Cross-domain recommendation method and system based on federated learning and self-distillation

Abstract

The invention discloses a cross-domain recommendation method and system based on federated learning and self-distillation, and the method comprises the steps: carrying out the multi-dimensional semantic analysis and enhancement of an article in each participation domain, and generating an article semantic enhancement representation fused with world knowledge; and jointly training a global representation mapping function through a federated learning framework, and mapping semantic representation of each domain to a unified collaborative representation space. And carrying out sequence modeling in each domain based on a user historical interaction sequence to generate a user representation. Meanwhile, a teacher-student matching relationship is constructed based on semantic similarity by identifying long-tail and hot articles, and knowledge of the hot articles is migrated to the long-tail articles by using a self-distillation mechanism. And finally, each domain independently constructs a recommendation model for matching scoring and sorting based on the user representation and the optimized article representation, and a personalized recommendation list is generated. Safe migration of cross-domain knowledge and effective relieving of intra-domain data imbalance are achieved, the recommendation effect is improved, and the application range is widened.

Description

Technical Field

[0001] This invention relates to the field of cross-domain recommendation technology, and in particular to a cross-domain recommendation method and system based on federated learning and self-distillation. Background Technology

[0002] Cross-domain recommendation systems alleviate data sparsity by leveraging knowledge transfer and collaboration across different domains, playing a crucial role in multi-domain scenarios such as e-commerce and content platforms. Their core objective is to achieve effective transfer and sharing of cross-domain knowledge. However, traditional cross-domain recommendation methods face three key challenges: First, insufficient semantic understanding depth. Existing methods often employ simple ID embedding or basic text encoding, lacking a deep understanding of multi-dimensional semantic information about items and heavily relying on overlapping users or items for inter-domain knowledge bridging, limiting the applicability and transfer effectiveness of cross-domain recommendations. Second, serious cross-domain information leakage. Existing learning methods require uploading parameter information containing user behavior patterns, posing a potential risk of indirect privacy leakage through gradient attacks; or constructing a central code table containing item semantic information, posing a serious risk of direct leakage of core trade secrets. Third, uneven data distribution within the domain. Existing methods lack effective semantic guidance mechanisms when dealing with uneven data distribution within the domain, resulting in long-tail items struggling to learn effective representations due to sparse interactions, while popular items face the risk of overfitting, severely impacting overall recommendation performance.

[0003] Therefore, this invention proposes a cross-domain recommendation method and system based on federated learning and self-distillation, thereby overcoming the above-mentioned defects and achieving efficient, safe, and balanced cross-domain recommendation. Summary of the Invention

[0004] This invention overcomes the shortcomings of existing technologies and provides a cross-domain recommendation method and system based on federated learning and self-distillation. Its main objective is to achieve efficient, secure, and balanced cross-domain recommendation.

[0005] To achieve the above objectives, the first aspect of this invention provides a cross-domain recommendation method based on federated learning and self-distillation, comprising: For each participating recommendation domain, multi-dimensional semantic parsing is performed using a large language model based on the basic attribute information and text description information of the item. The semantics of the item are then enhanced according to the results of the semantic parsing, generating a set of semantically enhanced representations of the item that integrates world knowledge. Each participating recommendation domain jointly trains a global representation mapping function based on the set of item semantic enhancement representations using a federated learning framework. By sharing parameter gradients, the semantic enhancement representations of items in each domain are mapped to a shared collaborative representation space, generating a cross-domain aligned set of item collaborative recommendation representations. Within each participating recommendation domain, based on the collaborative recommendation representation of items corresponding to the user's historical interaction sequence, sequence modeling of user interests is performed to generate user representations that reflect the user's dynamic interest preferences. Each participating recommendation domain independently identifies local long-tail items and popular items, and matches at least one semantically similar popular item as a teacher for each long-tail item based on the semantically enhanced representation of the items. The representational knowledge of the popular items is transferred to the long-tail items through a self-distillation learning mechanism to obtain an optimized set of collaborative item recommendation representations. Each participating recommendation domain independently constructs a recommendation prediction model based on user representations that reflect users' dynamic interests and preferences and optimized collaborative item recommendation representations. A personalized recommendation list is generated by calculating the matching score between users and items.

[0006] In this scheme, for each participating recommendation domain, multi-dimensional semantic parsing is performed using a large language model based on the item's basic attribute information and text description information. Based on the semantic parsing results, item semantic enhancement is performed to generate a set of item semantic enhancement representations that integrates world knowledge. Specifically, this includes: Within each participating recommendation domain, the basic attribute information and text description information corresponding to the item to be processed are obtained from the local item information database, and the original item description text is generated. The basic attribute information includes category information, function information, and specification information, and the text description information includes item title information and text description information. The pre-configured large language model application interface is invoked, and the model is guided to perform multi-dimensional semantic parsing on the original item description text through preset prompt engineering instructions, generating structured discrete semantic features with item ID, dimension type, feature label and confidence score as four-tuples; The structured discrete semantic features and the original item description text are input into a pre-trained RoBERTa text encoder. The Transformer network structure is used to model each word and its complex relationships in the input text, and finally, an enhanced semantic representation of the item that integrates world knowledge is generated. Each participating domain will associate and bind the generated structured discrete semantic features with the enhanced semantic representation of the item through the item ID, and persistently store them in the locally built semantic representation library.

[0007] In this scheme, each participating recommendation domain jointly trains a global representation mapping function based on the set of item semantic enhancement representations using a federated learning framework. By sharing parameter gradients, the semantic enhancement representations of items from each domain are mapped to a shared collaborative representation space, generating a cross-domain aligned set of item collaborative recommendation representations. Specifically, this includes: In the local mapping learning submodule of each participating recommendation domain, a multilayer perceptron network with the same structure and consistent with the dimension of the item semantic enhancement representation is initialized as the local mapping model, and the local set of item semantic enhancement representations is used as input for model training. During the training of the local mapping model, each participating recommendation domain uses items that the user has interacted with as positive samples and selects uninteracted items as negative samples to build a training sample set through a negative sampling strategy. Each participating recommendation domain independently executes the forward propagation process of its local model based on its corresponding training sample set to calculate the mapped collaborative recommendation representation of items. The reconstruction loss and downstream task association loss are defined as training objectives, and the parameter gradient of the current local mapping model is calculated through the backpropagation algorithm. After completing the local model parameter gradient calculation, each participating recommendation domain only uploads the calculated model parameter gradient to the parameter aggregation submodule in the federated learning framework, and does not upload any original item semantic enhancement representation data or user behavior data. After collecting the gradients of all local model parameters uploaded by participating recommendation domains, the parameter aggregation submodule uses the federated averaging algorithm to perform a weighted average operation on the collected local model parameter gradients, and obtains the global model parameter gradient update amount that integrates the knowledge distribution of all participating recommendation domains. The parameter aggregation submodule sends the aggregated global model parameter gradient update back to each participating recommendation domain. After receiving the global model parameter gradient update, each participating recommendation domain updates the model parameters of its local mapping model, completing one round of joint training. Through repeated iterative joint training, when the preset convergence condition is met, the parameter aggregation submodule outputs a global representation mapping function and distributes it to each participating recommendation domain. Each participating recommendation domain maps the semantic enhancement representations of all local items to a collaborative representation space shared by all participating domains based on the global representation mapping function, generating a cross-domain aligned collaborative recommendation representation set of items.

