Federated recommendation cold start method, system and medium based on attribute representation learning

Abstract

The present invention discloses a federated recommendation cold start method, system and medium based on attribute representation learning, which relates to the field of artificial intelligence technology. User representation and attribute information of cold products are input into a trained federated recommendation model, and the user's preference probability for cold products is output. The training process of the federated recommendation model is as follows: a training data set is constructed based on the acquired hot products; in the current training round, each client inputs the user representation and local hot product representation into a scoring model for training; all clients upload the updated local hot product representation obtained in the current training round to the server, and thereby obtain the global hot product representation and universal hot product representation fed back by the server. In the next training round, the global hot product representation is used as the local hot product representation of the client. The federated recommendation cold start method, system and medium effectively solve the cold start problem in the federated recommendation scenario.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a federated recommendation cold start method, system, and medium based on attribute representation learning. Background Art

[0002] Recommender systems are an important tool for filtering redundant information from massive amounts of data and providing personalized services to users. They are crucial for finding content that truly interests users and alleviating information overload. Cold-start recommendations, a long-standing core issue in recommendation systems, require the system to effectively predict recommendations for newly added products with no historical interaction history. To address this cold-start recommendation problem, integrating attribute information from historically interacted products into the model to assist the model in understanding new products has become a mainstream technical approach. However, with the increasing emphasis on user privacy protection across society, providing high-quality recommendation services while preventing the leakage of user privacy data has become a critical issue that needs to be addressed.

[0003] Federated learning, an emerging distributed privacy-preserving framework, was first proposed by Google in 2016. Recommendation frameworks incorporating federated learning have emerged as a result. Its core mechanism is to deploy recommendation models on the client. Each client trains the model locally based on locally stored private data, uploading only the model parameters to the server for secure global aggregation before distributing them to the client. This approach effectively prevents the server from directly collecting raw user data and is an effective solution to the risk of privacy data leakage in recommendation systems. However, in actual recommendation scenarios, the challenge of quickly starting (i.e., cold starting) newly added products remains a prominent issue. Because these products have not participated in the historical federated training process, they cannot obtain the global product representation trained using the federated learning mechanism, resulting in a significant decline in the quality of recommendations for these products.

[0004] The recent boom in large language models has driven breakthroughs in deep semantic understanding technology. The mapping of natural language text content to semantic representation vector spaces is becoming increasingly mature, such as OpenAI's official text embedding model (text-embedding-3-small). These pre-trained large language models can encode product attribute text information into fixed-length representation vectors in a semantic space, significantly reducing the cost of understanding unused products and providing new insights into solving the cold-start problem in federated recommendations.

[0005] However, there is currently little research on combining the deep semantic understanding capabilities of large language models with the cold start of federated recommendation, and the cold start problem still exists in federated recommendation scenarios. Summary of the Invention

[0006] Based on the technical problems existing in the background technology, the present invention proposes a federated recommendation cold start method, system and medium based on attribute representation learning, which effectively solves the cold start problem in the federated recommendation scenario.

[0007] The proposed federated recommendation cold start method based on attribute representation learning inputs user representation and attribute information of cold products into a trained federated recommendation model and outputs the user's preference probability for cold products.

[0008] The training process of the federated recommendation model is as follows:

[0009] Building a training dataset based on the acquired hot commodities;

[0010] In the current training round, each client inputs user representation and local hot product representation into the scoring model built by the multi-layer perceptron for training;

[0011] All clients upload the updated local hot product representations obtained in the current training round to the server, which then receives feedback on the global and universal hot product representations. In the next training round, the global hot product representations are used as the client's local hot product representations, and the universal hot product representations are used as regularized constraints for local hot product representation learning.

[0012] In the current training round, the server aggregates all updated local hot product representations into a global hot product representation, and uses the hot product attribute information text to construct a universal hot product representation;

[0013] Construct a total loss function to adjust the trainable parameters in the federated recommendation model.

[0014] Furthermore, before all clients upload the updated local hot product representations obtained in the current training round to the server, each client randomly adds Laplace distribution noise to the updated local hot product representations.

[0015] Furthermore, on the server side, the hot commodity attribute information text is used to construct a universal hot commodity representation, specifically:

[0016] The attribute information text of the hot products is encoded by the pre-trained large language model to obtain the initial hot product representation;

[0017] The initial hot product representation is input into the representation alignment model constructed by the multi-layer perceptron, the mapping relationship between the initial hot product representation and the global hot product representation is learned, and the universal hot product representation is output.

