A method and system for personalizing multi-user software
By generating LoRA parameters through a hypernetwork model, the problems of rapid adaptation and structured mapping in the personalization of multi-user software are solved, achieving second-level personalized adaptation and interpretable parameter adjustment, and supporting knowledge sharing among users.
Patent Information
- Application Number
- CN202611211783.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, multi-user software personalization methods struggle to achieve rapid personalized adaptation of software model parameters without requiring iterative optimization for each user. Furthermore, they lack a structured mapping relationship from user preferences to the functional modules of the software model, resulting in time-consuming personalization adaptation, a lack of interpretability in parameter adjustments, and the inability to share and reuse knowledge among users.
By employing a hypernetwork model, LoRA parameters are directly generated by acquiring user feature data and utilizing response map prediction and parameter generation networks. This enables the correspondence between low-dimensional response maps and the basic software model, supporting rapid personalized adaptation. Furthermore, weight updates are performed using a low-rank decomposition matrix, ensuring that the parameter generation process has clear structured guidance and interpretability.
It has improved the speed of personalized adaptation from minutes/hours to seconds, meeting the real-time interaction requirements of software. The parameter generation process is interpretable and the knowledge between users can be reused, solving the cold start problem.
Smart Images

Figure CN123114423A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and software engineering technology, specifically relating to a method and system for personalized customization of multi-user software. Background Technology
[0002] With the increasing application of deep learning technology in large-scale software systems, providing differentiated services to different users based on a unified basic model has become an important requirement in the field of software engineering. When different users use the same software, their operating habits, functional preferences, and interaction styles vary significantly. How to efficiently generate software models that adapt to the personalized needs of each user is a key technical problem currently facing multi-user software systems.
[0003] In existing technologies, software personalization mainly adopts the following methods: One method is a rule-based configuration method, where the software system pre-sets several optional configuration items, and users manually set preferences to achieve customization. In this method, users need to understand the meaning of each configuration item and set them one by one, which is cumbersome and the customization granularity is limited by the granularity of the pre-set options, making it difficult to adapt to users' fine-grained and dynamically changing needs. Another method is a user profile-based recommendation method, where the system constructs a user profile vector by collecting users' historical usage data, and then recommends functions or ranks content based on this vector. This method has the following shortcomings: there is no direct mapping channel between the user profile and the software model parameters, and the preference information in the profile vector is difficult to effectively transmit to the model weight adjustment process, resulting in the customization effect being limited by the abstract level of the profile expression. In recent years, Efficient parameter fine-tuning techniques such as LoRA are applied to the personalized adaptation of large models. In existing technologies, a set of LoRA parameters is usually trained independently for each user's personalized needs. This approach has the following problems: First, each user's LoRA parameters need to undergo an independent iterative optimization process, which is time-consuming and difficult to meet the rapid response requirements of real-time interactive software scenarios. Second, the LoRA parameters of each user are independent of each other, and the intermediate information generated during training cannot be shared and reused among different users, resulting in low resource utilization. Third, in the existing method of directly generating LoRA parameters through hypernetworks, the mapping process from user features to model parameters lacks interpretable structured constraints, and the correspondence between the generated parameters and specific functional modules in the software model is unclear, leading to insufficient accuracy and controllability of parameter adjustment.
[0004] Therefore, a method and system for multi-user software personalization customization has emerged. Summary of the Invention
[0005] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a multi-user software personalization customization method and system to solve the technical problems in the prior art, which are: difficulty in achieving rapid personalized adaptation of software model parameters without the need for independent iterative optimization for each user; difficulty in establishing a structured mapping relationship from user preferences to various functional modules of the software model during parameter generation; and a lack of a software model personalization customization scheme that can both reuse personalized knowledge among users and ensure that the parameter adjustment process has a clear functional module correspondence. This leads to technical problems such as long personalization adaptation time, lack of interpretability of parameter adjustment, and inability to share and reuse knowledge among users.
[0006] To address the above problems, a first aspect of the present invention provides a method for personalized customization of multi-user software, comprising the following steps: S1: Obtain user characteristic data of the target user, including user behavior sequence data and user preference description data; S2: Input the user feature data into the hypernetwork model, which includes a response graph prediction network and a parameter generation network; S3: The response graph prediction network extracts user pattern features from the user feature data and outputs a low-dimensional response graph. The response graph is a low-rank matrix, and each dimension corresponds one-to-one with the weight matrix of each functional module in the basic software model. The parameter generation network takes the response graph as a priori input and sequentially performs dimensionality-upgrading mapping on the response graph through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix corresponding to the layer to be adapted in the basic software model. The low-rank decomposition matrix contains a first matrix and a second matrix. The product of the first matrix and the second matrix is calculated as the weight update matrix. The weight update matrix is added to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user. S4: In response to the target user's software usage request, output software interaction content through the personalized software model.
[0007] Preferably, in step S1, obtaining the user characteristic data of the target user includes the following steps: Collect the target user's original operation logs within a preset time window, including the call timestamps, call frequency, call duration, and call return status codes of each functional module; The original operation logs are arranged into an operation sequence in chronological order. An operation vector is extracted for each operation event in the operation sequence. The operation vector includes a functional module identifier, an operation type code, and a duration normalization value. Each operation vector in the operation sequence is input into the behavior encoder in time step order, and a fixed-dimensional behavior sequence latent vector is output as user behavior sequence data. Obtain the natural language customized text input by the target user, input the natural language customized text into the text encoder, and output the preference description latent vector as the user preference description data; The latent vector of the behavior sequence is used as the query vector, and the latent vector of the preference description is used as the key vector and value vector. They are input into the cross-attention fusion layer. The cross-attention fusion layer calculates the attention weight matrix between the features of each time step in the latent vector of the behavior sequence and the features of each semantic unit in the latent vector of the preference description. The latent vector of the preference description is weighted and aggregated according to the attention weight matrix, and the fused user feature data is output.
[0008] Preferably, in step S2, the training method of the hypernetwork model includes the following steps: A training dataset is obtained, which includes the sample user features corresponding to several sample users and LoRA parameters of each group of samples. The LoRA parameters of each group of samples are obtained by iteratively optimizing the basic software model for the corresponding sample users. The objective function of the iterative optimization is to minimize the prediction loss of the basic software model on the labeled data of the corresponding sample users. For each group of samples, calculate the change matrix of LoRA parameters relative to the original parameters of the basic software model. Perform singular value decomposition on each change matrix and extract the first few elements of each change matrix. The left and right singular vectors corresponding to the 1st maximal singular value are used to calculate a low-rank approximation matrix as the sample response map of the corresponding sample user. The preset low-rank dimension value; The sample user features of each sample user are used as input, and the corresponding sample response map is used as supervision label to train the response map prediction network. The parameters of the trained response map prediction network are fixed, and the sample user features of each sample user are input into the trained response map prediction network to output the corresponding predicted response map. The predicted response maps of each sample user are used as input, and the corresponding LoRA parameters of each group of samples are used as supervision labels to train the parameter generation network.
