Adaptive Control Knowledge Tracking Method Based on Dual-Channel Modeling

By employing dual-channel modeling and adaptive control methods, this approach addresses the issues of fixed metrics and lack of unified adaptive control in existing knowledge tracing methods, achieving clear representation and stable prediction of learners' recent behavior and long-term cognition.

CN123113345APending Publication Date: 2026-11-17GUANGXI NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611320466.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing knowledge tracing methods tend to amplify the impact of accidental interactions when processing historical interaction information by using fixed metrics, and lack a unified adaptive control mechanism, resulting in unstable prediction results and insufficient probabilistic reliability, especially when historical evidence is sparse or interaction noise is high.

Method used

An adaptive control method based on dual-channel modeling is adopted. By constructing short-term behavioral channels and long-term cognitive channels, learners’ recent response changes and long-term knowledge evolution are modeled respectively. Shared evidence generation units and gated fusion prediction main paths are introduced to dynamically adjust the historical retrieval and state fusion processes. Adaptive retrieval distance is constructed using squared-Wasserstein distance and Fisher–Rao-inspired dimension-by-dimensional approximation distance.

Benefits of technology

It improves the clarity of knowledge state representation and the stability of prediction, reduces the impact of accidental interactions, enhances the adaptability of the model under different learning stages and evidence conditions, and improves the accuracy and reliability of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113345A_ABST
    Figure CN123113345A_ABST
Patent Text Reader

Abstract

This invention discloses an adaptive control knowledge tracing method based on dual-channel modeling. It models short-term behavioral changes and long-term knowledge evolution through parallel, functionally heterogeneous short-term behavioral channels and long-term cognitive channels, respectively. Then, in the gated fusion prediction main path, the fusion weights of the short-term behavioral state and the long-term cognitive state are determined, generating the final control state for fusion control and outputting the probability of the next response. This invention can dynamically adjust the historical retrieval process and the dual-channel fusion process based on current evidence information, improving the stability, robustness, and probabilistic reliability of knowledge tracing prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of knowledge tracing (KT) technology, specifically to an adaptive control knowledge tracing method based on dual-channel modeling. Background Technology

[0002] Knowledge tracking is a fundamental technology in intelligent tutoring, personalized practice recommendations, and learning diagnosis. Its purpose is to infer a learner's current knowledge mastery based on their historical interaction records over a period of time and to predict their subsequent performance. The learning process itself is characterized by both significant short-term fluctuations and long-term evolution. Relying solely on a single state to compress all historical information can easily lead to the mixed encoding of recent behavioral changes with stable knowledge trajectories over a longer period, thus affecting the accuracy and stability of knowledge state estimation.

[0003] While existing methods have made progress in multi-timescale state modeling and uncertainty-aware prediction, a unified evidence modifier is still lacking for historical retrieval and state fusion. Specifically, this is manifested in the following ways:

[0004] 1) Existing knowledge tracing methods use fixed similarity metrics or fixed distribution distances to match current interactions with historical interactions when processing historical interaction information. While these methods can adjust retrieval weights based on input changes, their underlying matching rules often remain unchanged, making it difficult to dynamically adjust the matching geometry based on the strength of the learner's current evidence, the richness of historical interactions, and different learning stages. When historical evidence is scarce or interaction noise is high, fixed metrics can easily amplify the influence of accidental interactions; when historical evidence is abundant, it may be difficult to identify more subtle state differences, thus affecting the quality of long-term cognitive state formation.

[0005] 2) Existing knowledge tracing methods typically lack a unified adaptive control mechanism when fusing short-term behavioral information with long-term cognitive information. Some methods only introduce multi-scale modeling in the state representation stage, while others only perform probability calibration in the output stage. However, they fail to simultaneously utilize evidence information in both the historical retrieval stage and the state fusion stage, making it difficult to dynamically adjust the historical retrieval method and fusion weights based on the current evidence state. This deficiency is more pronounced in scenarios with sparse historical data, long-tailed knowledge points, or significant fluctuations in response behavior, easily leading to unstable prediction results, insufficient probabilistic reliability, and excessive reliance on short-term noise.

[0006] Therefore, historical matching based on fixed metrics may amplify the impact of accidental interactions when evidence is sparse or noisy, while state fusion without evidence makes it difficult for the model to reasonably allocate dependencies when gating signals conflict, thereby weakening prediction stability and probabilistic reliability. Summary of the Invention

[0007] The present invention addresses the problems of existing knowledge tracing methods, such as mixed modeling of short-term behavioral states and long-term cognitive states, fixed historical retrieval metrics, and lack of unified adaptive control for state fusion. It provides an adaptive control knowledge tracing method based on dual-channel modeling.

[0008] To solve the above problems, the present invention is achieved through the following technical solution:

[0009] The adaptive control knowledge tracing method based on dual-channel modeling includes the following steps:

[0010] Step 1: Construct a knowledge tracking network. This knowledge tracking network consists of an embedded coding unit, a short-term behavior channel, a long-term cognitive channel, and a gated fusion prediction main path. The input of the embedded coding unit forms the input of the knowledge tracking network. The output of the embedded coding unit is connected to the input of the short-term behavior channel and the long-term cognitive channel. The output of the short-term behavior channel and the long-term cognitive channel is connected to the input of the gated fusion prediction main path. The output of the gated fusion prediction main path forms the output of the knowledge tracking network.

[0011] Step 2: Obtain the learner's historical interaction sequence and feed the historical interaction sequence into the knowledge tracing network to train it, thereby obtaining the knowledge tracing model;

[0012] Step 3: Obtain the learner's question to be predicted and feed it into the knowledge tracking model to obtain the learner's predicted answer to the question.

[0013] In the above scheme, the short-term behavior channel includes a short-term attenuation unit and a GRU unit; the input of the short-term attenuation unit forms the input of the short-term behavior channel, the output of the short-term attenuation unit is connected to the input of the GRU unit, and the output of the GRU unit forms the output of the short-term behavior channel.

