Collaborative transmission method based on multi-platform group cognition rule

Abstract

This invention discloses a collaborative propagation method based on the cognitive patterns of multiple platforms, comprising: for each current item, retrieving the item and its nearest neighbor item nodes in the knowledge graph, obtaining offline features of the current item and the retrieved nearest neighbor items, and performing attention fusion on these features to generate knowledge transfer features based on KG; for each user, emphasizing historical interaction information, obtaining features of the target user's historical interaction items after knowledge transfer, and performing attention fusion on these features to obtain features related to historical interactions; through domain-level attention, combining KG-based embedding and user item embedding in the target domain, obtaining features that integrate auxiliary multi-platform information for final prediction. This invention compensates for the information loss or insufficiency caused by a single auxiliary domain and improves the security of user data privacy.

Description

Technical Field

[0001] This invention relates to the field of cross-domain collaborative communication, and in particular to a collaborative communication method based on the cognitive patterns of multi-platform groups with privacy protection. Background Technology

[0002] The problem of information overload has spurred the development of recommendation systems, which aim to automatically analyze and estimate users' cognitive patterns and then recommend one or more uninterrupted items to users based on their interests. Recommendation systems have proven to play a significant role in many fields, including film, music, news, and education. [1] .

[0003] Collaborative filtering is one of the mainstream recommendation methods. [2] This approach originates from methods that utilize the collaborative behavior of all users to predict the behavior of a target user. Its power has recently been fully demonstrated on complex neural networks. A major type of research focuses on facilitating the modeling of feature representations of users and items. Many studies use users' historical interaction behavior to better model user representations, rather than modeling users as one-hot vectors of their IDs. Collaborative filtering techniques have been proposed, based on utilizing the similarity between users or items to be recommended, recommending similar items to users or recommending certain items to similar users based on existing user-item interaction history. Matrix factorization techniques have also been proposed based on collaborative filtering. [3] Matrix factorization maps users and items to implicit vectors used to calculate recommendation scores.

[0004] However, existing collaborative propagation approaches still face two long-standing obstacles: First, some datasets contain limited interaction information, resulting in insufficiently rich learned features and hindering effective recommendations—a problem known as data sparsity. Second, new users or items have limited information, impacting feature modeling of user or product information and making personalized recommendations difficult—a problem known as the cold start problem, which significantly limits performance. With the introduction of more domains, users have more options to interact with items of interest. The idea of ​​leveraging information collected from other domains to address the issues of data sparsity and cold start in the target domain has been proposed. In recent years, increasing efforts have combined cross-domain approaches with collaborative propagation, leading to the emergence and significant development of cross-domain recommendation (CDR).

[0005] Noting the cold start and data sparsity problems in Single Domain Propagation Recommendation (SDR), Cross Domain Recommendation benefits from other domains by transferring useful knowledge. Early propagation recommendation methods did not rely on deep learning, focusing primarily on neighborhood-based solutions. [4][5]Due to the remarkable capabilities demonstrated by deep learning in recent years, numerous deep learning-based algorithms have been proposed. Cross-domain scenarios with overlapping users or items are more practical and worthy of study, and common cross-domain propagation recommendation methods are discussed in various ways depending on whether users and items overlap between domains. This involves using shared users or items as a bridge connecting different domains for knowledge transfer, and these two cases are typically handled in a similar symmetrical manner. A widely used approach is to combine embeddings of overlapping users or items from different domains to achieve a unified fusion of embeddings from both domains. [6][7] Although cross-domain methods have developed well and are becoming increasingly popular, the issue of privacy leakage has been largely overlooked. To protect user privacy across different domains, privacy-preserving cross-domain recommendation methods have been proposed. These methods transmit the embeddings of shared items while keeping non-shared items unchanged. This way, user characteristics are not transferred between the two domains; instead, item features containing the cognitive patterns of the user group are transmitted. This avoids the leakage of user feature information and thus protects user privacy. However, when the number of shared items is small, the improvement is limited. Summary of the Invention

[0006] This invention provides a privacy-preserving collaborative propagation method for multi-platform group cognitive patterns. It establishes multi-source auxiliary domains to compensate for information gaps or deficiencies caused by a single auxiliary domain. Considering user data privacy and aiming to utilize auxiliary information as much as possible, this invention constructs a privacy-preserving Knowledge Graph (KG). This KG models the correlations between items from different auxiliary domains without any explicit user item interaction. Simultaneously, this invention proposes a KG-based knowledge transfer method to address the current situation where cross-domain knowledge transfer suffers from a lack of interpretability, as detailed below:

[0007] A collaborative dissemination method based on multi-platform group cognitive patterns, the method comprising:

[0008] For each current project, retrieve the nearest neighbor project nodes in the project and its auxiliary domain from the knowledge graph, obtain the offline features of the current project and the retrieved nearest neighbor projects, and perform attention fusion on these features to generate knowledge transfer features based on KG.

