A multi-granularity knowledge tracking method and system for a cold start scenario

Abstract

This invention relates to the field of knowledge tracing technology, and more particularly to a multi-granularity knowledge tracing method and system for cold-start scenarios. The method first acquires the historical interaction data of the target learner to construct a behavioral feature vector, and matches the target learner to the corresponding learner group category. Subsequently, through global knowledge modeling, group knowledge modeling, and individual knowledge modeling, a global knowledge state representation, a group knowledge state representation, and an individual knowledge state representation at the current moment are generated and weighted and fused. Based on the accumulation degree of the target learner's historical interaction data, the fusion weights of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation are dynamically determined to obtain the target knowledge state representation and complete knowledge tracing. This invention effectively solves the prediction instability problem caused by data sparsity in the cold-start stage through multi-granularity knowledge flow fusion and dynamic weight scheduling.

Description

Technical Field

[0001] This invention relates to the field of knowledge tracing technology, and in particular to a multi-granularity knowledge tracing method and system for cold start scenarios. Background Technology

[0002] With the widespread adoption of smart devices, online education has become an important direction for educational development. Knowledge tracking is a crucial part of online education. By analyzing learners' historical learning behaviors and building cognitive models, learners can be dynamically assessed to dynamically evaluate their knowledge mastery. This is a vital foundation for personalized exercise recommendations and learning path planning. However, in practical applications, many new learners using online education platforms for the first time have limited or no historical learning data. This makes it difficult for existing knowledge tracking models to effectively model their cognitive state, resulting in low prediction accuracy and unstable prediction results.

[0003] The initial launch and acquisition of initial users and feedback for a new system, course, or platform, without sufficient user behavior data, market awareness, or interactive feedback, is known as the cold start phase. The core challenge of this phase is the lack of historical data, making accurate recommendations, personalized services, and effective promotion difficult. Therefore, other strategies are needed to validate value, accumulate initial users, and achieve initial growth. During the cold start phase, if the model has limited data, it may be unable to reasonably assess learners' knowledge levels, leading to a mismatch between personalized recommendations and learners' actual abilities, impacting learning outcomes and ultimately damaging the learner's experience. Therefore, the predictive stability and accuracy during the cold start phase have become a technical bottleneck restricting the practical deployment of knowledge tracing models.

[0004] Existing deep learning-based knowledge tracking models rely heavily on learners' historical interaction data in their structural design. When the amount of data is small, the model often has to rely on random initialization for prediction, resulting in low accuracy and poor stability. It is also difficult to achieve a smooth transition as data accumulates, which severely limits the model's practicality in real-world educational scenarios. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] Therefore, the first objective of this invention is to propose a multi-granularity knowledge tracing method for cold start scenarios, the steps of which include:

[0007] S1, acquire the historical interaction data of the target learner, construct a behavioral feature vector based on the historical interaction data, and match the target learner to the corresponding learner group category; S2 generates the global knowledge state representation, group knowledge state representation, and individual knowledge state representation at the current moment through global knowledge modeling, group knowledge modeling, and individual knowledge modeling, respectively. S3, the global knowledge state representation, group knowledge state representation and individual knowledge state representation are weighted and fused based on the fusion weights, and the fusion weights of the global knowledge state representation, group knowledge state representation and individual knowledge state representation are dynamically determined according to the historical interaction data accumulation degree of the target learner, so as to obtain the target knowledge state representation to complete knowledge tracking.

[0008] In one embodiment of the present invention, S1 further includes: S11, extract relevant features of the target learner’s learning behavior from historical learning logs, and construct a behavioral feature vector containing various types of learning behavior features; S12, an unsupervised clustering algorithm is used to perform cluster analysis on the learners' behavioral feature vectors, to divide the learners into clusters, and to obtain the center vector of each cluster of learners. S13. Based on the center vectors of learners in each cluster, a distance-based soft matching method is used to calculate the probability that the target learner belongs to each cluster, and the target learner is matched to the learner group category with the highest probability.

[0009] In one embodiment of the present invention, S2 further includes: S21, Train a global knowledge representation model based on the historical interaction data of all learners, and use the global knowledge modeling function to process the current question, knowledge points and answer results to obtain the knowledge state representation of the global flow at time ; S22, construct a corresponding clustering knowledge representation model for each cluster to obtain the knowledge state representation of the cluster at a certain time, and represent the clustering flow knowledge state of the target learner at a certain time by the weighted sum of the knowledge states of each cluster; S23. Based on the target learner's own interaction sequence, the individual knowledge modeling function is used to process the individual flow knowledge state representation at the previous moment and the interaction data at the current moment to obtain the knowledge state representation output by the individual flow at that moment as the individual knowledge state representation.