[0008] In this scheme, within each participating recommendation domain, based on the collaborative recommendation representation of items corresponding to the user's historical interaction sequence, sequence modeling of user interests is performed to generate user representations reflecting the user's dynamic interest preferences. Specifically, this includes: Within each participating recommendation domain, the target user's historical interaction item ID sequence sorted by timestamp is extracted from the local user interaction database. Based on the interaction item ID identifier, the item collaborative recommendation representation corresponding to each interaction item is retrieved from the cross-domain aligned item collaborative recommendation representation set, and an item representation sequence is generated. Based on the item representation sequence, user interests are sequence modeled. For each position in the sequence, a unique position code is generated using sine and cosine functions. The item collaborative recommendation representation at each position is added element by element to the corresponding position code vector to obtain an enhanced representation sequence that integrates item semantic information and temporal information. The augmented sequence representation sequence is input into a Transformer sequence encoder consisting of multiple stacked coding layers. In each coding layer, the input augmented sequence representation sequence is transformed into a query vector, a key vector, and a value vector through linear projection. The dot product of the query vector and all key vectors is calculated and scaled, and the attention weights are obtained through the Softmax function. The value vector is weighted and summed using the obtained attention weights to obtain the output of each attention head. The outputs of all attention heads are concatenated and linearly transformed, and then nonlinear enhancement is performed using a feedforward neural network to output the sequence context representation of the current layer. Through iterative interaction and nonlinear transformation of multiple coding layers, a contextualized representation that integrates the user's full sequence context information is obtained. The attention pooling mechanism is used to aggregate all sequence context representations, and finally, a user representation that reflects the user's dynamic interests and preferences is output.

[0009] In this scheme, each participating recommendation domain independently identifies local long-tail items and popular items. Based on the semantically enhanced representation of items, at least one semantically similar popular item is matched as a teacher for each long-tail item. The representational knowledge of popular items is transferred to long-tail items through a self-distillation learning mechanism to obtain an optimized set of collaborative item recommendation representations, specifically including: Each participating recommendation domain calculates the cumulative interaction frequency of each item based on historical interaction records in the local user interaction database, and divides the items into a long-tail item set and a popular item set according to a preset frequency threshold. Using the semantic augmentation representations of items stored in the local semantic representation library, the cosine similarity between each long-tail item and all popular items in the semantic augmentation representation space is calculated based on the cosine metric algorithm. Based on the calculated cosine similarity value, the top-K most popular items with the highest similarity are selected for each long-tail item, and a semantic similarity matching relationship mapping table is constructed with long-tail items as students and popular items as teachers. A self-distillation learning framework is constructed and a dual loss function is set, which includes recommendation task loss and teacher knowledge distillation loss. The teacher knowledge distillation loss is defined by the mean square error between the mean of the current item collaborative recommendation representation of the long-tail item and the mean of the item collaborative recommendation representation of the matching teacher item. During the local model training process, the gradient descent algorithm is used to jointly optimize the dual loss function and perform multiple rounds of iterative training. This allows the collaborative recommendation representation of long-tail items to continuously absorb collaborative signals from popular teacher items with similar semantics while maintaining the original semantic characteristics. When the preset convergence condition is met, the self-distillation learning process terminates, and each domain integrates the latest collaborative recommendation representations of the original popular items and the optimized long-tail items, and outputs the optimized item collaborative recommendation representation set. Establish a dynamic update mechanism. When new user interaction data appears in the recommendation domain, re-identify long-tail items and popular items based on the new user interaction data, update the teacher-student relationship mapping table, and perform knowledge self-distillation again.

[0010] In this scheme, each participating recommendation domain independently constructs a recommendation prediction model based on user representations reflecting users' dynamic interests and preferences and optimized item collaborative recommendation representations. A personalized recommendation list is generated by calculating the matching score between users and items, specifically including: Each participating recommendation domain will perform feature fusion processing on user representations reflecting users' dynamic interests and preferences and optimized item collaborative recommendation representations. The user representation vector and item representation vector will be combined into a user-item interaction feature vector through vector concatenation operations. Based on user-item interaction feature vectors, recommendation prediction models based on the Wide & Deep architecture are independently constructed for each domain. The Wide part uses a linear model to process and remember the sparse cross features between users and items, while the Deep part uses a deep neural network to perform deep nonlinear transformations and abstractions to capture the complex implicit interaction patterns between users and items. During the model training phase, each participating recommendation domain uses historical behavior data from the local user interaction database as a supervision signal. The model calculates the interaction prediction probability of user-item pairs through forward propagation, and the parameters of the recommendation prediction model are iteratively updated by the backpropagation algorithm and gradient descent optimizer until the model reaches a convergent state. During the model inference phase, the interaction prediction probability between the target user's user representation and the optimized collaborative recommendation representation of all items in the candidate item set is calculated as a matching score to analyze the user's preference for each candidate item. Finally, based on all the calculated matching scores, the candidate items are sorted in descending order and N items with scores greater than the filtering threshold are selected, and a personalized Top-N recommendation list for the target user is output.