[0009] For each user, we focus on historical interaction information, obtain the characteristics of the target user's historical interaction items after knowledge transfer, and then perform attention fusion to obtain the characteristics of historical interactions.

[0010] By focusing on domains and combining KG-based embeddings with user item embeddings in the target domain, features that incorporate auxiliary information from multiple platforms are obtained and ultimately used for prediction.

[0011] Specifically, for each current item, retrieving the item and its neighboring items in the knowledge graph from the knowledge graph involves:

[0012] For each item v in the target domain j Retrieve the corresponding node V in KG j Retrieve its top-n most relevant neighbor nodes;

[0013] The retrieved relevant nodes are represented as {V} j→n The retrieved node V j→n The project embedding also collects knowledge learned from the auxiliary domain, represented as g. j→n .

[0014] Items in the target domain are fed into the proposed privacy-preserving knowledge graph structure to retrieve their corresponding nodes V. j and closely related nodes V from different auxiliary sources j→n1 V j→n2 Used for knowledge fusion.

[0015] The beneficial effects of the technical solution provided by this invention are:

[0016] 1. This invention first constructs a knowledge graph of node relationships between different domains, which effectively protects user privacy from being leaked, and can be extended to multiple auxiliary domains;

[0017] 2. When performing feature modeling, this invention integrates node information in the graph structure and node information of its neighboring nodes, thereby utilizing the auxiliary information of multiple auxiliary domains to enhance the richness of the modeling and avoiding the information loss problem caused by a single auxiliary domain.

[0018] 3. This invention enables personalized cognitive pattern learning through historical interaction and makes cross-domain knowledge transfer from auxiliary domains more interpretable. Attached Figure Description

[0019] Figure 1 A schematic diagram illustrating the construction of a knowledge graph that transfers knowledge from multiple auxiliary domains;

[0020] Figure 2 This is a schematic diagram of an information modeling framework for a collaborative dissemination method of multi-platform group cognitive patterns. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0022] To enrich the connections between users or projects, many people have proposed using graph structures for information modeling in networks. [8]Knowledge graphs can assemble various types of data (e.g., item attributes and external knowledge) and relationships (e.g., interactions), thus making them widely applicable in recommendation systems. Through graph representation learning, multiple triples connected to the current user / item are found to learn the neighborhood features of the entity, thereby integrating information from different domains.

[0023] To address the problems existing in the background technology, this invention, starting from privacy protection, constructs a knowledge graph for a project to collect implicit user project cognitive patterns from different domains. As shown below, through the proposed knowledge graph structure, knowledge from multiple domains is integrated into a new KG-based feature to obtain an overall user project model.

[0024] Multi-Platform Knowledge Graph Construction: The mainstream graph structure of collaborative propagation recommender systems typically models the interactions between users and items. User-item interactions are often used as edge information between user nodes and item nodes. However, as mentioned above, it can be difficult to access user-item interactions across all domains, as sharing them may violate the privacy policies of the auxiliary domains. Therefore, this project utilizes only offline-learned item embeddings from the auxiliary domains. The offline embedding features of items contain cognitive patterns of user groups within a fixed domain, and their usability is reliable because embedding-based propagation recommender solutions are applied in the industry. This breaks down the privacy restrictions on sharing user-related data, allowing the utilization of knowledge from multiple domains. In this project, a knowledge graph was constructed that comprehensively utilizes offline-learned embeddings of related items from multiple auxiliary domains. It does not require explicit sharing of user-item interactions, thus protecting user privacy across multiple auxiliary domains.

[0025] To protect user privacy, the knowledge graph node entities only involve items of the same or different kinds from multiple auxiliary domains, while the edge uses user-item interactions in the target domain to measure the relevance between nodes and the similarity in item metadata. For each item, this embodiment first retrieves it and its neighboring items in the knowledge graph and its auxiliary domains. Then, their embeddings are fused with attention to achieve a more comprehensive KG-based embedding, thereby generating KG-based knowledge transfer. Finally, cross-domain fusion and prediction are performed. For each user, historical interaction information is emphasized, and their KG-based embedding is combined with item embeddings through attention. Then, through domain-level attention, combined with KG-based embeddings, the information modeling of users and items is enriched and ultimately used for prediction.

[0026] Example 1

[0027] A collaborative propagation method for multi-platform group cognitive patterns based on privacy protection, see [link to relevant documentation]. Figure 1 and Figure 2 The method includes the following steps:

[0028] 101: Represent user item interactions in the target domain as an implicit matrix and obtain the binary value at each interaction.