[0010] In one embodiment of the present invention, S3 further includes: S31, multiply the knowledge state representation output by the global stream at time t, the knowledge state representation output by the clustering stream at time t, and the knowledge state representation output by the individual stream at time t by their respective fusion weights to obtain the fused learner knowledge state representation; S32, construct a gated network, and use the learner's interaction feature vector at this moment to adaptively generate global flow weights, cluster flow weights and individual flow weights through the softmax function. The interaction feature vector is a feature vector that reflects the degree of historical interaction data accumulation of the target learner. S33 imposes constraints on weight allocation when the learner is in the cold start phase, such that the individual flow weight is less than the sum of the cluster flow weight and the global flow weight.

[0011] In one embodiment of the present invention, the calculation formula for the soft matching method in step S13 is as follows: ; in, Let s be the behavioral feature vector of learner s. For the first k Cluster centers, Number of learner clusters.

[0012] In one embodiment of the present invention, the clustering flow knowledge state expression of the target learner in step S22 is: ; in, The cluster where the target learners are located at time The knowledge state representation, For target learners s Belongs to clustering k The probability, Number of learner clusters.

[0013] In one embodiment of the present invention, the formula for representing the individual knowledge state in step S23 is: ; in, This represents a global knowledge modeling function. The question for the learner to answer at time t. For the question The corresponding knowledge points For the answer result, .

[0014] In one embodiment of the present invention, S33 further includes: As the amount of learner interaction data increases, the system will relax the weight allocation constraints and increase the weight of individual streams.

[0015] To achieve the above objectives, a second aspect of the present invention proposes a multi-granularity knowledge tracing system for cold start scenarios, comprising: The data acquisition module is used to acquire the historical interaction data of the target learner and construct a behavioral feature vector based on the historical interaction data; The clustering flow construction module is used to match the target learner to the corresponding learner group category and obtain the current group knowledge state representation through group knowledge modeling. The individual flow construction module is used to obtain the current individual knowledge state representation based on the learner's own interaction sequence through individual knowledge modeling. The global flow construction module is used to train a global knowledge representation model based on the historical interaction data of all learners, perform global knowledge modeling, and obtain the current global knowledge state representation. The three-stream fusion module performs weighted fusion of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation based on the fusion weights, and dynamically determines the fusion weights of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation according to the degree of historical interaction data accumulation of the target learner, so as to obtain the target knowledge state representation to complete knowledge tracking.

[0016] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0017] The method, system, and storage medium of this invention effectively solve the problem of unstable knowledge state prediction for new learners due to insufficient historical interaction data, significantly improving the prediction accuracy during the cold start phase. Furthermore, through a dynamic weight fusion mechanism, a smooth transition from group knowledge modeling to individual personalized modeling is achieved.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a multi-granularity knowledge tracing method for cold start scenarios according to an embodiment of the present invention; Figure 2 This is a structural diagram of a multi-granularity knowledge tracing system for cold start scenarios according to an embodiment of the present invention; Figure 3 This is a model diagram of a multi-granularity knowledge tracing system for cold start scenarios according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes a multi-granularity knowledge tracing method and system for cold start scenarios proposed according to embodiments of the present invention.

[0023] Example 1 Figure 1 This is a flowchart of a multi-granularity knowledge tracing method for cold start scenarios according to an embodiment of the present invention.

[0024] like Figure 1 As shown, the multi-granularity knowledge tracing method for cold start scenarios includes the following steps: S1, acquire the target learner's historical interaction data, construct a behavioral feature vector based on the historical interaction data, and match the target learner to the corresponding learner group category.

[0025] Specifically, the system first collects interaction data from the target learners, extracting multidimensional features that characterize learning habits and aptitudes to construct behavioral feature vectors. Then, based on a pre-defined learner group classification model, it calculates the correlation between the target learner and each group category, completing category matching. This matching process is not limited to hard classification; probability distributions can also be used to represent the affiliation relationship.

[0026] Specifically, the system constructs a learning behavior feature vector for each learner from historical learning logs. This feature vector is used to describe the overall learning behavior pattern of learners. As one implementation method, the types of learning behavior features include statistical features of correct answer rates for knowledge points, error recovery speed and continuous error distribution features, learning interval time structure features, knowledge transfer ability features, and forgetting speed. This feature vector includes at least one or more types of features.

