CN-121980514-A - Improved sequence data state tracking and predicting method and system

CN121980514ACN 121980514 ACN121980514 ACN 121980514ACN-121980514-A

Abstract

The invention discloses an improved sequence data state tracking and predicting method and system, which comprise the following steps of obtaining item feature representation and result enhancement feature representation corresponding to a current item based on sequence data, respectively generating a first state vector and a second state vector of the current identifier by using an identified exercise time interval based on the item feature representation and the result enhancement feature representation, extracting a learnable attenuation rate parameter from the item feature representation, respectively attenuating the first state vector and the second state vector by combining the exercise time interval, dynamically fusing the attenuated first state vector and the attenuated second state vector through a gating mechanism to obtain a current comprehensive state, encoding by adopting an attention mechanism based on a historical comprehensive state sequence and the current comprehensive state to obtain a high-level cognitive state vector, and predicting the correct response probability of a user based on the high-level cognitive state vector and the item feature representation of the current item.

Inventors

ZHOU GUANGYOU
TAN LIHUA
XIE ZHIWEN

Assignees

华中师范大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (8)

1. An improved sequence data state tracking and prediction method, comprising the steps of: Acquiring historical interaction sequence data of a user, wherein the data at least comprises an interaction item identifier, an interaction result and an interaction time stamp; generating a project characteristic representation based on the current interaction project identifier; generating a result enhanced feature representation based on the item feature representation and the current interaction result; generating a first state vector and a second state vector corresponding to the identifier respectively based on the item characteristic representation and the result enhanced characteristic representation by utilizing the time interval between the current interaction and the last interaction with the same identifier; calculating a first attenuation factor and a second attenuation factor based on attenuation parameters extracted from the item feature representation and the time interval, and respectively attenuating the first state vector and the second state vector; fusing the attenuated first state vector and the attenuated second state vector through a gating unit to obtain a current fused state vector; Based on the historical fusion state vector sequence and the current fusion state vector, coding is carried out through the attention neural network, and a context characteristic representation is obtained; Based on the contextual feature representation and the item feature representation, a probability of the user producing a correct result for the current interaction is predicted.
2. The improved sequential data state tracking and prediction method of claim 1, wherein generating a project feature representation based on a current interaction project identification comprises: representing the difficulty of the project by adopting a learnable project difficulty vector; representing the dynamic characteristics of the mark by adopting a learnable mark dynamic change vector; Acquiring the single-hot coding vector of the mark; multiplying the project difficulty vector and the identification dynamic change vector element by element to obtain a first intermediate vector; And adding the first intermediate vector and the identified one-hot encoding vector to obtain the project characteristic representation.
3. The improved sequential data state tracking and prediction method of claim 2, wherein generating a result enhanced feature representation based on the item feature representation and current interaction results comprises: Differentiating the influence of the interaction result by adopting a learnable identification-result interaction vector; Acquiring a joint embedding vector of the identification and the interaction result; multiplying the learnable difficulty vector in the project characteristic representation with the identifier-result interaction vector element by element to obtain a second intermediate vector; and adding the second intermediate vector and the joint embedded vector to obtain the result enhanced feature representation.
4. The improved sequential data state tracking and prediction method of claim 1, wherein generating first and second state vectors corresponding to the identity, respectively, based on the item feature representation and the resulting enhanced feature representation, and utilizing the time interval, comprises: assigning an independent time constant to each of the identifications; calculating a smoothing coefficient based on the time interval and the time constant; Updating the first state vector by adopting an exponential moving average method based on the smoothing coefficient and the first state vector at the last moment; And performing difference between the current input and the updated first state vector to obtain a second state vector.
5. The improved sequential data state tracking and prediction method of claim 4 wherein the process of calculating a first attenuation factor and a second attenuation factor from the extracted attenuation parameters from the item feature representation and the time interval, and attenuating the first state vector and the second state vector, respectively, comprises: Extracting a first decay rate parameter and a second decay rate parameter from the item feature representation; calculating a first attenuation factor by an exponential function based on the first attenuation rate parameter and the time interval; Calculating a second attenuation factor by an exponential function based on the second attenuation rate parameter and the time interval; Multiplying the first state vector element by using the first attenuation factor; And multiplying the second state vector element by adopting the second attenuation factor.
6. The improved sequential data state tracking and prediction method of claim 1, wherein fusing the attenuated first state vector and the attenuated second state vector by a gating unit to obtain a current fused state vector, comprising: splicing the attenuated first state vector and the attenuated second state vector to obtain a spliced vector; performing linear transformation on the spliced vector and activating the spliced vector through an S-shaped function to generate a gating vector; And based on the gating vector, carrying out weighted summation on the attenuated first state vector and the attenuated second state vector to obtain the current fusion state vector.
7. The improved sequential data state tracking and prediction method of claim 1, wherein the encoding by the attention neural network based on the historical fused state vector sequence and the current fused state vector results in a contextual feature representation, comprising: encoding the history fusion state vector sequence through a self-attention mechanism to obtain a global history state; and focusing the current fusion state vector on the global history state based on the content of the current fusion state vector through a cross attention mechanism to obtain a fused representation serving as the context characteristic representation.
8. An improved sequential data state tracking and prediction system for implementing the method of claim 1, said system comprising: The input representation module is used for acquiring historical interaction sequence data of a user, generating item feature representations based on the current interaction item identifiers, and generating result enhancement feature representations based on the item feature representations and the current interaction results; the state separation and updating module is used for generating a first state vector and a second state vector corresponding to the identifier respectively based on the item characteristic representation and the result enhanced characteristic representation by utilizing the time interval between the current interaction and the last interaction with the same identifier; The time interval perception processing module is used for calculating attenuation factors based on attenuation parameters extracted from the project characteristic representation and the time interval, and carrying out attenuation and fusion on the first state vector and the second state vector to obtain a current fusion state vector; The context coding module is used for coding through the attention neural network based on the historical fusion state vector sequence and the current fusion state vector to obtain a context characteristic representation; and the performance prediction module is used for predicting the probability of generating a correct result for the current interaction by the user based on the context characteristic representation and the item characteristic representation.

