CN-122001666-A - User behavior completion and prediction method based on end cloud cooperation

CN122001666ACN 122001666 ACN122001666 ACN 122001666ACN-122001666-A

Abstract

The invention discloses a user behavior completion and prediction method based on end cloud cooperation, which comprises the steps of collecting user behavior data in a mobile terminal and encrypting to generate an encrypted user behavior log, analyzing the encrypted user behavior log by a cloud end and generating a differential behavior packet through a behavior completion model, constructing an enhanced user behavior sequence based on the differential behavior packet, performing embedding mapping processing on the enhanced user behavior sequence to generate a behavior input embedded tensor, generating a predicted event list through an improved FEDformer network, constructing a weighted loss function based on the predicted event list, updating an improved FEDformer network, and updating the behavior completion model. The invention improves the accuracy of user behavior completion and prediction and the response capability of sudden behavior through the improved FEDformer network.

Inventors

HUANG WEI
CHEN JUNNAN
HU SHUANGHONG
LIU YU
HE LU

Assignees

深圳市启创数智科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260313

Claims (9)

1. A user behavior completion and prediction method based on end cloud cooperation is characterized by comprising the following steps: collecting user behavior data in a mobile terminal, generating a user behavior log sequence through time alignment processing, and encrypting the user behavior log sequence to generate an encrypted user behavior log; Secondly, analyzing the encrypted user behavior log by the cloud to generate a user-level behavior sequence, and carrying out complement prediction on the user-level behavior sequence through a behavior complement model to generate a differential behavior packet; Step three, the mobile terminal executes space-time matching judgment on the difference behavior packet and fuses the difference behavior packet with the user behavior log sequence to generate an enhanced user behavior sequence; step four, carrying out embedding mapping processing on the enhanced user behavior sequence to generate a behavior input embedding tensor; inputting the behavior input embedded tensor to an improved FEDformer network deployed on the mobile terminal, and executing behavior time sequence feature modeling and probability prediction decoding to output a predicted event list, wherein the improved FEDformer network comprises a trend-period-jump decomposition module, a space grid modulation module, a time sequence fusion module and a prediction decoding module; Step six, after the actual implementation is that the event occurs, calculating a prediction error and constructing a weighted loss function based on the matching relation between a predicted event list and a real behavior event sequence, and updating the improved FEDformer network; and step seven, the mobile terminal encrypts the prediction errors and then uploads the encrypted prediction errors to the cloud terminal, and the cloud terminal updates the behavior completion model based on the prediction errors of the plurality of mobile terminals to generate a new version of differential behavior package.
2. The method for user behavior completion and prediction based on end cloud cooperation of claim 1, wherein the first step specifically comprises: in the mobile terminal, collecting user behavior data generated by a user in a preset time window based on an operating system bottom layer interface, and generating a user identifier associated with the user behavior data based on the identifier of the mobile terminal; The user behavior data comprises a spatial position point identifier, a behavior event type, a behavior confidence coefficient and a behavior time stamp; Sequencing the user behavior data according to the sequence of the behavior time stamps to obtain an original behavior log sequence; Setting a time step sequence, and resampling and aligning the original behavior log sequence according to the time step sequence to obtain a user behavior log sequence; and carrying out symmetric encryption processing on the user behavior log sequence through an AES encryption algorithm to generate an encrypted user behavior log, and storing the encrypted user behavior log in a local encryption log bin of the mobile terminal.
3. The method for user behavior completion and prediction based on end cloud cooperation of claim 1, wherein the step two specifically comprises: uploading encrypted user behavior logs of a plurality of mobile terminals to a cloud, decrypting and analyzing the encrypted user behavior logs, extracting a spatial position point identifier, a behavior event type, a behavior confidence coefficient and a behavior time stamp according to user identifiers, and generating a user-level behavior sequence; Setting a prediction time and the number of candidate space position points, and constructing a candidate space position set according to the number of candidate space position points based on a user-level behavior sequence at the prediction time; constructing a behavior complement model by adopting a pre-trained transducer network, carrying out joint modeling on a user-level behavior sequence and a candidate space position set, outputting a prediction score of each candidate space position at a prediction moment, and carrying out Softmax normalization processing on the prediction score to obtain a prediction weight; Distributing an effective time window and a space grid identifier for each candidate space position, and forming a complement behavior quadruple by the candidate space position, the prediction weight, the effective time window and the space grid identifier; And packaging all the complement behavior quadruples into differential behavior packages and issuing the differential behavior packages to corresponding mobile terminals.
4. The method for user behavior completion and prediction based on end cloud cooperation of claim 1, wherein the third step specifically comprises: in a set updating period, the mobile terminal receives a differential behavior packet corresponding to the user identifier from the cloud; The space grid identification of the current time of the mobile terminal is obtained, and space-time matching judgment is carried out on each complement action quadruple in the differential action packet, wherein the space grid identification is screened to be the same as the space grid identification in the complement action quadruple, and the current time does not exceed the complement action quadruple of the effective time window, so as to form a hit complement set; Sorting the four complement behavior tuples in the hit complement set in descending order according to the predicted weight, and writing the four complement behavior tuples into a behavior cache pool of the mobile terminal to form a cache complement behavior sequence; Carrying out structure standardization processing on the cache completion behavior sequence, namely taking a candidate space position as a predicted space position point identifier, assigning a preset event label as a predicted behavior event type, taking a predicted weight as a predicted behavior confidence coefficient, and taking the minimum value of the current time of the mobile terminal and the upper limit of a corresponding effective time window as a predicted behavior time stamp; And reading the user behavior log sequence from the mobile terminal, and carrying out alignment fusion on the cache completion behavior sequence and the user behavior log sequence according to the time stamp sequence to obtain an enhanced user behavior sequence, wherein the enhanced user behavior sequence comprises a plurality of behavior events, and each behavior event consists of a spatial position point identifier, a behavior event type, a behavior confidence coefficient and a behavior time stamp.