[0014] In the above scheme, the short-term decay unit uses an exponential decay function for rapid decay, while the long-term decay unit uses a power-law decay function for slow decay.

[0015] In the above scheme, the long-term cognitive channel includes a long-term decay unit, a long-term normal-inverse gamma distribution evidence unit, a long-term uncertainty unit, a prefix center unit, and a CAM-Transformer encoder unit. The input of the long-term decay unit forms the input of the long-term cognitive channel, the output of the long-term decay unit is connected to the input of the long-term normal-inverse gamma distribution evidence unit, the output of the long-term normal-inverse gamma distribution evidence unit is connected to the input of the long-term uncertainty unit, and the output of the long-term uncertainty unit is connected to the input of the prefix center unit. The outputs of the long-term decay unit and the prefix center unit are connected to the input of the CAM-Transformer encoder unit, and the output of the CAM-Transformer encoder unit forms the output of the long-term cognitive channel.

[0016] In the above scheme, the short-term decay unit uses an exponential decay function for rapid decay, while the long-term decay unit uses a power-law decay function for slow decay.

[0017] In the above scheme, the CAM-Transformer encoder unit consists of multiple encoding layers connected in series. Each encoding layer consists of a CAM module, a feedforward network, and two residual connection and normalization modules. The input of the CAM module forms the input of the encoding layer, the output of the CAM module is connected to the input of the first residual connection and normalization module, the output of the first residual connection and normalization module is connected to the input of the feedforward network, the output of the feedforward network is connected to the second residual connection and normalization module, and the output of the second residual connection and normalization module forms the output of the encoding layer. The input of the first encoding layer forms one input of the entire CAM-Transformer encoder unit, the output of the last encoding layer forms the output of the entire CAM-Transformer encoder unit, and the other input of the CAM module of each encoding layer forms the other input of the entire CAM-Transformer encoder unit.

[0018] In the above scheme, the CAM module first maps the long-term decay embedding representation output by the long-term decay unit into a mean vector and a scale vector, respectively, and constructs a diagonal Gaussian retrieval distribution based on the mean vector and the scale vector; then it calculates the square between the diagonal Gaussian retrieval distribution of the last interaction and the previous interaction. -Wasserstein distance and Fisher–Rao-inspired dimensionally approximate distance; then squared The original CAM distance is generated by weighting the Wasserstein distance and the Fisher–Rao-inspired dimension-wise approximate distance.

[0019] In the above scheme, the gated fusion prediction main path consists of a gated normal-inverse gamma distribution evidence unit, a gated uncertainty unit, a gated variance unit, a gated weight unit, a gated fusion unit, a response predictor unit, and a normalization unit. The short-term input of the gated fusion unit forms one input of the gated fusion prediction main path, and the long-term input of the gated fusion unit and the input of the gated normal-inverse gamma distribution evidence unit form another input of the gated fusion prediction main path. The output of the gated normal-inverse gamma distribution evidence unit is connected to the input of the gated uncertainty unit, the output of the gated uncertainty unit is connected to the input of the gated variance unit, the outputs of the gated uncertainty unit and the gated variance unit are connected to the input of the gated weight unit, the output of the gated weight unit is connected to the weight input of the gated fusion unit, the output of the gated fusion unit is connected to the input of the response predictor unit, the output of the response predictor unit is connected to the input of the normalization unit, and the output of the normalization unit forms the output of the gated fusion prediction main path.

[0020] Compared with the prior art, the present invention has the following characteristics:

[0021] 1. By constructing short-term behavioral channels and long-term cognitive channels and dividing their functions, learners can model recent changes in responses and long-term knowledge evolution separately, thereby reducing the mixed encoding of information at different time scales by a single state and improving the clarity of knowledge state representation.

[0022] 2. By introducing a shared evidence generation unit, an adaptive control signal is generated before historical retrieval, and the matching geometry between historical interactions and current interactions is dynamically adjusted accordingly. This enables the historical retrieval process to adaptively adjust according to the strength of current evidence, avoiding the over-amplification of accidental interactions by fixed metrics when evidence is sparse or noise is high.

[0023] 3. By representing current and historical interactions as probability distributions and constructing an adaptive retrieval distance by combining the squared 2-Wasserstein distance and the Fisher-Rao-inspired dimension-wise approximation distance, historical retrieval not only considers the overall differences in distribution center and scale, but also takes into account the local differences in standardized parameters, thereby improving the precision of long-term cognitive state construction.

[0024] 4. By reusing the same shared evidence generation unit before and after historical retrieval, the control signal before retrieval is used to regulate historical retrieval, and the evidence signal after retrieval is used to regulate dual-channel fusion. This organizes historical retrieval and state fusion into a unified adaptive control process, improving the consistency and interpretability of the control logic.

[0025] 5. By adaptively determining the fusion weights of short-term behavioral states and long-term cognitive states based on the final control state, the model can dynamically adjust its dependence on short-term and long-term information under different learning stages, different historical lengths, and different evidence conditions, thus avoiding the problem of insufficient adaptability caused by a fixed fusion ratio.

[0026] 6. By introducing a question difficulty bias in the final prediction layer and limiting it to the output stage, without participating in the construction of the preceding interactive representation and the historical retrieval process, the question difficulty factor can make necessary corrections to the prediction results, while reducing interference with the knowledge state modeling process. Attached Figure Description

[0027] Figure 1 This is a flowchart of an adaptive control knowledge tracking method based on dual-channel modeling.

[0028] Figure 2 The diagrams show the structure of a standard Transformer encoder unit and a CAM-Transformer encoder unit. (a) Standard Transformer encoder unit, (b) CAM-Transformer encoder unit.

[0029] Figure 3 The CAM metric weights vary with the CAM control signal. The changes.

[0030] Figure 4 Long-term uncertainty score after retrieval The relationship between gating fusion weights (bin statistics).

[0031] Figure 5 The results are for DECA, AKT, simpleKT, and UKT in historical cold start scenarios.

[0032] Figure 6 The results for DECA, AKT, simpleKT, and UKT on low-frequency knowledge points and sparse question subsets are given. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific examples.