Description

Improved sequence data state tracking and predicting method and system Technical Field The invention belongs to the technical field of education data mining and intelligent education, and particularly relates to an improved sequence data state tracking and predicting method and system. Background Currently, the identification tracking model is mainly based on deep learning and other technologies, and models a historical interaction sequence of a user so as to predict the result of subsequent interaction behaviors. These approaches have made some progress in processing time series with a single evolution pattern, but have still suffered from significant drawbacks in processing sequence data with the following complex structure: The existing method generally encodes all interaction information in a sequence into a single continuous hidden state, and lacks a mechanism for independently modeling information corresponding to different conceptual labels in the sequence. This results in a model that has difficulty in distinguishing the characteristic evolution process of different concepts, and information between different concepts is easy to interfere with each other, thereby affecting the prediction accuracy of the related behavior of a specific concept. Most models are simple to handle for time intervals in the sequence, using only time stamps as assist features or employing a fixed decay function. The method is difficult to effectively capture the evolution speed difference of different concepts at different time intervals, cannot finely model the time dynamics specific to the concepts, and limits the prediction performance and generalization capability of the model in a real scene. While existing models are able to learn complex patterns from historical sequences, their internal representation and decision processes tend to be black-boxed. This makes it difficult for the model to interpret whether a particular prediction result is primarily affected by historical accumulated information, or changes in characteristics due to recent interactions or long time intervals, thereby limiting the reliability of the model and its applicability in practical systems. The essential reason for these deficiencies is that the existing model fails to fully consider two key structural characteristics of multi-concept interleaving and irregular time interval distribution which are commonly existing in sequence data, and fails to design a corresponding processing mechanism at the model architecture level. Accordingly, the present invention is directed to an improved sequential data state tracking and prediction method and system. Disclosure of Invention In order to solve the above technical problems, the present invention provides an improved method and system for tracking and predicting the state of sequence data, so as to solve the problems of the prior art. To achieve the above object, the present invention provides an improved sequence data state tracking and predicting method, comprising: Acquiring historical interaction sequence data of a user, wherein the data at least comprises an interaction item identifier, an interaction result and an interaction time stamp; generating a project characteristic representation based on the current interaction project identifier; generating a result enhanced feature representation based on the item feature representation and the current interaction result; generating a first state vector and a second state vector corresponding to the identifier respectively based on the item characteristic representation and the result enhanced characteristic representation by utilizing the time interval between the current interaction and the last interaction with the same identifier; calculating a first attenuation factor and a second attenuation factor based on attenuation parameters extracted from the item feature representation and the time interval, and respectively attenuating the first state vector and the second state vector; fusing the attenuated first state vector and the attenuated second state vector through a gating unit to obtain a current fused state vector; Based on the historical fusion state vector sequence and the current fusion state vector, coding is carried out through the attention neural network, and a context characteristic representation is obtained; Based on the contextual feature representation and the item feature representation, a probability of the user producing a correct result for the current interaction is predicted. Optionally, based on the current interaction item identification, the process of generating the item feature representation includes: representing the difficulty of the project by adopting a learnable project difficulty vector; representing the dynamic characteristics of the mark by adopting a learnable mark dynamic change vector; Acquiring the single-hot coding vector of the mark; multiplying the project difficulty vector and the identification dynamic c