[0018] Furthermore, on the server side, the initial hot product representation is input into the representation alignment model constructed by the multi-layer perceptron to learn the mapping relationship between the initial hot product representation and the global hot product representation, and output a universal hot product representation, specifically:

[0019] In the current training round, the hot product attribute information text is spliced ​​into the text to be encoded, input into the pre-trained large language model, and the initial hot product representation vector in the semantic vector space is extracted;

[0020] Input the initial hot product representation vector into the representation alignment model, use the aggregated global hot product representation as the target of the representation alignment model, learn the mapping relationship between the initial hot product representation vector and the global hot product representation, input the initial hot product representation into the representation alignment model updated after the current training round, and estimate the universal hot product representation;

[0021] In the next training round, each client uploads the updated local hot product representation, the server re-aggregates securely, and updates the global hot product representation. The initial hot product representation vector obtained by encoding and transforming the pre-trained large language model remains unchanged and continues to be input into the representation alignment model for a new round of training.

[0022] Furthermore, a round of data interaction between the client and the server is regarded as a training round. In each training round, the client and the server respectively guide the adjustment of the trainable parameters in the scoring model, user representation, local hot product representation and representation alignment model through the total loss function.

[0023] Furthermore, in the total loss function, the loss function in the client as follows:

[0024] ;

[0025] in, is the binary cross entropy loss function, is the supervision loss function, is the balance coefficient.

[0026] Furthermore, in constructing the total loss function, the server-side alignment loss function as follows:

[0027] ;

[0028] in, Indicates the number of hot products participating in the training. For the Global hot product representation of hot products, For the A universal hot commodity representation for each hot commodity.

[0029] Furthermore, the reasoning phase is performed on cold products based on the trained federated recommendation model, as follows:

[0030] On the server side, the attribute information of cold goods is spliced ​​into the text to be encoded and input into the pre-trained large language model to obtain the initial cold goods representation vector in the semantic vector space;

[0031] The initial cold product representation vector is input into the representation alignment model, and a universal cold product representation is estimated and sent to the client. In the client, the universal cold product representation and the user representation are input into the scoring model together, and the obtained prediction score is used as the user's preference probability for cold products.

[0032] The federated recommendation cold start system based on attribute representation learning inputs user representation and cold product attribute information into the trained federated recommendation model and outputs the user's preference probability for cold products.

[0033] The training process of the federated recommendation model includes a dataset construction module, a client training module, a server training module and a loss construction module;

[0034] The data set construction module is used to construct a training data set based on the acquired hot commodities;

[0035] In the client training module, in the current training round, each client inputs the user representation and local hot product representation into the scoring model constructed by the multi-layer perceptron for training; all clients upload the updated local hot product representation obtained in the current training round to the server, and obtain the global hot product representation and universal hot product representation fed back by the server. In the next training round, the global hot product representation is used as the local hot product representation of the client, and the universal hot product representation is used as the regularization constraint supervision item for the learning of the local hot product representation;

[0036] In the server-side training module, in the current training round, the server-side averagely aggregates all updated local hot product representations into a global hot product representation, and constructs a universal hot product representation using hot product attribute information text;

[0037] The loss building module is used to build a total loss function for adjusting trainable parameters in the federated recommendation model.

[0038] A computer-readable storage medium stores a plurality of classification programs, wherein the plurality of classification programs are used to be called by a processor and execute the above-mentioned federated recommendation cold start method.

[0039] Those skilled in the art will understand that all or part of the steps of implementing the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk, etc. Various media that can store program codes.

[0040] The advantages of the federated recommendation cold start method, system, and medium based on attribute representation learning provided by the present invention are as follows: while protecting user privacy, it innovatively utilizes the deep understanding ability of a pre-trained large language model on product attribute information, uses product attribute information text to generate initial hot product representations, and designs a representation alignment model to learn the mapping relationship from the initial hot product representations to the global hot product representations in the federated recommendation scenario; this mechanism can effectively give newly added cold products high-quality global product representations, thereby significantly improving the ability to predict user preferences for cold products in the federated recommendation scenario; the system supports length clipping of hot product representations generated by the pre-trained large language model, which can significantly reduce the consumption of model training resources and transmission load; in addition, by introducing a configurable noise mechanism, it can effectively balance data privacy and recommendation efficiency, allowing for trade-offs based on actual application scenarios. Therefore, the applicable scenarios of this embodiment can be improved by adjusting a small number of parameters; the mechanism using product attribute information as a unified semantic bridge can effectively connect the representation learning process of hot and cold products. This idea can be well integrated into other federated recommendation frameworks and is easy to promote on a large scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the process of the present invention;

