Cross-domain sequence recommendation method and system based on privacy computing and decoupled information fusion

Abstract

The present invention relates to the field of privacy computing. Disclosed are a cross-domain sequence recommendation method and system based on privacy computing and decoupled information fusion. Specifically, in the cross-domain sequence recommendation method based on privacy computing and decoupled information fusion, a data requester acquires auxiliary information from multiple parties by means of a homomorphic encryption method and performs privacy computing on the auxiliary information without disclosing privacy; and then, by using a decoupled auxiliary information fusion method, a fusion process is shifted from the input to an attention layer, enhancing the modeling capability of an attention mechanism, and avoiding unnecessary attention randomness caused by mixed correlations of heterogeneous embeddings, and flexible gradients are supported to adaptively learn various auxiliary information in different scenarios; and finally, a prediction task is completed using an auxiliary attribute predictor. According to the cross-domain sequence recommendation method based on privacy computing and decoupled information fusion of the present invention, various auxiliary information can be effectively utilized for sequence recommendation tasks, achieving higher attention representation capability and flexibility to learn the relative importance of the auxiliary information.

Description

Cross-domain sequence recommendation method and system based on privacy computing and decoupled information fusion Technical Field

[0001] The present invention belongs to the field of big data technology, and specifically relates to a cross-domain sequence recommendation method and system based on privacy computing and decoupled information fusion. Background Art

[0002] Sequential recommendation aims to model users' dynamic preferences based on their historical behavior and suggest the next item to recommend. However, effectively incorporating auxiliary information into the recommendation process remains a challenging open problem. Numerous research efforts have focused on incorporating auxiliary information at various stages of recommendation. Despite significant improvements, current solutions based on early ensembles still suffer from several drawbacks. For example, the rank of the attention matrix in existing solutions is inherently limited by the size of the projection under the multi-head query key, which is often smaller than the matrix can achieve. This means that integrating embeddings before the attention layer encounters a ranking bottleneck in the attention matrix, resulting in poor representation of attention scores. Second, focusing on the composite embedding space can lead to random interference, where mixed embeddings from various information sources inevitably focus on irrelevant information. Third, because the integrated embeddings remain non-negligible throughout the entire attention block, early ensembles force models to develop complex and cumbersome ensemble solutions and training schemes to provide flexible gradients for various auxiliary information. Using simple fusion solutions, such as additive fusion, where embeddings share the same training gradient, limits the model's ability to learn the relative importance of the auxiliary information encoding relative to the item embeddings.

[0003] Disentangled auxiliary information fusion for sequential recommendation shifts the fusion process from the input to the attention layer. By generating a key and query for each attribute and item in the attention layer, it decouples various auxiliary information and item embeddings, and then fuses all attention matrices using a fusion function. This simple and effective solution overcomes the ranking bottleneck, enhances the modeling capabilities of the attention mechanism, and avoids unnecessary attention randomness caused by the mixed correlations of heterogeneous embeddings. Furthermore, it supports flexible gradients to adaptively learn various auxiliary information in different scenarios. In multi-task training schemes, lightweight auxiliary attribute predictors are used to better activate auxiliary information, thereby beneficially influencing the final learned representation. So, how can privacy computing and disentangled information fusion be combined?

[0004] Summary of the Invention

[0005] Based on the above background and problems existing in the prior art, the present invention adopts the following technical solutions: First, a cross-domain sequence recommendation method based on privacy computing and decoupled information fusion is provided, which can utilize multi-party data through homomorphic encryption method and decoupled auxiliary information fusion method and complete computing tasks efficiently and accurately while protecting privacy.

[0006] A cross-domain sequence recommendation method based on privacy computing and decoupled information fusion includes the following steps:

[0007] Obtain data from data demanders and encrypted auxiliary information obtained based on homomorphic encryption, where the auxiliary information comes from various participants in different fields;

[0008] Inputting the data demander's data into the project embedding layer to obtain project embedding, and inputting the encrypted auxiliary information into the attribute embedding layer to obtain auxiliary information embedding;

[0009] Decoupling the item embedding and the auxiliary information embedding and fusing the auxiliary information to obtain an updated item representation;

[0010] The updated item representation is input into an auxiliary attribute predictor for prediction, thereby obtaining an optimized prediction result of the interaction between the user and the item.