5. The method for user behavior completion and prediction based on end cloud cooperation of claim 1, wherein the fourth step specifically comprises: performing an embedding mapping process on each behavioral event in the enhanced user behavioral sequence: mapping the space position point mark into an initial position vector through a trainable position embedding lookup table, and mapping the initial position vector into a set dimension through linear transformation to generate a position embedding vector; Performing independent thermal coding on the behavior event type to generate a behavior type independent thermal vector, and performing matrix multiplication on the behavior type independent thermal vector and a trainable behavior type embedding matrix to obtain a behavior type embedding vector; Generating a Time embedding vector by using a Time2Vec coding mode through the action Time stamp; mapping the behavior confidence coefficient into a confidence coefficient embedded vector through a multi-layer perceptron structure; Performing feature stitching fusion on the position embedded vector, the behavior type embedded vector, the time embedded vector and the confidence embedded vector to obtain a joint behavior embedded vector; and stacking the joint behavior embedding vectors of all the behavior events according to the time stamp sequence to construct a behavior input embedding tensor.
6. The method for user behavior completion and prediction based on end cloud cooperation of claim 1, wherein the fifth step specifically comprises: taking each behavior event as a time step; The trend-period-jump decomposition module generates a stable trend component, a period disturbance component and a jump event component respectively through trend, period and jump residual decomposition; In the space grid modulation module, mapping a space grid mark of the current time of the mobile terminal into a space grid embedded vector through a trainable space grid embedded lookup table; broadcasting the space grid embedded vector as a matrix along the time step dimension to obtain a space grid modulation matrix; multiplying the space grid modulation matrix with the periodic disturbance component element by element to obtain a frequency domain modulation characteristic tensor; in a time sequence fusion module, performing time-step inverse discrete Fourier transform on the frequency domain modulation characteristic tensor to obtain a time sequence modulation disturbance tensor; adding the stable trend component and the time sequence modulation disturbance tensor element by element on each time step and each channel dimension to generate a fusion time sequence characteristic tensor; The prediction decoding module performs category space mapping and probability normalization on the fusion time sequence feature tensor, generates a behavior-time-probability tensor and outputs a prediction event list, and specifically comprises the following steps: Constructing a candidate spatial location category space based on the number of candidate spatial location points; Projecting the fusion time sequence characteristic tensor to a candidate space position category space through linear mapping to generate a prediction score tensor; carrying out Softmax normalization on the prediction fractional tensor in the dimension of the candidate space position point and the time step to generate a behavior-time-probability tensor, wherein the behavior-time-probability tensor consists of a prediction probability value of each time step and each candidate space position point; Setting a confidence threshold, and carrying out threshold screening and event construction on the behavior-time-probability tensor: If the predicted probability value of the current time step and the current candidate space position point is larger than or equal to the set confidence threshold value, the current time step, the candidate space position point mark and the predicted probability value form a predicted event, and the predicted events are ordered according to the descending order of the predicted probability values to form a predicted event list.
7. The method for user behavior completion and prediction based on end cloud coordination according to claim 6, wherein the trend-period-jump decomposition module generates a stable trend component, a period disturbance component and a jump event component respectively through trend, period and jump residual decomposition, and specifically comprises: Carrying out local sliding convolution on the behavior input embedded tensor through one-dimensional convolution with the convolution kernel size of 7, and carrying out nonlinear mapping through a Tanh activation function to obtain a stable trend component; calculating residual errors of the behavior input embedded tensor and the steady trend component to obtain a periodic input component; the periodic input components are respectively subjected to short-time Fourier transformation with the time window length of 16, 32 and 64 time steps to obtain a first frequency domain characteristic component, a second frequency domain characteristic component and a third frequency domain characteristic component; Setting three groups of trainable weight coefficients, and carrying out weighted fusion on the first frequency domain characteristic component, the second frequency domain characteristic component and the third frequency domain characteristic component to obtain a periodic disturbance component; calculating residual errors of the periodic disturbance input component and the periodic disturbance component to obtain a jump residual error component; calculating the residual module length of each time step in the jump residual component through the L2 norm; Setting a jump judgment threshold value, and constructing a jump mask vector, wherein if the residual error module of the current time step is larger than the jump judgment threshold value, the element value of the time step corresponding to the jump mask vector is equal to 1, otherwise, the element value of the time step corresponding to the jump mask vector is equal to 0; the jump mask vector is multiplied element by element with the jump residual component to obtain a jump event component.
8. The method for user behavior completion and prediction based on end cloud cooperation according to claim 1, wherein the sixth step specifically comprises: After the user actually generates the behavior event, the actual behavior position identification and the actual behavior confidence coefficient of each time step are extracted from the actual behavior event sequence recorded by the mobile terminal, and are compared with the predicted event list: if the time step of the predicted event is equal to the time step of the real behavior event and the candidate spatial position point mark is equal to the real behavior position mark, the predicted event is used as an effective predicted event; Obtaining a prediction probability value from the effective prediction event, and calculating an absolute difference value between the actual behavior confidence coefficient and the prediction probability value to obtain a prediction error; Calculating a jump event module length corresponding to the time step through an L2 norm based on the jump event component; Multiplying the jump event module length by the prediction error to obtain a weighted prediction error, and accumulating the weighted prediction errors of all effective prediction events to obtain a weighted loss function; Based on the weighted loss function, a back propagation operation is performed on the modified FEDformer network.
9. The method for user behavior completion and prediction based on end cloud cooperation of claim 1, wherein the step seven specifically comprises: Acquiring a real behavior position identifier of a real behavior event corresponding to each prediction error; Calculating the cross entropy between the real behavior position identification and the prediction weight, and multiplying the cross entropy with the prediction error to obtain a weighted cross entropy error; And accumulating weighted cross entropy errors corresponding to all the prediction errors to obtain a cross entropy loss function, updating the behavior completion model based on the cross entropy loss function, and generating a new version of differential behavior packet.