[0042] Figure 2 Schematic diagram of the training phase of the federated recommendation model;

[0043] Figure 3 Schematic diagram of the inference phase of the federated recommendation model. DETAILED DESCRIPTION

[0044] The technical solutions of the present invention are described in detail below through specific embodiments. Numerous specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0045] like Figures 1 to 3 As shown in the figure, the federated recommendation cold start method based on attribute representation learning proposed in the present invention inputs user representation and cold products into the trained federated recommendation model and outputs the user's preference probability for cold products;

[0046] like Figure 2 As shown in FIG, the training process of the federated recommendation model is as follows:

[0047] Step 1: Build a training dataset based on the acquired hot products;

[0048] Step 2: In the current training round, each client inputs the user representation and local hot product representation into the scoring model built by the multi-layer perceptron for training;

[0049] Step 3: All clients upload the updated local hot product representations obtained in the current training round to the server, and then obtain the global hot product representations and universal hot product representations fed back by the server. In the next training round, the global hot product representations are used as the local hot product representations of the clients, and the universal hot product representations are used as the regularization constraint supervision item for the learning of local hot product representations.

[0050] Step 4: In the current training round, the server averages and aggregates all updated local hot product representations into a global hot product representation, and constructs a universal hot product representation using the hot product attribute information text;

[0051] Step 4: Construct a total loss function to adjust the trainable parameters in the federated recommendation model.

[0052] This example designs a representation alignment model to learn the mapping relationship between the initial hot product representations and the global hot product representations in the federated recommendation scenario. This mechanism effectively assigns high-quality global product representations to newly added cold products, significantly improving the ability to predict user preferences for cold products in the federated recommendation scenario.

[0053] It should be noted that the hot products in this example refer to products with historical user interaction records in the federated recommendation system. Their global product representations are obtained by aggregating the local hot product representations uploaded and maintained by all participating clients on the server side, and they contain the preferences of all users for these products. The cold products in this example refer to products that have newly joined the federated recommendation system and have not yet had any user interaction. Because they have not participated in the historical federated training process and lack effective global product representations, the system struggles to accurately predict user preferences.

[0054] Specifically, the product attribute information and its associated functions described in this embodiment are explained as follows. Product attribute information refers to textual data describing the core characteristics of a product (e.g., product category, color, material, brand, etc.). Product attribute information is the semantic link between popular and unpopular products. If the attribute information of different products is highly semantically related, the distance between the product representations expressing their characteristics should also be close, and users' preference patterns for these products are also expected to be similar.

[0055] Example: Consider a product (Category: Lipstick, Color: Tomato Red), Product (Category: Sweater, Color: White), Product (Category: Lip Glaze, Color: Maple Red). Product and products The semantics of the usage category (lip makeup) and color attribute (high saturation) are highly similar, and its attribute correlation is much stronger than that of the product. Therefore, in the federated recommendation scenario, users’ and products The preference pattern is expected to be higher than the user's preference for the product and products (or commodity and products ) are closer. Accordingly, the product and products The distance between the global product representations in the vector space should also be closer.

[0056] In addition, the correspondence between the obtained hot products and the local hot product representations is that if hot products want to participate in training, they need a hot product representation (a vector of a certain length) to represent the hot products; therefore, the relationship between the two is that the hot products on the client participate in training, and the hot products are represented from the client's perspective.

[0057] It can be understood that a round of data interaction between the client and the server is regarded as a training round. In each training round, the client and the server respectively guide the adjustment of the trainable parameters in the scoring model, user representation, local hot product representation and representation alignment model through the total loss function.

[0058] In one embodiment, the client training process is as follows:

[0059] The client mainly sets up user representation , local hot commodity characterization and scoring models In federated learning, due to the existence of multiple clients, the user representation and rating model of each client are independent and do not affect each other. The local hot product representation is replaced by the global hot product representation sent by the server in each round of training. Respectively represent The user representation of the user and the The local hot product representation of a hot product is a representation vector used to characterize user characteristics and product characteristics respectively. In this embodiment, , Indicates the number of clients participating in training. , Indicates the number of hot products participating in the training, which is a common recommendation scenario. , vector length , randomly initialized.

[0060] Scoring Model It is composed of MLP (Multilayer Perceptron, a feedforward neural network model), which is responsible for fitting users' preferences for products.

