General cognitive diagnosis model architecture searching method based on evolutionary multi-task optimization

Abstract

The invention provides a general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization. The general cognitive diagnosis model architecture search method comprises the following steps: S1, modeling an architecture search problem of a generalized cognitive diagnosis model into a multi-objective optimization problem; s2, defining a search space of the cognitive diagnosis model, and generating a model initial population; s3, constructing a collaborative optimization framework of the main task and the auxiliary task, performing optimization through a target evolutionary algorithm, and iteratively updating a population; s4, designing a multi-stage knowledge migration strategy, and obtaining dynamically adjusted main task and auxiliary task populations; s5, ensuring that a population conforming to semantic constraints is generated in genetic manipulation through a semantic maintenance strategy; s6, judging whether convergence is met or not; if yes, outputting a Pareto optimal model; and if not, returning to S3. According to the method, by integrating evolutionary multi-task optimization, a multi-stage knowledge transfer strategy and a semantic constraint mechanism, the generalization ability and interpretation of a cognitive diagnosis model on an expensive heterogeneous data set are improved, and the convergence efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart education cognitive diagnosis, and specifically to a general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization. Background Art

[0002] Cognitive diagnostic models (CDMs) are a type of statistical and machine learning model that analyzes students' response data to infer their knowledge status (e.g., skill mastery and cognitive deficits). They are a core tool in the field of educational technology. This technology is widely used in online education platforms to support personalized learning recommendations and optimize teaching strategies. Their core goal is to model students' response behavior and infer their proficiency in various knowledge concepts, thereby providing data-driven decision support for educators.

[0003] With the increasing diversity and complexity of educational data, designing highly generalizable and interpretable cognitive diagnostic models has become a research hotspot. Existing CDMs can be categorized into three main approaches to improving model generalization: CDMs based on high interpretability, CDMs based on neural network-enhanced fitting capabilities, and CDMs based on neural architecture search technology. Highly interpretable CDMs leverage educational psychology theory to improve model generalization, with representative models including DINA, IRT, and MIRT. CDMs based on neural network-enhanced fitting capabilities utilize neural networks to improve model fitting capabilities, thereby enhancing generalization. Representative models include NCD, KaNCD, KSCD, and RCD. CDMs based on neural architecture search technology utilize NAS techniques to improve model generalization for CD tasks, with the EMO-NAS-CD being a typical example. However, research has found that the generalization capabilities of CDMs based on high interpretability and CDMs based on neural network-enhanced fitting capabilities are still limited on datasets containing unobserved interaction data. To address this issue and construct more generalizable CDMs, NAS-GCD proposes the NAS method.

[0004] However, current NAS methods reduce the search space to simplify the MOP, resulting in the neglect of many potentially excellent models. More importantly, current MOPs are computationally expensive and highly heterogeneous, leading to slow convergence and unbalanced resource allocation. Furthermore, the models automatically generated by current NAS methods are highly complex, violating educational psychology theory and reducing the credibility of diagnostic results. Therefore, we propose a general cognitive diagnostic model architecture search method based on evolutionary multi-task optimization. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization, which can solve the existing problems.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] The present invention is implemented through the following technical solution: a general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization, comprising the following steps:

[0008] S1. Model the architecture search problem of the generalized cognitive diagnosis model as a multi-objective optimization problem;

[0009] S2. defining a search space for a cognitive diagnosis model based on the tree structure of the multi-objective optimization problem and generating an initial population of the model;

[0010] S3. Constructing a collaborative optimization framework for the main task and the auxiliary task based on the multi-objective optimization problem and the initial population of the model, performing optimization using a multi-objective evolutionary algorithm and a single-objective evolutionary algorithm, and iteratively updating the population;

[0011] S4. Design a multi-stage knowledge transfer strategy to obtain dynamically adjusted main task and auxiliary task populations through population knowledge transfer between the main task and auxiliary tasks;

[0012] S5. Based on the search space and the multi-objective evolutionary algorithm and the single-objective evolutionary algorithm, a semantic maintenance strategy is used to ensure that a population that meets the semantic constraints is generated in a genetic operation;

[0013] S6: Based on the optimization results of S3 and the semantic constraints of S5, determine whether convergence is satisfied;

[0014] If so, output the Pareto optimal model to complete the entire search method;

[0015] If not, return to S3.

[0016] Furthermore, the architecture search problem of the generalized cognitive diagnosis model is modeled as a multi-objective optimization problem, including:

[0017] Enter student answer records i ,e j ,r ij and the Q matrix, where s i Indicates students, e j Indicates practice, r ij ∈0,1} represents student s i In practice j The answer result on r ij =1 means the answer is correct, r ij =0 means wrong answer; Indicates practice e j Is the knowledge concept c related? k , Indicates association, Indicates no correlation; cognitive diagnosis is based on students’ answer records and Q matrix, and a cognitive diagnosis model is built to predict students’ answers, thereby exploring students’ knowledge proficiency, which is recorded as:

[0018]

[0019] In formula (1) and Students i and practice e j The student-related vector is obtained through two embedding modules Embed1(·) and Embed2(·) and practice related vectors Extract vectors related to knowledge concepts from Q Representation Model The predicted results; After training, h S Used to indicate students’ knowledge proficiency. Varies across different cognitive diagnostic models.