[0009] Preferably, in step S3, the response map prediction network includes a temporal attention encoding unit and a response mapping unit, comprising the following steps: The temporal attention encoding unit performs self-attention encoding on the latent vector of the behavior sequence in the user feature data, and uses the features of each time step in the latent vector of the behavior sequence as query vector, key vector and value vector respectively, calculates the attention weight matrix between each time step, and outputs temporal enhancement features. The response mapping unit takes the temporal enhancement features as input and maps them into a low-dimensional response map through multiple fully connected layers. The number of neurons in the output layer of the response mapping unit is equal to the total number of functional modules in the basic software model. Among them, the output layer of the response mapping unit is the first The output value of the nth neuron is used as the nth neuron in the low-dimensional response map. The numerical value of the i-th dimension, the i-th dimension of the low-dimensional response map The first dimension index and basic software model The weight matrices of each functional module have the same position index in the model parameter storage structure. The response mapping unit outputs the first functional module according to the position index. The output value of the nth neuron is written into the first low-dimensional response map. Each dimension.
[0010] Preferably, in step S3, the parameter generation network uses the response map as prior input and sequentially performs dimensionality-upgrading mapping on the response map through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix, including the following steps: The mapping layer inputs the low-dimensional response map to the first linear transformation unit. The output dimension of the first linear transformation unit is equal to the original dimension of the weight matrix of the layer to be adapted, and outputs the intermediate latent vector. The reconstruction layer will adjust the intermediate latent vectors according to the output dimension of the layer to be adapted. and input dimensions Dimensional splitting is performed, and the first part after splitting is mapped to the first matrix through the second linear transformation unit, and the second part after splitting is mapped to the second matrix through the third linear transformation unit; During training, the reconstruction layer applies column orthogonality constraints to the first matrix. Apply row orthogonality constraints to the second matrix. ,in, for An identity matrix of order 1. For the first matrix, For the second matrix, It is the transpose vector.
[0011] Preferably, in step S3, adding the weight update matrix to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user includes the following steps: Obtain the original weight matrix of each layer to be adapted in the basic software model. ,in, For the first in the basic software model One layer to be adapted; The parameter generation network outputs the first matrix for each layer to be adapted. Second matrix Calculate the weight update matrix for each layer to be adapted. ; Update the weight matrix of each layer to be adapted. With the corresponding original weight matrix By adding each element, we obtain the personalized weight matrix for each layer to be adapted. ; The personalized weight matrix of each layer to be adapted By backfilling according to the original position of each layer to be adapted in the basic software model, the complete weight matrix of the personalized software model corresponding to the target user is obtained.
[0012] Preferably, the method further includes step S5: dynamically updating the weight matrix of the personalized software model corresponding to the target user. During the process of target users using the personalized software model, continuously collect new operation logs and new customized instructions; The newly added operation logs are arranged into a new operation sequence according to time order. The new operation vector of each new operation event is extracted. The new operation vectors are input into the behavior encoder in time step order and the new behavior latent vector is output. The new customized instruction is input into the text encoder and the new preference latent vector is output. The newly added behavior latent vector and the newly added preference latent vector are input into the cross-attention fusion layer, and the updated user feature data is output. The updated user feature data is input into the hypernetwork model, the response map prediction network outputs the updated low-dimensional response map, the parameter generation network takes the updated low-dimensional response map as the prior input, outputs the updated first matrix and the updated second matrix, and calculates the updated weight update matrix. Calculate the difference norm between the updated weight update matrix and the weight update matrix currently loaded in the personalized software model. When the difference norm is greater than a preset difference threshold, replace the currently loaded weight update matrix with the updated weight update matrix and add it to the corresponding weight matrix in the basic software model to obtain the weight matrix of the updated personalized software model.
[0013] A second aspect of the present invention provides a multi-user software personalization system, comprising the following modules: User Feature Awareness Module: Acquires user feature data of target users, including user behavior sequence data and user preference description data; Hypernetwork processing module: Inputs the user feature data into the hypernetwork model, which includes a response graph prediction network and a parameter generation network; Personalized weight generation module: The response graph prediction network extracts user pattern features from the user feature data and outputs a low-dimensional response graph. The response graph is a low-rank matrix, and each dimension corresponds one-to-one with the weight matrix of each functional module in the basic software model. The parameter generation network uses the response graph as prior input and sequentially performs dimensionality-up mapping on the response graph through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix corresponding to the layer to be adapted in the basic software model. The low-rank decomposition matrix contains a first matrix and a second matrix. The product of the first matrix and the second matrix is calculated as the weight update matrix. The weight update matrix is added to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user. Interactive reasoning module: In response to the target user's software usage request, it outputs software interaction content through the personalized software model.
[0014] The beneficial effects of this invention are: This invention uses a hypernetwork model to directly generate LoRA parameters based on user feature data in a feedforward manner, eliminating the need for independent gradient iteration optimization for each user. This improves the speed of personalized adaptation from minutes / hours to seconds, meeting the real-time interaction requirements of software. This invention introduces a low-dimensional response map as an intermediate representation. Each dimension of the map corresponds one-to-one with the weight matrix of each functional module in the basic software model. The parameter generation network uses the response map as a priori input to perform dimensional reconstruction, which gives the parameter generation process clear structured guidance. Each dimension carries the adaptation information of the corresponding functional module, making the parameter adjustment process interpretable and traceable. This invention extracts low-rank response maps from the LoRA parameters independently optimized by each user through SVD decomposition as supervision labels. It adopts a two-stage decoupled training method to encode the parameter adjustment direction information of each user into reusable structured knowledge, realizing the sharing and reuse of personalized knowledge among users through response maps. New users can refer to the response maps of existing users as prior constraints to solve the cold start problem. At the same time, the response map always maintains the correspondence with the functional modules. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the module flow of the present invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] Please see Figure 1 As shown, the present invention is a method for personalizing multi-user software, comprising the following steps: S1: Obtain user characteristic data of the target user, including user behavior sequence data and user preference description data; S2: Input the user feature data into the hypernetwork model, which includes a response graph prediction network and a parameter generation network; S3: The response graph prediction network extracts user pattern features from the user feature data and outputs a low-dimensional response graph. The response graph is a low-rank matrix, and each dimension corresponds one-to-one with the weight matrix of each functional module in the basic software model. The parameter generation network takes the response graph as a priori input and sequentially performs dimensionality-upgrading mapping on the response graph through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix corresponding to the layer to be adapted in the basic software model. The low-rank decomposition matrix contains a first matrix and a second matrix. The product of the first matrix and the second matrix is calculated as the weight update matrix. The weight update matrix is added to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user. S4: In response to the target user's software usage request, output software interaction content through the personalized software model.