[0034] An adaptive control knowledge tracking method based on dual-channel modeling, such as Figure 1 As shown, the specific steps include:

[0035] Step 1: Construct a knowledge tracking network.

[0036] The constructed knowledge tracking network comprises an embedded coding unit, a short-term behavior channel, a long-term cognitive channel, and a gated fusion prediction main path. The input of the embedded coding unit forms the input of the knowledge tracking network, the output of the embedded coding unit connects to the input of the short-term behavior channel and the long-term cognitive channel, the output of the short-term behavior channel and the long-term cognitive channel connects to the input of the gated fusion prediction main path, and the output of the gated fusion prediction main path forms the output of the knowledge tracking network.

[0037] The aforementioned short-term behavior channel includes a short-term attenuation unit and a GRU unit; the input of the short-term attenuation unit forms the input of the short-term behavior channel, the output of the short-term attenuation unit is connected to the input of the GRU unit, and the output of the GRU unit forms the output of the short-term behavior channel.

[0038] The aforementioned long-term cognitive channel includes a long-term decay unit, a long-term normal-inverse gamma distribution evidence unit, a long-term uncertainty unit, a prefix center unit, and a CAM-Transformer encoder unit. The input of the long-term decay unit forms the input of the long-term cognitive channel, the output of the long-term decay unit is connected to the input of the long-term normal-inverse gamma distribution evidence unit, the output of the long-term normal-inverse gamma distribution evidence unit is connected to the input of the long-term uncertainty unit, and the output of the long-term uncertainty unit is connected to the input of the prefix center unit. The outputs of the long-term decay unit and the prefix center unit are connected to the input of the CAM-Transformer encoder unit, and the output of the CAM-Transformer encoder unit forms the output of the long-term cognitive channel.

[0039] The aforementioned gated fusion prediction main path comprises a gated normal-inverse gamma distribution evidence unit, a gated uncertainty unit, a gated variance unit, a gated weight unit, a gated fusion unit, a response predictor unit, and a normalization unit. The short-term input of the gated fusion unit forms one input to the gated fusion prediction main path. The long-term input of the gated fusion unit and the input of the gated normal-inverse gamma distribution evidence unit form another input to the gated fusion prediction main path. The output of the gated normal-inverse gamma distribution evidence unit is connected to the input of the gated uncertainty unit. The output of the gated uncertainty unit is connected to the input of the gated variance unit. The outputs of the gated uncertainty unit and the gated variance unit are connected to the input of the gated weight unit. The output of the gated weight unit is connected to the weight input of the gated fusion unit. The output of the gated fusion unit is connected to the input of the response predictor unit. The output of the response predictor unit is connected to the input of the normalization unit. The output of the normalization unit forms the output of the gated fusion prediction main path.

[0040] Step 2: Obtain the learner's historical interaction sequence and feed it into the knowledge tracing network to train the knowledge tracing model.

[0041] (1) Obtaining historical interaction sequences.

[0042] Step 21: Obtain the number of interactions to be predicted for the learner. Previously, up to the [number]th The historical interaction sequence of this interaction .

[0043] ;

[0044] in, , Indicates the total number of interactions; Indicates the first The topic of this interaction; Indicates the first The questions in this interaction relate to specific knowledge points; Indicates the first The results of the questions answered in this interaction. This represents the result of an incorrect answer. This represents the correct answer. Indicates the first The interval between the positions of the same knowledge point, i.e., the first in the historical interaction sequence. Subtract the position from the first position The question is related to the same knowledge point, and its position in the historical interaction sequence is considered. If the knowledge point has not appeared before, then... .

[0045] (2) Embedded encoding.

[0046] Step 22: Embedding coding units for historical interaction sequences Embedding encoding is performed to obtain the historical interaction sequence embedding representation. .

[0047] ;

[0048] in, Indicates the first The question-related knowledge points are embedded in the representation of the interaction. Indicates the first The positional interval of the same knowledge point in the next interaction is embedded. Indicates the first Question embedding representation based on the response result of the second interaction:

[0049] ;

[0050] in, For the first The question embedding representation of the interaction. and These represent the globally learnable response vectors corresponding to correct and incorrect answers, respectively, and are updated through backpropagation during training.

[0051] After encoding is completed, this invention constructs two functionally heterogeneous state modeling channels, where the short-term behavior channel is used to represent the learner's short-term behavioral changes and the long-term cognitive channel is used to represent the learner's long-term cognitive evolution, and adjusts the retrieval and fusion process of the two channels through an adaptive control mechanism.

[0052] (3) Short-term behavior channel.

[0053] Step 23: Embedding representation of historical interaction sequences using short-term behavioral channels Short-term behavioral states are extracted to obtain the short-term behavioral states of historical interaction sequences. .

[0054] Given that short-term behavioral states focus on capturing recent response fluctuations, the short-term behavioral channel (System-1) proposed in this invention includes short-term decay units and GRU units. Its innovation lies in combining a fast position gap decay mechanism with a lightweight cyclic update mechanism, so that the path focuses more on capturing recent behavioral fluctuations, rather than treating all historical interactions as having equal persistence.

[0055] Step 231: Embedding short-term decay units into historical interaction sequences Short-term decay is achieved by employing an exponential decay function to obtain the short-term decay embedding representation of the historical interaction sequence. .

[0056] ;

[0057] in, Indicates the first Short-term decay embedding representation of the sub-interaction:

[0058] ;

[0059] in, This is the learnable decay coefficient for short-term behavior channels.

[0060] The short-term decay unit uses an exponential decay function. The short-term decay channel uses an exponential decay kernel defined on logarithmic positional intervals to decay earlier interactions. As the interval between sequences of the same knowledge point increases, the direct impact of old responses on the current state weakens rapidly, making the short-term behavior of historical interaction sequences more pronounced. It more effectively reflects consecutive errors, recent mastery changes, and local behavioral shifts. Exponential decay and recursive state updates together enhance the model's response to recent response changes and local behavioral patterns.