[0011] The second aspect of the present invention provides a cross-domain recommendation system based on federated learning and self-distillation, for implementing the cross-domain recommendation method based on federated learning and self-distillation as described in any of the above claims. The system includes: a basic data module, a large model semantic enhancement module, a federated mapping learning module, a self-distillation representation learning module, and a recommendation service module. The basic data module includes a user interaction database, an item information database, and a semantic representation database, which are used to store, manage, and provide data locally in each participating recommendation domain; The large model semantic enhancement module includes a large model semantic parsing module and a semantic encoding sub-module, which are used to perform deep and multi-dimensional semantic understanding and representation enhancement of items in each participating recommendation domain. The federated mapping learning module includes a mapping learning submodule and a parameter aggregation submodule, which are used to coordinate the joint learning of cross-domain semantic alignment mapping functions by participating recommendation domains under the premise of strictly protecting the data privacy of each domain; The self-distillation representation learning module includes a long-tail item identification submodule, a teacher selection submodule, and a distillation optimization submodule, which are used to solve the problem of uneven data distribution within each participating recommendation domain and improve the representation quality of long-tail items through knowledge transfer. The recommendation service module includes a user modeling submodule, a feature fusion submodule, and a recommendation prediction submodule, which are used to complete the final personalized recommendation task within each participating recommendation domain.

[0012] This invention provides a cross-domain recommendation method and system based on federated learning and self-distillation, which effectively incorporates world knowledge to achieve multi-dimensional deep semantic understanding of items, supports accurate cross-domain knowledge transfer, and enables efficient cross-domain collaborative recommendation under strict privacy protection. Simultaneously, it achieves knowledge transfer from popular items to long-tail items based on semantic similarity, effectively solving the problem of uneven data distribution within the domain, thus adapting to the practical needs of cross-company and cross-industry recommendations. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0014] Figure 1 A flowchart of a cross-domain recommendation method based on federated learning and self-distillation is provided for one embodiment of the present invention; Figure 2 A block diagram of a cross-domain recommendation system based on federated learning and self-distillation is provided as an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0017] Figure 1 A flowchart of a cross-domain recommendation method based on federated learning and self-distillation is provided for one embodiment of the present invention; like Figure 1 As shown, this invention provides a flowchart of a cross-domain recommendation method based on federated learning and self-distillation, including: S102, for each participating recommendation domain, multi-dimensional semantic parsing is performed using a large language model based on the basic attribute information and text description information of the item. The semantics of the item are enhanced according to the semantic parsing results, and a set of semantically enhanced representations of the item that integrates world knowledge is generated. S104, Each participating recommendation domain jointly trains a global representation mapping function based on the set of item semantic enhancement representations using a federated learning framework. By sharing parameter gradients, the semantic enhancement representations of items in each domain are mapped to a shared collaborative representation space, generating a cross-domain aligned set of item collaborative recommendation representations. S106, Within each participating recommendation domain, based on the item collaborative recommendation representation corresponding to the user's historical interaction sequence, sequence modeling of user interests is performed to generate user representations that reflect the user's dynamic interest preferences. S108, each participating recommendation domain independently identifies local long-tail items and popular items, and matches at least one semantically similar popular item as a teacher for each long-tail item based on the semantically enhanced representation of the items. The representational knowledge of the popular items is transferred to the long-tail items through a self-distillation learning mechanism to obtain an optimized set of collaborative item recommendation representations. S110, each participating recommendation domain independently constructs a recommendation prediction model based on user representations reflecting users' dynamic interests and preferences and optimized item collaborative recommendation representations, and generates a personalized recommendation list by calculating the matching score between users and items.

[0018] Furthermore, in a preferred embodiment of the present invention, for each participating recommendation domain, multi-dimensional semantic parsing is performed using a large language model based on the basic attribute information and textual description information of the item. Based on the semantic parsing results, item semantic enhancement is performed to generate a set of item semantic enhancement representations that integrates world knowledge. Specifically, this includes: Within each participating recommendation domain, the basic attribute information and text description information corresponding to the item to be processed are obtained from the local item information database, and the original item description text is generated. The basic attribute information includes category information, function information, and specification information, and the text description information includes item title information and text description information. The pre-configured large language model application interface is invoked, and the model is guided to perform multi-dimensional semantic parsing on the original item description text through preset prompt engineering instructions, generating structured discrete semantic features with item ID, dimension type, feature label and confidence score as four-tuples; The structured discrete semantic features and the original item description text are input into a pre-trained RoBERTa text encoder. The Transformer network structure is used to model each word and its complex relationships in the input text, and finally, an enhanced semantic representation of the item that integrates world knowledge is generated. Each participating domain will associate and bind the generated structured discrete semantic features with the enhanced semantic representation of the item through the item ID, and persistently store them in the locally built semantic representation library.

[0019] It should be noted that this method aims to address the lack of deep semantic understanding in item representations within traditional recommendation systems. By fully leveraging the rich world knowledge inherent in large language models, it performs multi-dimensional deep semantic parsing and enhancement on items, generating high-quality semantic representations that integrate world knowledge. This method overcomes the limitations of traditional item representations based on IDs and simple features, providing a precise semantic alignment foundation for cross-domain knowledge transfer and significantly improving the recommendation system's ability to understand and represent items. First, basic attributes (such as "Brand: XX" and "Memory: 8GB" for electronic products) and text descriptions (such as product titles and details) are extracted from the local item information database and combined into a complete natural language description text. Then, the large language model interface is invoked through a pre-set prompt word template. For example, inputting "Please analyze the following product description from four dimensions: functional attributes, application scenarios, cultural background, and industry knowledge:..." will output structured semantic features. For example, for "a certain brand of electric toothbrush," discrete features such as "functional attributes: sonic vibration, wireless charging" and "application scenarios: daily home care, travel carrying" might be analyzed. Next, these discrete features are appended to the original description text in natural language form to form enhanced text. This text is then input into a pre-trained RoBERTa encoder, which uses a multi-layer self-attention mechanism of Transformer to deeply model the word relationships in the text, ultimately outputting a high-dimensional semantic vector normalized by L2. Finally, each participating domain associates the generated discrete features with the continuous vector through the item ID and stores them in a local semantic representation database. For example, for an electric toothbrush with item ID "P123", an index is established to associate its semantic feature quadruple with a 256-dimensional semantic vector, providing data support for subsequent processing.