[0018] In one embodiment of the present invention, step S1, obtaining user characteristic data of the target user, includes the following steps: Collect the target user's original operation logs within a preset time window, including the call timestamps, call frequency, call duration, and call return status codes of each functional module; The original operation logs are arranged into an operation sequence in chronological order. An operation vector is extracted for each operation event in the operation sequence. The operation vector includes a functional module identifier, an operation type code, and a duration normalization value. Each operation vector in the operation sequence is input into the behavior encoder in time step order, and a fixed-dimensional behavior sequence latent vector is output as user behavior sequence data. Obtain the natural language customized text input by the target user, input the natural language customized text into the text encoder, and output the preference description latent vector as the user preference description data; The latent vector of the behavior sequence is used as the query vector, and the latent vector of the preference description is used as the key vector and value vector. They are input into the cross-attention fusion layer. The cross-attention fusion layer calculates the attention weight matrix between the features of each time step in the latent vector of the behavior sequence and the features of each semantic unit in the latent vector of the preference description. The latent vector of the preference description is weighted and aggregated according to the attention weight matrix, and the fused user feature data is output.
[0019] Specifically, the system collects raw operation logs of the target user within a preset time window. This preset time window is pre-set by the system administrator; in this embodiment, it is set to the most recent 30 days. Each record in the raw operation log contains the following fields: the call timestamp of the functional module, the cumulative call frequency of that functional module within the preset time window, the duration of each call, and the call return status code (used to indicate whether the call was successful). The call return status code is an integer value, where 200 indicates a successful call, and anything other than 200 indicates a failed call or an error. The collected raw operation logs are arranged into an operation sequence according to the order of the call timestamps, resulting in a sequence of length [length missing]. operation sequence ,in, The total number of operation events within a preset time window; for each operation event in the operation sequence. Extract operation vector , operation vector For a three-dimensional vector, it is represented as: ,in, This is a functional module identifier bit, using one-hot encoding or integer index encoding, used to identify the functional module to which the operation event belongs; The operation type is encoded to identify the type of operation event, such as "open", "close", "edit", "export", "query", etc., and each category is pre-assigned a different integer value; This is a duration-normalized value, obtained by dividing the call duration by the maximum call duration of all operation events within a preset time window. The range of values is ; The operation vectors in the operation sequence are input into the behavior encoder in time step order. This behavior encoder is a time series coding network based on a Gated Recurrent Unit (GRU). The GRU behavior encoder is constructed as follows: the hidden state dimension of the GRU is set to... (In this embodiment) The network has a single layer, and the operation vectors at each time step are... (With a dimension of 3) as input to the GRU, the hidden state is updated step-by-step according to the following update method: ; ;; ;in, To update the gate vector, To reset the gate vector, This is a candidate hidden state. This represents the hidden state at the current time step. This is the hidden state of the previous time step. , , , ,and The weight matrix is a learnable matrix. , and For learnable bias vectors, It is the Sigmoid activation function. The hyperbolic tangent activation function is used. For element-wise multiplication; the last time step Hidden state As the encoding result of the entire operation sequence, the output is a fixed-dimensional latent vector of the behavior sequence. The training process of the GRU behavior encoder is carried out jointly with the hypernetwork model. The system acquires customized natural language text input from the target user. This text is a description of the user's personalized needs submitted through the software interface in natural language, such as "I prefer to place frequently used functions on the left side of the interface, and charts should be displayed primarily as line graphs." This customized natural language text is then input into a text encoder, which is a pre-trained language model based on the Transformer architecture. In this embodiment, the BERT-based model is used as the text encoder. During encoding, the customized natural language text is segmented and then input into the BERT model. The vector corresponding to the [CLS] marker in the BERT model output is taken as the encoding result of the entire customized natural language text, and the latent vector describing the preference is output. ; implicit vector of behavior sequence As a query vector, the implicit vector describing preferences will be used. Simultaneously serving as both a key vector and a value vector, the input vectors are fed into the cross-attention fusion layer; specifically, the cross-attention fusion layer first maps the input vectors to the same feature space through three linear transformations: Where is the query vector. For key vectors, For value vectors, , and The attention weight matrix is a learnable weight matrix; the attention weight matrix between the latent vectors of the behavior sequence and the latent vectors of the preference description is calculated. ,in, Obtain the similarity score vector. The function normalizes the vector into an attention weight distribution; it then performs weighted aggregation on the latent vectors describing preferences based on the attention weight matrix to obtain the fused user feature data. That is, the output dimension is The fusion of user feature vectors; the linear transformation matrices in the cross-attention fusion layer. , , The training and hypernetwork model are carried out together, and the fused user feature data The user feature data serves as the input to the hypernetwork model in step S2.
[0020] In one embodiment of the present invention, step S2, the training method of the hypernetwork model includes the following steps: A training dataset is obtained, which includes the sample user features corresponding to several sample users and LoRA parameters of each group of samples. The LoRA parameters of each group of samples are obtained by iteratively optimizing the basic software model for the corresponding sample users. The objective function of the iterative optimization is to minimize the prediction loss of the basic software model on the labeled data of the corresponding sample users. For each group of samples, calculate the change matrix of LoRA parameters relative to the original parameters of the basic software model. Perform singular value decomposition on each change matrix and extract the first few elements of each change matrix. The left and right singular vectors corresponding to the 1st maximal singular value are used to calculate a low-rank approximation matrix as the sample response map of the corresponding sample user. The preset low-rank dimension value; The sample user features of each sample user are used as input, and the corresponding sample response map is used as supervision label to train the response map prediction network. The parameters of the trained response map prediction network are fixed, and the sample user features of each sample user are input into the trained response map prediction network to output the corresponding predicted response map. The predicted response maps of each sample user are used as input, and the corresponding LoRA parameters of each group of samples are used as supervision labels to train the parameter generation network.
