Learning path identification coevolution method based on differentiation transfer strategy

Abstract

The invention discloses a learning path identification coevolution method based on a differentiation transfer strategy. The method comprises the following steps: accelerating an algorithm convergence speed by using a subgroup generation mechanism based on an environmental pressure factor; the populations are divided into a superior population, an exploration population and an elimination population according to fitness, and the superior population, the exploration population and the elimination population respectively undertake tasks of storing an optimal individual set, maintaining population diversity and eliminating poorer individuals; individuals in the exploration population are differentiated and transferred after executing a positive and negative feedback mechanism, so that the individuals are regenerated on the basis of storing a part of key information, and the search performance is enhanced. Compared with a plurality of comparison methods such as GEO, MSEA and the like, the method provided by the invention has better optimization ability, and can achieve a good effect of obtaining an optimal learning path in the process of recognizing students to learn knowledge points.

Description

Technical Field

[0001] The invention relates to the technical field of learning path planning, and in particular to a knowledge point analysis learning path identification co-evolution framework method based on a differentiation transfer strategy. Background Art

[0002] At present, with the rapid development of information technology and the continuous progress of social economy, the field of education has also ushered in unprecedented changes and opportunities. The traditional education model faces many challenges, including low personalized learning, insufficient efficiency of knowledge acquisition, and uneven distribution of teaching resources. At the same time, with the increasing maturity and popularization of artificial intelligence technology, it has brought new possibilities and prospects to the field of education. Therefore, how to obtain the best real learning path has become one of the important topics in intelligent education research. However, to achieve the goal of intelligent education, a complete learning path system must be established. The learning path is a path planning for students to acquire knowledge and skills in a certain order and steps during the learning process. Traditional learning paths are often fixed and unified, and cannot meet the personalized learning needs of different students. There is a causal relationship between different knowledge points, which can generally be represented by a causal graph (Causal Graphs). The commonly used causal graph type is a directed acyclic graph, which consists of nodes and directed relationships. The causal relationship between knowledge points can be visualized by pointing to each other between nodes, that is, from cause to effect. The structural causal model (SCM) proposed by Judea Pearl is a commonly used framework in causal inference research. The best learning path for students can be identified through graph structure search. Graphs are a structure that can well reflect causal relationships. Graph structures can represent the interactive relationships between nodes and are widely used in education, medical care, engineering and other fields. Because of the widespread existence of graph structures, they have become the support of many systems. Graphs store information about entities with interactive relationships and specific interaction information.

[0003] Since obtaining accurate causal results from data is an NP-hard problem, how to efficiently learn causal relationships from data has been a research focus in recent decades. As a result, many causal search methods have emerged. For example, scoring-based causal structure search algorithms can simplify calculations using scoring algorithms, but their disadvantage is that they must traverse all possible graph structures, which makes it easy to obtain local optimal solutions. For example, the GES algorithm uses the BDeu scoring function for evaluation and uses greedy search to obtain local optimal solutions.

[0004] In summary, the problems existing in the prior art are:

[0005] (1) In the LPR problem, the learning path problem composed of a large number of knowledge points may make the optimization process of the population extremely difficult, and the abnormal data input to the problem may make the population optimization impossible.

[0006] (2) The complex structure composed of knowledge points of different sizes cannot be effectively identified. The optimization time may be greatly increased due to different algorithm parameters during the optimization process. Although the identification steps of each learning path are the same, different optimization methods may still lead to huge differences in the identification of the optimal learning path.

[0007] (3) The compatibility of currently used technologies with different evaluation algorithms varies greatly, and they can often only be combined with one evaluation algorithm, and their efficiency is not high. Summary of the invention

[0008] In view of the deficiencies in the prior art, the present invention provides a learning path identification co-evolution method based on the differentiation transfer strategy to solve the technical problems in the prior art that data anomalies may lead to the inability to perform population optimization, there are huge differences in the identification of the optimal learning path, and poor compatibility.

[0009] The present invention provides a learning path identification co-evolution method based on differentiation transfer strategy, comprising the following steps:

[0010] Step 1: According to the learning path information of the knowledge points of the students' learning knowledge, the initial population is randomly initialized, and each individual only stores the randomly generated learning path information;

[0011] Step 2: Identify incorrectly generated learning paths in the population and fix them;

[0012] Step 3: The parent population is transformed into a child population through a differential evolution algorithm. Each individual in the child population generates a new learning path under the influence of the parent population.