[0020] Furthermore, in a preferred embodiment of the present invention, each participating recommendation domain jointly trains a global representation mapping function based on the set of item semantic enhancement representations using a federated learning framework. By sharing parameter gradients, the semantic enhancement representations of items from each domain are mapped to a shared collaborative representation space, generating a cross-domain aligned set of item collaborative recommendation representations. Specifically, this includes: In the local mapping learning submodule of each participating recommendation domain, a multilayer perceptron network with the same structure and consistent with the dimension of the item semantic enhancement representation is initialized as the local mapping model, and the local set of item semantic enhancement representations is used as input for model training. During the training of the local mapping model, each participating recommendation domain uses items that the user has interacted with as positive samples and selects uninteracted items as negative samples to build a training sample set through a negative sampling strategy. Each participating recommendation domain independently executes the forward propagation process of its local model based on its corresponding training sample set to calculate the mapped collaborative recommendation representation of items. The reconstruction loss and downstream task association loss are defined as training objectives, and the parameter gradient of the current local mapping model is calculated through the backpropagation algorithm. After completing the local model parameter gradient calculation, each participating recommendation domain only uploads the calculated model parameter gradient to the parameter aggregation submodule in the federated learning framework, and does not upload any original item semantic enhancement representation data or user behavior data. After collecting the gradients of all local model parameters uploaded by participating recommendation domains, the parameter aggregation submodule uses the federated averaging algorithm to perform a weighted average operation on the collected local model parameter gradients, and obtains the global model parameter gradient update amount that integrates the knowledge distribution of all participating recommendation domains. The parameter aggregation submodule sends the aggregated global model parameter gradient update back to each participating recommendation domain. After receiving the global model parameter gradient update, each participating recommendation domain updates the model parameters of its local mapping model, completing one round of joint training. Through repeated iterative joint training, when the preset convergence condition is met, the parameter aggregation submodule outputs a global representation mapping function and distributes it to each participating recommendation domain. Each participating recommendation domain maps the semantic enhancement representations of all local items to a collaborative representation space shared by all participating domains based on the global representation mapping function, generating a cross-domain aligned collaborative recommendation representation set of items.

[0021] It should be noted that this method aims to achieve efficient cross-domain collaborative recommendation under strict privacy protection. By constructing a federated learning cross-domain representation mapping architecture and designing a semantic mapping mechanism, each domain only shares mapping parameters without involving user behavior information. This fully leverages the semantic enhancement capabilities of large models to achieve cross-domain knowledge transfer while strictly protecting user privacy. This method effectively addresses the dual challenges of data silos and privacy leaks in cross-domain recommendation, significantly improving the performance and practicality of cross-domain recommendation.

[0022] Each participating domain first initializes the same multilayer perceptron network locally as a mapping model. The input dimension of this network is aligned with the semantic augmentation representation of the items, while the output dimension corresponds to the shared collaborative representation space. For example, when the semantic augmentation representation is 256-dimensional, the mapping network can be designed as a three-layer structure with 256 nodes in the input layer, 128 nodes in the hidden layer, and 64 nodes in the output layer, ensuring that the output representations of each domain have a unified dimension. During training, each domain constructs positive and negative sample pairs based on local user behavior. For example, items clicked by the user are used as positive samples, and unclicked items are randomly sampled as negative samples to form a supervision signal. The mapped collaborative representation is calculated through forward propagation, and the reconstruction loss and task relevance loss are jointly optimized. The gradient of the local model parameters is calculated using backpropagation.

[0023] To protect data privacy, each domain only uploads the calculated parameter gradients to the parameter aggregation submodule, strictly avoiding the transmission of raw semantic data and user behavior information. The parameter aggregation submodule employs a federated averaging algorithm, weighting the uploaded gradients based on the data volume of each domain to obtain a global gradient update that integrates multi-domain knowledge distribution. This update is then securely distributed to each domain to update its local mapping model parameters. Through multiple rounds of iterative "local computation-gradient upload-aggregation distribution," the mapping models of each domain gradually converge to a consistent state, resulting in a globally shared representation mapping function. Finally, each domain applies this function to map its local item semantic enhancement representations to a unified collaborative representation space, generating cross-domain aligned item collaborative recommendation representations. For example, "basketball shoes" from the e-commerce domain and "sports instruction videos" from the content domain, although belonging to different domains, can be positioned close in the same representation space through this mapping, thus providing a comparable semantic basis for subsequent cross-domain recommendations. This process achieves effective cross-domain knowledge transfer and deep alignment of the semantic space without sharing raw data.

[0024] Furthermore, in a preferred embodiment of the present invention, the step of performing sequence modeling of user interests based on the item collaborative recommendation representation corresponding to the user's historical interaction sequence within each participating recommendation domain, and generating a user representation reflecting the user's dynamic interest preferences, specifically includes: Within each participating recommendation domain, the target user's historical interaction item ID sequence sorted by timestamp is extracted from the local user interaction database. Based on the interaction item ID identifier, the item collaborative recommendation representation corresponding to each interaction item is retrieved from the cross-domain aligned item collaborative recommendation representation set, and an item representation sequence is generated. Based on the item representation sequence, user interests are sequence modeled. For each position in the sequence, a unique position code is generated using sine and cosine functions. The item collaborative recommendation representation at each position is added element by element to the corresponding position code vector to obtain an enhanced representation sequence that integrates item semantic information and temporal information. The augmented sequence representation sequence is input into a Transformer sequence encoder consisting of multiple stacked coding layers. In each coding layer, the input augmented sequence representation sequence is transformed into a query vector, a key vector, and a value vector through linear projection. The dot product of the query vector and all key vectors is calculated and scaled, and the attention weights are obtained through the Softmax function. The value vector is weighted and summed using the obtained attention weights to obtain the output of each attention head. The outputs of all attention heads are concatenated and linearly transformed, and then nonlinear enhancement is performed using a feedforward neural network to output the sequence context representation of the current layer. Through iterative interaction and nonlinear transformation of multiple coding layers, a contextualized representation that integrates the user's full sequence context information is obtained. The attention pooling mechanism is used to aggregate all sequence context representations, and finally, a user representation that reflects the user's dynamic interests and preferences is output.