[0021] Specifically, in the offline training preparation phase of the actual model, the system first needs to construct a representative user profile dataset, which can be randomly sampled from historical logs in practice. Using active users as a sample, for the first Sample users Perform the following operations: Step 1: Obtain sample user characteristics, collect the first... The original operation logs and customized natural language text of a sample user within a preset time window are processed by a behavior encoder, a text encoder, and a cross-attention fusion layer to obtain the sample user features. Step 2: Obtain the sample LoRA parameters corresponding to the sample users, and then apply the basic software model to the sample users in the first step. Iterative optimization is performed on the labeled data of a sample user to obtain the LoRA parameters corresponding to that sample user. The labeled data includes input-output pairs generated during the sample user's historical usage, including input operation commands and corresponding expected software interface feedback. The objective function of the iterative optimization is to minimize the prediction loss of the basic software model on the labeled data of that sample user. Specifically, the original weight matrix of the basic software model is denoted as... , The number of the layer to be adapted. , The LoRA parameter represents the total number of layers to be adapted in the basic software model, and includes the low-rank decomposition matrix corresponding to each layer to be adapted. and During the iterative optimization process, the weight matrices of each layer to be adapted in the basic software model remain fixed, and only the LoRA parameters are adjusted. and Perform gradient updates; for the ... Layer, personalized weight matrix is represented as ,in, The iterative optimization uses the Adam optimizer with a learning rate of 1e-4. Optimization stops when the prediction loss on the validation set no longer decreases after 10 consecutive training epochs. After the above iterative optimization, all sample LoRA parameters corresponding to the i-th sample user are obtained, denoted as the sample LoRA parameter set corresponding to that sample user. ,in, Repeat steps 1 and 2 above until all are obtained. The characteristics of each sample user and sample LoRA parameter set , constitute the training dataset ; For each group of samples, calculate the change matrix of LoRA parameters relative to the original parameters of the basic software model. For the first group... The first sample user One layer to be adapted, its change matrix Singular value decomposition is performed on each transformation matrix. Specifically, for the first... The first sample user Layer change matrix Decompose it into: ,in, Let be a left singular vector matrix, whose column vectors are left singular vectors. It is a diagonal matrix whose diagonal elements are singular values arranged in descending order. Let be a right singular vector matrix, whose column vectors are right singular vectors. This is the transpose vector; extract the first part of each transformation matrix. The left and right singular vectors corresponding to the maximal singular values, where... The preset low-rank dimension value, The value is set by the system administrator based on the balance requirements between accuracy and compression ratio; a low-rank approximation matrix is calculated as the sample response spectrum of the corresponding sample user. Specifically, for the _____, The first sample user Layer, before extraction The left singular vector corresponding to the maximal singular value and right singular vector Construct the sample response map matrix of this layer. ; will the first All sample users Sample response spectrum matrix of each layer to be adapted By splicing or overlaying the data, a complete sample response map corresponding to that sample user can be constructed. In this embodiment, the element-wise average value of the sample response spectrum matrix of each layer is used as the sample response spectrum of that layer; The Response Graph Prediction Network is used to learn the mapping relationship from sample user features to sample response graphs. The network is constructed as follows: the input dimension of the Response Graph Prediction Network is equal to the user feature data. The dimensions (256 in this embodiment, corresponding to those in Embodiment 1) The network structure is a multilayer perceptron (MLP), containing three hidden layers and one output layer: the first hidden layer has 512 neurons, the second hidden layer has 512 neurons, the third hidden layer has 256 neurons, and the number of neurons in the output layer is equal to the total number of functional modules in the basic software model, denoted as . The activation function for each hidden layer is... Function; based on the sample user characteristics of each sample user As input, the corresponding complete sample response map As supervisory labels, the response map prediction network is trained, and the training loss function is the mean squared error (MSE) loss. ,in, In response to the forward propagation mapping function of the graph prediction network, For the network to the first The predicted response map output by each sample user The square of the Frobenius norm. The total number of sample users; training uses the Adam optimizer with a learning rate of 1e-3 and a batch size of 32. Training stops when the loss value on the validation set no longer decreases for 10 consecutive training epochs; after training, the trained response graph prediction network is obtained. Its parameters are fixed and no longer participate in gradient updates during subsequent training; Sample user characteristics of each sample user Input into the trained response map prediction network Output the corresponding predicted response map At this point, the parameters of the response map prediction network are fixed and no gradient updates are performed. The parameter generation network is used to learn the mapping relationship from the predicted response map to the sample LoRA parameters. This network consists of a mapping layer and a reconstruction layer. The mapping layer is the first linear transformation unit, and its input dimension is equal to the dimension of the predicted response map (i.e., the total number of functional modules). The output dimension is equal to the sum of the original dimensions of the weight matrices of each layer to be adapted. Specifically, for the first layer... There are one layer to be adapted (its weight matrix dimension is...). The dimension of the intermediate latent vector output by the mapping layer that corresponds to that layer is... ,all The intermediate latent vectors corresponding to the layer are concatenated and output through this linear transformation unit. The linear transformation of the mapping layer is denoted as: ,in, This is the weight matrix of the mapping layer. For bias vectors, The intermediate latent vectors are used as the intermediate latent vectors; the reconstruction layer reconstructs the intermediate latent vectors according to the output dimensions of each layer to be adapted. and input dimensions Perform dimensional splitting, for the first Layer, from Extract the part corresponding to this layer ,Will After being stretched into a one-dimensional vector in column-major order, it is mapped to the first and second matrices through the second and third linear transformation units, respectively. The vectorized result is then reshaped into matrix form according to the corresponding dimensions. During training, the predicted response maps of each sample user are used. The LoRA parameters of each group of samples, as input. As a supervisory label, the parameter generation network is trained, and the training loss function is the mean squared error loss between the weight update matrix reconstructed from the low-rank factorization matrix output by the parameter generation network and the sample LoRA parameters: ,in, For the parameter generation network, targeting the first The first sample user The weight update matrix output by the layer; training uses the Adam optimizer with a learning rate of 1e-4 and a batch size of 32. Training stops when the loss value on the validation set no longer decreases for 10 consecutive training epochs. After training, the trained parameter generator network is obtained. ; During the two-stage training process, the response map prediction network and the parameter generation network are trained independently. After the first stage of training is completed, the parameters of the response map prediction network are fixed before entering the second stage. The training data sources for the two stages are the same, but the supervision labels are different, so they do not interfere with each other.