[0061] Step 232: The GRU unit uses a recurrent neural network to embed a short-term decaying representation of the historical interaction sequence. Sequence modeling is performed to obtain the short-term behavioral states of historical interaction sequences. .

[0062] ;

[0063] (4) Long-term cognitive pathway

[0064] Step 24: Long-term cognitive channels embedding representations of historical interaction sequences Long-term cognitive state extraction is performed to obtain the long-term cognitive states of historical interaction sequences. .

[0065] Given that long-term cognitive states require the aggregation of information from all effective interactions through historical retrieval, the long-term cognitive pathway (System-2) proposed in this invention includes a long-term decay unit, a long-term normal-inverse gamma distribution evidence unit, and a CAM-Transformer unit. Its innovation lies in the fact that long-term knowledge is not obtained solely through slow decay, but is formed by combining long-term memory with evidence-based historical relevant interaction retrieval.

[0066] Step 241: Long-term decay units embed historical interaction sequences into representations. A power-law decay function is used for long-term decay to obtain the long-term decay embedding representation of the historical interaction sequence. .

[0067] ;

[0068] in, Indicates the first Long-term decay embedding representation of the sub-interaction:

[0069] ;

[0070] in, This represents the learnable decay coefficient of the long-term cognitive channel.

[0071] The long-term decay unit uses a power-law decay function. The decay is gradual. Long-term cognitive channels use a power-law kernel with a slower decay to construct long-term inputs. Compared to the decay kernel of short-term decay channels, the power-law form allows earlier practice of the same knowledge points to maintain its influence over a longer sequence, so that longevity does not come solely from the decay function.

[0072] Step 242: Long-term normal-inverse gamma distribution evidence units are used to represent the long-term decaying embedding of historical interaction sequences. Evidence extraction was performed to obtain long-term evidence parameters. .

[0073] ;

[0074] in, Indicates the first Long-term evidence parameters for the second interaction:

[0075] ;

[0076] in, Indicates the first Long-term evidence parameters for the confidence level of the control mean estimate of the sub-interactions. Indicates the first Long-term evidence parameters for the confidence level of the control variance estimate of the sub-interaction. Indicates the first Long-term evidence parameters for the scale of control variance of secondary interactions.

[0077] The Normal-Inverse Gamma Distribution Evidence Unit (NIGHead) is a learnable mapping function. The normal-inverse gamma parameterization method used generates three parameters related to uncertainty calculation at each interaction location based on the input representation. , and This allows for parametric modeling of the cognitive state and its uncertainties inherent in the current interaction. The normal-inverse gamma distribution evidence unit is specifically implemented as a fully connected layer, or a lightweight network composed of multiple stacked fully connected layers. It follows the standard evidence regression formula and uses a SoftPlus-based transformation to enforce the condition. , and Long-term normal-inverse gamma distribution evidence units are used to control the CAM retrieval geometry.

[0078] Step 243: Long-term uncertainty unit based on long-term evidence parameters Calculate the long-term uncertainty score .

[0079] ;

[0080] in, Indicates the first Long-term uncertainty score of the sub-interaction:

[0081] ;

[0082] in, Represents the digamma function. This represents the Gamma function.

[0083] Long-term evidence parameters Summarized as long-term uncertainty fractions in the form of differential entropy This is used to characterize the overall dispersion of the current predicted distribution. Since differential entropy fractions can take negative values, their absolute sign does not constitute a fixed threshold for "certainty" or "uncertainty." Under the same parameterization, larger... This typically corresponds to a more dispersed forecast distribution and higher overall uncertainty.

[0084] Step 244: Prefix central unit for long-term uncertainty score Perform causal prefix masking to obtain the CAM control signal. .

[0085] ;

[0086] in, For the effective location mask, for the effective observation location, For the fill position, ; Small constants to ensure numerical stability.

[0087] To eliminate baseline differences among students and maintain the causality of subsequent predictions, it is necessary to utilize long-term uncertainty scores. Generate CAM control signals This is used to adjust the distance combination weights during historical retrievals, thereby achieving adaptive control of the retrieval process. The CAM control signal for the historical interaction sequence... Only by The effective prefix position is calculated, and future positions are not included in the centering. CAM control signal. It does not represent the absolute amount of evidence, but rather the original uncertainty score of the current position. The deviation relative to the student's observed effective prefix baseline. When When >0, the current position Higher than the average level of previous effective positions; when When <0, it is below the average level; when When = 0, it equals the average level. CAM maps this relative deviation from the learnable input weights to adjust the combination of the squared 2-Wasserstein distance and the Fisher–Rao-inspired dimension-wise approximation distance. Therefore, the role of centralization is to provide a relative scale of control within the student, varying with the historical interaction sequence, rather than artificially setting an absolute uncertainty threshold shared across students. Effective location masks and causal prefix statistics simultaneously exclude the influence of filler locations on future interactions.

[0088] Step 245, CAM-Transformer encoder unit in CAM control signal Long-term decay embedding representation of historical interaction sequences under control Modeling is performed to obtain the long-term cognitive state of historical interaction sequences. .

[0089] See Figure 2 The CAM-Transformer encoder unit is an improvement upon the standard Transformer encoder unit. It replaces the multi-head attention module (for query-key fixed dot product similarity calculation) with a CAM module (cognitive adaptive distance measurement), while retaining the rest. The CAM-Transformer encoder unit consists of multiple sequentially cascaded encoding layers. Each encoding layer comprises a CAM module, a feedforward network, and two residual connection and normalization modules. The input to the CAM module forms the input to the encoding layer. The output of the CAM module is connected to the input of the first residual connection and normalization module. The output of the first residual connection and normalization module is connected to the input of the feedforward network. The output of the feedforward network is connected to the second residual connection and normalization module. The output of the second residual connection and normalization module forms the output of the encoding layer. The input of the first encoding layer forms one input to the entire CAM-Transformer encoder unit, and the output of the last encoding layer forms the output of the entire CAM-Transformer encoder unit. The other input of the CAM module in each encoding layer forms the other input to the entire CAM-Transformer encoder unit.