[0020] Furthermore, the search space of the cognitive diagnosis model is defined according to the tree structure of the multi-objective optimization problem, including:

[0021] The preset computation tree consists of several leaf nodes and internal computation nodes. The leaf nodes represent the input of the model and are selected from {h S ,h E ,h C}, the internal computing node selects from 15 candidate operators. With the support of this search space, the cognitive diagnosis model is rewritten as:

[0022]

[0023] In formula (2), the first part It is used to process h S ,h E and h C The computation tree of , whose output is a scalar or vector; the second part The final prediction result is further generated based on the output of the first part, and all cognitive diagnosis models in the search space are represented by tree coding.

[0024] Furthermore, the candidate operators include:

[0025] The formula x+y is written as Add;

[0026] The formula x·yorx⊙y is denoted as Mul;

[0027] Formula -x, denoted as Neg;

[0028] Formula |x|, denoted as Abs;

[0029] Formula 1 / (x+∈),∈=1e -6 , recorded as Inv;

[0030] Formula x 2 , recorded as Square;

[0031] formula Denoted as Sqrt;

[0032] formula Recorded as Sum;

[0033] formula Recorded as Mean;

[0034] The formula tanh(x) is denoted as Tanh;

[0035] Formula sigmoid(x), denoted as Sigmoid;

[0036] Formula softplus(x), denoted as Softplus;

[0037] formula Denoted as Concat;

[0038] formula Denoted as FFN;

[0039] formula Denoted as FFND.

[0040] Furthermore, the initial population of generative models includes half existing models and half randomly generated models.

[0041] Furthermore, the collaborative optimization framework of the main task and the auxiliary task is constructed, including:

[0042] First, the iteration counters g and g of the main task and each auxiliary task are i Initialized to 0;

[0043] Second, a model population is initialized for all tasks;

[0044] Finally, the main task and each auxiliary task are executed simultaneously through different threads;

[0045] Specific steps of the main task: Main population P m Execute the "mating pool selection", "genetic operation" and "deduplication" steps in sequence to generate the offspring population Q mIf the auxiliary task has not yet completed its previous η×Gen generation evolution, the main task will not receive any knowledge from the auxiliary task; at this time, the main task will complete the evaluation of Q m After that, directly from P m and Q m The main task will call the knowledge transfer strategy in the later stage to obtain potentially useful individuals from the auxiliary tasks and store them in P t ; Then for P t Evaluate and compare with P m and Q m Compare together to get a new main population; when the number of iterations g exceeds the set upper limit Gen, and the main population P m When the hypervolume of does not change in five consecutive generations, the main task stops;

[0046] The specific steps of auxiliary tasks are the same as those of the main task, but each auxiliary task has a flag i Used to record the convergence status of the auxiliary task population, flag i Initially 0, it increases automatically when the auxiliary population converges, but does not exceed 2. The multi-stage knowledge transfer strategy will switch the knowledge transfer stage according to this flag;

[0047] When the main task is completed, the main population P is output m The set of non-dominated individuals P in non .

[0048] Furthermore, the main task objective function is expressed as:

[0049]

[0050] The problem requires maximizing the model A over multiple observed data sets. Performance on

[0051] There are Num auxiliary tasks, where the objective function of the i-th auxiliary task is:

[0052]

[0053] In formulas (3) and (4), Represents the search space. The constructed multi-objective optimization problem has Num objectives, each of which For model A in the dataset After training, on the validation set The AUC on .

[0054] Furthermore, the multi-stage knowledge transfer strategy includes:

[0055] Input main population P m , auxiliary population {Pi |1≤i≤Num}, population ID, convergence status flag;

[0056] Initialize the output population P o Is an empty set; judge whether all auxiliary tasks have been traversed, if yes, get the population P for knowledge transfer o , No, continue to supplement P o ;

[0057] From the population P of the i-th auxiliary task i Select the top K individuals with the best performance, denoted as T i , from the main task population P m Select the top K individuals with the best performance on the i-th target, denoted as Tm i ; Through the formula Calculate the fitness fh of the auxiliary task i And the fitness of the main task fm i , fh i T i The average AUC of fm i Tm i The average AUC of ; 2K individuals R are randomly selected from the i-th auxiliary task i , randomly select 2K individuals Rm from the main task i ;

[0058] When the ID is the main task ID and flag i = 0, the first stage of knowledge transfer begins, transferring useful individuals in the auxiliary task to the main task, and using the rapid convergence of the auxiliary task to accelerate the optimization of the main task; when fm i >fh i , R i Join P o , to increase the diversity of the main task population; otherwise, it means that the auxiliary task performs better on this goal, and T i Join P o ;

[0059] When the ID is an auxiliary task ID and flag i = 1, the second stage of knowledge transfer begins, the auxiliary task absorbs useful individuals from the main task, and when the auxiliary task falls into the local optimum, the diversity of the main task is used to help it jump out of the local optimum; when fm i <fh i , Rm i Join P o , in order to increase the diversity of the auxiliary task population; otherwise, it means that the main task performs better on this goal. In order to take into account both convergence and diversity in the non-fast convergence stage, Tm i and Rmi Join P o ;

[0060] When the ID is the main task ID and flag i = 2, the third stage of knowledge transfer begins. When the auxiliary task converges again with the help of the diversity of the main task, its optimized individuals are transferred back to the main task to form a circular optimization mechanism. i >fh i , R i Join P o , to increase the diversity of the main task population; at the same time flag i Self-decrement makes the knowledge transfer strategy return to the second stage. When the main task and the auxiliary task are close in performance, the evolution of the auxiliary task is promoted again; otherwise, it means that the auxiliary task performs better on this goal, and T i and R i Join P o .