[0025] It should be noted that, firstly, the historical item interaction sequences of the target user, sorted chronologically, are extracted from the local user interaction records of each participating domain. Based on the item collaborative recommendation representation set generated by the previous cross-domain alignment, discrete item IDs are mapped to continuous semantic vectors, thus constructing an item representation sequence containing temporal information. To capture the evolution of user interests over time, sine and cosine functions are used to generate corresponding positional encoding vectors for each position in the sequence, which are then added to the item semantic vectors to form an enhanced sequence representation that simultaneously includes item semantics and interaction temporal sequence. This representation is then input into a sequence encoder based on the Transformer architecture. This encoder, through its core multi-head self-attention mechanism, dynamically calculates the association weights between different positions in the sequence, thereby modeling the dependency relationship between long-term user preferences and recent behaviors. For example, a strong association may be established between "hiking shoes" purchased early on and "camping gear" frequently viewed recently, reflecting the continuous interest theme of "outdoor activities." After layer-by-layer extraction and fusion of the sequence context through multiple Transformer blocks, the semantic information of the entire sequence is adaptively aggregated through attention pooling operations, outputting a dense vector that integrates the user's full-cycle behavioral patterns, i.e., the user representation. This representation not only summarizes the user's historical interest distribution but also reflects the dynamic changes and focusing trends of their interests, providing a deeper user profile basis for subsequent accurate matching with item representations.

[0026] Furthermore, in a preferred embodiment of the present invention, each participating recommendation domain independently identifies local long-tail items and popular items, and based on the semantically enhanced representation of items, matches at least one semantically similar popular item as a teacher for each long-tail item. Through a self-distillation learning mechanism, the representational knowledge of popular items is transferred to long-tail items to obtain an optimized set of collaborative item recommendation representations, specifically including: Each participating recommendation domain calculates the cumulative interaction frequency of each item based on historical interaction records in the local user interaction database, and divides the items into a long-tail item set and a popular item set according to a preset frequency threshold. Using the semantic augmentation representations of items stored in the local semantic representation library, the cosine similarity between each long-tail item and all popular items in the semantic augmentation representation space is calculated based on the cosine metric algorithm. Based on the calculated cosine similarity value, the top-K most popular items with the highest similarity are selected for each long-tail item, and a semantic similarity matching relationship mapping table is constructed with long-tail items as students and popular items as teachers. A self-distillation learning framework is constructed and a dual loss function is set, which includes recommendation task loss and teacher knowledge distillation loss. The teacher knowledge distillation loss is defined by the mean square error between the mean of the current item collaborative recommendation representation of the long-tail item and the mean of the item collaborative recommendation representation of the matching teacher item. During the local model training process, the gradient descent algorithm is used to jointly optimize the dual loss function and perform multiple rounds of iterative training. This allows the collaborative recommendation representation of long-tail items to continuously absorb collaborative signals from popular teacher items with similar semantics while maintaining the original semantic characteristics. When the preset convergence condition is met, the self-distillation learning process terminates, and each domain integrates the latest collaborative recommendation representations of the original popular items and the optimized long-tail items, and outputs the optimized item collaborative recommendation representation set. Establish a dynamic update mechanism. When new user interaction data appears in the recommendation domain, re-identify long-tail items and popular items based on the new user interaction data, update the teacher-student relationship mapping table, and perform knowledge self-distillation again.

[0027] It should be noted that this method specifically addresses the problem of insufficient representation learning for long-tail items, a common issue in recommender systems. By constructing a self-distillation learning mechanism based on semantically enhanced similarity, semantically related popular items are mined as teacher signals to guide the representation learning of long-tail items, achieving effective knowledge transfer from popular items to long-tail items. This method can effectively alleviate the representation learning bias caused by uneven data distribution within the domain, significantly improving the recommendation effect of long-tail items and the balance of overall recommendation performance.

[0028] First, based on the historical interaction frequency of local users, a threshold is set to distinguish between long-tail items and popular items. For example, in an e-commerce domain, "smartphones" with extremely high sales might be classified as popular items, while "handmade wooden phone stands" with unique designs but low sales might be identified as long-tail items. Then, using semantically enhanced representations of items, cosine similarity is calculated to find the most semantically similar popular item as a "teacher" for each long-tail item. For example, "handmade wooden phone stands" might have high semantic similarity to popular items like "solid wood desks" or "Nordic-style home decor paintings" in terms of material, design style, or usage scenario, thus establishing a teacher-student matching relationship. This matching is not based on co-occurrence behavior but on deep semantic understanding, ensuring the accuracy and interpretability of knowledge transfer. After establishing the matching relationship, self-distillation learning is initiated. During training, optimization for long-tail items not only relies on their own sparse user interaction signals (through recommendation task loss) but also introduces distillation loss based on matching teacher items. This loss prompts the collaborative representation of long-tail items to converge towards the mean representation of multiple semantically similar popular teacher items. Through multiple rounds of joint optimization, long-tail items absorb well-trained representation patterns and generalized knowledge from popular items, thereby effectively alleviating the problem of insufficient learning caused by sparse interaction data.

[0029] It's worth noting that a dynamic update mechanism was designed to adapt to real-time data changes. When new user interaction data is added to the domain, the system automatically recalculates item popularity, updates the division between long-tail and popular items, recalculates semantic matching relationships, and initiates a new round of self-distillation fine-tuning. This mechanism ensures that the recommendation system can continuously adapt to changes in item popularity and maintain its ability to optimize the representation of long-tail items. Ultimately, the self-distillation-optimized set of collaborative recommendation representations significantly improves the representation quality of long-tail items, laying a solid foundation for achieving more balanced and accurate personalized recommendations.

[0030] Furthermore, in a preferred embodiment of the present invention, each participating recommendation domain independently constructs a recommendation prediction model based on user representations reflecting users' dynamic interests and preferences and optimized item collaborative recommendation representations. A personalized recommendation list is generated by calculating the matching score between users and items, specifically including: Each participating recommendation domain will perform feature fusion processing on user representations reflecting users' dynamic interests and preferences and optimized item collaborative recommendation representations. The user representation vector and item representation vector will be combined into a user-item interaction feature vector through vector concatenation operations. Based on user-item interaction feature vectors, recommendation prediction models based on the Wide & Deep architecture are independently constructed for each domain. The Wide part uses a linear model to process and remember the sparse cross features between users and items, while the Deep part uses a deep neural network to perform deep nonlinear transformations and abstractions to capture the complex implicit interaction patterns between users and items. During the model training phase, each participating recommendation domain uses historical behavior data from the local user interaction database as a supervision signal. The model calculates the interaction prediction probability of user-item pairs through forward propagation, and the parameters of the recommendation prediction model are iteratively updated by the backpropagation algorithm and gradient descent optimizer until the model reaches a convergent state. During the model inference phase, the interaction prediction probability between the target user's user representation and the optimized collaborative recommendation representation of all items in the candidate item set is calculated as a matching score to analyze the user's preference for each candidate item. Finally, based on all the calculated matching scores, the candidate items are sorted in descending order and N items with scores greater than the filtering threshold are selected, and a personalized Top-N recommendation list for the target user is output.