[0011] As an implementable method, the encrypted auxiliary information obtained based on homomorphic encryption includes the following steps:

[0012] Generate a public key and a private key based on a key generation function, and send the public key to each participant;

[0013] Based on a given public key, the auxiliary information corresponding to each participant is encrypted to obtain the encrypted auxiliary information and send it to the data demander. The auxiliary information includes user information, project information, and behavioral attribute information that provide additional information for prediction.

[0014] As an implementation method, the decoupling of the project embedding and the auxiliary information embedding and the fusion of the auxiliary information to obtain an updated project representation include the following steps:

[0015] The item embedding and auxiliary information embedding are input into the attention layer to calculate the attention scores, and the item representation attention score and the attribute representation attention score are obtained respectively. The attention layer includes multiple layers, and the relationships between different orders are obtained through multiple attention layers;

[0016] Inputting the item representation attention score and the attribute representation attention score into an aggregation function for aggregation to obtain a decoupled representation of the attention matrix;

[0017] Based on the decoupled representation of the attention matrix, the above process is repeated until the item representation no longer changes, and an updated item representation is obtained.

[0018] As an implementation method, inputting the updated item representation into an auxiliary attribute predictor for prediction to obtain an optimized prediction result of the user-item interaction includes the following steps:

[0019] Inputting the updated item representation into the auxiliary attribute predictor, and predicting the item and each auxiliary information corresponding to the updated item representation based on the auxiliary attribute predictor to obtain a predicted value for the item and a predicted value for each auxiliary information respectively;

[0020] Use the cross entropy function to calculate the item loss between the item prediction value and the true value and the information loss between the auxiliary information preset value and the true value respectively;

[0021] By jointly training the item loss and information loss, an optimized prediction result of the interaction between the user and the item is obtained.

[0022] A cross-domain sequential recommendation system based on privacy computing and decoupled information fusion, including a data acquisition module, an embedding module, a decoupled auxiliary information fusion module, and a prediction module;

[0023] The data acquisition module acquires data from the data demander and encrypted auxiliary information obtained based on homomorphic encryption, wherein the auxiliary information comes from various participants in different fields;

[0024] The embedding module inputs the data demander's data into the project embedding layer to obtain project embedding, and inputs the encrypted auxiliary information into the attribute embedding layer to obtain auxiliary information embedding;

[0025] The decoupled auxiliary information fusion module performs decoupled auxiliary information fusion on the project embedding and the auxiliary information embedding to obtain an updated project representation;

[0026] The prediction module inputs the updated item representation into the auxiliary attribute predictor for prediction, thereby obtaining an optimized prediction result of the interaction between the user and the item.

[0027] As an implementable embodiment, the data acquisition module is configured to:

[0028] Generate a public key and a private key based on a key generation function, and send the public key to each participant;

[0029] Based on a given public key, the auxiliary information corresponding to each participant is encrypted to obtain the encrypted auxiliary information and send it to the data demander. The auxiliary information includes user information, project information, and behavioral attribute information that provide additional information for prediction.

[0030] As an implementable method, the decoupling auxiliary information fusion module is configured as follows:

[0031] The item embedding and auxiliary information embedding are input into the attention layer to calculate the attention scores, and the item representation attention score and the attribute representation attention score are obtained respectively. The attention layer includes multiple layers, and the relationships between different orders are obtained through multiple attention layers;

[0032] Inputting the item representation attention score and the attribute representation attention score into an aggregation function for aggregation to obtain a decoupled representation of the attention matrix;

[0033] Based on the decoupled representation of the attention matrix, the above process is repeated until the item representation no longer changes, and an updated item representation is obtained.

[0034] As an implementation method, the prediction module is configured as follows:

[0035] Inputting the updated item representation into the auxiliary attribute predictor, and predicting the item and each auxiliary information corresponding to the updated item representation based on the auxiliary attribute predictor to obtain a predicted value for the item and a predicted value for each auxiliary information respectively;

[0036] Use the cross entropy function to calculate the item loss between the item prediction value and the true value and the information loss between the auxiliary information preset value and the true value respectively;

[0037] By jointly training the item loss and information loss, an optimized prediction result of the interaction between the user and the item is obtained.