[0061] Furthermore, the semantic maintenance strategy includes:

[0062] First, the shape of the tree is constrained by presetting the maximum depth Dep and the maximum width Wid to directly limit the depth and width of the generated tree; the width of the tree is constrained by limiting the number of leaf nodes it contains; to ensure the rationality of the model, the student-related leaf nodes h S Leaf node h related to the exercise E Contained in the tree;

[0063] Secondly, we use the idea of semantic approximation and semantic information to guide tree generation, and design two "semantic-aware" sub-operations, Semantic-Exchange and Semantic-Replace. The core idea of Semantic-Exchange is to exchange two semantically similar subtrees; the core idea of Semantic-Replace is to replace a randomly selected node with a semantically similar node in the candidate operator set. The 15 candidate operators are divided into four categories according to their semantics: the first category includes Add, Mul, and Concat; the second category includes Tanh, Sigmoid, and Softplus; the third category includes Sum, Mean, FFN, and FFN_D; and the fourth category includes Neg, Abs, Inv, Square, and Sqr.

[0064] The genetic operation includes:

[0065] Continuously randomly select one of the four sub-operations {Semantic-Exchange, Semantic-Replace, Delete, Insert}, operate on two parent individuals, and generate two offspring individuals until the size of the offspring population Q reaches the requirement; before adding the generated individuals to Q, it is necessary to first check whether their tree shape meets the pre-set depth and width restrictions. Semantic-Exchange represents semantic exchange, Semantic-Replace represents semantic replacement, Delete represents random deletion of a node, and Insert represents random insertion of a node.

[0066] The present invention also provides a computer-readable storage medium, wherein the storage medium stores a program, and the program is executed by a processor to implement any one of the above methods.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] The present invention provides a general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization,

[0069] The present invention regards the MOP as the main task and the optimization of each objective in the MOP as a single task, and introduces the idea of evolutionary multi-task optimization to solve these tasks simultaneously, so as to more effectively optimize the original complete MOP. It avoids the problem of traditional NAS methods that the search space is limited due to the simplification of the multi-objective optimization problem (MOP) and the model's insufficient generalization on new data sets. The performance of the model on unobserved data sets is improved, thereby improving the generalization ability of the model.

[0070] 2. The present invention adopts a multi-stage knowledge transfer strategy and utilizes the rapid convergence of auxiliary tasks to accelerate the optimization of the main task. Combined with the asynchronous evolution mechanism, it avoids the problem of wasting computing resources caused by traditional single-task optimization or synchronous evolution mechanism, reduces the evaluation cost of expensive and heterogeneous data sets, and improves convergence efficiency.

[0071] 3. The present invention designs a semantic maintenance strategy to constrain the maximum depth and width of the spanning tree, and groups operators by semantics, thus avoiding the problems of high model structure complexity, violation of educational psychology theory, and poor interpretability automatically generated by the traditional NAS method. This reduces the model complexity, makes the diagnostic logic more consistent with the cognition of educational experts, and improves the model interpretability.

[0072] The present invention improves the generalization ability and interpretability of cognitive diagnosis models on expensive heterogeneous datasets by integrating evolutionary multi-task optimization, multi-stage knowledge transfer strategy and semantic constraint mechanism, while also improving convergence efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0074] Figure 1 This is a flow chart of a general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to the present invention;

[0075] Figure 2 It is a schematic diagram of a partial framework in an embodiment of the present invention;

[0076] Figure 3 A flowchart of a multi-stage knowledge transfer strategy in an implementation case of the present invention;

[0077] Figure 4 Schematic diagram of the tree-based search space used in the implementation case of the present invention. DETAILED DESCRIPTION

[0078] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0079] The present invention provides a general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization, the process of which is as follows: Figure 1 As shown, the following steps are included:

[0080] S1. Model the architecture search problem of the generalized cognitive diagnosis model as a multi-objective optimization problem;

[0081] Specifically, the first step is to define the cognitive diagnosis task and model the architecture search problem of the generalized cognitive diagnosis model as a multi-objective optimization problem (MOP), where the goal of the MOP is to maximize the performance of the model on multiple cognitive diagnosis datasets. It should be noted that the requirements of the generalized cognitive diagnosis model include but are not limited to prediction accuracy, model interpretability, and computational efficiency. The cognitive diagnosis model contains students' exercise response logs, exercise-concept relationship matrices, and exercise information, and can predict the answer results of any student to any question, thereby knowing the student's level of mastery of the concept.