Description

User behavior completion and prediction method based on end cloud cooperation Technical Field The invention relates to the technical field of mobile intelligent terminal behavior modeling and privacy protection computing, in particular to a user behavior complement and prediction method based on end cloud cooperation. Background Along with the continuous improvement of the intelligent level of the mobile terminal operating system, the predictive reasoning and personalized service technology oriented to user behavior modeling is widely focused. The existing user behavior modeling method mainly relies on cloud centralized modeling, and by uploading an original behavior log of a terminal to a server, offline training is performed by using sequence models such as LSTM or a Transformer, and a prediction result is issued to the terminal. However, the following problems are common in practical applications: The user behavior data has a strong privacy attribute, the original position track and behavior event usually comprise a space movement mode, living habit and use preference of the user, and the user behavior data is directly uploaded to the cloud to easily expose individual sensitive information, so that the risk of data leakage and abuse exists. In a new device or a cold start user scenario, the terminal lacks enough historical behavior data, which results in extremely low or even no operation of behavior prediction accuracy, affecting the usability of the system-level scenario awareness function. The traditional behavior complement strategy mostly adopts single-point signal matching or static rules, ignores the time sequence characteristics of behavior events and the multi-scale behavior mode of the user, and has insufficient complement precision and higher error complement rate. The cloud model is usually trained according to a T+1 period, is difficult to respond to the current intention change of a user in real time, has the problems of lag of a behavior prediction result, poor timeliness and the like, and limits the application range of the cloud model in intelligent recommendation and resource scheduling of an operating system level. Therefore, how to provide a user behavior complement and prediction method based on end cloud collaboration is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a user behavior completion and prediction method based on end cloud cooperation, which fully integrates user behavior log encryption modeling, differential behavior completion distribution, improved FEDformer network prediction and feedback closed loop optimization technology, and details the modeling and prediction of a multi-time scale user behavior sequence under the condition of ensuring privacy compliance, and has the advantages of no data output domain, availability for cold start, high prediction precision and strong response time. According to the embodiment of the invention, the user behavior complement and prediction method based on the end cloud cooperation comprises the following steps: collecting user behavior data in a mobile terminal, generating a user behavior log sequence through time alignment processing, and encrypting the user behavior log sequence to generate an encrypted user behavior log; Secondly, analyzing the encrypted user behavior log by the cloud to generate a user-level behavior sequence, and carrying out complement prediction on the user-level behavior sequence through a behavior complement model to generate a differential behavior packet; Step three, the mobile terminal executes space-time matching judgment on the difference behavior packet and fuses the difference behavior packet with the user behavior log sequence to generate an enhanced user behavior sequence; step four, carrying out embedding mapping processing on the enhanced user behavior sequence to generate a behavior input embedding tensor; inputting the behavior input embedded tensor to an improved FEDformer network deployed on the mobile terminal, and executing behavior time sequence feature modeling and probability prediction decoding to output a predicted event list, wherein the improved FEDformer network comprises a trend-period-jump decomposition module, a space grid modulation module, a time sequence fusion module and a prediction decoding module; Step six, after the actual implementation is that the event occurs, calculating a prediction error and constructing a weighted loss function based on the matching relation between a predicted event list and a real behavior event sequence, and updating the improved FEDformer network; and step seven, the mobile terminal encrypts the prediction errors and then uploads the encrypted prediction errors to the cloud terminal, and the cloud terminal updates the behavior completion model based on the prediction errors of the plurality of mobile terminals to generate a new version of differential behavior package. Option