[0022] In one embodiment of the present invention, step S3, the response graph prediction network includes a temporal attention encoding unit and a response mapping unit, and includes the following steps: The temporal attention encoding unit performs self-attention encoding on the latent vector of the behavior sequence in the user feature data, and uses the features of each time step in the latent vector of the behavior sequence as query vector, key vector and value vector respectively, calculates the attention weight matrix between each time step, and outputs temporal enhancement features. The response mapping unit takes the temporal enhancement features as input and maps them into a low-dimensional response map through multiple fully connected layers. The number of neurons in the output layer of the response mapping unit is equal to the total number of functional modules in the basic software model. Among them, the output layer of the response mapping unit is the first The output value of the nth neuron is used as the nth neuron in the low-dimensional response map. The numerical value of the i-th dimension, the i-th dimension of the low-dimensional response map The first dimension index and basic software model The weight matrices of each functional module have the same position index in the model parameter storage structure. The response mapping unit outputs the first functional module according to the position index. The output value of the nth neuron is written into the first low-dimensional response map. Each dimension.
[0023] Specifically, the temporal attention encoding unit performs self-attention encoding on the latent vectors of the behavior sequence in the user feature data. These latent vectors of the behavior sequence are denoted as... , The The line represents the first The behavioral characteristics at each time step will The features at each time step are used as the query vector, key vector, and value vector, respectively. Specifically, they are transformed using three learnable linear transformation matrices. Mapping to the same feature space yields the query vector matrix. Key vector matrix Sum value vector matrix : ; ; ;in, , and The weight matrix is learnable; the attention weight matrix between time steps is calculated. , Multiplying the two together yields A similarity score matrix of dimension 1. The function normalizes the matrix row by row, so that the sum of the attention weights in each row is 1, and then divides by... This is a scaling factor used to prevent the dot product result from becoming too large. Gradient saturation; weighted aggregation of value vectors based on the attention weight matrix to output temporal enhancement features. : Among them, the weight matrix in the temporal attention encoding unit , and The training of this network is performed simultaneously with the training of the response map prediction network in Example 2. That is, these weight matrices are optimized using gradient descent, and the loss function is described in Example 2. ; Response mapping unit with temporal enhancement features As input, it is mapped to a low-dimensional response map through multiple fully connected layers. The network structure of the response mapping unit is as follows: the input layer receives... The global temporal features are obtained by performing average pooling along the time step dimension. , ;Will The system sequentially passes through two fully connected hidden layers and an output layer. The first hidden layer has 512 neurons, and the second hidden layer has 256 neurons. The activation function for both is ReLU. The number of neurons in the output layer equals C, which is the total number of functional modules in the basic software model. The output layer is a fully connected layer without an activation function, and its output value is a real number. The response mapping unit's output layer contains the first... The output value of the nth neuron is used as the nth neuron in the low-dimensional response map. Numerical values for each dimension: To establish the correspondence between each dimension and the weight matrices of each functional module in the basic software model, a one-to-one correspondence is achieved using the "same position index" method. Specifically, in the parameter storage structure of the basic software model, the weight matrices of each functional module are stored in order according to the functional module numbers (1, 2, ..., C). The output layer of the response mapping unit also arranges neurons in the same number order. The output value of the nth neuron corresponds to the nth neuron. The functional module, the response mapping unit, writes the output value of the i-th neuron in the output layer into the i-th low-dimensional response map according to the position index. The correspondence is achieved through the following method: During system initialization, the parameter storage structure of the basic software model is read, the arrangement order of the weight matrices of each functional module in the storage structure is obtained, a mapping table of "functional module number → storage location index" is established, the output layer of the response mapping unit generates output values in the same numbering order, and the first dimension is set as the first dimension. The output value of the nth neuron is written into the low-dimensional response map vector. The write operation directly assigns the output value to the element at the corresponding index of the low-dimensional response map vector, without additional calculation. The weight and bias parameters of each fully connected layer in the response mapping unit are trained synchronously with the response map prediction network in Example 2, i.e., optimized through gradient descent. The loss function is shown in Example 2. ; The first of the low-dimensional response maps The first dimension index and basic software model The weight matrices of each functional module have the same position index in the model parameter storage structure. This is achieved as follows: During system initialization, the parameter reading interface of the basic software model is called to sequentially obtain the start and end addresses of the weight matrices of each functional module in the parameter storage structure according to the functional module number order (1, 2, ..., C), generating a functional module position index table. This index table is a key-value pair data structure, where the key is the functional module number (1 to C), and the value is the position index of the module's weight matrix in the parameter storage structure (also an integer value from 1 to C, representing the order of arrangement). The output layer of the response mapping unit configures neurons according to the functional module number order (1, 2, ..., C), that is, the first neuron of the output layer corresponds to functional module number 1, the second neuron of the output layer corresponds to functional module number 2, and so on. During forward propagation, the first neuron of the output layer... The activation value of the nth neuron is directly assigned to the nth neuron in the low-dimensional response map vector. The element; since the order of the output layer neurons and the order of the functional modules in the parameter storage structure are both determined by the same number (1, 2, ..., C), the first element of the low-dimensional response map is... The dimension and the first The weight matrices of each functional module have the same position index in the parameter storage structure; this correspondence does not need to be obtained through training, but is predetermined in advance through sequential consistency during the network construction stage.
[0024] In one embodiment of the present invention, in step S3, the parameter generation network uses the response map as prior input and sequentially performs dimensionality-upgrading mapping on the response map through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix, including the following steps: The mapping layer inputs the low-dimensional response map to the first linear transformation unit. The output dimension of the first linear transformation unit is equal to the original dimension of the weight matrix of the layer to be adapted, and outputs the intermediate latent vector. The reconstruction layer will adjust the intermediate latent vectors according to the output dimension of the layer to be adapted. and input dimensions Dimensional splitting is performed, and the first part after splitting is mapped to the first matrix through the second linear transformation unit, and the second part after splitting is mapped to the second matrix through the third linear transformation unit; During training, the reconstruction layer applies column orthogonality constraints to the first matrix. Apply row orthogonality constraints to the second matrix. ,in, for An identity matrix of order 1. For the first matrix, For the second matrix, It is the transpose vector.