[0027] The learner's behavioral feature vector is represented as follows: ; in This represents the learning behavior characteristics of the m-th class.

[0028] Subsequently, clustering was performed on each learner, and the system generated behavioral feature vectors for all learners. Unsupervised clustering algorithms are used for cluster analysis to divide learners into groups, each group corresponding to a typical learning behavior pattern. After clustering, the system obtains the center vector of each cluster. .

[0029] After completing the learner clustering analysis, for each new learner entering the system, the system calculates the probability of them belonging to each cluster based on their initial interaction behavior, using a distance-based soft matching method. The calculation formula is as follows: ; in, Let s be the behavioral feature vector of learner s. For the first k Cluster centers, Number of learner clusters.

[0030] By constructing behavioral feature vectors and matching them with group categories, new learners can reuse the knowledge patterns of similar groups when they lack individual historical data. This significantly improves the stability of knowledge state prediction during the cold start phase and provides a reliable initial prior for subsequent personalized modeling, avoiding prediction fluctuations caused by random initialization.

[0031] S2 generates the global knowledge state representation, group knowledge state representation, and individual knowledge state representation at the current moment through global knowledge modeling, group knowledge modeling, and individual knowledge modeling, respectively.

[0032] Further, step S2 includes: S21. Train a global knowledge representation model based on the historical interaction data of all learners, and use the global knowledge modeling function to process the question, knowledge points and answer results at the current moment to obtain the knowledge state representation of the global flow at the current moment.

[0033] At any moment The knowledge state output by the global stream is represented as follows: ; in This represents a global knowledge modeling function. The question for the learner to answer at time t. For the question The corresponding knowledge points For the answer result, .

[0034] S22, construct a corresponding clustering knowledge representation model for each cluster to obtain the knowledge state representation of the cluster at a certain time, and represent the clustering flow knowledge state of the target learner at a certain time by the weighted sum of the knowledge states of each cluster.

[0035] Specifically, the clustering flow knowledge state expression for the target learner is: ; in, The cluster where the target learners are located at time The knowledge state representation, For target learners s Belongs to clustering k The probability, Number of learner clusters.

[0036] S23. Based on the target learner's own interaction sequence, the individual knowledge modeling function is used to process the individual flow knowledge state representation at the previous moment and the interaction data at the current moment to obtain the knowledge state representation output by the individual flow at that moment as the individual knowledge state representation.

[0037] The formula for representing an individual's knowledge state is: ; in, This represents the global knowledge modeling function.

[0038] This step separates the knowledge modeling process at different granularities, enabling the model to maintain predictive stability by relying on global and collective knowledge when data is sparse, while ensuring personalized expression capabilities through individual knowledge flow when data is abundant. This achieves a balance between robustness and adaptability in knowledge tracing under cold start scenarios.

[0039] S3, the global knowledge state representation, group knowledge state representation and individual knowledge state representation are weighted and fused based on the fusion weights, and the fusion weights of the global knowledge state representation, group knowledge state representation and individual knowledge state representation are dynamically determined according to the historical interaction data accumulation degree of the target learner, so as to obtain the target knowledge state representation to complete knowledge tracking.

[0040] S31, multiply the knowledge state representation output by the global stream at time t, the knowledge state representation output by the clustering stream at time t, and the knowledge state representation output by the individual stream at time t by their respective fusion weights to obtain the fused learner knowledge state representation.

[0041] Specifically, the formula for representing a learner's knowledge state is: ; in, For global flow weights, This represents the knowledge state output by the global stream at time t. For cluster flow weights, This represents the knowledge state representation output by the clustering stream at time t. For individual flow weights, This represents the knowledge state output by an individual stream at time t.

[0042] S32. Construct a gated network and use the learner's interaction feature vector at that moment to adaptively generate global flow weights, cluster flow weights, and individual flow weights through the softmax function. The interaction feature vector is a feature vector that reflects the degree of historical interaction data accumulation of the target learner.

[0043] The expression for the gated network is: ; in, Let W represent the learner's interaction feature vector at time t, where W and b are learnable parameters.

[0044] S33 imposes constraints on weight allocation when the learner is in the cold start phase, such that the individual flow weight is less than the sum of the cluster flow weight and the global flow weight.

