CN-121997291-A - Student performance prediction method and system based on Mamba state space model

CN121997291ACN 121997291 ACN121997291 ACN 121997291ACN-121997291-A

Abstract

The invention relates to the technical field of data processing, and particularly provides a student performance prediction method and system based on Mamba state space model, wherein the method comprises the steps of preprocessing and vectorizing the original characteristic data of a multi-source student, mapping different types of characteristics into characteristic vectors with uniform dimensions based on a characteristic type perception strategy, and obtaining a characteristic sequence; inputting Mamba the feature sequence into a state space model, processing the feature sequence through Mamba the state space model to obtain a hidden state vector containing comprehensive information of all the features, constructing a regression prediction model, and inputting the hidden state vector into the regression prediction model to obtain a student performance prediction result. According to the invention, complex interactions between features are effectively captured through sequence modeling, so that better performance can be obtained in prediction accuracy, parameter scale and calculation complexity are greatly reduced, and the model is less prone to fitting and more robust in training on a small data set.

Inventors

CAI YUELIANG
ZHAO YUQI
YIN JIAMIN
WANG JINHENG

Assignees

广州理工学院

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (9)

1. The student performance prediction method based on Mamba state space model is characterized in that the student performance prediction method based on Mamba state space model comprises the following steps: The method comprises the steps of obtaining original characteristic data of a multi-source student, preprocessing and vectorizing the original characteristic data of the multi-source student, mapping different types of characteristics into characteristic vectors with unified dimensions based on a characteristic type sensing strategy, and obtaining a characteristic sequence by combining a position coding weight capable of being learned; Constructing Mamba a state space model, inputting the feature sequence into the Mamba state space model, and processing the feature sequence through the Mamba state space model to obtain a hidden state vector containing all feature comprehensive information; And constructing a regression prediction model, and inputting the hidden state vector into the regression prediction model to obtain a student performance prediction result.
2. A student performance prediction method based on Mamba state space model as claimed in claim 1, wherein the specific method for processing the feature sequence comprises the following steps: Gradually carrying out hidden state updating and information aggregation processing on the feature sequence through a Mamba state space coding module, and generating a hidden state vector containing all feature comprehensive information at the tail end of the feature sequence; Wherein the Mamba state space model includes a Mamba state space encoding module.
3. The student performance prediction method based on Mamba state space model as claimed in claim 2, wherein the obtaining of the feature sequence comprises calculating the feature representation of the fused position information based on the following formula : Where k is the index of the position of the feature in the sequence, Is the embedded vector of the kth feature, The vector is encoded for the position of the kth position, The method is a learnable weight parameter and is used for dynamically adjusting the contribution degree of the position codes to the characteristic representation; Representing the features of all the fused position information The feature sequences are arranged in a predetermined order.
4. A student performance prediction method based on Mamba state space model as defined in claim 3, wherein the specific method for obtaining the student performance prediction result comprises the following steps: And (3) taking the fully-connected neural network layer as an output prediction head, inputting the hidden state vector into the fully-connected neural network layer, and obtaining a student performance prediction result through an activation function or linear transformation.
5. The student performance prediction method based on Mamba state space model as claimed in claim 1, wherein the specific method for obtaining the feature sequence comprises the following steps: Selecting a corresponding coding mode according to the type of original characteristic data of the multi-source students, and carrying out characteristic identical coding pretreatment; And acquiring an embedding weight and an embedding bias term, and mapping the preprocessed original characteristic data of the multi-source student according to the embedding weight and the embedding bias term to acquire a characteristic sequence. The method comprises the steps of encoding numerical type features by adopting a normalization mode, encoding category type features by adopting a One-hot encoding mode, constructing a feature type sensing function, dynamically adjusting a mapping strategy according to feature types, and mapping preprocessed feature data into feature embedding vectors with uniform dimensions through leachable embedding weights and embedding bias items.
6. A student performance prediction system based on Mamba state space model for implementing a student performance prediction method based on Mamba state space model according to any one of claims 1 to 5, wherein the student performance prediction system based on Mamba state space model comprises: The feature embedding and serializing module is used for acquiring original feature data of the multi-source students, preprocessing and vectorizing the original feature data of the multi-source students, mapping different types of features into feature vectors with uniform dimensions based on a feature type sensing strategy, and acquiring feature sequences by combining the learnable position coding weights; Mamba a state space model, which is used for processing the feature sequence to obtain a hidden state vector containing all feature comprehensive information; And the regression prediction model is used for obtaining student performance prediction results according to the hidden state vector.
7. The student performance prediction system based on Mamba state space model as defined in claim 6, wherein said Mamba state space model comprises: and Mamba the state space coding module is used for gradually carrying out hidden state updating and information aggregation processing on the feature sequence and generating a hidden state vector containing all feature comprehensive information at the tail end of the feature sequence.
8. The student performance prediction system based on Mamba state space model as claimed in claim 7, wherein said feature embedding and serialization module comprises: the embedding module is used for carrying out embedding processing on the original characteristic data of the multi-source students and converting different types of characteristics into vector representations with uniform dimensions; and the arrangement module is used for arranging all the features according to a preset sequence and constructing a feature sequence.
9. The student performance prediction system based on a Mamba state space model of claim 8, wherein the regression prediction model comprises: And the fully-connected neural network layer is used for carrying out normalization processing on the hidden state vector to obtain a student performance prediction result.