[0090] First, the CAM module uses independent mean projection and scale projection to project the first... Long-term decay embedding representation of sub-interactions Mapped to mean vectors respectively and scale vector :

[0091] ;

[0092] ;

[0093] in, and The projection matrix is ​​learnable. and It is a learnable bias vector; It is represented as a very small number to ensure numerical stability and prevent the scale vector from reaching zero. The Softplus activation function is defined as follows: This is used to map the input to positive values, ensuring... All dimensions of the scale are positive.

[0094] Therefore, a diagonal Gaussian retrieval distribution is constructed. :

[0095] ;

[0096] in, Denotes the dimension of the Gaussian distribution. This represents the normal distribution (Gaussian distribution). Indicates a diagonal Gaussian distribution. This indicates element-wise operations.

[0097] Then, the CAM module is controlled by CAM control signals. Generate based on square -A hybrid weight of Wasserstein distance (second-order Wasserstein distance) and Fisher–Rao-inspired dimension-wise approximate distance.

[0098] For interactive locations Other interactive locations ,in Calculate the diagonal Gaussian retrieval distribution respectively. and The square of the difference between the distribution mean and the scale -Wasserstein distance and Fisher–Rao-inspired dimension-wise approximate distance based on differences in standardized distribution parameters:

[0099] square -Wasserstein distance for:

[0100] ;

[0101] Fisher–Rao-inspired dimension-wise approximation distance for:

[0102] ;

[0103] in, The dimension index representing the Gaussian distribution. Denotes the dimension of the Gaussian distribution. This represents the Euclidean norm.

[0104] The 2-squared Wasserstein distance characterizes the overall difference between the distribution center and scale; the Fisher–Rao-inspired dimensionally approximate distance characterizes the local differences in the normalized distribution parameters. The CAM module does not consistently use either distance, but rather selects based on the current CAM control signal. Generate mixed weights And calculate the original CAM distance. :

[0105] ;

[0106] ;

[0107] in, express function, This represents the clipping function. , and Represents the learnable weights and biases. and These represent the upper and lower boundary parameters used to limit degradation in the early stages of training. This ensures that both distances maintain a non-zero contribution during training, preventing complete degradation to a single distance in the early stages of training.

[0108] The original CAM distance is obtained by weighted combination of the squared 2-Wasserstein distance and the Fisher–Rao-inspired dimension-wise approximation distance. For historical interaction sequences, the same set of mixed weights is used to calculate the distance between the sequence and all historical keys, with the specific distance depending on the context. This design allows the model to learn obvious distance shifts, and also allows for a smooth mixture that maintains a certain distance bias across some data distributions.

[0109] Finally, the feedforward network and residual connection and normalization module perform the original operations to obtain the long-term cognitive state of the historical interaction sequence. :

[0110]

[0111]

[0112]

[0113] in, This represents the standardized distance score (which is the z-score of the original CAM distance within the set of valid historical keys of the current query, used to measure the relative distance between each historical position and the current position). and Representing positions respectively All corresponding valid historical locations of The mean and standard deviation are used to standardize the distance. -score; It is a very small constant used to ensure the stability of the denominator value; The scaling parameter for distance scores controls the sharpness of the retrieval distribution; For the mask item, for satisfying and The effective historical location value is Regarding future location ( ), fill position ( The values ​​for invalid positions are: This ensures that the retrieval process utilizes only historical information and satisfies causal constraints. The sequence is represented by long-term decay embedding. The Value matrix obtained by linear mapping is used to carry the retrieved and aggregated historical information.

[0114] The original CAM distance is normalized within the valid key set of each historical interaction sequence to obtain the distance score. Standardization does not change the distance order of historical positions within the same historical interaction sequence; the smaller the original CAM distance, the higher the retrieval score for the corresponding effective key. Softmax is calculated along the key dimension, converting the retrieval score into normalized weights. (Value matrix) The long-term cognitive representation is obtained by linear projection of the System-2 input and then weighted and aggregated according to these weights. .

[0115] In summary, the heterogeneity between the short-term behavioral channel and the long-term cognitive channel is reflected in both the time decay dynamics and the information aggregation mechanism. The former forms a recent behavioral summary through recursive updates, while the latter forms a long-term cognitive representation through evidence-conditional historical retrieval using CAM.

[0116] (5) Gated fusion prediction of main path

[0117] Step 25: Determine the short-term behavioral state of the historical interaction sequence. Long-term cognitive state and historical interaction sequence The input is fed into the main path of the fusion prediction to obtain the prediction probability. .

[0118] Step 251: Gated Normal-Inverse Gamma Distribution Evidence Units' Long-Term Cognitive State of Historical Interaction Sequences Evidence extraction was performed to obtain gated evidence parameters. .

[0119] ;

[0120] in, The gated evidence parameter represents the confidence level of the control mean estimate. The gated evidence parameter represents the confidence level of the control variance estimate. Gated evidence parameters that represent the scale controlling for variance.

[0121] The Normal-Inverse Gamma Distribution Evidence Unit (NIGHead) is a learnable mapping function. The normal-inverse gamma parameterization method used generates three parameters related to uncertainty calculation at each interaction location based on the input representation. , and This allows for parametric modeling of the cognitive state and its uncertainties inherent in the current interaction. The normal-inverse gamma distribution evidence unit is specifically implemented as a fully connected layer, or a lightweight network composed of multiple stacked fully connected layers. It follows the standard evidence regression formula and uses a SoftPlus-based transformation to enforce the condition. , and Gated normal-inverse gamma distribution evidence units are used to control the fusion of short-term and long-term states.

[0122] Step 252: The gating uncertainty unit is based on gating evidence parameters. Calculate the gating uncertainty score .

[0123] ;

[0124] in, Represents the digamma function. This represents the Gamma function.