[0045] When learners are in the cold start phase, the system imposes the following constraints on weight allocation: ; As the amount of learner interaction data increases, the system gradually relaxes the above constraints, continuously increasing the weight of individual flows. This enables the system to ensure prediction stability by utilizing global and group common knowledge when data is sparse, and to improve accuracy by utilizing individual knowledge when data is abundant.

[0046] Compared with existing technologies, this invention introduces global flow and cluster flow as the main prediction basis in the cold start phase. Before individual flows have sufficient learning data, predictions are made using common knowledge patterns of all learners and prior knowledge of similar learner groups, thus significantly improving the stability and accuracy of knowledge state prediction in the cold start phase. By introducing a dynamic weight scheduling mechanism, the weights of global flow, cluster flow, and individual flow continuously change with the amount of learner interaction data, achieving a smooth transition from "group knowledge-dominated" to "individual knowledge-dominated" and avoiding prediction discontinuity problems caused by model switching. A parallel three-type knowledge flow modeling structure is constructed. As learner interaction data gradually becomes richer, the weight of individual flow in the fusion result is automatically increased, enabling the model to gradually strengthen its ability to characterize learners' personalized knowledge states, achieving a dynamic balance between stability and personalization. The knowledge tracking process is divided into three independent but collaborative modeling modules: global flow, cluster flow, and individual flow. Each module has clear functional boundaries, can be trained independently, or jointly optimized, facilitating engineering deployment in actual teaching systems or online learning platforms. Furthermore, the three-stream fusion structure of this invention has good scalability, and new knowledge modeling modules or specific implementation models can be introduced or replaced without changing the overall framework, adapting to different teaching scenarios and data scale requirements.

[0047] Example 2 like Figure 2 As shown, this invention proposes a multi-granularity knowledge tracing system 10 for cold start scenarios, comprising: The data acquisition module 100 is used to acquire the historical interaction data of the target learner and construct a behavioral feature vector based on the historical interaction data. The clustering flow construction module 200 is used to match the target learner to the corresponding learner group category and obtain the current group knowledge state representation through group knowledge modeling. The individual flow construction module 300 is used to obtain the current individual knowledge state representation based on the learner's own interaction sequence through individual knowledge modeling. The global flow construction module 400 is used to train a global knowledge representation model based on the historical interaction data of all learners, perform global knowledge modeling, and obtain the current global knowledge state representation. The three-stream fusion module 500 performs weighted fusion of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation based on the fusion weights, and dynamically determines the fusion weights of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation according to the degree of historical interaction data accumulation of the target learner, so as to obtain the target knowledge state representation to complete knowledge tracking.

[0048] Example 3 To further illustrate the system proposed in this invention, the following description uses a specific multi-granularity knowledge tracing system model for cold start scenarios.

[0049] like Figure 3 As shown, the model achieves adaptive knowledge tracing for cold starts based on global flow, cluster flow, and individual flow.

[0050] First, the historical interaction data of the target learner is acquired using the data acquisition module 100, and a behavioral feature vector is constructed based on the historical interaction data.

[0051] Subsequently, the learner data was processed using the clustering flow construction module 200, the individual flow construction module 300, and the global flow construction module 400, respectively. Specifically, In the clustering stream construction module 200, learner data is processed by the spectral clustering algorithm module to construct a group knowledge space, resulting in a knowledge space construction result. This knowledge space construction result is then input into the clustering stream to obtain the clustering stream processing result.

[0052] In the individual flow construction module 300, the learner data is divided into two groups. One group is processed by the genetic algorithm module to obtain the genetic algorithm processing result. The other group is further divided into two groups and processed by the forgetting perception bias algorithm module and the interval repetition enhancement mechanism module, respectively. Attention weights are calculated on the processing results of the two groups to obtain the calculation results. The attention weight calculation results and the genetic algorithm processing results are input into the individual flow module to obtain the individual flow processing result.

[0053] In the global stream construction module 400, the global knowledge construction module constructs global general knowledge from learner data, inputs the construction results into the global stream, and obtains the global stream processing results.

[0054] Finally, using the dynamic gating network in the three-stream fusion module 500, the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation are weighted and fused based on the fusion weights. The fusion weights of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation are dynamically determined according to the historical interaction data accumulation of the target learner, so as to obtain the target knowledge state representation and complete knowledge tracking.

[0055] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described multi-granularity knowledge tracing method for cold start scenarios.