[0082] Cognitive diagnostic tasks are defined in this application and include:

[0083] Enter student answer records i ,e j ,r ij and the Q matrix, where s i Indicates students, ej Indicates practice, r ij ∈0,1} represents student s i In practice j The answer result on r ij =1 means the answer is correct, r ij =0 means wrong answer; Indicates practice e j Is the knowledge concept c related? k , Indicates association, Indicates no correlation; cognitive diagnosis is based on students’ answer records and Q matrix, and a cognitive diagnosis model is built to predict students’ answers, thereby exploring students’ knowledge proficiency, which is recorded as:

[0084]

[0085] In formula (1) and Students i and practice e j The student-related vector is obtained through two embedding modules Embed1(·) and Embed2(·) and practice related vectors Extract vectors related to knowledge concepts from Q Representation Model The predicted results; After training, h S It can be used to indicate the students’ knowledge proficiency. The architecture of different cognitive diagnostic models affects the generalization ability and interpretability of the model.

[0086] It should be noted that the Q matrix is usually annotated by experts or automatically generated, and it is pointed out that its quality affects the effect of cognitive diagnosis.

[0087] It should be noted that the dimension D of each model is set to the number of concepts K.

[0088] S2. defining a search space for a cognitive diagnosis model based on the tree structure of the multi-objective optimization problem and generating an initial population of the model;

[0089] Specifically,

[0090] Defining the search space of the cognitive diagnosis model according to the tree structure of the multi-objective optimization problem includes:

[0091] The computation tree consists of several leaf nodes and internal computation nodes. The leaf nodes represent the input of the model and are selected from {h S ,h E ,h C}, the internal computing node selects from 15 candidate operators. With the support of this search space, the cognitive diagnosis model can be rewritten as:

[0092]

[0093] In formula (2), the first part It is used to process h S ,h E and h C The computation tree of , whose output can be a scalar or a vector; the second part The final prediction result is further generated based on the output of the first part. All cognitive diagnosis models in the search space are represented by tree coding. The process is as follows: Figure 4 shown.

[0094] It should be noted that in this embodiment The dimensions of the fully connected (FC) layers are set to 512, 256, and 1.

[0095] The candidate operators mentioned above include:

[0096] The formula x+y is denoted as Add; the formula x·yorx⊙y is denoted as Mul; the formula -x is denoted as Neg; the formula |x| is denoted as Abs; the formula 1 / (x+∈),∈=1e -6 , denoted as Inv; Formula x 2 , denoted as Square; formula Denoted as Sqrt; formula Denoted as Sum; formula The formula tanh(x) is denoted as Tanh; the formula sigmoid(x) is denoted as Sigmoid; the formula softplus(x) is denoted as Softplus; the formula Denoted as Concat; formula Denoted as FFN; formula Denoted as FFND.

[0097] The initial population of generative models includes half existing models and half randomly generated models;

[0098] Specifically, half of the models are existing ones, which can help the algorithm find relatively high-quality solutions faster in the early stages; the other half are randomly generated, which helps maintain diversity and prevents falling into local optimality too early.

[0099] Exemplary models include NCD (Nonparametric Cognitive Diagnosis), RCD (Rule-based Cognitive Diagnosis), IRT (Item Response Theory), and MIRT (Multidimensional IRT).

[0100] Specific methods for randomly generating models include:

[0101] 1. Modular random combination

[0102] Cognitive diagnosis models are usually composed of multiple functional modules (such as Q matrix modeling layer, student ability estimation layer, interaction function layer, etc.). When randomly generated:

[0103] Steps: Randomly select components from a predefined module library (e.g., Table 1 below);

[0104] These modules are legally connected in a tree structure (refer to S1 for modeling).

[0105] Table 1

[0106]

[0107]

[0108] Example random model:

[0109] Model A: Expert-defined Q matrix + Probit capability estimation + graph neural network interaction + probability output.

[0110] Model B: Data-driven Q matrix + Logistic capability estimation (random slope = 1.2) + linear interaction + binary classification output.

[0111] 2. Parameter random initialization

[0112] Perform parameter perturbations on variants of existing models (such as IRT, MIRT):

[0113] Method: Based on the existing model structure, key parameters (such as discrimination and the range of guessed parameters) are randomly adjusted.

[0114] Exemplary:

[0115] A MIRT model is randomly generated, but the number of capability dimensions is randomly selected from 1 to 5, and the item parameters are sampled from a uniform distribution.

[0116] 3. Random Architecture Based on Generative Rules

[0117] Define generation rules through domain knowledge to ensure that the random model makes sense:

[0118] Rule example:

[0119] “At least one interpretable interaction function (such as the DINA rule) must be included in the diagnostic model.”

[0120] “Pure black box structures (such as multi-layer neural networks) are prohibited as core diagnostic layers.”

[0121] Generation process:

[0122] Randomly select whether to include the Q matrix (70% probability of inclusion);

[0123] Random selection power estimation methods (Logistic / Probit);

[0124] Randomly add auxiliary modules (such as 30% probability of adding time decay factor).