[0125] Gated evidence parameters Summarized as long-term uncertainty fractions in the form of differential entropy This is used to characterize the overall dispersion of the current predicted distribution. Since differential entropy fractions can take negative values, their absolute sign does not constitute a fixed threshold for "certainty" or "uncertainty." Under the same parameterization, larger... This typically corresponds to a more dispersed forecast distribution and higher overall uncertainty.

[0126] Step 253: Gated variance unit based on gated uncertainty score Calculate the variance of random uncertainty and cognitive uncertainty variance .

[0127] Random uncertainty variance for:

[0128] ;

[0129] Cognitive uncertainty variance for:

[0130] ;

[0131] Random uncertainty variance This reflects observational noise in the NIG prediction model; cognitive uncertainty variance. The value decreases as the strength of evidence increases, thus reflecting the uncertainty caused by limited evidence.

[0132] Step 254: The gating weight unit is based on the gating uncertainty score. Random uncertainty variance and cognitive uncertainty variance Generate fusion weights .

[0133] variance of random uncertainty and cognitive uncertainty variance With gate uncertainty score Together they constitute the final evidence state vector Since all three components are uncertainty descriptors derived from NIG, they are not additional observational evidence. This vector The gate fusion weights are obtained by using them as input for dual-channel fusion at the gate. .

[0134] ;

[0135] The gated weighted unit uses an MLP to transform the three-dimensional evidence vector The mapping is performed as a scalar logit, which is then converted into fusion weights for the gated behavior channels via a sigmoid function. This mapping does not presuppose a fixed monotonic relationship between uncertainty and a particular system: the same level of uncertainty may correspond to recent behavioral changes or short-term noise in different data distributions. Therefore, Gate learns soft arbitration with evidence conditions, rather than hard switching that is manually specified.

[0136] Step 255: The fusion unit is based on the fusion weight. Short-term behavior state of historical interaction sequences Long-term cognitive state and historical interaction sequence The fusion process is performed to obtain a fused state. .

[0137] ;

[0138] Step 256: Merge state The data is fed into the response predictor to obtain a comprehensive prediction score. .

[0139] ;

[0140] in, This represents a response predictor, specifically implemented as a fully connected network consisting of one or more fully connected layers (i.e., a form of multilayer perceptron MLP), used to map fused states to scalar predicted values. This serves as a learnable difficulty bias for the next question. This bias is only introduced in the final prediction layer to make necessary difficulty adjustments to the final prediction results, and it does not participate in interactive representation construction or CAM retrieval.

[0141] Step 257: Merge state The data is fed into a normalization unit for normalization to obtain the predicted probability. .

[0142] ;

[0143] Thus, DECA has completed the entire calculation process from the observed interaction sequence to the probability of the next response: the short-term behavior channel (System-1) forms the short-term behavior state, the long-term cognition channel (System-2) retrieves and aggregates relevant historical information through CAM to form the long-term cognition state, the two are modeled in parallel, and the gated fusion channel (Gate) adaptively adjusts the relative effects of the two based on the final evidence state.

[0144] DECA uses masked binary cross-entropy loss (BCE) to optimize the prediction of the next response. Let represent the set containing learner-interaction position pairs with valid next-round response labels.

[0145] ;

[0146] in, Indicate learner At time step The observed response value at that time, This represents the probability of a correct prediction. Pre-retrieval control signal. Status of evidence after retrieval Both are constructed from the stage-specific outputs of the shared NIGHead and participate in CAM history retrieval and Gate dual-channel fusion, respectively. The fused state is converted into the probability of the next response by a binary predictor. The binary cross-entropy loss (BCE) objective can propagate the prediction supervision signal along the CAM retrieval path and the Gate fusion path, and optimize the shared NIGHead and related representation parameters end-to-end.

[0147] Step 3: Obtain the learner's question to be predicted, i.e., the question number... The topic of the second interaction The question to be predicted is then fed into the knowledge tracking model to obtain the learner's predicted answer to the question.

[0148] The performance of the present invention will be illustrated by the following experiments:

[0149] The experiments used four public knowledge tracking datasets: Algebra2005, EdNet-KT1 (EdNet for short), Junyi2015, and BD2006. As shown in Table 1, the four datasets differ significantly in size, sequence length, and response distribution: Junyi2015 provides large-scale interactive scenarios, Algebra2005 and BD2006 mainly feature longer learning sequences, while EdNet presents shorter sequences and lower response accuracy. These differences provide complementary evaluation scenarios for examining the prediction performance of DECA under different data distributions.

[0150] Table 1. Statistics of the four datasets after preprocessing

[0151]

[0152] Accuracy refers to the accuracy of responses in the data, rather than the model evaluation result.

[0153] To examine the effectiveness of DECA from different modeling dimensions, this invention selects representative knowledge tracing methods covering sequence modeling, memory enhancement, attention modeling, multi-timescale state modeling, and uncertainty-aware representation. DKT, as a classic recurrent knowledge tracing model, is used to examine the overall benefit of DECA relative to basic sequence state modeling. DKVMN, representing external memory enhancement, is used to compare the role of explicit knowledge storage mechanisms. SAKT and AKT represent self-attention knowledge tracing methods, with AKT further incorporating knowledge difficulty and temporal relevance, serving as an important reference for comparing long-range interactive modeling with DECA. simpleKT, as a concise and competitive strong baseline, is used to examine whether the benefit of DECA exceeds the simple attention structure itself. HiTSKT, representing hierarchical and multi-timescale state modeling, is most relevant to DECA's heterogeneous dual-channel design. UKT, representing uncertainty-aware distributed knowledge tracing, is used to compare the differences between DECA and existing probability distribution representations and Wasserstein matching mechanisms.

[0154] All datasets employ shifted next-response alignment, where the t-th position encodes only the interactions observed up to that position and predicts the next response; future positions and padding positions are not included in prefix statistics or training loss. All models use the same data preprocessing, student-level splitting, shifted next-response input and prediction target, maximum sequence length handling, and test events to avoid incomparable results or label leakage due to differences in preprocessing or evaluation protocols. Data is split by student: 80% of students are divided into 5 fixed folds, with fold 0 used for validation and folds 1–4 used for training; the remaining 20% ​​of students constitute an independent test set.

