Table 4 shows the results of a sample run of the EA-based rule generator.

[0234] The quality of data refinement and enhancement is assessed using cross-validation:

[0235] The original profile of learner i is in vector V i A and V i R Represent the enhanced and improved features of the learner respectively. i A ∈VA,V i R ∈VR For each learner’s record in the dataset, binary-valued “view” and “exploration” fields are predicted for accuracy evaluation; the value of view is 0 or 1, if the learner has less than 1 interaction in the course, the value of view is 0; the value of exploration is 0 or 1, if the learner has less than half of the interaction or viewing of the course content, the value of view is 0. Rule-based prediction is abandoned because the rules of the evolutionary generator refine and enhance the data. Therefore, we adopt basic logistic regression (LR) and linear discriminant analysis (LDA) to predict the target field of view and exploration speed. Figure 2-5 The performance of the selected methods is compared in both cases. The ascending scale of the learner set is shown along the X-axis, with a scale of 1000 learners per unit, and the prediction accuracy is on the Y-axis. The learner vector contains only the original data, represented by the VO line, VR represents the refined data, VA represents the augmented data, and V-RA means that the learner vector contains both refined and augmented data.

[0236] After executing two cycles of the evolutionary rule generator to drive feature generation, each learner profile is refined into 11 features summarized from high-dimensional data by assigning redundant information. Figures 2 to 5 , which shows that each line oscillates gently without showing a significant difference between its maximum and minimum values; this can be attributed to the fact that the performance of feature generation does not completely depend on the size of the learner ensemble.

[0237] By evaluating the quality of predictions made by the four types of data, the line representing the accuracy of the predictions generated by the original data (i.e., VO) always appears at the bottom, and the gap between VO and the other three lines in each graph is representable.

[0238] Despite Figure 4 and Figure 5There is almost no intersection of VR and VA trend lines in the , but the only adoption of augmented data generally outperforms the exact data by predicting the higher accuracy of views and explorations. However, since V-RA achieves the best prediction performance in terms of both method metrics and target metrics, the approach combining data refinement and augmentation outperforms in prediction performance. Therefore, this scheme can significantly improve the quality of behavioral data.

[0239] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A micro-learning data refinement and enhancement method for improving evolutionary rule generator, characterized in that: include Build and improve the evolutionary rule generator; An improved evolutionary rule generator is used to extract association rules from micro-learning data; According to association rules, micro-learning data is refined and enhanced. The improved evolutionary rule generator includes: an outer loop and an inner loop; The outer loop runs a genetic algorithm for association rule generation, and the mutation process of the genetic algorithm is driven by the inner loop to optimize individuals; the inner loop runs an artificial bee colony algorithm to integrate micro-learning data augmentation into the reproduction of association rules.

2. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 1 is characterized in that: The genetic algorithm for generating association rules comprises: S101. Generate frequent itemsets and rare itemsets using the entire micro-learning dataset and minimum support; initialize iteration conditions; S102, coding individuals using the Michigan method; S103, using the set probability to select individuals from the entire population as ancestors; S104, iteratively performing crossover and mutation on the ancestors to generate offspring; S105, determine whether the iteration termination condition is met, if so, execute S1010; if not, execute S106; S106. For each offspring O i ∈O, randomly select two individuals and compare their fitness; S107. Use individuals with higher fitness to produce offspring through crossover and mutation; S108, training and evaluating offspring using a comprehensive fitness function; S109, putting the offspring into the population, replacing the individual with the lowest fitness as the ancestor, updating the number of iterations, and executing S104; S1010: Output the generated association rules.

3. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 2 is characterized in that: The comprehensive fitness function is: Where Sup(O i ) for the offspring O i The support degree of Conf(O i ) for the offspring O i Confidence, Int(O i ) for the offspring O i Interest level, Com(O i ) for the offspring O i understandability; w1, w2, w3, and w4 are the weights corresponding to the four measurements; For θ i Normalize it so that its value is in the range of [0,1].

4. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 3 is characterized in that: n association rules are expressed in the form of A→B, where A and B are discontinuous item sets; Both A and B belong to the entire data source I; Support and confidence are used to measure the strength of association rules, which are expressed as: The interest level is the lift, expressed as: The measure of comprehensibility is expressed as:

5. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 4 is characterized in that: The artificial bee colony algorithm comprises: S401, scanning all offspring O generated during the crossover and mutation process of the genetic algorithm i ; Initialize iteration conditions; S402, for each offspring O i , check O i Whether it includes: MRI, MFI or MZG; S403, if O i Including MRI, the rare item set and rare association rule discovery method are used to generate rare association rules, and its closure C i ; S404: If Oi contains a negative item set, the negative association rule discovery method is used to generate the negative association rule to generate the enhanced item set A. i , and evaluate C according to the comprehensive fitness function i or A i Adaptability; S405. If the fitness meets the preset fitness, C i or A i Added as a starter solution; S406, flip C i or A i The itemsets in , generate optimized offspring; S407, replace the current offspring with the optimized offspring; update NO i for S408, determine whether the iteration termination condition is met, if yes, execute S4016; if no, execute S409; S409, selecting a new solution according to probability; S4010, Flip C i or A i The itemsets in , generate optimized offspring; S4011, evaluate the fitness of the optimized offspring; S4012, replacing the old offspring with the optimized offspring; S4013, if the fitness of the updated offspring is not improved compared with the old offspring, then new offspring are generated according to the preset model; i.e., the wandering bee stage; S4014, evaluate the fitness of new offspring; S4015, replace the old offspring with the new offspring, update the number of iterations, and execute S408; #Update NO i ; S4016. Treat the new offspring as offspring produced by mutation.

6. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 5, characterized in that: The following formula is used in S406 to flip C i or A i Itemsets in : The S4010 adopts the following flip C i or A i , include:

7. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 5, characterized in that: In S409, a new solution is selected using the probability generated by the following formula: where i∈{1,2,…,N}, N is the size of the population; the population is the set of all solutions.

8. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 5, characterized in that: The preset model is: in and are the upper and lower limits of the neighborhood search, and rand(0,1) is a random value between 0 and 1.

9. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 5, characterized in that: The rare item set and rare association rule discovery method comprises: Scan the entire microlearning dataset; According to the support, the itemsets are divided into non-existing itemsets, frequent itemsets, and rare itemsets; When the sets of frequent itemsets and rare itemsets are not empty, repeat the following steps: Generate item sets; Merge all rare itemsets to form R; For each item set R j : ScanR j all subsets of ; If R j All subsets of are frequent itemsets, then R j Set as MRI; For each MRI: Find the closure of MRI, denoted by Generate rare association rules 10. The micro-learning data refinement and enhancement method for improving the evolutionary rule generator according to claim 5, characterized in that: The negative association rule discovery method comprises: Frequent itemsets and rare itemsets generated using the entire micro-learning dataset and minimum support; For each frequent itemset F i : For F i For every pair of X and Y in and If Confidence(X→Y)≥Minimum Confidence&&Lift(X→Y)≥1, then output rule X→Y and record it as the positive association rule PAR; where Confidence is the confidence, Minimum Confidence is the minimum confidence, and Lift is the lift; On the contrary, if Output rules Record as negative association rule; like Output rules Record as negative association rule; like Output rules Record as negative association rule; For each rare itemset R i : For R i For every pair of X and Y in R, X∩Y=R i and At the same time, Support(X)>MinimumSupport, Support(Y)>MinimumSupport: where is the support; If Confidence(X→Y)≥Minimum Confidence&&Lift(X→Y)≥1, then output rule X→Y and record it as a positive association rule; On the contrary, if Output rules Record as negative association rule; like Output rules Record as negative association rule; like Output rules Record as negative association rule; Output positive association rules and negative association rules.

Images