[0013] Step 4: Sort the offspring population internally into winning population, exploration population, and elimination population according to individual fitness values;

[0014] Step 5: Perform a positive feedback mechanism on the winning population, analyze the newly added learning path and combine it with the positive feedback parameter PF; perform a negative feedback mechanism on the eliminated population, analyze the newly added learning path and combine it with the negative feedback parameter NF;

[0015] Step 6: Based on the results of the positive and negative feedback mechanisms, a differentiation and transfer strategy is implemented for the exploration population, adding and deleting learning paths in the exploration population;

[0016] Step 7: Sort the fitness of the offspring population after differentiation and transfer with the parent population, and perform cross-merging;

[0017] Step 8: Divide the population after cross sorting into a new winning population and an eliminated population, discard the individuals in the eliminated population, and let the individuals in the winning population enter the next cycle to generate a new parent population;

[0018] Step 9: Repeat steps 3-8 until the function evaluation times are reached, and take the best individual in the population as the optimal knowledge point learning path.

[0019] Further, in step 2, the erroneously generated learning path includes: a learning path that forms a loop;

[0020] The method for repairing a learning path that forms a loop is to delete the last knowledge point pointing edge that causes the learning path to form a loop.

[0021] Furthermore, the step 2 includes: detecting loops in the learning path through depth-first search and path stack.

[0022] Furthermore, in step 3, the formula for obtaining the offspring population is as follows:

[0023]

[0024] In the formula, is the first p*Np individuals in the fitness ranking of a random population, Np is the population size, and p is the given ratio; is a randomly selected individual from the population; Individuals are randomly selected from the current population and the external archive collection.

[0025] Furthermore, in step 4, the formula for internal sorting of individual fitness values ​​is:

[0026]

[0027] Where con is the constraint violation value of the solution, that is, the upper and lower limits of the learning path value; Obj is the target value of the solution, that is, the target value of the learning path predicted by the individuals in the population.

[0028] Furthermore, in step 5, the formula for combining the newly added learning path with the positive feedback parameter PF and the negative feedback parameter NF is:

[0029]

[0030] count=count*NF

[0031] In the formula, the initial value of count is 0; NewEdge is the number of paths added for each individual in the population.

[0032] Beneficial effects of the present invention:

[0033] The present invention introduces a multi-subgroup collaborative graph structure learning strategy. In each iteration process, the generated offspring population is divided into three populations from high to low fitness levels: a winning population, an exploration population, and an eliminated population. On the one hand, the individuals of the winning population are combined with a positive feedback mechanism, and the individuals of the eliminated population are combined with a negative feedback mechanism, thereby optimizing the entire offspring population; on the other hand, the differentiation and transfer population can maintain the diversity of the population. By evaluating and ranking the optimized offspring population and the current population by fitness value, the population with a high fitness value directly enters the next iteration, and the population with a low fitness value is eliminated, which can better balance the superiority of the individuals and the diversity of the population.

[0034] The present invention utilizes positive and negative feedback mechanisms to evaluate and compare the generated offspring with the current population individuals through different evaluation algorithms passed in, and optimizes and reorganizes the learning path in the graph structure by changing the probability values ​​of the positive and negative feedback mechanisms, thereby quickly obtaining the optimal learning path.

[0035] The present invention repairs the possible directed cyclic graph into a directed acyclic graph by repairing the abnormal graph structure data, so that the algorithm can correctly identify the learning path. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0037] Figure 1 is a method flow chart of a specific embodiment of the present invention;

[0038] Figure 2 An operation structure diagram of the DE-TS method provided by an embodiment of the present invention;

[0039] Figure 3 Method evolution curves of CSSF and CSSF(-BFM) provided by the embodiments of the present invention under different test problems;

[0040] Figure 4 Confusion matrices of CSSF and comparison methods provided by the embodiment of the present invention under different test problems;

[0041] Figure 5 Box plots of CSSF under different test questions provided by an embodiment of the present invention;

[0042] Figure 6 Friedman ranking of CSSF and comparison methods provided by the embodiment of the present invention on real data sets;

[0043] Figure 7Friedman ranking of the CSSF and the comparison method provided in the embodiment of the present invention on the generated data set. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0045] The present invention is further illustrated below in conjunction with specific embodiments. Those skilled in the art should understand that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention, and modifications to various equivalent forms of the present invention fall within the scope defined by the appended claims of this application.

