CN-122020070-A - Memory state prediction method based on contrast learning and stabilization long-term and short-term memory network and application thereof

CN122020070ACN 122020070 ACN122020070 ACN 122020070ACN-122020070-A

Abstract

The application relates to the technical field of deep learning, in particular to a memory state prediction method based on a contrast learning and stabilization long-term and short-term memory network and application thereof. In the data preprocessing stage, three disturbance types including label overturning, time disturbance and probability deviation are respectively carried out on the sequence of the review sequence data, so that the sequence data has robustness when facing noise labels, generalization capability under different review time deviation and stable convergence when observation errors and recording deviations occur, characteristic extraction is carried out on positive and negative samples after the review sequence and data enhancement processing by utilizing a stabilized long-short-term memory network, the stability of a numerical value during long-sequence modeling is ensured, and loss calculation and model optimization are carried out on the positive and negative sample characteristics and the integral characteristics of the sequence in a combined mode, so that the model can pull the distance between the positive sample pair and the distance between the negative sample pair in an embedded space. The method aims at solving the problem of how to realize the memory state prediction with noise robustness and stable time sequence modeling capability.

Inventors

WANG JUN
TIAN QIAO
LI ZIJIE
GAN JIANHOU
WEN BIN
LIU SANNVYA

Assignees

云南师范大学

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. The memory state prediction method based on the contrast learning and stabilization long-term and short-term memory network is characterized by comprising the following steps of: s10, acquiring a preprocessed review sequence data set, wherein the review sequence data set is composed of a plurality of review sequences comprising recall results, review time intervals and historical recall probabilities; S20, taking part of the review sequences in the review sequence data set, and respectively executing recall result disturbance enhancement, review time interval disturbance enhancement and historical recall probability disturbance enhancement to obtain a first positive sample, a second positive sample and a third positive sample, and taking part of other review sequences which do not belong to the positive samples in the review sequence data set as negative samples; S30, extracting sequence characterization in the review sequence data set based on a stabilized long-term memory network, extracting the first positive sample, the second positive sample and the third positive sample to obtain positive sample characterization, and extracting negative sample characterization corresponding to the negative sample; s40, constructing a loss function based on the sequence characterization, the positive sample characterization and the negative sample characterization to guide a preset large model to train; S50, obtaining the to-be-predicted review sequence data, and performing memory state prediction on the to-be-predicted review sequence data based on the trained large model.
2. The method for predicting the memory state of a long-term and short-term memory network based on contrast learning and stabilization as claimed in claim 1, wherein said S20 comprises: S21, extracting the review sequence data set according to a first preset proportion to obtain a first target review sequence, and performing label turning enhancement on recall results in the first target review sequence to obtain a first positive sample; s22, extracting the review sequence data set according to a second preset proportion to obtain a second target review sequence, and shortening/prolonging the corresponding review time interval according to the value of the recall result in the second target review sequence to obtain a second positive sample; S23, extracting the review sequence data set according to a third preset proportion to obtain a third target review sequence, and injecting Gaussian noise and random offset into the history recall probability of the third target review sequence to obtain a third positive sample; S24, taking the parts of the review sequences except the first positive sample, the second positive sample and the third positive sample in the review sequence data set as negative samples.
3. The method for predicting the memory state of a long-term and short-term memory network based on contrast learning and stabilization as set forth in claim 2, wherein S21 specifically includes: s211, performing label turning enhancement on the first target review sequence by adopting the following formula: ; In the formula, In order to uniformly distribute the random variable(s), For the preset probability of the flip-flop, In order to enhance the recall result after the enhancement, Recall the results for the initial; s212, taking the first target review sequence after label overturning enhancement as a first positive sample.
4. The method for predicting the memory state of a long-term and short-term memory network based on contrast learning and stabilization as set forth in claim 2, wherein S22 specifically includes: S221, summing a first preset increment and a corresponding review time interval to shorten the review time interval when the recall result value of the second target review sequence meets a first preset value, wherein the first preset increment is a negative value; S222, summing a second preset increment and a corresponding review time interval to prolong the review time interval when the recall result value of the second target review sequence meets a second preset value, wherein the second preset increment is a positive value; s223, a second target review sequence for shortening/lengthening the review time interval is taken as a second positive sample.
5. The method for predicting the memory state of a long-term and short-term memory network based on contrast learning and stabilization as set forth in claim 2, wherein S23 specifically includes: s231, gaussian noise and random offset are injected into the historical recall probability of the third target review sequence: ; In the formula, To recall the probability of disturbance history after injection of gaussian noise, Recall probabilities for an initial history; to introduce random perturbation terms to the historical recall probabilities, Indicating compliance with mean 0 and variance 0 Is a gaussian random noise of (a); s232, disturbance history recall probability Is limited to the interval Inner: ; S233, taking the third target review sequence subjected to Gaussian noise injection and random offset as a third positive sample.
6. The method for predicting the memory state of a long-term and short-term memory network based on contrast learning and stabilization as set forth in claim 1, wherein S40 specifically includes: S41, predicting the sequence representation by adopting a multi-layer perceptron regression network Corresponding memory half-life value : ; In the formula, 、 Respectively representing the learnable weight parameters of each layer in the multi-layer perceptron regression network; is an activation function; 、 respectively representing the corresponding learnable bias parameters of each layer; S42, according to the prediction half-life And the first in the current training batch True memory half-life tag of individual review sequences Determining a mean square error loss function : ; In the formula, The number of samples participating in regression loss calculation in the current training batch is calculated; s43, determining sequence characterization Characterization with positive samples, respectively Characterization of negative samples Cosine similarity between 、 According to cosine similarity 、 Determining a contrast learning loss function : ; In the formula, In order to be of a batch size, Is a temperature parameter; s44, according to the mean square error loss function Contrast learning loss function Is used to determine the loss function : ; In the formula, , Respectively controlling the weights of half-life regression loss and contrast loss as super parameters; s45, based on loss function And carrying out back propagation training on the trainable parameters in the preset large model.
7. The method for predicting the memory state of a long-term and short-term memory network based on contrast learning and stabilization as claimed in claim 1, wherein in S10, the step of preprocessing the review sequence data set comprises: S11, setting the ith review action of the same user on the same knowledge item Sequencing all review behaviors in time sequence to obtain an initial review sequence S: ; In the formula, Representing recall results; Representing the time interval of adjacent review; the probability of a history recall is represented, Representing the sequence length; s12, the time interval in the initial review sequence S Performing scaling normalization to obtain a transformed time interval Recall probabilities of history Performing centering and scale transformation to obtain transformation history recall probability : ; ; In the formula, 、、 The change coefficients of the tested values are taken; S13, constructing a recall result Time interval of conversion And transform history recall probabilities And a composed pre-processed review sequence data set : 。
8. The method for predicting the memory state of a long-term and short-term memory network based on contrast learning and stabilization as claimed in claim 1, wherein in S50, the memory state comprises a half-life prediction value And recall probability values The step of executing the memory state prediction on the to-be-predicted review sequence data based on the trained preset model comprises the following steps: S51, based on trained large model Calculating review sequence data to be predicted Half-life prediction value of (2) : ; S52, according to half-life prediction value And review sequence data to be predicted Time interval in (a) Calculating recall probability values : 。
9. A computer system comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program when executed by the processor implementing the steps of the method for predicting a memory state based on a contrast learning and stabilizing long-short memory network according to any one of claims 1 to 8.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the steps of the memory state prediction method based on a contrast learning and stabilizing long-short-term memory network according to any one of claims 1 to 8.