[0125] S3. Constructing a collaborative optimization framework for the main task and the auxiliary task based on the multi-objective optimization problem and the initial population of the model, performing optimization using a multi-objective evolutionary algorithm and a single-objective evolutionary algorithm, and iteratively updating the population;

[0126] Specifically, a collaborative optimization framework for the main task and multiple auxiliary tasks is constructed, including:

[0127] First, the iteration counters g and g of the main task and each auxiliary task are i Initialize to 0, then initialize a model population for all tasks. Next, the main task and each auxiliary task are executed simultaneously through different threads;

[0128] Specific steps of the main task: Main population P m Execute the "mating pool selection", "genetic operation" and "deduplication" steps in sequence to generate the offspring population Q m In order to maintain good population quality and diversity in the early iterations, if the auxiliary task has not completed its first η×Gen generation evolution, the main task will not receive any knowledge from the auxiliary task; at this time, the main task will evaluate Q m After that, directly from P m and Q m To obtain faster and better convergence, the main task will call the knowledge transfer strategy in the later stage to obtain potentially useful individuals from the auxiliary tasks and store them in P t ; Then for P t Evaluate and compare with P m and Q mCompare together to get a new main population. When the number of iterations g exceeds the set upper limit Gen, and the main population P m When the hypervolume of does not change in five consecutive generations, the main task stops;

[0129] The specific steps of auxiliary tasks are the same as those of the main task, but each auxiliary task has a flag i Used to record the convergence status of the auxiliary task population, flag i Initially 0, it increases automatically when the auxiliary population converges, but does not exceed 2. The multi-stage knowledge transfer strategy will switch the knowledge transfer stage according to this flag;

[0130] When the main task is completed, output the main population P m The set of non-dominated individuals P in non .

[0131] Illustratively, in the implementation case of the present invention, during the search process of each task, each student's answer record is randomly divided into 70%, 10%, and 20%, which are used for training, verification, and testing, respectively.

[0132] Exemplarily, in the embodiment of the present invention, the Adam optimizer is used to minimize the cross entropy loss between the predicted value and the true value, the learning rate is 0.001, the batch size is 128, and the number of training rounds is 30.

[0133] For example, in the embodiment of the present invention, the population size is 100, the maximum number of iterations is 100, and the η parameter is set to 1 / 5.

[0134] The objective function of each task is defined in the present invention, including:

[0135] The main task objective function is expressed as:

[0136]

[0137] The problem requires maximizing the model A over multiple observed data sets. Among them, Represents the search space. The constructed multi-objective optimization problem has Num objectives, each of which For model A in the dataset After training, on the validation set The AUC on .

[0138] There are Num auxiliary tasks, where the objective function of the i-th auxiliary task is:

[0139]

[0140] S4. Design a multi-stage knowledge transfer strategy to obtain dynamically adjusted main and auxiliary task populations through population knowledge transfer between the main and auxiliary tasks. This strategy aims to accelerate the convergence of the main task and improve the diversity of the auxiliary tasks.

[0141] Specifically, multi-stage knowledge transfer strategies, such as Figure 3 Shown, including:

[0142] Input main population P m , auxiliary population {P i |1≤i≤Num}, population ID, convergence status flag;

[0143] Initialize the output population P o Is an empty set; judge whether all auxiliary tasks have been traversed, if yes, get the population P for knowledge transfer o , No, continue to supplement P o ;

[0144] From the population P of the i-th auxiliary task i Select the top K individuals with the best performance, denoted as T i , from the main task population P m Select the top K individuals with the best performance on the i-th target, denoted as Tm i ; Through the formula Calculate the fitness fh of the auxiliary task i And the fitness of the main task fm i , fh i T i The average AUC of fm i Tm i The average AUC of ; 2K individuals R are randomly selected from the i-th auxiliary task i , randomly select 2K individuals Rm from the main task i ;

[0145] When the ID is the main task ID and flag i = 0, the first stage of knowledge transfer begins, transferring useful individuals in the auxiliary task to the main task, and using the rapid convergence of the auxiliary task to accelerate the optimization of the main task; when fm i >fh i , R i Join P o , to increase the diversity of the main task population; otherwise, it means that the auxiliary task performs better on this goal, and T i Join P o ;

[0146] When the ID is an auxiliary task ID and flag i= 1, the second stage of knowledge transfer begins, the auxiliary task absorbs useful individuals from the main task, and when the auxiliary task falls into a local optimum, the diversity of the main task is used to help it jump out of the local optimum. i <fh i , Rm i Join P o , in order to increase the diversity of the auxiliary task population; otherwise, it means that the main task performs better on this goal. In order to take into account both convergence and diversity in the non-fast convergence stage, Tm i and Rm i Join P o ;

[0147] When the ID is the main task ID and flag i = 2, the third stage of knowledge transfer begins. When the auxiliary task converges again with the help of the diversity of the main task, its optimized individuals are transferred back to the main task to form a circular optimization mechanism. i >fh i , R i Join P o , to increase the diversity of the main task population; at the same time flag i Self-decrement makes the knowledge transfer strategy return to the second stage. When the main task and the auxiliary task are close in performance, the evolution of the auxiliary task is promoted again; otherwise, it means that the auxiliary task performs better on this goal, and T i and R i Join P o .

[0148] For example, in the implementation case of the present invention, K is set to 10 and Num is set to 2.