[0046] like Figure 1 As shown, the present invention provides a learning path identification collaborative evolution method (CSSF) based on differentiation transfer strategy, comprising the following steps:

[0047] Step 1: According to the knowledge point learning path information contained in the real student learning knowledge, the initial population is randomly initialized, and each individual in the infancy of the initial population only stores the randomly generated learning path information, wherein the optimal path is obtained according to the learning path information by the method provided by the present invention;

[0048] Step 2: Identify the incorrectly generated learning paths in the population and repair them. The loops in the learning paths are detected through depth-first search and path stack. When repairing, the last knowledge point pointing edge that causes the learning path to generate a loop is deleted. The specific steps are as follows:

[0049] Step 21: Check whether the learning path included in the generated individual contains a loop, and perform a depth-first search on the learning path corresponding to each unvisited vertex. When the currently visited node has not been visited before, mark this node as visited and add the current vertex to the path stack;

[0050] Step 22: When the path stack is not empty, get the top element of the stack and find the neighbor nodes of the current top element of the stack; find all neighbor nodes that have not been visited, add the neighbor nodes to the path stack, and when all neighbor nodes have been visited, remove the top element of the stack and determine whether there is a loop; when there is no loop, return to step 2.1 to continue the determination; when there is a loop, end the search;

[0051] Step 23: When there is a loop in the path, obtain the vertex information on the loop and remove duplicates, and return the processed learning path;

[0052] Step 3: Obtain the offspring population from the parent population through the differential evolution algorithm. The specific formula is as follows:

[0053]

[0054] In the formula, is an individual randomly selected from the top p*Np individuals in terms of fitness in a population. Np is the population size, and p is a given ratio; is an individual randomly selected from the population; is an individual randomly selected from the current population and the external archive set,

[0055] Each individual in the offspring population generates a new learning path under the influence of the parent population;

[0056] Step 4: Each time the fitness value of an individual is calculated, it consumes one function evaluation. Compare the already consumed function evaluation times FE with the preset maximum function evaluation times maxFE: when FE < maxFE, continue to run; when FE ≥ maxFE, jump to end the run, and take the optimal individual in the population as the optimal knowledge point learning path;

[0057] Step 5: Sort the offspring population internally according to the individual fitness values into a superior population, an exploration population, and an elimination population. Among them, the internal sorting is carried out through the following formula:

[0058]

[0059] In the formula, con is the constraint violation value of the solution, that is, the upper and lower limit values of the learning path value; Obj is the objective value of the solution, that is, the learning path objective value predicted by the population individual;

[0060] Step 6: Implement a positive feedback mechanism for the superior population, analyze the newly added learning paths and combine them with the positive feedback parameter PF; implement a negative feedback mechanism for the elimination population, analyze the newly added learning paths and combine them with the negative feedback parameter NF. The specific combination formula is as follows:

[0061]

[0062] count = count * NF

[0063] In the formula, the initial value of count is 0; NewEdge is the number of paths added by each population individual. Finally, the value of count is the cumulative sum of NewEdge and multiplied by the value of NF or PF to obtain the final result.

[0064] Step 7: Based on the results of the positive and negative feedback mechanisms, a differentiation and transfer strategy is implemented for the exploration population, specifically adding and deleting learning paths in the exploration population;

[0065] Step 8: Sort the fitness of the offspring population after differentiation and transfer with the parent population, and perform cross-merging;

[0066] Step 9: Divide the cross-sorted population into a new winning population and an eliminated population, discard the individuals in the eliminated population, and let the individuals in the winning population enter the next cycle to generate a new parent population.

[0067] like Figure 2 As shown, the method of the present invention is mainly based on individual optimization and differentiation transfer strategy, which is marked by the gray area, including the generation of offspring population, positive and negative feedback mechanism, and cross selection. During the algorithm execution process, after judging that FE has not reached the maximum number of times, the operation of generating offspring population is performed, the offspring population is divided into a winning subgroup, an exploration subgroup, and an eliminated subgroup, and the positive and negative feedback mechanism is completed. At this time, the generated offspring individuals are subjected to individual optimization operations. Subsequently, the optimized offspring population and the current population are sorted according to fitness and merged. Finally, the current population is sorted according to the specified optimization direction and divided into a winning subgroup and an eliminated subgroup. The winning subgroup is retained as the parent population of the next iteration, and the eliminated subgroup is discarded.

[0068] like Figure 3 As shown, the algorithm evolution curves of CSSF and CSSF (-BFM) provided by the embodiment of the present invention under different test problems. CSSF (-BFM) is a variant of CSSF that removes the positive and negative feedback mechanisms. CSSF with the two-way feedback mechanism retained is compared with CSSF (-BFM) without the two-way feedback mechanism. Initially, the loss values ​​of the two methods remain consistent. As the number of method runs increases, the loss value of CSSF is always lower than that of CSSF (-BFM) without the two-way feedback mechanism. The final result shows that the loss value of CSSF is significantly lower. In the LPR-RWD problem, the final loss value of CSSF is approximately 11% lower than that of CSSF (-BFM) after running 5000 times. In the LPR-RWD1 problem, the final loss value of CSSF is approximately 9% lower than that of CSSF (-BFM) after running 5000 times. In the LPR-RWD2 problem, the final loss value of CSSF is approximately 5% lower than that of CSSF (-BFM) after running 5000 times. It is proved that the two-way feedback mechanism has a significant effect in judging the learning path in the real education data set.