[0038] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following method:

[0039] Obtain data from data demanders and encrypted auxiliary information obtained based on homomorphic encryption, where the auxiliary information comes from various participants in different fields;

[0040] Inputting the data demander's data into the project embedding layer to obtain project embedding, and inputting the encrypted auxiliary information into the attribute embedding layer to obtain auxiliary information embedding;

[0041] Decoupling the item embedding and the auxiliary information embedding and fusing the auxiliary information to obtain an updated item representation;

[0042] The updated item representation is input into an auxiliary attribute predictor for prediction, thereby obtaining an optimized prediction result of the interaction between the user and the item.

[0043] A cross-domain sequence recommendation device based on privacy-preserving computing and decoupled information fusion includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, the following method is implemented:

[0044] Obtain data from data demanders and encrypted auxiliary information obtained based on homomorphic encryption, where the auxiliary information comes from various participants in different fields;

[0045] Inputting the data demander's data into the project embedding layer to obtain project embedding, and inputting the encrypted auxiliary information into the attribute embedding layer to obtain auxiliary information embedding;

[0046] Decoupling the item embedding and the auxiliary information embedding and fusing the auxiliary information to obtain an updated item representation;

[0047] The updated item representation is input into an auxiliary attribute predictor for prediction, thereby obtaining an optimized prediction result of the interaction between the user and the item.

[0048] (1) A cross-domain sequence recommendation method and system based on privacy computing and decoupled information fusion is proposed. By using the homomorphic encryption method of privacy computing, the data owner can send the data to the privacy computing participants for arbitrary processing without worrying about the leakage of the original information of the data;

[0049] (2) A cross-domain sequence recommendation method and system based on privacy computing and decoupled information fusion is proposed. By using the decoupled auxiliary information fusion method, various auxiliary information can be effectively utilized for sequence recommendation tasks to obtain higher representation capabilities, avoid mixed correlations, and use flexible training gradients to learn the relative importance of auxiliary information.

[0050] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] FIG1 is a schematic diagram of the steps of a cross-domain sequence recommendation method based on privacy computing and decoupled information fusion according to the present invention;

[0052] FIG2 is a flow chart of a cross-domain sequence recommendation method based on privacy computing and decoupled information fusion according to the present invention;

[0053] FIG3 is a schematic structural diagram of the system of the present invention. DETAILED DESCRIPTION

[0054] In order to clearly explain the present invention and make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described below in combination with the drawings in the embodiments of the present invention, so that those skilled in the art can implement them according to the text of the description.

[0055] In the context of the present invention, a trusted execution environment (TEE) provides an isolated runtime environment from the perspective of the underlying hardware and operating system, protecting the code and data running within it from external attacks, including attacks from the operating system, hardware, and other applications. This technology has been used in some fields to achieve the objectives described above, and some of the basic principles of this technology are also known to those skilled in the art. However, after reading this application, those skilled in the art will understand how to apply this technology in this context and will clearly understand the novelty of this technology when combined with other features in a specific context.

[0056] The technology of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1:

[0058] The present invention is a cross-domain sequence recommendation method based on privacy computing and decoupled information fusion. The steps of the invention are shown in Figure 1. The specific steps are as follows:

[0059] S100: Acquire data from a data demander and encrypted auxiliary information obtained based on homomorphic encryption, wherein the auxiliary information comes from various participants in different fields;

[0060] S200: Input the data demander's data into the project embedding layer to obtain project embedding, and input the encrypted auxiliary information into the attribute embedding layer to obtain auxiliary information embedding;

[0061] S300, decoupling the project embedding and the auxiliary information embedding and fusing the auxiliary information to obtain an updated project representation;

[0062] S400: Input the updated item representation into an auxiliary attribute predictor for prediction, to obtain an optimized prediction result of the interaction between the user and the item.

[0063] (1) The data demander uses homomorphic encryption to obtain encrypted auxiliary information, including the following steps:

[0064] The data requester generates a public key and a private key based on a key generation function, and sends the public key to each participant;

[0065] Based on a given public key, each participant encrypts their own auxiliary information to obtain encrypted auxiliary information and sends it to the data demander. The data demander obtains the encrypted auxiliary information, where the auxiliary information includes attributes of users, items, and behaviors that provide additional information for prediction.