[0031] It's important to note that to achieve accurate and effective recommendations for users, user representations and item representations are first organically combined through feature fusion. Specifically, the user representation vector reflecting dynamic user interests and preferences is concatenated with an optimized item-based collaborative recommendation representation vector to form a comprehensive user-item interaction feature vector. This concatenation operation preserves both user-specific preferences and item characteristics, laying the foundation for subsequent accurate predictions. For example, concatenating a user representation of someone who likes technology products with the representation of the latest smartphone results in a feature vector that simultaneously reflects the user's interests and the product's core attributes. Based on the fused feature vector, each domain independently constructs a Wide & Deep hybrid recommendation model. The Wide part uses a linear model to handle sparse feature interactions, such as combining user age features with item category features to effectively capture explicit association patterns like "young female users and beauty products." The Deep part uses a multi-layer neural network for deep feature transformation, automatically learning complex non-linear relationships between users and items, such as discovering implicit associations like "users who buy professional cameras may subsequently be interested in photography tutorials." This dual-path architecture ensures both the model's memory capacity and its generalization ability.

[0032] During model training, historical interaction data is used as a supervision signal to optimize model parameters by minimizing the cross-entropy loss between predicted and true values. Backpropagation is used to calculate gradients during training, combined with gradient descent for parameter updates. For example, when the model's predicted probability of a user's purchased items is too low, gradient backpropagation is used to adjust model parameters to more accurately fit the user's true preferences. In the inference phase, the matching score between the target user and all candidate items is calculated. To improve computational efficiency, optimization algorithms such as approximate nearest neighbor search are often used in actual deployments to quickly filter the most relevant items from a massive pool of candidates. Finally, a personalized recommendation list is generated based on the matching score. Notably, a dynamic update mechanism is established. When a user generates new interactions, the system updates the user representation and recalculates the recommendation results in real time. For example, when a user recently favorites a product, the system immediately adjusts the recommendation strategy to ensure that the recommendation results always reflect the user's latest interests and preferences. This end-to-end processing flow guarantees the accuracy and real-time performance of the recommendation system.

[0033] Figure 2 This invention provides a cross-domain recommendation system based on federated learning and self-distillation, used to implement the cross-domain recommendation method based on federated learning and self-distillation described in any of the above claims. The system includes: a basic data module, a large model semantic enhancement module, a federated mapping learning module, a self-distillation representation learning module, and a recommendation service module. The basic data module includes a user interaction database, an item information database, and a semantic representation database, which are used to store, manage, and provide data locally in each participating recommendation domain; The large model semantic enhancement module includes a large model semantic parsing module and a semantic encoding sub-module, which are used to perform deep and multi-dimensional semantic understanding and representation enhancement of items in each participating recommendation domain. The federated mapping learning module includes a mapping learning submodule and a parameter aggregation submodule, which are used to coordinate the joint learning of cross-domain semantic alignment mapping functions by participating recommendation domains under the premise of strictly protecting the data privacy of each domain; The self-distillation representation learning module includes a long-tail item identification submodule, a teacher selection submodule, and a distillation optimization submodule, which are used to solve the problem of uneven data distribution within each participating recommendation domain and improve the representation quality of long-tail items through knowledge transfer. The recommendation service module includes a user modeling submodule, a feature fusion submodule, and a recommendation prediction submodule, which are used to complete the final personalized recommendation task within each participating recommendation domain.

[0034] It should be noted that the specific functions of the user interaction database, item information database, and semantic representation database included in the basic data module are as follows: The user interaction database is responsible for storing and querying real-time and historical behavior data of local users, recording interaction logs including user ID, item ID, interaction type, timestamp, and rating value, and outputting structured user historical interaction sequences and item interaction frequency statistics, providing a data foundation for subsequent user modeling and long-tail identification.

[0035] The item information database is responsible for maintaining the basic attributes and metadata information of items, storing item IDs, titles, descriptions, categories, and attributes, and providing standardized item description text to the large model semantic enhancement module as input for semantic parsing.

[0036] The semantic representation library adopts vector database technology, which is responsible for storing and indexing the structured discrete semantic features and high-dimensional continuous semantic vectors generated by the large model semantic enhancement module, establishing the mapping relationship between item ID and its multi-dimensional semantic enhancement representation, and providing efficient semantic retrieval services for downstream mapping learning and teacher selection.

[0037] It should be noted that the specific functions of the large model semantic parsing module and the semantic encoding submodule included in the large model semantic enhancement module are as follows: The large model semantic enhancement module automatically parses the item description text by calling the unified configured large language model application interface and using prompting engineering instructions. It extracts structured discrete semantic features from multiple dimensions such as functional attributes, scenario applications, cultural background and industry knowledge, and outputs them in a four-tuple format with confidence.

[0038] The semantic encoding submodule receives discrete features and original description text of items from the semantic parsing submodule, and uses a pre-trained deep text encoding model to fuse and encode multimodal semantic information into a unified, dense, and normalized high-dimensional continuous semantic vector, namely the semantic enhancement representation of items, providing a semantic foundation rich in world knowledge for cross-domain alignment.

[0039] It should be noted that the specific functions of the mapping learning submodule and the parameter aggregation submodule included in the federated mapping learning module are as follows: The mapping learning submodule is deployed locally in each participating domain. It is responsible for initializing and maintaining a local mapping model (such as a multilayer perceptron network), training it using local item semantic augmentation representations, and calculating the gradient of the model parameters.

[0040] The parameter aggregation submodule is responsible for securely collecting the gradients of the mapping model parameters uploaded from each domain, weighting and aggregating them using a federated averaging algorithm, generating a global gradient update, and distributing it to each domain. Through multiple rounds of iteration, it drives the local mapping models of each domain to converge collaboratively, thereby indirectly obtaining a globally shared global representation mapping function that maps the semantic representations of each domain to a unified collaborative space.

[0041] It should be noted that the specific functions of the long-tail item recognition submodule, teacher selection submodule, and distillation optimization submodule included in the self-distillation characterization learning module are as follows: The long-tail item identification submodule dynamically identifies and divides the long-tail item set and the popular item set based on the local interaction frequency statistics of the items and according to a preset threshold.