[0061] Characterize users , local hot commodity characterization Input to the scoring model:

[0062] ;

[0063] in, Indicates that and Splicing, Represents the prediction score output by the scoring model.

[0064] The client rating model, user representation, and local hot product representation are trained by locally stored data, and their loss function It includes two parts. One part is based on whether there is an intersection between users and products, and combines the predicted scores to construct a binary cross entropy loss function. The other part is that the client's scoring model learns a unique local hot product representation for each product to characterize the product characteristics. At the same time, the server also learns a universal hot product representation through the representation alignment model, and sends the universal hot product representation to the client as an additional supervision signal to constrain the learning of local hot products. Therefore, the client constructs a supervision loss function based on the local hot product representation and the universal hot product representation learned from the server. .

[0065] Therefore, the loss function in the client as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] in, is the balance coefficient, used to control and The proportion of n represents the number of hot products participating in the training. It is A universal hot commodity representation of a hot commodity, Represents the set of interactions between users and hot products. Represents the set of whether users interact with hot products; when ( When users and Hot products have interacted, and the real sample label is 1. Similarly, when ( When users and There is no interaction with the hot products, and the real sample label is 0.

[0070] It should be noted that after the client completes a round of training, the updated local hot product representation maintained by each client will be uploaded to the server. and scoring models The data is stored locally, preventing user privacy leaks while preserving the client's ability to personalize preferences. The newly aggregated global popular product representations sent by the server serve as local popular product representations, combined with the user representations updated in the previous training round, and fed into the scoring model for the next round of training.

[0071] Finally, in order to further enhance the privacy protection capability of the system and prevent the server from inferring the client's privacy content through the received local hot product representation, the uploaded local hot product representation is integrated with local differential privacy technology to enhance privacy protection. Add noise as follows:

[0072] ;

[0073] in, For The representation obtained by adding noise is uploaded to the server. is the Laplace distributed noise, Indicates the noise intensity. High noise reduces the risk of privacy leakage, but also affects the recommendation performance. To balance the two effects, this embodiment sets .

[0074] By balancing the preferences for data privacy and recommendation efficiency and setting a reasonable noise intensity, the applicable scenarios of this embodiment can be improved by adjusting a small number of parameters.

[0075] In one embodiment, the server-side training process is as follows:

[0076] After receiving the local hot product representations uploaded by all clients, the server averages and aggregates them into a global hot product representation, which describes the general product characteristics preferred by all users. In this embodiment, the average aggregation method is used, which does not require additional computational overhead:

[0077] ;

[0078] in, Indicates uploading The number of clients that represent the local hot product of a hot product, Indicates the Hot products in Each client learns the local hot product representation, , Indicates the Global hot product representation of each hot product.

[0079] The global hot product representation is distributed back to each client to convey the overall product characteristics and is used as the initial value of the local hot product representation in the client in the next round of training.

[0080] The server uses the hot item attribute information to construct a universal hot item representation and sends it to the client as a regularized constraint for local hot item representation learning. Specifically, the hot item attribute information is encoded using a pre-trained large language model to generate an initial hot item representation. This initial hot item representation is then input into a representation alignment model built using a multi-layer perceptron to learn the mapping from the initial hot item representation to the global hot item representation, resulting in a universal hot item representation as output.

[0081] Specifically, to protect data privacy, only rich attribute information for popular and unpopular items is stored on the server. Items with similar attribute text have similar initial product representation vectors after encoding and conversion using the pre-trained large language model. This distance between initial product representation vectors can be used to establish connections between items. Using a representation alignment model, we first establish an understanding of the mapping pattern from initial hot product representations to global hot product representations in federated recommendation scenarios. This understanding is then transferred to estimate unpopular items, effectively addressing the cold start problem.

[0082] This embodiment uses a text encoding representation vector based on a pre-trained Large Language Model (LLM) as the initial hot product representation for understanding hot products; specifically, first, the hot product attribute information text is spliced ​​into the text to be encoded :

[0083] Will Input to the pre-trained large language model In the example, the encoding is converted into the initial hot product representation vector in the semantic vector space:

[0084] ;

[0085] in, To pre-train a large language model, we use the text-embedding-3-small text vectorization model released by OpenAI. It takes the input text to be encoded and outputs a representation vector after deep semantic understanding. For the initial hot product representation, the pre-trained large language model generally defaults to outputting the representation vector length .