Description

Student performance prediction method and system based on Mamba state space model Technical Field The invention relates to the technical field of data processing, in particular to a student performance prediction method and system based on Mamba state space models. Background The existing student performance prediction method mainly comprises two major categories, namely a traditional machine learning model and a deep learning model. Traditional methods such as linear regression, logistic regression, etc., and tree models such as decision trees, random forests, etc., have been widely adopted. These methods are easy to implement on small and medium-scale data sets with some interpretability, but they have limitations in modeling complex relationships. In recent years, there have been studies on attempts to apply a deep model of tabular data (including models based on a self-attentive mechanism such as a transducer) to student performance prediction tasks, and it is desired to improve prediction accuracy by utilizing a strong representation learning ability of deep learning. However, the data available in the educational field tends to be limited in size, limited in feature dimension, and small in number of training samples. In the face of the 'small data' scene, the existing methods have the defects that the linear model is limited in prediction performance due to the fact that the nonlinear interaction relation between features is difficult to capture due to simplicity, the tree model is easy to be subjected to overfitting when the data volume is small, the problem that calculation cost is improved is also caused by the increase of the complexity of the model, and the large-scale deep models such as a Transformer are large in parameter scale although excellent in large data tasks, are easy to be subjected to overfitting on small data sets, are prone to occurrence of parameter redundancy and unstable in training process, and meanwhile, the performance improvement matched with the cost is difficult to be obtained in educational prediction tasks due to high calculation complexity. Therefore, how to consider model performance and calculation efficiency in a small data scene becomes a technical problem to be solved in student performance prediction. Disclosure of Invention Based on the problem, in order to solve the problem that the model performance and the calculation efficiency are difficult to be achieved in a small data scene in the prior art, the invention provides a student performance prediction method and a student performance prediction system based on Mamba state space model, and the specific technical scheme is as follows: a student performance prediction method based on Mamba state space model comprises the following steps: The method comprises the steps of obtaining original characteristic data of a multi-source student, preprocessing and vectorizing the original characteristic data of the multi-source student, mapping different types of characteristics into characteristic vectors with unified dimensions based on a characteristic type sensing strategy, and obtaining a characteristic sequence by combining a position coding weight capable of being learned; Constructing Mamba a state space model, inputting the feature sequence into the Mamba state space model, and processing the feature sequence through the Mamba state space model to obtain a hidden state vector containing all feature comprehensive information; And constructing a regression prediction model, and inputting the hidden state vector into the regression prediction model to obtain a student performance prediction result. According to the student performance prediction method based on Mamba state space model, complex interaction between features is effectively captured through sequence modeling, so that better performance can be obtained in prediction accuracy, parameter scale and calculation complexity are greatly reduced, the model is not easy to fit and train more stably on a small dataset, and the problem that model performance and calculation efficiency are difficult to complete in a small data scene in the prior art is solved. Preferably, the specific method for processing the feature sequence comprises the following steps: Gradually carrying out hidden state updating and information aggregation processing on the feature sequence through a Mamba state space coding module, and generating a hidden state vector containing all feature comprehensive information at the tail end of the feature sequence; Wherein the Mamba state space model includes a Mamba state space encoding module. Preferably, the acquiring the feature sequence specifically includes calculating a feature representation of the fused position information based on the following formula: Where k is the index of the position of the feature in the sequence,Is the embedded vector of the kth feature,The vector is encoded for the position of the kth position,The method is a learnabl