[0066] After the encryption parameter λ is given, the public key pk and private key sk are generated based on the key generation function keyGen(), and the public key is sent to each participant;

[0067] Key generation function: keyGen(λ)→(pk,sk)

[0068] Based on the given public key pk, each participant encrypts their own data pt through the encryption function Encrypt() to obtain the ciphertext ct;

[0069] Encryption function: Encrypt(pt,pk)→ct

[0070] Each participant sends their own encrypted auxiliary information to the data demander.

[0071] (2) Input the data demander's data into the project embedding layer, and input the encrypted auxiliary information into the attribute embedding layer to obtain project embedding and auxiliary information embedding respectively. The flowchart is shown in FIG2 and includes the following steps:

[0072] Let I and U represent the item set and user set respectively. For user u∈U, the user’s historical interaction can be expressed as: S u =[v1,v2,…,v n ], where the term v i represents the i-th interaction in a chronological sequence. Auxiliary information can be attributes of users, items, and behaviors that provide additional information for prediction. Auxiliary information includes item-related information (e.g., brand, category) and behavior-related information (e.g., location, rating). Each interaction can be represented as: in represents the auxiliary information of the i-th interaction in the j-th type sequence, I i Indicates the item ID of the i-th interaction.

[0073] Input sequence S u =[v1,v2,…,v n ] is input into the item embedding layer and various attribute embedding layers to obtain the item embedding E ID and auxiliary information embedding E f1 ,…,Ef p： E ID =ε id ([I1,I2,…,I n ]), …

[0074] Where ε represents the corresponding embedding layer that encodes items and different item attributes into vectors. The embedding matrix can be expressed as M id ∈R |I|×d , and |·| represent the total number of corresponding different items and various auxiliary information, and d and d f1 ,…,d fp Represents the dimension of the item and various auxiliary information embedding. Then the embedding module obtains the output embedding E ID ∈R n×d ,

[0075] (3) The project embedding and the auxiliary information embedding are decoupled and fused to obtain an updated project representation, including the following steps:

[0076] The item embedding and auxiliary information embedding are input into the attention layer to calculate the attention score, and the attention scores of the item representation and attribute representation are obtained;

[0077] Given an input length n, a project hidden size d, and a multi-head query key projection size d h , we have Indicates the project represents R (ID) ∈R n×d h head (d h =d / h) of the query, key, and value projection matrices. The attention scores of the item representations are then calculated:

[0078] Generate a multi-head attention matrix for each attribute, where the attribute embedding is Note that we have df j≤ d,j∈[p] to avoid over-parameterization and reduce computational overhead. Then we have the corresponding Indicates h head (d hj =d fj / h)'s query, key, and value projection matrices: …,

[0079] In our proposed solution, all attributes generate their own disentangled attention matrices, which are then fused into the final attention matrix. The disentangled attention computation is performed by breaking the head projection size d h The ranking bottleneck of the restricted attention matrix improves the expressiveness of the model. It also avoids inflexible gradients and uncertain cross-correlations between different attributes and items to achieve reasonable and stable self-attention.

[0080] Input the attention scores of the item representation and the attribute representation into the aggregation function for aggregation to obtain a decoupled representation of the attention matrix;

[0081] Then our decoupled auxiliary information fusion attention layer aggregates all attention matrices through the aggregation function F, including addition, concatenation and gating, and obtains the output of each head as:

[0082] Finally, the outputs of all attention heads are concatenated and fed into the feed-forward layer.

[0083] Based on the decoupled representation of the attention matrix, multiple attention layers are used to repeat the above process to obtain updated item representations.

[0084] The decoupled side information fusion module contains several stacked blocks of consecutively combined decoupled side information fusion attention layers and feed-forward layers. Each decoupled side information fusion attention layer block takes two types of inputs, namely the current item representation and the auxiliary side information embedding, and then outputs the updated item representation. The auxiliary side information embedding is not updated at each layer to save computation and avoid overfitting. Let represents the input item of block i. The process can be expressed as:

[0085] Where FFN represents a fully connected feedforward network and LN represents layer normalization.

[0086] (4) Inputting the updated item representation into the auxiliary attribute predictor for prediction to obtain an optimized prediction result of the user-item interaction, including the following steps:

[0087] Predict the project and each auxiliary attribute to obtain the predicted value of the project and each auxiliary attribute;

[0088] The final representation of encoding sequence information using auxiliary information We use The last element of To estimate the probability that user u interacts with each item in the item vocabulary. The item prediction layer can be expressed as:

[0089] in represents the |i|-dimensional probability, M id ∈R |I|×d is the item embedding table in the embedding layer.