[0086] The length of the initial hot product representation is selected based on the resource pressure and efficiency of the recommendation system training:

[0087] ;

[0088] in, The length of the initial hot commodity representation after cutting is . Indicates clipping, which is done by taking the front of the initial hot product characterization vector Position, then pass The system can customarily trim the length of the initial hot product representation to reduce training resource consumption and transmission load.

[0089] The representation alignment model on the server learns the representation mapping pattern from the pruned initial hot product representation to the global hot product representation. The server then sends the universal hot product representation learned by the representation alignment model in the current training round to the client as a supplementary supervisory signal to enhance the training of the local hot product representation on the client.

[0090] ;

[0091] in, To characterize the alignment model, characterize the alignment model It is composed of a multilayer perceptron (MLP, a feedforward neural network model), which is responsible for learning the mapping from the initial hot product representation to the global hot product representation in the recommendation scenario.

[0092] The global hot product representation and the universal hot product representation output by the representation alignment model are aligned and learned, so the alignment loss function in the server side is as follows:

[0093] ;

[0094] in, Indicates the number of hot products participating in the training. For the Global hot product representation of hot products, For the A universal hot commodity representation for each hot commodity.

[0095] This embodiment innovatively utilizes the pre-trained large language model's ability to deeply understand product attribute information, aiming to solve the problem of how to obtain accurate global representation of newly added cold products in the federated recommendation system, thereby achieving effective user preference prediction.

[0096] This example builds on the pre-trained large language model's understanding of the textual semantics of product attribute information to learn universal product representations suitable for recommendation scenarios. Its core approach is to use product attribute information as a bridge to connect the two, establishing semantic associations between popular products participating in federated training and newly added cold products, effectively addressing the product cold start problem in privacy-preserving federated recommendation scenarios.

[0097] The pre-trained large language model uses product attribute information to establish semantic associations between popular and unpopular products. Leveraging its deep semantic understanding capabilities, gained through pre-training on a massive text corpus, the pre-trained large language model encodes product attribute text into a unified semantic vector space, generating semantically rich initial product representations. In this semantic space, products with semantically similar attribute information have smaller distances between their generated initial representation vectors, while products with significant semantic differences have larger distances between their initial representation vectors.

[0098] In particular, the above-mentioned client and server are trained repeatedly until the rating model, user representation, local hot product representation on the client and the representation alignment model on the server converge, thereby obtaining a trained federated recommendation model.

[0099] In this embodiment, if Figure 3 As shown in the figure, the inference phase is performed on cold products based on the trained federated recommendation model. The server and client inferences are used to obtain the cold start product estimates, as follows:

[0100] (a1) On the server side, the attribute information text of the cold goods is concatenated into the text to be encoded and input into the pre-trained large language model to obtain the initial cold goods representation vector in the semantic vector space;

[0101] (a2) Customize and trim the initial cold product representation vector, input the trimmed initial cold product representation vector into the trained representation alignment model, estimate the universal cold product representation and send it to the client. In any client, input the universal cold product representation and the user representation in the current client into the trained scoring model, and use the obtained prediction score as the user's preference probability for cold products.

[0102] While protecting user privacy, this embodiment combines a pre-trained large language model with product attribute text information to generate initial hot product representations. Using a representation alignment model, the mapping from the initial hot product representations to global hot product representations is learned. The attribute information of cold products is encoded using the pre-trained large language model to generate initial cold product representations. This is then input into the alignment model to estimate the universal product representation, which is then sent to the client for cold product preference estimation, effectively solving the cold start problem in federated recommendation scenarios.

[0103] Moreover, this embodiment is an idea of ​​connecting hot and cold products by using product attribute information as a bridge. This idea can be well integrated into other federated recommendation frameworks and is easy to promote on a large scale.

[0104] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A federated recommendation cold start method based on attribute representation learning, characterized by: Input the user representation and the attribute information of cold goods into the trained federated recommendation model, and output the user's preference probability for cold goods. Specifically, the model includes: on the server side, concatenating the attribute information of cold goods into text to be encoded and inputting it into the pre-trained large language model to obtain an initial cold goods representation vector in the semantic vector space; inputting the initial cold goods representation vector into the representation alignment model, estimating a universal cold goods representation and sending it to the client; on the client side, inputting the universal cold goods representation and the user representation into the scoring model together, and using the obtained prediction score as the user's preference probability for the cold goods; The training process of the federated recommendation model is as follows: Building a training dataset based on the acquired hot commodities; In the current training round, each client inputs user representation and local hot product representation into the scoring model built by the multi-layer perceptron for training; All clients upload the updated local hot product representations obtained in the current training round to the server, which then receives feedback on the global and universal hot product representations. In the next training round, the global hot product representations are used as the client's local hot product representations, and the universal hot product representations are used as regularized constraints for local hot product representation learning. In the current training round, the server aggregates all updated local hot product representations into a global hot product representation, and uses the hot product attribute information text to construct a universal hot product representation; Construct a total loss function to adjust the trainable parameters in the federated recommendation model.