[0025] Specifically, the mapping layer will map the low-dimensional response map. The input is fed into the first linear transformation unit, which is a fully connected layer with an input dimension of . The output dimension is equal to the sum of the original dimensions of all the weight matrices of the layers to be adapted in the basic software model. Let the basic software model have a total of The first layer to be adapted, the... The layer weight matrix has a dimension of ,in, For output dimensions, Given the input dimension, the output dimension of the first linear transformation unit is... In this embodiment, (That is, the basic software model has 12 Transformer layers of self-attention and feedforward network modules as layers to be adapted); the forward computation process of the first linear transformation unit is as follows: ,in, For the row vectors of the low-dimensional response map, The weight matrix of the first linear transformation unit. For bias vectors, The intermediate implicit vectors of the output; in the first linear transformation unit and As trainable parameters, they are optimized through gradient descent during the parameter generation network training phase in Example 2; The reconstruction layer will use the intermediate hidden vectors According to the output dimensions of each layer to be adapted and input dimensions Perform dimensional splitting, specifically, for the ... One layer to be adapted, from the intermediate latent vector Extract the part corresponding to this layer The extraction method is as follows: according to the order of each layer to be adapted in the basic software model, ... Cut into in sequence The nth consecutive segment, the nth The starting position of each segment is The end position is Due to the differences between the segments The elements are arranged consecutively and do not overlap, so extraction can be completed by index slicing; after extraction, ... Restructured according to row priority The 3D matrix form is denoted as ,Will According to the output dimension of the layer to be adapted and input dimensions The parts are split into two, and the first part after splitting is... The former The second part is (corresponding to the output dimension direction). The remaining part is divided as follows: Let It is a preset low-rank value, and ,Will Split into two matrices: (Split by row direction), where, for The former OK, for The remaining columns; in this embodiment, the following split mapping method is adopted: Mapped to the first matrix by the second linear transformation unit The third linear transformation unit maps it to the second matrix. The second and third linear transformation units are both fully connected layers, wherein: ; ,in, This is the reshaped intermediate latent vector segment. and The weight matrix is a learnable matrix. and For the corresponding bias vector, It is the ReLU activation function. For the first matrix, Let be the second matrix, where, and Dimensions and original weight matrix The dimensions are consistent; the second linear transformation unit and the third linear transformation unit , , and As trainable parameters, they are optimized through gradient descent during the parameter generation network training phase in Example 2; The reconstruction layer modifies the first matrix during training. Apply column orthogonality constraints to the second matrix. Row orthogonality constraints are applied, which are achieved by adding a regularization term to the training loss function. Column orthogonality constraints require that the first matrix... The columns are pairwise orthogonal, that is: The row orthogonality constraint requires that each row of the second matrix B_l be pairwise orthogonal, i.e. The orthogonal constraints mentioned above are added as regularization terms to the training loss function of the parameter generation network, defining the orthogonal regularization loss. as follows: ,in, Let Frobenius be the square of the norm, i.e., the sum of the squares of all elements in the matrix. The first term penalizes the degree of non-orthogonality between the columns of the first matrix, and the second term penalizes the degree of non-orthogonality between the rows of the second matrix. When the column orthogonality constraint is strictly satisfied, The first term is 0; when When strictly satisfying row orthogonality constraints, The second term is 0; the network reconstruction loss is generated by combining the parameters in Example 2. The complete training loss function is: ,in, is the regularization coefficient, used to control the strength of orthogonal constraints. In this embodiment, The acquisition method is as follows: In the early stages of training (the first 10 training rounds), set... The network is initially optimized primarily using reconstruction loss; starting from the 11th training epoch, the loss is gradually increased. value, until the maximum value is reached. It remains unchanged afterward. The value is determined through the following steps: on the validation set... Ten candidate values are sampled at logarithmic intervals within the range [0.001, 0.5], and trained for the same number of epochs. The value that minimizes the reconstruction loss of the validation set is selected. The value, ultimately determined to be 0.1 in this embodiment, is achieved by applying orthogonal constraints to the first and second matrices, resulting in a low-rank decomposition matrix. The rank is maintained during training. The weight update matrix always maintains a preset low-rank structure.
[0026] In one embodiment of the present invention, step S3 involves adding the weight update matrix to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user, including the following steps: Obtain the original weight matrix of each layer to be adapted in the basic software model. ,in, For the first in the basic software model One layer to be adapted; The parameter generation network outputs the first matrix for each layer to be adapted. Second matrix Calculate the weight update matrix for each layer to be adapted. ; Update the weight matrix of each layer to be adapted. With the corresponding original weight matrix By adding each element, we obtain the personalized weight matrix for each layer to be adapted. ; The personalized weight matrix of each layer to be adapted By backfilling according to the original position of each layer to be adapted in the basic software model, the complete weight matrix of the personalized software model corresponding to the target user is obtained.
[0027] Specifically, the original weight matrices of each layer to be adapted in the basic software model are obtained. The basic software model is a pre-trained deep learning model whose parameters are stored in a model file. In this embodiment, the basic software model is a software interaction model based on the Transformer architecture, including... One layer to be adapted, For the first For each layer to be adapted, the original weight matrix of that layer is read from the parameter storage structure of the basic software model, denoted as... ,in, This is the output dimension of this layer. For the input dimension of this layer, each layer and The value is determined by the network structure of the basic software model; the reading method is: calling the parameter loading interface of the basic software model, through layer indexing. Locate the corresponding weight tensor, load it into memory, and store it in tensor format in PyTorch or TensorFlow. Among them, the personalized weight matrix of each layer to be adapted Backfilling is performed according to the original position of each layer to be adapted in the basic software model to obtain the complete weight matrix of the personalized software model corresponding to the target user. The backfilling method is as follows: For each layer to be adapted... , calculate Write back to the storage location of the original weight matrix of this layer in the basic software model parameter storage structure, overwriting the original weights of this layer. Specifically, this involves calling the parameter update interface of the basic software model and passing in the layer index. and the updated weight matrix The interface writes the new weights to the corresponding location in the model parameter file or memory. During the model inference phase, a parameter replacement rather than parameter merging loading method is used. Specifically, the basic software model loads the original weights during initialization. Before inference, for each layer to be adapted... Read from the personalized weight matrix storage structure Call the model parameter update interface to... Directly write the weight tensor to the storage location of the corresponding layer in the model, replacing the original one. No additional processing or merging of the weight update matrix is required. Each layer to be adapted is replaced sequentially according to its number, from smallest to largest. After replacement, the weights of each layer in the basic software model have been replaced with their corresponding personalized weights, resulting in the complete weight matrix of the personalized software model corresponding to the target user. This model can be directly used for inference without reloading the entire model file. The refilled personalized software model is deployed to the inference environment. When the target user initiates a software usage request, the system calls this personalized software model for forward inference, outputting software interaction content that matches the target user's preferences. Since the weights of each layer to be adapted have been replaced with those containing personalized information... During the inference process, the model will automatically generate output results that adapt to the user's operating habits and preferences.