[0090] The prediction of attribute j can be expressed as:

[0091] in represents the |fj|-dimensional probability, and b fj ∈R |fj|×1 is a learnable parameter and σ is the simoid function.

[0092] Use the cross entropy function to calculate the item loss and auxiliary attribute information loss;

[0093] We use the cross entropy function to calculate the item loss L id To measure the prediction The difference between the true value y is:

[0094] We use binary cross entropy to calculate the auxiliary information loss L of the j-th type fj To support multi-label attributes:

[0095] By jointly training the item loss and the auxiliary attribute information loss, an optimized prediction result of the interaction between the user and the item is obtained.

[0096] The combined loss function with the balance parameter α can be expressed as:

[0097] In summary, the present invention provides a cross-domain sequence recommendation method and system based on privacy computing and decoupled information fusion. In this method, the data demander obtains auxiliary information from multiple parties through a homomorphic encryption method, and performs privacy computing on the auxiliary information without leaking privacy. Then, the decoupled auxiliary information fusion method is used to transfer the fusion process from the input to the attention layer, thereby enhancing the modeling capability of the attention mechanism, avoiding unnecessary attention randomness caused by mixed correlations of heterogeneous embeddings, and supporting flexible gradients to adaptively learn various auxiliary information in different scenarios. Finally, an auxiliary attribute predictor is used to complete the prediction task. The cross-domain sequence recommendation method based on privacy computing and decoupled information fusion of the present invention can effectively utilize various auxiliary information for sequence recommendation tasks, and has higher attention representation capabilities and flexibility to learn the relative importance of auxiliary information.

[0098] Example 2:

[0099] A cross-domain sequential recommendation system based on privacy computing and decoupled information fusion, as shown in FIG3 , includes a data acquisition module 100 , an embedding module 200 , a decoupled auxiliary information fusion module 300 , and a prediction module 400 ;

[0100] The data acquisition module 100 acquires data from the data demander and encrypted auxiliary information obtained based on homomorphic encryption, wherein the auxiliary information comes from various participants in different fields;

[0101] The embedding module 200 inputs the data demander's data into the project embedding layer to obtain project embedding, and inputs the encrypted auxiliary information into the attribute embedding layer to obtain auxiliary information embedding;

[0102] The decoupled auxiliary information fusion module 300 performs decoupled auxiliary information fusion on the project embedding and the auxiliary information embedding to obtain an updated project representation;

[0103] The prediction module 400 inputs the updated item representation into an auxiliary attribute predictor for prediction, thereby obtaining an optimized prediction result of the interaction between the user and the item.

[0104] Various changes and modifications can be made without departing from the spirit and scope of the present invention, and all equivalent technical solutions also fall within the scope of the present invention.

[0105] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0106] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, apparatus, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device produce a device for implementing the functions specified in one or more processes in the flowcharts and / or one or more blocks in the block diagrams.

[0108] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0109] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce computer-implemented processing, so that the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0110] It should be noted that:

[0111] References in this specification to "one embodiment" or "an embodiment" mean that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Therefore, appearances of the phrases "one embodiment" or "an embodiment" in various places throughout this specification do not necessarily refer to the same embodiment.

[0112] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention should fall within the scope of protection of the present invention.

Claims

1. A cross-domain sequential recommendation method based on privacy computing and decoupled information fusion, characterized in that It includes the following steps: Obtain the data of the data requester and the encrypted auxiliary information obtained based on homomorphic encryption, where the auxiliary information comes from various participants in different fields; Input the data of the data requester into the project embedding layer to obtain a project embedding, and input the encrypted auxiliary information into the attribute embedding layer to obtain an auxiliary information embedding; Perform decoupled auxiliary information fusion on the project embedding and the auxiliary information embedding to obtain an updated project representation; Input the updated project representation into the auxiliary attribute predictor for prediction to obtain an optimized prediction result of the user's interaction with the project.