[0042] The teacher selection submodule calculates the semantic similarity (such as cosine similarity) between long-tail items and popular items based on the semantic representation library of items semantically enhanced representations, and matches several popular items with the most semantic similarity for each long-tail item to establish a "teacher-student" pairing relationship.

[0043] The distillation optimization submodule uses a self-distillation learning framework and matched teacher item representations as supervision signals to guide student representations of long-tail items to learn. By optimizing the joint loss, it achieves knowledge transfer from popular items to long-tail items, thereby obtaining balanced and optimized collaborative recommendation representations of items.

[0044] It should be noted that the specific functions of the user modeling submodule, feature fusion submodule, and recommendation prediction submodule included in the recommendation service module are as follows: The user modeling submodule is based on the user's historical interaction sequence and its corresponding item collaborative recommendation representation. It uses sequence modeling techniques (such as the Transformer encoder) to model the user's dynamic interest preferences and generate personalized user representation vectors.

[0045] The feature fusion submodule integrates the user representation output by the user modeling submodule with the item collaborative recommendation representation optimized by the self-distillation representation learning module to form a joint feature vector that comprehensively reflects user preferences and item characteristics.

[0046] The recommendation prediction submodule calculates the matching score or interaction probability between the user and the candidate items based on the fused joint features and a locally trained recommendation prediction model (such as the Wide & Deep model). It then sorts all candidate items according to the score and finally generates and outputs a personalized Top-N recommendation list for the target user.

[0047] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0048] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0049] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0050] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0052] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cross-domain recommendation method based on federated learning and self-distillation, characterized in that, include: For each participating recommendation domain, multi-dimensional semantic parsing is performed using a large language model based on the basic attribute information and text description information of the item. The semantics of the item are then enhanced according to the results of the semantic parsing, generating a set of semantically enhanced representations of the item that integrates world knowledge. Each participating recommendation domain jointly trains a global representation mapping function based on the set of item semantic enhancement representations using a federated learning framework. By sharing parameter gradients, the semantic enhancement representations of items in each domain are mapped to a shared collaborative representation space, generating a cross-domain aligned set of item collaborative recommendation representations. Within each participating recommendation domain, based on the collaborative recommendation representation of items corresponding to the user's historical interaction sequence, sequence modeling of user interests is performed to generate user representations that reflect the user's dynamic interest preferences. Each participating recommendation domain independently identifies local long-tail items and popular items, and matches at least one semantically similar popular item as a teacher for each long-tail item based on the semantically enhanced representation of the items. The representational knowledge of the popular items is transferred to the long-tail items through a self-distillation learning mechanism to obtain an optimized set of collaborative item recommendation representations. Each participating recommendation domain independently constructs a recommendation prediction model based on user representations that reflect users' dynamic interests and preferences and optimized collaborative item recommendation representations. A personalized recommendation list is generated by calculating the matching score between users and items.

2. The cross-domain recommendation method based on federated learning and self-distillation according to claim 1, characterized in that, For each participating recommendation domain, multi-dimensional semantic parsing is performed using a large language model based on the item's basic attribute information and text description information. Based on the semantic parsing results, item semantic enhancement is performed to generate a set of item semantic enhancement representations that integrates world knowledge. Specifically, this includes: Within each participating recommendation domain, the basic attribute information and text description information corresponding to the item to be processed are obtained from the local item information database, and the original item description text is generated. The basic attribute information includes category information, function information, and specification information, and the text description information includes item title information and text description information. The pre-configured large language model application interface is invoked, and the model is guided to perform multi-dimensional semantic parsing on the original item description text through preset prompt engineering instructions, generating structured discrete semantic features with item ID, dimension type, feature label and confidence score as four-tuples; The structured discrete semantic features and the original item description text are input into a pre-trained RoBERTa text encoder. The Transformer network structure is used to model each word and its complex relationships in the input text, and finally, an enhanced semantic representation of the item that integrates world knowledge is generated. Each participating domain will associate and bind the generated structured discrete semantic features with the enhanced semantic representation of the item through the item ID, and persistently store them in the locally built semantic representation library.

3. The cross-domain recommendation method based on federated learning and self-distillation according to claim 1, characterized in that, Each participating recommendation domain, based on its set of item semantic enhancement representations, jointly trains a global representation mapping function using a federated learning framework. By sharing parameter gradients, the semantic enhancement representations of items from each domain are mapped to a shared collaborative representation space, generating a cross-domain aligned set of collaborative item recommendation representations. Specifically, this includes: In the local mapping learning submodule of each participating recommendation domain, a multilayer perceptron network with the same structure and consistent with the dimension of the item semantic enhancement representation is initialized as the local mapping model, and the local set of item semantic enhancement representations is used as input for model training. During the training of the local mapping model, each participating recommendation domain uses items that the user has interacted with as positive samples and selects uninteracted items as negative samples to build a training sample set through a negative sampling strategy. Each participating recommendation domain independently executes the forward propagation process of its local model based on its corresponding training sample set to calculate the mapped collaborative recommendation representation of items. The reconstruction loss and downstream task association loss are defined as training objectives, and the parameter gradient of the current local mapping model is calculated through the backpropagation algorithm. After completing the local model parameter gradient calculation, each participating recommendation domain only uploads the calculated model parameter gradient to the parameter aggregation submodule in the federated learning framework, and does not upload any original item semantic enhancement representation data or user behavior data. After collecting the gradients of all local model parameters uploaded by participating recommendation domains, the parameter aggregation submodule uses the federated averaging algorithm to perform a weighted average operation on the collected local model parameter gradients, and obtains the global model parameter gradient update amount that integrates the knowledge distribution of all participating recommendation domains. The parameter aggregation submodule sends the aggregated global model parameter gradient update amount back to each participating recommendation domain. After receiving the global model parameter gradient update amount, each participating recommendation domain updates the model parameters of its local mapping model and completes one round of joint training. Through repeated iterative joint training, when the preset convergence condition is met, the parameter aggregation submodule outputs a global representation mapping function and distributes it to each participating recommendation domain. Each participating recommendation domain maps all local semantic enhancement representations of items to a collaborative representation space shared by all participating domains based on the global representation mapping function, generating a cross-domain aligned set of collaborative recommendation representations of items.