Description

Memory state prediction method based on contrast learning and stabilization long-term and short-term memory network and application thereof Technical Field The application relates to the technical field of deep learning, in particular to a memory state prediction method based on a contrast learning and stabilization long-term and short-term memory network and application thereof. Background The interval repetition technology improves the long-term keeping effect by arranging review at proper time according to the forgetting curve rule. For example, in a language learning scenario, a learner needs to review the same word multiple times to develop long-term memory. The interval repetition does not repeatedly present the word in a short time, but rather disperses multiple reviews to different points in time, gradually lengthening the time interval between adjacent reviews, thereby enhancing long-term memory retention while reducing ineffective repetition. Deep learning models generally require predicting future memory states based on learner's historical review behavior to generate personalized review plans. The input learner's historical recall behavior generally includes recall results, recall time intervals, and historical recall probabilities, and the output future memory state, including memory half-life and recall probabilities, is used to characterize the rate of memory decay and determine recall timing. In the related technical scheme, the conventional method relies on heuristic rules or traditional regression models, noise and time fluctuation in learning behaviors are difficult to process, prediction accuracy is insufficient, gradient instability is easy to occur in a depth sequence model in long sequence modeling, and an effective anti-noise mechanism is lacked. In view of this, the present application proposes a memory state prediction method with both noise robustness and stable timing modeling capability. Disclosure of Invention The application mainly aims to provide a memory state prediction method based on a contrast learning and stabilizing long-term and short-term memory network, which aims to solve the problem of how to realize memory state prediction with noise robustness and stable time sequence modeling capability. In order to achieve the above object, the present application provides a memory state prediction method based on a contrast learning and stabilization long-short-term memory network, the method comprising: s10, acquiring a preprocessed review sequence data set, wherein the review sequence data set is composed of a plurality of review sequences comprising recall results, review time intervals and historical recall probabilities; S20, taking part of the review sequences in the review sequence data set, and respectively executing recall result disturbance enhancement, review time interval disturbance enhancement and historical recall probability disturbance enhancement to obtain a first positive sample, a second positive sample and a third positive sample, and taking part of other review sequences which do not belong to the positive samples in the review sequence data set as negative samples; S30, extracting sequence characterization in the review sequence data set based on a stabilized long-term memory network, extracting the first positive sample, the second positive sample and the third positive sample to obtain positive sample characterization, and extracting negative sample characterization corresponding to the negative sample; s40, constructing a loss function based on the sequence characterization, the positive sample characterization and the negative sample characterization to guide a preset large model to train; S50, obtaining the to-be-predicted review sequence data, and performing memory state prediction on the to-be-predicted review sequence data based on the trained large model. Optionally, the S20 includes: S21, extracting the review sequence data set according to a first preset proportion to obtain a first target review sequence, and performing label turning enhancement on recall results in the first target review sequence to obtain a first positive sample; s22, extracting the review sequence data set according to a second preset proportion to obtain a second target review sequence, and shortening/prolonging the corresponding review time interval according to the value of the recall result in the second target review sequence to obtain a second positive sample; S23, extracting the review sequence data set according to a third preset proportion to obtain a third target review sequence, and injecting Gaussian noise and random offset into the history recall probability of the third target review sequence to obtain a third positive sample; S24, taking the parts of the review sequences except the first positive sample, the second positive sample and the third positive sample in the review sequence data set as negative samples. Optionally, S21 specifically includes: s211,