[0056] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0057] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A multi-granularity knowledge tracing method for cold start scenarios, characterized by the following steps: include: S1, acquire the historical interaction data of the target learner, construct a behavioral feature vector based on the historical interaction data, and match the target learner to the corresponding learner group category; S2 generates the global knowledge state representation, group knowledge state representation, and individual knowledge state representation at the current moment through global knowledge modeling, group knowledge modeling, and individual knowledge modeling, respectively. S3, the global knowledge state representation, group knowledge state representation and individual knowledge state representation are weighted and fused based on the fusion weights, and the fusion weights of the global knowledge state representation, group knowledge state representation and individual knowledge state representation are dynamically determined according to the historical interaction data accumulation degree of the target learner, so as to obtain the target knowledge state representation to complete knowledge tracking.

2. The method as described in claim 1, characterized in that, S1 further includes: S11, extract relevant features of the target learner’s learning behavior from historical learning logs, and construct a behavioral feature vector containing various types of learning behavior features; S12, an unsupervised clustering algorithm is used to perform cluster analysis on the learners' behavioral feature vectors, to divide the learners into clusters, and to obtain the center vector of each cluster of learners. S13. Based on the center vectors of learners in each cluster, a distance-based soft matching method is used to calculate the probability that the target learner belongs to each cluster, and the target learner is matched to the learner group category with the highest probability.

3. The method as described in claim 1, characterized in that, S2 further includes: S21, Train a global knowledge representation model based on the historical interaction data of all learners, and use the global knowledge modeling function to process the current question, knowledge points and answer results to obtain the knowledge state representation of the global flow at time ; S22, construct a corresponding clustering knowledge representation model for each cluster to obtain the knowledge state representation of the cluster at a certain time, and represent the clustering flow knowledge state of the target learner at a certain time by the weighted sum of the knowledge states of each cluster; S23. Based on the target learner's own interaction sequence, the individual knowledge modeling function is used to process the individual flow knowledge state representation at the previous moment and the interaction data at the current moment to obtain the knowledge state representation output by the individual flow at that moment as the individual knowledge state representation.

4. The method as described in claim 1, characterized in that, S3 further includes: S31, multiply the knowledge state representation output by the global stream at time t, the knowledge state representation output by the clustering stream at time t, and the knowledge state representation output by the individual stream at time t by their respective fusion weights to obtain the fused learner knowledge state representation; S32, construct a gated network, and use the learner's interaction feature vector at this moment to adaptively generate global flow weights, cluster flow weights and individual flow weights through the softmax function. The interaction feature vector is a feature vector that reflects the degree of historical interaction data accumulation of the target learner. S33 imposes constraints on weight allocation when the learner is in the cold start phase, such that the individual flow weight is less than the sum of the cluster flow weight and the global flow weight.

5. The method as described in claim 2, characterized in that, The calculation formula for the soft matching method in step S13 is as follows: ； in, Let s be the behavioral feature vector of learner s. For the first k Cluster centers, Number of learner clusters.

6. The method as described in claim 3, characterized in that, The clustering flow knowledge state expression for the target learner in step S22 is: ； in, The cluster where the target learners are located at time The knowledge state representation, For target learners s Belongs to clustering k The probability, Number of learner clusters.

7. The method as described in claim 3, characterized in that, The formula for representing the individual knowledge state in step S23 is: ； in, This represents a global knowledge modeling function. The question for the learner to answer at time t. For the question The corresponding knowledge points For the answer result, .

8. The method as described in claim 4, characterized in that, S33 further includes: As the amount of learner interaction data increases, the system will relax the weight allocation constraints and increase the weight of individual streams.

9. A multi-granularity knowledge tracing system for cold start scenarios, characterized in that, include: The data acquisition module is used to acquire the historical interaction data of the target learner and construct a behavioral feature vector based on the historical interaction data; The clustering flow construction module is used to match the target learner to the corresponding learner group category and obtain the current group knowledge state representation through group knowledge modeling. The individual flow construction module is used to obtain the current individual knowledge state representation based on the learner's own interaction sequence through individual knowledge modeling. The global flow construction module is used to train a global knowledge representation model based on the historical interaction data of all learners to perform global knowledge modeling and processing, and obtain the current global knowledge state representation. The three-stream fusion module performs weighted fusion of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation based on the fusion weights, and dynamically determines the fusion weights of the global knowledge state representation, the group knowledge state representation, and the individual knowledge state representation according to the degree of historical interaction data accumulation of the target learner, so as to obtain the target knowledge state representation to complete knowledge tracking.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as claimed in any one of claims 1-8.

Images