[0069] like Figure 4 As shown in FIG. 1 , the confusion matrix of the CSSF provided by the embodiment of the present invention and the comparison method under different test problems. Figure 4In the confusion matrix, the learning paths generated by each method are divided into four categories according to the learner's actual learning situation and the model's prediction. In the confusion matrix, matrix (1,1) represents TP, matrix (1,0) represents NF, matrix (0,0) represents TN, and matrix (0,1) represents NT. Figure 4 It can be seen from the confusion matrix in that on the real data set LPR-RWD, the number of TP and TN corresponding to the confusion matrix generated by CSSF proposed in the present invention is similar to the number of TP and TN in the confusion matrix generated by EESHHO and GEO, indicating that the three methods have similar effects on the LPR-RWD problem, and the number of TP and TN corresponding to the confusion matrix generated by CSSF is greater than the number of TP and TN in the confusion matrix generated by MFOSPEA2 and MSEA. This shows that CSSF is better than the above two methods in the LPR-RWD problem. In the LPR-RWD1-2 problem, the number of TP and TN corresponding to the confusion matrix generated by CSSF is more than the number of TP and TN corresponding to the confusion matrix generated by the other four comparison methods, indicating that CSSF is better than other comparison methods in the LPR-RWD1-2 problem. The above results are highly consistent with the results of the above-mentioned comparative experiments.

[0070] like Figure 5 As shown, box plots of CSSF provided by an embodiment of the present invention under different test problems. Figure 5 The convergence curves of CSSF on the three problems in the LPR-RWD series are shown in Figure 1. Then, a box plot is drawn after CSSF is run 30 times on the three problems in the LPR-RWD series, and the median of the results is marked. The distribution of the results can be clearly shown from the box plot. Figure 5 From the results in , we can see that the average loss of CSSF on the three problems in the LPR-RWD series are 33.43%, 27.62% and 23.30% respectively. The results show that the experimental results are highly close, which verifies that CSSF has good convergence to a certain extent.

[0071] like Figure 6 , 7 As shown, the CD diagrams of the CSSF provided by the embodiment of the present invention and the comparison method under different data sets can draw the conclusion that the CSSF effect is due to the comparison algorithm, as shown in the Friedman test results of the real data set and the synthetic data set in Table 1 below:

[0072]

[0073] Table 1

[0074] In Table 1, the fireman test ranking results show that the CSSF algorithm proposed in the present invention always ranks first in both real data sets and synthetic data sets, and has the best effect compared with other comparison algorithms. It also shows that the proposed algorithm has a very strong ability to identify real learning paths.

[0075] The present invention is subjected to simulation experiments as follows:

[0076] The simulation experiments of the present invention are all carried out on the Learning Path Recognition series (LPR) data sets. All methods use loss as the evaluation index in order to improve the fairness of the experiment. In this stage, all simulation experiments are carried out on a device equipped with an Intel i5-13600KF@3.50GHz dual-core processor. On the one hand, the hardware environment for the algorithm to run also includes a running memory of 32G and a graphics card of NVIDIA RTX 4070; on the other hand, the software platform for the algorithm to run is Matlab2022b and PlatEMO (Platform for EvolutionaryMulti-Objective Optimization) platform.

[0077] Part of the visualization work in this invention is based on the tool on chiplot.online. It should be noted that, except for the algorithm proposed in this invention, the parameter combinations used by all the comparison methods are from their original papers, or the default parameter combinations provided by the PlatEMO platform are called. The settings of the hyperparameters of the comparison methods used are shown in Table 2 below:

[0078]

[0079] Table 2

[0080] The simulation experiment of the present invention implements the LPR series test problems based on the PlatEMO platform. In order to ensure the fairness of the experiment, the maximum number of evaluations of all methods is set to 10,000 times.