[0149] Illustratively, in the implementation case of the present invention, the main population ID is Main, and the auxiliary population ID is serial number i.

[0150] S5, based on the search space and the multi-objective evolutionary algorithm and the single-objective evolutionary algorithm, a semantic maintenance strategy is used to ensure that a population that meets the semantic constraints is generated in the genetic operation; to ensure the interpretability of the final model; wherein the processes of S3, S4, and S5 are as follows Figure 2 shown.

[0151] Specifically, semantic maintenance strategies include:

[0152] First, the shape of the tree is constrained by presetting the maximum depth Dep and the maximum width Wid to directly limit the depth and width of the generated tree; the width of the tree is constrained by limiting the number of leaf nodes it contains; to ensure the rationality of the model, the student-related leaf nodes h S Leaf node h related to the exercise E Must be included in the tree;

[0153] Secondly, based on the idea of semantic approximation, semantic information is used to guide tree generation, and two "semantic-aware" sub-operations, Semantic-Exchange and Semantic-Replace, are designed. The core idea of Semantic-Exchange is to exchange two subtrees with similar semantics; the core idea of Semantic-Replace is to replace a randomly selected node with a node in the candidate operator set that is semantically similar to it; the 15 candidate operators are divided into four categories according to their semantics. The first category includes Add, Mul, Concat; the second category includes Tanh, Sigmoid, Softplus; the third category includes Sum, Mean, FFN, FFN_D; the fourth category includes Neg, Abs, Inv, Square, Sqrt.

[0154] For example, the depth (Dep) and width (Wid) are set to 10.

[0155] Specifically, the genetic operations include:

[0156] Continuously randomly select one of the four sub-operations {Semantic-Exchange, Semantic-Replace, Delete, Insert} and perform the operation on two parent individuals to generate two offspring individuals until the size of the offspring population Q reaches the required level. Before adding the generated individuals to Q, it is necessary to first check whether the tree shape meets the pre-defined depth and width constraints. Semantic-Exchange represents semantic exchange, Semantic-Replace represents semantic replacement, Delete represents the random deletion of a node, and Insert represents the random insertion of a node.

[0157] S6: Based on the optimization results of S3 and the semantic constraints of S5, determine whether convergence is satisfied;

[0158] If so, output the Pareto Optimality model to complete the entire search method;

[0159] If not, return to S3.

[0160] Exemplarily, the convergence conditions include: super volume improvement rate < 1% for 5 generations;

[0161] All Pareto solutions pass the semantic verification of S5; maximum number of iterations.

[0162] All experiments are performed on an NVIDIA RTX 4090 GPU.

[0163] This example uses accuracy (ACC), area under the ROC curve (AUC), and root mean square error (RMSE) as evaluation criteria.

[0164] In this example, seven methods were selected to compare their effects with the method of the present invention. The selected methods were DINA, IRT, MIRT, NCD, KaNCD, RCD-native, and NAS-GCD. EMT-AS-GCD-best was the method of the present invention. The experimental results are shown in Table 1:

[0165] Table 2 Performance comparison of the proposed method and the comparison algorithm on the ASSISTments2009, JunYi, e-Math and ASSISTments2012 datasets

[0166]

[0167]

[0168] Table 2 shows that EMT-AS-GCD-best achieves the best performance in almost all metrics on the public datasets ASSISTments2009, JunYi, e-Math, and ASSISTments2012, with only the RMSE on ASSISTments2009 being slightly inferior. Experiments demonstrate that the model found by this invention has good generalization capabilities.

[0169] Some of the data in the above formula are calculated by removing the dimensions and taking their numerical values. The formula is a formula that is closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0170] Working principle of the present invention:

[0171] The present invention defines a cognitive diagnosis task and models the architecture search problem of a generalized cognitive diagnosis model as a multi-objective optimization problem (MOP), wherein the goal of the MOP is to maximize the performance of the model on multiple cognitive diagnosis data sets; defines the search space of the cognitive diagnosis model based on a tree structure; generates an initial population of the model; constructs a collaborative optimization framework for a main task and multiple auxiliary tasks, and defines the objective function of each task, wherein the main task directly solves the MOP through a multi-objective evolutionary algorithm, and each auxiliary task optimizes a single objective in the MOP through a single-objective evolutionary algorithm; and improves the convergence efficiency through an asynchronous evolution mechanism; designs a multi-stage knowledge transfer strategy, and accelerates the convergence of the main task and improves the diversity of the auxiliary tasks through population knowledge transfer between the main task and the auxiliary tasks; adopts a semantic maintenance strategy to constrain the semantic similarity and topological complexity of the tree structure in genetic operations to ensure the interpretability of the generated model.

[0172] The present invention can achieve the following effects:

[0173] 1. This paper regards the MOP as the main task and the optimization of each objective in the MOP as a separate task. It introduces the idea of evolutionary multi-task optimization to solve these tasks simultaneously, so as to more effectively optimize the original complete MOP. It avoids the problem of traditional NAS methods that the search space is limited due to the simplification of the multi-objective optimization problem (MOP) and the model's insufficient generalization on new datasets. It improves the performance of the model on unobserved datasets and enhances the generalization ability of the model.