[0028] In one embodiment of the present invention, step S5 is further included: dynamically updating the weight matrix of the personalized software model corresponding to the target user. During the process of target users using the personalized software model, continuously collect new operation logs and new customized instructions; The newly added operation logs are arranged into a new operation sequence according to time order. The new operation vector of each new operation event is extracted. The new operation vectors are input into the behavior encoder in time step order and the new behavior latent vector is output. The new customized instruction is input into the text encoder and the new preference latent vector is output. The newly added behavior latent vector and the newly added preference latent vector are input into the cross-attention fusion layer, and the updated user feature data is output. The updated user feature data is input into the hypernetwork model, the response map prediction network outputs the updated low-dimensional response map, the parameter generation network takes the updated low-dimensional response map as the prior input, outputs the updated first matrix and the updated second matrix, and calculates the updated weight update matrix. Calculate the difference norm between the updated weight update matrix and the weight update matrix currently loaded in the personalized software model. When the difference norm is greater than a preset difference threshold, replace the currently loaded weight update matrix with the updated weight update matrix and add it to the corresponding weight matrix in the basic software model to obtain the weight matrix of the updated personalized software model.
[0029] Specifically, new behavioral latent vectors will be added. As a query vector, a new implicit preference vector will be added. As key and value vectors, they are input to the cross-attention fusion layer, which has the same network structure as the cross-attention fusion layer in Example 1, and shares the same weight matrix parameters; specifically, it follows the cross-attention fusion method in Example 1: ; ; ; ; ;in, For query vector, The learnable weight matrix for the query vector. For key vectors, Let be the learnable weight matrix of the key vectors. For value vectors, The learnable weight matrix is the value vector. This is the attention weight matrix, where each element represents the degree of correlation between the newly added behavioral latent vector and each semantic unit in the newly added preference description. It is a transpose vector. For normalized exponential functions, The updated user feature data is the final output after the new behavioral features and new preference features are fused through cross-attention, and it serves as the input to the hypernet model. Updated user feature data The input is fed into the trained hypernetwork model, where the response graph prediction network is used as... As input, following the forward propagation method in Example 3, the updated low-dimensional response map is output. , ,in, The mapping function for the trained response map prediction network; where the parameter generation network uses... As prior input, following the forward propagation method in Example 4, the matrix passes sequentially through the mapping layer and the reconstruction layer, outputting the updated first matrix. and the updated second matrix ,in, ,in, Generate a mapping function for the trained parameters network, and calculate the updated weight update matrix for each layer to be adapted. ; Calculate the difference norm between the updated weight update matrix and the weight update matrix loaded in the current personalized software model. The currently loaded weight update matrix is denoted as... That is, the weight update matrix generated and loaded into the personalized software model in step S3, for each layer to be adapted. ,calculate and The difference in Frobenius norm between them: ,in, The Frobenius norm is used; the average of the difference norms of all layers to be adapted is used as the overall difference norm. Among them, the preset difference threshold During the system deployment phase, a batch of personalized adaptation records for users are randomly selected from historical user data. For each user, the difference norm of the weight update matrix between two consecutive updates is calculated to obtain the difference norm distribution. The 95th percentile of this distribution is then taken as the preset difference threshold. The initial value; where, when When it is determined that the user's behavioral preferences have changed significantly, a model update needs to be triggered; when If the current difference is determined to be within an acceptable range, no update will be triggered, and the system will continue to use the current personalized software model. Among them, when the difference norm When the time comes, perform the following replacement operation: update the weights matrix with the updated weights. Replace the currently loaded weight update matrix The new current weight update matrix is stored in memory; following the method in Example 5, the following is performed: Corresponding to the original weight matrix in the basic software model By adding each element, we obtain the updated personalized weight matrix: ,Will Backfilling is performed according to the original position of each layer to be adapted in the basic software model, replacing the weights of the corresponding layers in the current model, and obtaining the weight matrix of the updated personalized software model. Specifically, a double-buffered seamless switching mechanism is adopted in the weight backfilling and replacement process, and the system allocates a new video memory area in the background to load the updated weights. After loading, the inference pointer of the personalized software model is atomically switched to the new video memory area, and then the memory resources occupied by the old weights are released, thus ensuring that the foreground interaction of the target user is not interrupted. After the update is completed, the system continues to provide services to the current user using the updated personalized software model, and at the same time, the updated version of the software model will be updated. The above dynamic update process, which is stored as a new "currently loaded weight update matrix" for the next calculation of the difference norm, runs continuously in the background of the system without interrupting the user's foreground interaction.
[0030] Please see Figure 2 As shown, the present invention is a multi-user software personalization customization system, comprising the following modules: User Feature Awareness Module: Acquires user feature data of target users, including user behavior sequence data and user preference description data; Hypernetwork processing module: Inputs the user feature data into the hypernetwork model, which includes a response graph prediction network and a parameter generation network; Personalized weight generation module: The response graph prediction network extracts user pattern features from the user feature data and outputs a low-dimensional response graph. The response graph is a low-rank matrix, and each dimension corresponds one-to-one with the weight matrix of each functional module in the basic software model. The parameter generation network uses the response graph as prior input and sequentially performs dimensionality-up mapping on the response graph through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix corresponding to the layer to be adapted in the basic software model. The low-rank decomposition matrix contains a first matrix and a second matrix. The product of the first matrix and the second matrix is calculated as the weight update matrix. The weight update matrix is added to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user. Interactive reasoning module: In response to the target user's software usage request, it outputs software interaction content through the personalized software model.
[0031] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A method for multi-user software personalization customization, characterized by, Includes the following steps: S1: Obtain user characteristic data of the target user, including user behavior sequence data and user preference description data; S2: Input the user feature data into the hypernetwork model, which includes a response graph prediction network and a parameter generation network; S3: The response graph prediction network extracts user pattern features from the user feature data and outputs a low-dimensional response graph. The response graph is a low-rank matrix, and each dimension corresponds one-to-one with the weight matrix of each functional module in the basic software model. The parameter generation network takes the response graph as a priori input and sequentially performs dimensionality-upgrading mapping on the response graph through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix corresponding to the layer to be adapted in the basic software model. The low-rank decomposition matrix contains a first matrix and a second matrix. The product of the first matrix and the second matrix is calculated as the weight update matrix. The weight update matrix is added to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user. S4: In response to the target user's software usage request, output software interaction content through the personalized software model.
2. The method of claim 1, wherein, In step S1, obtaining the user characteristic data of the target user includes the following steps: Collect the target user's original operation logs within a preset time window, including the call timestamps, call frequency, call duration, and call return status codes of each functional module; The original operation logs are arranged into an operation sequence in chronological order. An operation vector is extracted for each operation event in the operation sequence. The operation vector includes a functional module identifier, an operation type code, and a duration normalization value. Each operation vector in the operation sequence is input into the behavior encoder in time step order, and a fixed-dimensional behavior sequence latent vector is output as user behavior sequence data. Obtain the natural language customized text input by the target user, input the natural language customized text into the text encoder, and output the preference description latent vector as the user preference description data; The latent vector of the behavior sequence is used as the query vector, and the latent vector of the preference description is used as the key vector and value vector. They are input into the cross-attention fusion layer. The cross-attention fusion layer calculates the attention weight matrix between the features of each time step in the latent vector of the behavior sequence and the features of each semantic unit in the latent vector of the preference description. The latent vector of the preference description is weighted and aggregated according to the attention weight matrix, and the fused user feature data is output.