[0081] The simulation experiment results have been summarized in the following Table 3, as shown in the loss rate loss and its standard deviation. Among the comparative methods, the MSEA proposed by Y. Tian et al. is an algorithm that is constrained according to different evolutionary stages. It divides the optimization process into multiple stages to maintain better diversity of the population. The MFOSPEA2 proposed by R. Jiao et al. is a multi-form optimization framework that searches from the infeasible and feasible areas of the search space to quickly obtain the Pareto optimal solution. The Golden Eagle Optimization Algorithm (GEO) is an evolutionary algorithm that simulates the multi-stage hunting of the golden eagle. By simulating the behavior of the golden eagle to capture the best possible prey in the feasible area in the shortest possible time, the problem can achieve better global optimization. The Harris Hawk Optimization Algorithm based on the Elite Evolutionary Strategy is an improved evolutionary algorithm proposed in 2021 to simulate the Harris Hawk catching prey. The algorithm can quickly obtain the Pareto optimal solution using the elite non-dominated sorting mechanism.

[0082]

[0083] Table 3

[0084] Table 3 includes three test problems, LPR-RWD1-3, of which LPR-RWD1 and LPR-RWD2 respectively intercept the adjacency matrix data of the first 50 nodes and the adjacency matrix data of the 51st-100th nodes in the real data set, and LPR-RWD contains the adjacency matrix of the complete 116 nodes of the real data set. The purpose of intercepting part of the adjacency matrix data for comparison is to analyze whether the CSSF proposed in the present invention has performance superiority under real data of different dimensions. In Table 3, similar to the comparative experiment of the method for generating data sets, the effects of the two methods MSEA and MFOSPEA2 are similar, while the effects of the two methods GEO and EESHHO in the LPR-RWD problem containing a complete adjacency matrix are similar to the CSSF proposed in the present invention, but in the two problems LPR-RWD1-2, the effects of the first two methods are significantly lower than CSSF. Among all problems, CSSF has the best performance. It shows that the algorithm proposed in the present invention has the best effect on the real data set.

[0085] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A learning path identification co-evolution method based on differentiation transfer strategy, characterized in that: The steps include: Step 1: According to the learning path information of the knowledge points of the students' learning knowledge, the initial population is randomly initialized, and each individual only stores the randomly generated learning path information; Step 2: Identify incorrectly generated learning paths in the population and fix them; Step 3: The parent population is transformed into a child population through a differential evolution algorithm. Each individual in the child population generates a new learning path under the influence of the parent population. Step 4: Sort the offspring population internally into winning population, exploration population, and elimination population according to individual fitness values; Step 5: Perform positive feedback mechanism on the winning population, analyze the newly added learning path and combine it with the positive feedback parameter PF; perform negative feedback mechanism on the eliminated population, analyze the newly added learning path and combine it with the negative feedback parameter NF; Step 6: Based on the results of the positive and negative feedback mechanisms, a differentiation and transfer strategy is implemented for the exploration population, adding and deleting learning paths in the exploration population; Step 7: Sort the fitness of the offspring population after differentiation and transfer with the parent population, and perform cross-merging; Step 8: Divide the population after cross sorting into a new winning population and an eliminated population, discard the individuals in the eliminated population, and let the individuals in the winning population enter the next cycle to generate a new parent population; Step 9: Repeat steps 3-8 until the function evaluation times are reached, and take the best individual in the population as the optimal knowledge point learning path.

2. The learning path identification co-evolution method based on differentiation transfer strategy as claimed in claim 1, characterized in that: In step 2 shown, the learning paths generated by errors include: learning paths that form a loop; The method for repairing a learning path that forms a loop is to delete the last knowledge point pointing edge that causes the learning path to form a loop.

3. The learning path identification co-evolution method based on differentiation transfer strategy as claimed in claim 2, characterized in that: The step 2 includes: detecting loops in the learning path through depth-first search and path stack.

4. The learning path identification co-evolution method based on differentiation transfer strategy as claimed in claim 1, characterized in that: In step 3, the formula for obtaining the offspring population is as follows: In the formula, is the first p*Np individuals in the fitness ranking of a random population, Np is the population size, and p is the given ratio; is a randomly selected individual from the population; Individuals are randomly selected from the current population and the external archive collection.

5. The learning path identification co-evolution method based on differentiation transfer strategy as claimed in claim 1, characterized in that: In step 4, the formula for internal sorting of individual fitness values ​​is: Where con is the constraint violation value of the solution, that is, the upper and lower limits of the learning path value; Obj is the target value of the solution, that is, the target value of the learning path predicted by the individuals in the population.

6. The learning path identification co-evolution method based on differentiation transfer strategy as claimed in claim 1, characterized in that: In step 5, the formula for combining the newly added learning path with the positive feedback parameter PF and the negative feedback parameter NF is: count=count*NF In the formula, the initial value of count is 0; NewEdge is the number of paths added for each individual in the population.

Images