[0155] (1) RQ1 test of the overall predictive benefit of DECA under different data sizes, sequence lengths and response distributions;

[0156] Table 2 shows the overall predictive performance of DECA and seven baselines on four datasets.

[0157]

[0158] The results are derived from a specified random seed of 3407. Bold and underlined text indicate the best and second-best results, respectively. The percentage improvement in AUC and ACC, as well as the percentage decrease in RMSE, are calculated relative to the baseline that performs best in the corresponding dataset – metric combination.

[0159] Table 2 shows the overall performance comparison of DECA with seven baselines across four datasets. DECA achieved the best results in all 12 dataset-metric combinations. On the EdNet dataset, DECA improved AUC and ACC by 6.39% and 8.58% respectively compared to the best baseline, while reducing RMSE by 18.83%; on BD2006, AUC improved by 1.03%, ACC improved by 1.40%, and RMSE reduced by 0.75%. These results demonstrate that DECA achieves consistent predictive superiority across different data sizes, sequence lengths, and response distributions.

[0160] (2) RQ2 analyzes the independent contributions and complementary relationships of heterogeneous dual channels, CAM and Gate through component ablation and mechanism replacement;

[0161] Table 3 shows the ablation results of DECA components.

[0162]

[0163] w / o S1 means removing System-1 and keeping only System-2; w / o S2 means removing the entire System-2 branch and its CAM history retrieval, keeping only System-1; w / o Gate means replacing the evidence conditional Gate with fixed equal-weight fusion; w / o CAM means replacing CAM with dot product attention; w / o Rasch means removing the question difficulty bias in the output layer; DECA means full configuration. Bold text indicates the optimal value for each dataset-metric combination.

[0164] Table 4 shows the CAM distance and the mechanism replacement results for the System-1 encoder.

[0165]

[0166] W2 only and FR only represent retaining only the squared 2-Wasserstein distance and the Fisher-Rao heuristic dimension-by-dimensional approximation distance, respectively; in the complete DECA, CAM uses a combination of dual distances regulated by control signals, and System-1 uses a GRU encoder.

[0167] To verify the independent contributions of each component of DECA, ablation experiments were conducted. Table 3 shows that after removing the short-term behavior channel (System-1), long-term cognition channel (System-2), Gate, CAM, or Rasch term, the AUC and ACC of all four datasets decreased, indicating that the heterogeneous dual system, evidence-conditional history retrieval, adaptive state fusion, and question difficulty bias each play an irreplaceable role. Table 4 shows that the mechanism replacement experiment indicates that replacing the GRU of System-1 with LSTM or Transformer did not bring any improvement. The performance of using only the squared 2-Wasserstein distance or only the Fisher–Rao-inspired dimension-wise approximation distance was lower than that of the complete CAM, indicating that the two distances are complementary, and the combination of dynamic adjustment by the control signal is better than a fixed single distance.

[0168] (3) RQ3 uses internal weights and individual trajectory visualization to examine whether the shared NIGHead participates in CAM retrieval regulation and dual-channel fusion respectively;

[0169] Figure 3 The CAM metric weights vary with the CAM control signal. The changes. The red and blue lines represent... and The horizontal dashed lines represent equal weight states. Each subplot uses a dataset-related horizontal axis range, and the visual slope of the curves is not used for direct comparisons across datasets.

[0170] Figure 4Long-term uncertainty score after retrieval Relationship with gating fusion weights (binning statistics). Blue scatter plots represent test samples, red lines represent the mean of short-term behavior channel fusion weights after binning, and shading represents interquartile range; each subplot uses the dataset-related horizontal axis range.

[0171] Mechanism visualization analysis shows that ( Figure 3 , Figure 4 The retrieval distance combination weight of CAM changes dynamically with the control signal before retrieval, and the fusion weight of Gate is adaptively adjusted with the evidence status after retrieval, which verifies the design of the shared NIG Head participating in historical retrieval adjustment and dual-system fusion in this invention.

[0172] (4) RQ4 further examines the discriminative performance and probabilistic quality of DECA in both conventional and low-evidence scenarios. Therefore, the experiments cover overall effectiveness, component necessity, mechanism behavior, and low-evidence reliability. Evaluation metrics include discriminative performance, prediction error, and probabilistic quality to ensure a comprehensive assessment of the predictive benefits and reliability of DECA.

[0173] Figure 5 The results for DECA, AKT, simpleKT, and UKT in historical cold start scenarios are shown. The top row represents AUC, and the bottom row represents ECE; the AUC line and ECE points represent the mean of three runs, and the ECE error bars represent one standard deviation.

[0174] Figure 6 This section presents the results of DECA, AKT, simpleKT, and UKT on low-frequency knowledge points and sparse question subsets. The top row represents AUC, and the bottom row represents ECE; each point represents the mean of three runs, and the ECE error bars represent one standard deviation.

[0175] To further verify the performance of DECA under low-evidence conditions, a historical cold-start experiment and a physical long-tail experiment were designed. The historical cold-start experiment selected the earliest K valid prediction positions for each test student (K=5, 10, 20); the physical long-tail experiment constructed a low-frequency subset based on the frequency of occurrence of questions and knowledge concepts in the training set. Figure 5 and Figure 6 As shown, DECA achieved the highest average AUC in all 12 settings for historical cold starts and the lowest ECE in 9 of those settings; it also achieved the highest average AUC in all 8 settings for entity long tails and the lowest ECE in 7 of those settings. These results demonstrate that DECA maintains stable discriminative performance and good probability calibration even in low-evidence scenarios such as historical sparsity and entity long tails.

[0176] Unified experiments on four open knowledge tracking datasets show that, compared to representative baselines, DECA achieves a stable predictive advantage in regular tests as well as in low-evidence scenarios such as historical cold starts and long-tailed entities, and improves probability calibration in most low-evidence settings.