3. The method of claim 1, wherein, In step S2, the training method of the hypernetwork model includes the following steps: A training dataset is obtained, which includes the sample user features corresponding to several sample users and LoRA parameters of each group of samples. The LoRA parameters of each group of samples are obtained by iteratively optimizing the basic software model for the corresponding sample users. The objective function of the iterative optimization is to minimize the prediction loss of the basic software model on the labeled data of the corresponding sample users. respectively, singular value decompositions are performed on each of the change matrices, left singular vectors and right singular vectors corresponding to the first maximum singular values of each of the change matrices are extracted, and a low-rank approximation matrix is calculated as a sample response graph of a corresponding sample user, wherein, is a preset low-rank dimension value. The sample user features of each sample user are used as input, and the corresponding sample response map is used as supervision label to train the response map prediction network. The parameters of the trained response map prediction network are fixed, and the sample user features of each sample user are input into the trained response map prediction network to output the corresponding predicted response map. The predicted response maps of each sample user are used as input, and the corresponding LoRA parameters of each group of samples are used as supervision labels to train the parameter generation network.
4. The method for personalized customization of multi-user software according to claim 1, characterized in that, In step S3, the response graph prediction network includes a temporal attention encoding unit and a response mapping unit, and includes the following steps: The temporal attention encoding unit performs self-attention encoding on the latent vector of the behavior sequence in the user feature data, and uses the features of each time step in the latent vector of the behavior sequence as the query vector, key vector and value vector respectively, calculates the attention weight matrix between each time step, and outputs the temporal enhancement features. The response mapping unit takes the temporal enhancement features as input and maps them into a low-dimensional response map through multiple fully connected layers. The number of neurons in the output layer of the response mapping unit is equal to the total number of functional modules in the basic software model. Among them, the output layer of the response mapping unit is the first The output value of the nth neuron is used as the nth neuron in the low-dimensional response map. The numerical value of the i-th dimension, the i-th dimension of the low-dimensional response map The first dimension index and basic software model The weight matrices of each functional module have the same position index in the model parameter storage structure. The response mapping unit outputs the first functional module according to the position index. The output value of the nth neuron is written into the first low-dimensional response map. Each dimension.
5. The method for personalized customization of multi-user software according to claim 1, characterized in that, In step S3, the parameter generation network uses the response map as prior input and sequentially performs dimensionality-upgrading mapping on the response map through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix, including the following steps: The mapping layer inputs the low-dimensional response map to the first linear transformation unit. The output dimension of the first linear transformation unit is equal to the original dimension of the weight matrix of the layer to be adapted, and outputs the intermediate latent vector. The reconstruction layer will adjust the intermediate latent vectors according to the output dimension of the layer to be adapted. and input dimensions Dimensional splitting is performed, and the first part after splitting is mapped to the first matrix through the second linear transformation unit, and the second part after splitting is mapped to the second matrix through the third linear transformation unit; During training, the reconstruction layer applies column orthogonality constraints to the first matrix. Apply row orthogonality constraints to the second matrix. ,in, for An identity matrix of order 1. For the first matrix, For the second matrix, It is the transpose vector.
6. The method for personalizing multi-user software according to claim 1, characterized in that, In step S3, the weight update matrix is added to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user, including the following steps: Obtain the original weight matrix of each layer to be adapted in the basic software model. ,in, For the first in the basic software model One layer to be adapted; The parameter generation network outputs the first matrix for each layer to be adapted. Second matrix Calculate the weight update matrix for each layer to be adapted. ; Update the weight matrix of each layer to be adapted. With the corresponding original weight matrix By adding each element, we obtain the personalized weight matrix for each layer to be adapted. ; The personalized weight matrix of each layer to be adapted By backfilling according to the original position of each layer to be adapted in the basic software model, the complete weight matrix of the personalized software model corresponding to the target user is obtained.
7. A method for personalizing multi-user software according to claim 1, characterized in that, It also includes step S5: dynamically updating the weight matrix of the personalized software model corresponding to the target user: During the process of target users using the personalized software model, continuously collect new operation logs and new customized instructions; The newly added operation logs are arranged into a new operation sequence according to time order. The new operation vector of each new operation event is extracted. The new operation vectors are input into the behavior encoder in time step order and the new behavior latent vector is output. The new customized instruction is input into the text encoder and the new preference latent vector is output. The newly added behavior latent vector and the newly added preference latent vector are input into the cross-attention fusion layer, and the updated user feature data is output. The updated user feature data is input into the hypernetwork model, the response map prediction network outputs the updated low-dimensional response map, the parameter generation network takes the updated low-dimensional response map as the prior input, outputs the updated first matrix and the updated second matrix, and calculates the updated weight update matrix. Calculate the difference norm between the updated weight update matrix and the weight update matrix currently loaded in the personalized software model. When the difference norm is greater than a preset difference threshold, replace the currently loaded weight update matrix with the updated weight update matrix and add it to the corresponding weight matrix in the basic software model to obtain the weight matrix of the updated personalized software model.
8. A multi-user software personalization system, used to implement the multi-user software personalization method according to any one of claims 1-7, characterized in that, Includes the following modules: User Feature Awareness Module: Acquires user feature data of target users, including user behavior sequence data and user preference description data; Hypernetwork processing module: Inputs the user feature data into the hypernetwork model, which includes a response graph prediction network and a parameter generation network; Personalized weight generation module: The response graph prediction network extracts user pattern features from the user feature data and outputs a low-dimensional response graph. The response graph is a low-rank matrix, and each dimension corresponds one-to-one with the weight matrix of each functional module in the basic software model. The parameter generation network uses the response graph as prior input and sequentially performs dimensionality-up mapping on the response graph through a mapping layer and a reconstruction layer to generate a low-rank decomposition matrix corresponding to the layer to be adapted in the basic software model. The low-rank decomposition matrix contains a first matrix and a second matrix. The product of the first matrix and the second matrix is calculated as the weight update matrix. The weight update matrix is added to the corresponding weight matrix in the basic software model to obtain the weight matrix of the personalized software model corresponding to the target user. Interactive reasoning module: In response to the target user's software usage request, it outputs software interaction content through the personalized software model.