2. The federal recommendation cold start method according to claim 1, characterized in that: Before all clients upload the updated local hot product representations obtained in the current training round to the server, each client randomly adds Laplace distribution noise to the updated local hot product representations.

3. The federal recommendation cold start method according to claim 1, characterized in that: On the server side, the method of constructing a universal hot commodity representation using the hot commodity attribute information text is as follows: The attribute information text of the hot products is encoded by the pre-trained large language model to obtain the initial hot product representation; The initial hot product representation is input into the representation alignment model constructed by the multi-layer perceptron, the mapping relationship between the initial hot product representation and the global hot product representation is learned, and the universal hot product representation is output.

4. The federal recommendation cold start method according to claim 3, characterized in that: On the server side, the initial hot product representation is input into the representation alignment model constructed by the multi-layer perceptron, and the mapping relationship between the initial hot product representation and the global hot product representation is learned, and the universal hot product representation is output, which is specifically: In the current training round, the hot product attribute information text is spliced ​​into the text to be encoded, input into the pre-trained large language model, and the initial hot product representation vector in the semantic vector space is extracted; Input the initial hot product representation vector into the representation alignment model, use the aggregated global hot product representation as the target of the representation alignment model, learn the mapping relationship between the initial hot product representation vector and the global hot product representation, input the initial hot product representation into the representation alignment model updated after the current training round, and estimate the universal hot product representation; In the next training round, each client uploads the updated local hot product representation, the server re-aggregates securely, and updates the global hot product representation. The initial hot product representation vector obtained by encoding and transforming the pre-trained large language model remains unchanged and continues to be input into the representation alignment model for a new round of training.

5. The federal recommendation cold start method according to claim 1, characterized in that: A round of data interaction between the client and the server is considered as a training round. In each training round, the client and the server respectively use the total loss function to guide the adjustment of the trainable parameters in the scoring model, user representation, local hot product representation, and representation alignment model.

6. The federal recommendation cold start method according to claim 1, characterized in that: In the total loss function, the loss function in the client as follows: ； in, is the binary cross entropy loss function, is the supervision loss function, is the balance coefficient.

7. The federal recommendation cold start method according to claim 1, characterized in that: In constructing the total loss function, the server-side alignment loss function as follows: ； in, Indicates the number of hot products participating in the training. For the Global hot product representation of hot products, For the A universal hot commodity representation for each hot commodity.

8. A federated recommendation cold start system based on attribute representation learning, characterized by: Input the user representation and the attribute information of cold goods into the trained federated recommendation model, and output the user's preference probability for cold goods. Specifically, the model includes: on the server side, concatenating the attribute information of cold goods into text to be encoded and inputting it into the pre-trained large language model to obtain an initial cold goods representation vector in the semantic vector space; inputting the initial cold goods representation vector into the representation alignment model, estimating a universal cold goods representation and sending it to the client; on the client side, inputting the universal cold goods representation and the user representation into the scoring model together, and using the obtained prediction score as the user's preference probability for the cold goods; The training process of the federated recommendation model includes a dataset construction module, a client training module, a server training module and a loss construction module; The data set construction module is used to construct a training data set based on the acquired hot commodities; In the client training module, in the current training round, each client inputs the user representation and local hot product representation into the scoring model constructed by the multi-layer perceptron for training; all clients upload the updated local hot product representation obtained in the current training round to the server, and obtain the global hot product representation and universal hot product representation fed back by the server. In the next training round, the global hot product representation is used as the local hot product representation of the client, and the universal hot product representation is used as the regularization constraint supervision item for the learning of the local hot product representation; In the server-side training module, in the current training round, the server-side averagely aggregates all updated local hot product representations into a global hot product representation, and constructs a universal hot product representation using hot product attribute information text; The loss building module is used to build a total loss function for adjusting trainable parameters in the federated recommendation model.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of classification programs, which are used to be called by a processor and executed into the federal recommendation cold start method according to any one of claims 1 to 7.

Images