[0177] This invention proposes a knowledge tracing framework with an adaptive distributed retrieval mechanism, namely the Dual-Channel Evidence-Conditioned Architecture (DECA). DECA models short-term behavioral changes and long-term knowledge evolution through parallel, functionally heterogeneous short-term behavioral channels (System-1) and long-term cognitive channels (System-2), respectively. The short-term behavioral channel uses a fast decay mechanism and a recurrent neural network to generate short-term behavioral states, while the long-term cognitive channel uses a slow decay mechanism to construct long-term input representations. A shared evidence generation unit generates adaptive control signals for controlling historical retrieval, dynamically adjusting the matching geometry between historical and current interactions based on these control signals to perform adaptive distributed retrieval of historical interactions, forming long-term cognitive states. The shared evidence generation unit then generates the final control state for fusion control, and a gating module determines the fusion weights of the short-term behavioral and long-term cognitive states, outputting the probability of the next response. This invention can dynamically adjust the historical retrieval process and the dual-channel fusion process based on current evidence information, improving the stability, robustness, and probabilistic reliability of knowledge tracing prediction.

[0178] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. An adaptive control knowledge tracing method based on dual-channel modeling, characterized by: The steps include the following: Step 1: Construct a knowledge tracking network. This knowledge tracking network consists of an embedded coding unit, a short-term behavior channel, a long-term cognitive channel, and a gated fusion prediction main path. The input of the embedded coding unit forms the input of the knowledge tracking network. The output of the embedded coding unit is connected to the input of the short-term behavior channel and the long-term cognitive channel. The output of the short-term behavior channel and the long-term cognitive channel is connected to the input of the gated fusion prediction main path. The output of the gated fusion prediction main path forms the output of the knowledge tracking network. Step 2: Obtain the learner's historical interaction sequence and feed the historical interaction sequence into the knowledge tracing network to train it, thereby obtaining the knowledge tracing model; Step 3: Obtain the learner's question to be predicted and feed it into the knowledge tracking model to obtain the learner's predicted answer to the question.

2. The adaptive control knowledge tracking method based on dual-channel modeling according to claim 1, characterized in that, The short-term behavior channel includes short-term decay units and GRU units; The input of the short-term attenuation unit forms the input of the short-term behavior channel, the output of the short-term attenuation unit is connected to the input of the GRU unit, and the output of the GRU unit forms the output of the short-term behavior channel.

3. The adaptive control knowledge tracking method based on dual-channel modeling according to claim 2, characterized in that, The short-term decay unit uses an exponential decay function for rapid decay, while the long-term decay unit uses a power-law decay function for slow decay.

4. The adaptive control knowledge tracking method based on dual-channel modeling according to claim 1, characterized in that, The long-term cognitive pathway includes long-term decay units, long-term normal-inverse gamma distribution evidence units, long-term uncertainty units, prefix center units, and CAM-Transformer encoder units; The input of the long-term decay unit forms the input of the long-term cognitive channel; the output of the long-term decay unit is connected to the input of the long-term normal-inverse gamma distribution evidence unit; the output of the long-term normal-inverse gamma distribution evidence unit is connected to the input of the long-term uncertainty unit; and the output of the long-term uncertainty unit is connected to the input of the prefix center unit. The outputs of the long-term attenuation unit and the prefix center unit are connected to the input of the CAM-Transformer encoder unit, and the output of the CAM-Transformer encoder unit forms the output of the long-term cognitive channel.

5. The adaptive control knowledge tracking method based on dual-channel modeling according to claim 4, characterized in that, The short-term decay unit uses an exponential decay function for rapid decay, while the long-term decay unit uses a power-law decay function for slow decay.

6. The adaptive control knowledge tracking method based on dual-channel modeling according to claim 4, characterized in that, CA The M-Transformer encoder unit consists of multiple encoding layers connected in series. Each coding layer consists of a CAM module, a feedforward network, and two residual connection and normalization modules. The input of the CAM module forms the input of the coding layer, the output of the CAM module is connected to the input of the first residual connection and normalization module, the output of the first residual connection and normalization module is connected to the input of the feedforward network, the output of the feedforward network is connected to the second residual connection and normalization module, and the output of the second residual connection and normalization module forms the output of the coding layer. The input of the first encoding layer forms one input to the entire CAM-Transformer encoder unit, the output of the last encoding layer forms the output of the entire CAM-Transformer encoder unit, and another input of the CAM module of each encoding layer forms another input to the entire CAM-Transformer encoder unit.

7. The adaptive control knowledge tracking method based on dual-channel modeling according to claim 6, characterized in that, The CAM module first maps the long-term decay embedding representation output by the long-term decay unit into a mean vector and a scale vector, respectively, and constructs a diagonal Gaussian retrieval distribution based on the mean vector and the scale vector; then it calculates the square between the diagonal Gaussian retrieval distribution of the last interaction and the previous interaction. -Wasserstein distance and Fisher–Rao-inspired dimensionally approximate distance; then squared The original CAM distance is generated by weighting the Wasserstein distance and the Fisher–Rao-inspired dimension-wise approximate distance.

8. The adaptive control knowledge tracking method based on dual-channel modeling according to claim 1, characterized in that, The gated fusion prediction main path consists of a gated normal-inverse gamma distribution evidence unit, a gated uncertainty unit, a gated variance unit, a gated weight unit, a gated fusion unit, a response predictor unit, and a normalization unit; The short-term input of the gated fusion unit forms one input to the gated fusion prediction main path. The long-term input of the gated fusion unit and the input of the gated normal-inverse gamma distribution evidence unit form another input to the gated fusion prediction main path. The output of the gated normal-inverse gamma distribution evidence unit is connected to the input of the gated uncertainty unit. The output of the gated uncertainty unit is connected to the input of the gated variance unit. The outputs of the gated uncertainty unit and the gated variance unit are connected to the input of the gated weight unit. The output of the gated weight unit is connected to the weight input of the gated fusion unit. The output of the gated fusion unit is connected to the input of the response predictor unit. The output of the response predictor unit is connected to the input of the normalization unit. The output of the normalization unit forms the output of the gated fusion prediction main path.