4. The cross-domain recommendation method based on federated learning and self-distillation according to claim 1, characterized in that, Within each participating recommendation domain, based on the collaborative recommendation representation of items corresponding to the user's historical interaction sequence, sequence modeling of user interests is performed to generate a user representation reflecting the user's dynamic interest preferences, specifically including: Within each participating recommendation domain, the target user's historical interaction item ID sequence sorted by timestamp is extracted from the local user interaction database. Based on the interaction item ID identifier, the item collaborative recommendation representation corresponding to each interaction item is retrieved from the cross-domain aligned item collaborative recommendation representation set, and an item representation sequence is generated. Based on the item representation sequence, user interests are sequence modeled. For each position in the sequence, a unique position code is generated using sine and cosine functions. The item collaborative recommendation representation at each position is added element by element to the corresponding position code vector to obtain an enhanced representation sequence that integrates item semantic information and temporal information. The augmented sequence representation sequence is input into a Transformer sequence encoder consisting of multiple stacked coding layers. In each coding layer, the input augmented sequence representation sequence is transformed into a query vector, a key vector, and a value vector through linear projection. The dot product of the query vector and all key vectors is calculated and scaled, and the attention weights are obtained through the Softmax function. The value vector is weighted and summed using the obtained attention weights to obtain the output of each attention head. The outputs of all attention heads are concatenated and linearly transformed, and then nonlinear enhancement is performed using a feedforward neural network to output the sequence context representation of the current layer. Through iterative interaction and nonlinear transformation of multiple coding layers, a contextualized representation that integrates the user's full sequence context information is obtained. The attention pooling mechanism is used to aggregate all sequence context representations, and finally, a user representation that reflects the user's dynamic interests and preferences is output.

5. The cross-domain recommendation method based on federated learning and self-distillation according to claim 1, characterized in that, Each participating recommendation domain independently identifies local long-tail items and popular items, and based on the semantically enhanced representation of the items, matches at least one semantically similar popular item as a teacher for each long-tail item. Through a self-distillation learning mechanism, the representational knowledge of the popular items is transferred to the long-tail items, resulting in an optimized set of collaborative item recommendation representations, specifically including: Each participating recommendation domain calculates the cumulative interaction frequency of each item based on historical interaction records in the local user interaction database, and divides the items into a long-tail item set and a popular item set according to a preset frequency threshold. Using the semantic augmentation representations of items stored in the local semantic representation library, the cosine similarity between each long-tail item and all popular items in the semantic augmentation representation space is calculated based on the cosine metric algorithm. Based on the calculated cosine similarity value, the top-K most popular items with the highest similarity are selected for each long-tail item, and a semantic similarity matching relationship mapping table is constructed with long-tail items as students and popular items as teachers. A self-distillation learning framework is constructed and a dual loss function is set, which includes recommendation task loss and teacher knowledge distillation loss. The teacher knowledge distillation loss is defined by the mean square error between the mean of the current item collaborative recommendation representation of the long-tail item and the mean of the item collaborative recommendation representation of the matching teacher item. During the local model training process, the gradient descent algorithm is used to jointly optimize the dual loss function and perform multiple rounds of iterative training. This allows the collaborative recommendation representation of long-tail items to continuously absorb collaborative signals from popular teacher items with similar semantics while maintaining the original semantic characteristics. When the preset convergence condition is met, the self-distillation learning process terminates, and each domain integrates the latest collaborative recommendation representations of the original popular items and the optimized long-tail items, and outputs the optimized item collaborative recommendation representation set. Establish a dynamic update mechanism. When new user interaction data appears in the recommendation domain, re-identify long-tail items and popular items based on the new user interaction data, update the teacher-student relationship mapping table, and perform knowledge self-distillation again.

6. The cross-domain recommendation method based on federated learning and self-distillation according to claim 1, characterized in that, Each participating recommendation domain independently constructs a recommendation prediction model based on user representations reflecting users' dynamic interests and preferences and optimized item collaborative recommendation representations. It generates a personalized recommendation list by calculating the matching score between users and items, specifically including: Each participating recommendation domain will perform feature fusion processing on user representations reflecting users' dynamic interests and preferences and optimized item collaborative recommendation representations. The user representation vector and item representation vector will be combined into a user-item interaction feature vector through vector concatenation operations. Based on user-item interaction feature vectors, recommendation prediction models based on the Wide & Deep architecture are independently constructed for each domain. The Wide part uses a linear model to process and remember the sparse cross features between users and items, while the Deep part uses a deep neural network to perform deep nonlinear transformations and abstractions to capture the complex implicit interaction patterns between users and items. During the model training phase, each participating recommendation domain uses historical behavior data from the local user interaction database as a supervision signal. The model calculates the interaction prediction probability of user-item pairs through forward propagation, and the parameters of the recommendation prediction model are iteratively updated by the backpropagation algorithm and gradient descent optimizer until the model reaches a convergent state. During the model inference phase, the interaction prediction probability between the target user's user representation and the optimized collaborative recommendation representation of all items in the candidate item set is calculated as a matching score to analyze the user's preference for each candidate item. Finally, based on all the calculated matching scores, the candidate items are sorted in descending order and N items with scores greater than the filtering threshold are selected, outputting a personalized Top-N recommendation list for the target user.

7. A cross-domain recommendation system based on federated learning and self-distillation, characterized in that, To implement the cross-domain recommendation method based on federated learning and self-distillation as described in any one of claims 1-6, the system comprises: a basic data module, a large model semantic enhancement module, a federated mapping learning module, a self-distillation representation learning module, and a recommendation service module; The basic data module includes a user interaction database, an item information database, and a semantic representation database, which are used to store, manage, and provide data locally in each participating recommendation domain; The large model semantic enhancement module includes a large model semantic parsing module and a semantic encoding sub-module, which are used to perform deep and multi-dimensional semantic understanding and representation enhancement of items in each participating recommendation domain. The federated mapping learning module includes a mapping learning submodule and a parameter aggregation submodule, which are used to coordinate the joint learning of cross-domain semantic alignment mapping functions by participating recommendation domains under the premise of strictly protecting the data privacy of each domain; The self-distillation representation learning module includes a long-tail item identification submodule, a teacher selection submodule, and a distillation optimization submodule, which are used to solve the problem of uneven data distribution within each participating recommendation domain and improve the representation quality of long-tail items through knowledge transfer. The recommendation service module includes a user modeling submodule, a feature fusion submodule, and a recommendation prediction submodule, which are used to complete the final personalized recommendation task within each participating recommendation domain.

Images