2. The cross-domain sequence recommendation method based on privacy computing and decoupled information fusion according to claim 1, wherein The encrypted auxiliary information obtained based on homomorphic encryption includes the following steps: Generate a public key and a private key based on a key generation function, and send the public key to each participant; Based on the given public key, encrypt the auxiliary information corresponding to each participant to obtain encrypted auxiliary information and send it to the data requester, where the auxiliary information includes user information, project information, and behavioral attribute information that provide additional information for prediction.

3. The cross-domain sequence recommendation method based on privacy computing and decoupled information fusion according to claim 1, wherein The step of performing decoupled auxiliary information fusion on the project embedding and the auxiliary information embedding to obtain an updated project representation includes the following steps: Input the project embedding and the auxiliary information embedding into the attention layer to calculate attention scores, and respectively obtain the project representation attention score and the attribute representation attention score, where there are multiple attention layers, and the relationships between different orders are obtained through multiple attention layers; Input the project representation attention score and the attribute representation attention score into an aggregation function for aggregation to obtain a decoupled representation of the attention matrix; Based on the decoupled representation of the attention matrix, repeat the above process until the project representation no longer changes, then an updated project representation is obtained.

4. The cross-domain sequence recommendation method based on privacy computing and decoupled information fusion according to claim 1, characterized in that The step of inputting the updated project representation into the auxiliary attribute predictor for prediction to obtain an optimized prediction result of the user's interaction with the project includes the following steps: Input the updated project representation into the auxiliary attribute predictor, and based on the auxiliary attribute predictor, predict the project corresponding to the updated project representation and each auxiliary information, and respectively obtain a project prediction value and each auxiliary information prediction value; Use the cross-entropy function to calculate the project loss between the project prediction value and the true value and the information loss between the preset value of the auxiliary information and the true value respectively; Through joint training of the project loss and the information loss, an optimized prediction result of the user's interaction with the project is obtained.

5. A cross-domain sequential recommendation system based on privacy computing and decoupled information fusion, characterized in that, It includes a data acquisition module, an embedding module, a decoupled auxiliary information fusion module, and a prediction module; The data acquisition module obtains the data of the data requester and the encrypted auxiliary information obtained based on homomorphic encryption, where the auxiliary information comes from various participants in different fields; The embedding module inputs the data of the data requester into the project embedding layer to obtain a project embedding, and inputs the encrypted auxiliary information into the attribute embedding layer to obtain an auxiliary information embedding; The decoupled auxiliary information fusion module performs decoupled auxiliary information fusion on the project embedding and the auxiliary information embedding to obtain an updated project representation; The prediction module inputs the updated project representation into the auxiliary attribute predictor for prediction to obtain an optimized prediction result of the user's interaction with the project.

6. The cross-domain sequential recommendation system based on privacy computing and decoupled information fusion according to claim 5, wherein The data acquisition module is configured to: Generate a public key and a private key based on a key generation function, and send the public key to each participant; Encrypt the auxiliary information corresponding to each participant based on the given public key to obtain encrypted auxiliary information and send it to the data requester, where the auxiliary information includes user information, project information, and behavioral attribute information that provide additional information for prediction.

7. The cross-domain sequential recommendation system based on privacy computing and decoupled information fusion according to claim 5, wherein The decoupled auxiliary information fusion module is configured to: Input the project embedding and the auxiliary information embedding into the attention layer to calculate the attention scores, and obtain the project representation attention score and the attribute representation attention score respectively, where there are multiple attention layers, and the relationships between different orders are obtained through multiple attention layers; Input the project representation attention score and the attribute representation attention score into an aggregation function for aggregation to obtain a decoupled representation of the attention matrix; Based on the decoupled representation of the attention matrix, repeat the above process until the project representation no longer changes, then an updated project representation is obtained.

8. The cross-domain sequential recommendation system based on privacy computing and decoupled information fusion according to claim 5, wherein The prediction module is configured to: Input the updated project representation into the auxiliary attribute predictor, and based on the auxiliary attribute predictor, Predict the project corresponding to the updated project representation and each auxiliary information, and obtain the project prediction value and each auxiliary information prediction value respectively; Use the cross-entropy function to calculate the project loss between the project prediction value and the true value and the information loss between the preset value of the auxiliary information and the true value respectively; Through joint training of the project loss and the information loss, an optimized prediction result of user-project interaction is obtained.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 4.

10. A cross-domain sequential recommendation device based on privacy computing and decoupled information fusion, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 4.

Images