[0174] 2. The present invention adopts a multi-stage knowledge transfer strategy and utilizes the rapid convergence of auxiliary tasks to accelerate the optimization of the main task. Combined with the asynchronous evolution mechanism, it avoids the problem of wasting computing resources caused by traditional single-task optimization or synchronous evolution mechanism, reduces the evaluation cost of expensive and heterogeneous data sets, and improves convergence efficiency.

[0175] 3. The present invention designs a semantic maintenance strategy to constrain the maximum depth and width of the spanning tree, and groups operators by semantics, thus avoiding the problems of high model structure complexity, violation of educational psychology theory, and poor interpretability automatically generated by the traditional NAS method. This reduces the model complexity, makes the diagnostic logic more consistent with the cognition of educational experts, and improves the model interpretability.

[0176] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this document.

[0177] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0178] In the several embodiments provided herein, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices, or units, or can be an electrical, mechanical, or other form of connection.

[0179] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments herein.

[0180] In addition, the functional units in the various embodiments herein may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0181] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this article is essentially or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of this article. The aforementioned storage medium includes: various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0182] This article uses specific embodiments to illustrate the principles and implementation methods of this article. The description of the above embodiments is only used to help understand the methods and core ideas of this article. At the same time, for those skilled in the art, based on the ideas of this article, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation to this article.

Claims

1. A general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization, characterized by: The following steps are involved: S1. Model the architecture search problem of the generalized cognitive diagnosis model as a multi-objective optimization problem; S2. defining a search space for a cognitive diagnosis model based on the tree structure of the multi-objective optimization problem and generating an initial population of the model; S3. Constructing a collaborative optimization framework for the main task and the auxiliary task based on the multi-objective optimization problem and the initial population of the model, performing optimization using a multi-objective evolutionary algorithm and a single-objective evolutionary algorithm, and iteratively updating the population; S4. Design a multi-stage knowledge transfer strategy to obtain dynamically adjusted main task and auxiliary task populations through population knowledge transfer between the main task and auxiliary tasks; S5. Based on the search space and the multi-objective evolutionary algorithm and the single-objective evolutionary algorithm, a semantic maintenance strategy is used to ensure that a population that meets the semantic constraints is generated in a genetic operation; S6: Based on the optimization results of S3 and the semantic constraints of S5, determine whether convergence is satisfied; If so, output the Pareto optimal model to complete the entire search method; If not, return to S3.

2. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 1 is characterized by: The architecture search problem of the generalized cognitive diagnosis model is modeled as a multi-objective optimization problem, including: Enter student answer records (s i ,e j ,r ij ) and the Q matrix, where s i Indicates students, e j Indicates practice, r ij ∈{0,1} represents student s i In practice j The answer result on r ij =1 means the answer is correct, r ij =0 means wrong answer; Indicates practice e j Is the knowledge concept c related? k , Indicates association, Indicates no correlation; cognitive diagnosis is based on students’ answer records and Q matrix, and a cognitive diagnosis model is built to predict students’ answers, thereby exploring students’ knowledge proficiency, which is recorded as: In formula (1) and Students i and practice e j The student-related vector is obtained through two embedding modules Embed1(·) and Embed2(·) and practice related vectors Extract vectors related to knowledge concepts from Q Representation Model The predicted results; After training, h S Used to indicate students’ knowledge proficiency. Varies across different cognitive diagnostic models.

3. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 1 is characterized by: Defining the search space of the cognitive diagnosis model according to the tree structure of the multi-objective optimization problem includes: The preset computation tree consists of several leaf nodes and internal computation nodes. The leaf nodes represent the input of the model and are selected from {h S ,h E ,h C }, the internal computing node selects from 15 candidate operators. With the support of this search space, the cognitive diagnosis model is rewritten as: In formula (2), the first part It is used to process h S ,h E and h C The computation tree of , whose output is a scalar or vector; the second part The final prediction result is further generated based on the output of the first part, and all cognitive diagnosis models in the search space are represented by tree coding.

4. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 3 is characterized by: The candidate operators include: The formula x+y is written as Add; The formula x·yorx⊙y is denoted as Mul; Formula -x, denoted as Neg; Formula |x|, denoted as Abs; Formula 1 / (x+∈),∈=1e -6 , recorded as Inv; Formula x 2 , recorded as Square; formula Denoted as Sqrt; formula Recorded as Sum; formula Recorded as Mean; The formula tanh(x) is denoted as Tanh; Formula sigmoid(x), denoted as Sigmoid; Formula softplus(x), denoted as Softplus; formula Denoted as Concat; formula Denoted as FFN; formula Denoted as FFND.

5. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 1 is characterized by: The initial population of generative models includes half existing models and half randomly generated models.

6. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 1 is characterized by: The collaborative optimization framework for building the main task and the auxiliary task includes: First, the iteration counters g and g of the main task and each auxiliary task are i Initialized to 0; Second, a model population is initialized for all tasks; Finally, the main task and each auxiliary task are executed simultaneously through different threads; Specific steps of the main task: Main population P m Execute the "mating pool selection", "genetic operation" and "deduplication" steps in sequence to generate the offspring population Q m If the auxiliary task has not yet completed its previous η×Gen generation evolution, the main task will not receive any knowledge from the auxiliary task; at this time, the main task will complete the evaluation of Q m After that, directly from P m and Q m The main task will call the knowledge transfer strategy in the later stage to obtain potentially useful individuals from the auxiliary tasks and store them in P t ; Then for P t Evaluate and compare with P m and Q m Compare together to get a new main population; when the number of iterations g exceeds the set upper limit Gen, and the main population P m When the hypervolume of does not change in five consecutive generations, the main task stops; The specific steps of auxiliary tasks are the same as those of the main task, but each auxiliary task has a flag i Used to record the convergence status of the auxiliary task population, flag i Initially 0, it increases automatically when the auxiliary population converges, but does not exceed 2. The multi-stage knowledge transfer strategy will switch the knowledge transfer stage according to this flag; When the main task is completed, output the main population P m The set of non-dominated individuals P in non .

7. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 6 is characterized by: The main task objective function is expressed as: The problem requires maximizing the model A over multiple observed data sets. Performance on There are Num auxiliary tasks, where the objective function of the i-th auxiliary task is: In formulas (3) and (4), Represents the search space. The constructed multi-objective optimization problem has Num objectives, each of which For model A in the dataset After training, on the validation set The AUC on .

8. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 1 is characterized by: The multi-stage knowledge transfer strategy includes: Input main population P m , auxiliary population {P i |1≤i≤Num}, population ID, convergence status flag; Initialize the output population P o Is an empty set; judge whether all auxiliary tasks have been traversed, if yes, get the population P for knowledge transfer o , No, continue to supplement P o ; From the population P of the i-th auxiliary task i Select the top K individuals with the best performance, denoted as T i , from the main task population P m Select the top K individuals with the best performance on the i-th target, denoted as Tm i ; Through the formula Calculate the fitness fh of the auxiliary task i And the fitness of the main task fm i , fh i T i The average AUC of fm i Tm i The average AUC of ; 2K individuals R are randomly selected from the i-th auxiliary task i , randomly select 2K individuals Rm from the main task i ; When the ID is the main task ID and flag i = 0, the first stage of knowledge transfer begins, transferring useful individuals in the auxiliary task to the main task, and using the rapid convergence of the auxiliary task to accelerate the optimization of the main task; when fm i >fh i , R i Join P o , to increase the diversity of the main task population; otherwise, it means that the auxiliary task performs better on this goal, and T i Join P o ; When the ID is an auxiliary task ID and flag i = 1, the second stage of knowledge transfer begins, the auxiliary task absorbs useful individuals from the main task, and when the auxiliary task falls into the local optimum, the diversity of the main task is used to help it jump out of the local optimum; when fm i <fh i , Rm i Join P o , in order to increase the diversity of the auxiliary task population; otherwise, it means that the main task performs better on this goal. In order to take into account both convergence and diversity in the non-fast convergence stage, Tm i and Rm i Join P o ; When the ID is the main task ID and flag i = 2, the third stage of knowledge transfer begins. When the auxiliary task converges again with the help of the diversity of the main task, its optimized individuals are transferred back to the main task to form a circular optimization mechanism. i >fh i , R i Join P o , to increase the diversity of the main task population; at the same time flag i Self-decrement makes the knowledge transfer strategy return to the second stage. When the main task and the auxiliary task are close in performance, the evolution of the auxiliary task is promoted again; otherwise, it means that the auxiliary task performs better on this goal, and T i and R i Join P o .

9. The general cognitive diagnosis model architecture search method based on evolutionary multi-task optimization according to claim 1, characterized in that: The semantic maintenance strategy includes: First, the shape of the tree is constrained by presetting the maximum depth Dep and the maximum width Wid to directly limit the depth and width of the generated tree; the width of the tree is constrained by limiting the number of leaf nodes it contains; to ensure the rationality of the model, the student-related leaf nodes h S Leaf node h related to the exercise E Contained in the tree; Secondly, we leverage the idea of semantic approximation and use semantic information to guide tree generation. We design two "semantic-aware" sub-operations, Semantic-Exchange and Semantic-Replace. The core idea of Semantic-Exchange is to exchange two semantically similar subtrees; the core idea of Semantic-Replace is to replace a randomly selected node with a semantically similar node from the candidate operator set. The 15 candidate operators are divided into four categories based on semantics: the first category includes Add, Mul, and Concat; the second category includes Tanh, Sigmoid, and Softplus; the third category includes Sum, Mean, FFN, and FFN_D; and the fourth category includes Neg, Abs, Inv, Square, and Sqr. The genetic operation includes: Continuously randomly select one of the four sub-operations {Semantic-Exchange, Semantic-Replace, Delete, Insert}, operate on two parent individuals, and generate two offspring individuals until the size of the offspring population Q reaches the requirement; before adding the generated individuals to Q, it is necessary to first check whether their tree shape meets the pre-set depth and width restrictions. Semantic-Exchange represents semantic exchange, Semantic-Replace represents semantic replacement, Delete represents random deletion of a node, and Insert represents random insertion of a node.

10. A computer-readable storage medium, characterized in that: The storage medium stores a program, and the program is executed by a processor to implement the method according to any one of claims 1 to 9.

Images