[0029] Specifically, step 101 involves the following steps:

[0030] Suppose there are M users in the target domain (denoted as u = {u...}). i}, i∈[1,M]) and N items (represented as v={v j},j∈[1,N]),u i v represents each user j Each item represents a user item interaction in the target domain, represented by an implicit matrix Y∈R. M×N The binary value at each interaction point is defined as:

[0031]

[0032] The same applies to other domains, assuming their items are represented as follows: Where N a This indicates the number of items in the auxiliary domain, where 'a' represents different auxiliary domains. When there are overlapping items, these overlapping items in different domains are called bridging items, denoted as...

[0033] 102: Treat all items in the target domain and auxiliary domain as nodes, and merge each bridge item into a graph node;

[0034] In this embodiment of the invention, instead of directly obtaining user item interaction information from the auxiliary domain, the method selects to obtain items learned offline from the auxiliary domain. offline representation For ease of other analyses, the item representations of different auxiliary domains are identical, but the auxiliary items and embeddings may come from different domains and there may be domain offsets.

[0035] To construct the knowledge graph, this embodiment of the invention treats items in both the target domain and the auxiliary domain as nodes. Each bridge item is merged into a graph node, and the nodes in graph V are represented as follows:

[0036] V = v∩v a ={v j},j∈[1,K] (2)

[0037] Here, K represents the number of nodes in the constructed knowledge graph, including all non-overlapping project nodes. For each node in the graph, the metadata associated with the corresponding project (e.g., movie name, year, category) from different domains is unified. Then, Doc2vec is a method that converts words into user vectors for textual representation, used to encode metadata into feature embeddings m with fixed lengths. jAdditionally, each node v from the auxiliary domain j Each has offline features learned from user item interactions in the auxiliary domain. j .

[0038] 103: Obtain the similarity between the features learned offline by the target node and the auxiliary node, and construct edge information based on the similarity.

[0039] Regarding the edges in the graph, they are used to represent the relevance between items. Due to the domain differences between the target and auxiliary domains, the most intuitive approach is to calculate the similarity in terms of metadata as follows:

[0040] sim m (j, k) = sim(m) j m k (3)

[0041] Where sim(·,·) represents the normalized cosine similarity.

[0042] To further investigate item relationships within the target domain, we need to fully utilize user behavior data. If user u i Already with two projects v j and v k If two projects interact, they are likely related. Because of the focus on implicit feedback, the joint interaction count (or the number of users interacting with both projects) is considered useful for calculating their correlation. This correlation is denoted as:

[0043]

[0044] Where, α e To set the weight of this correlation, y *,j To be with project v j Interactive user group, y *,k To be with project v k Interactive user group.

[0045] Furthermore, edge information is constructed by discovering the similarity between features learned offline by the target node and the auxiliary node. These features are written as:

[0046] sim g (j, k) = cos(g) j g k (5)

[0047] Where cos(·,·) represents the normalized cosine similarity.

[0048] In summary, this invention constructs a graph structure with different domain items as nodes. Experimental testing shows that if the performance of calculating node edge information using three similarity formulas does not significantly improve the results, subsequent calculations are based on the edge information between nodes using the aforementioned similarity formulas. m and sim c Similarity is represented by the calculated combination.

[0049] sim(j,k)=sim m +sim c (6)

[0050] Since this graph structure does not involve users in each domain, user information will not be involved in subsequent operations on the graph, thus effectively protecting user privacy.

[0051] Example 2

[0052] Knowledge Transfer Information Modeling Based on Graph Structure: Given a Knowledge Group (KG) for projects in different domains, this invention transfers implicit knowledge from the auxiliary domain to obtain a more comprehensive KG-based information model for projects in the target domain. (Followed by...) Figure 2 As shown, it includes two steps: KG-based retrieval and KG-based knowledge fusion.

[0053] This invention extracts features from users and items in a target domain and an auxiliary domain using a specific network, and embeds the obtained offline user features and offline item features as features. The offline user features obtained under this method can be considered as information about the user's cognitive patterns in a specific domain, while the offline item features contain the group cognitive patterns of a user group within a specified domain. Feature g, described later... j ,p j ,q j These are all features obtained for this purpose.

[0054] 201: KG-based retrieval;

[0055] For each item v in the target domain j First, retrieve the corresponding node V in KG. j Next, retrieve its top-n most relevant neighbor nodes. The retrieved relevant nodes are represented as {V}. j→n In addition to collecting auxiliary embeddings as shared projects, the retrieved node V j→n The project embedding also collects knowledge learned from the auxiliary domain, which is represented as g. j→n .

[0056] 202: KG-based models can successfully transfer knowledge g through the item embedding of related nodes g. j→n ;

[0057] In this embodiment of the invention, the knowledge graph not only focuses on shared items but also on their nearest neighbor nodes to obtain a more comprehensive representation. This is in v j This is particularly useful when the project is not a bridge. In this case, the proposed KG-based model can successfully transfer knowledge g through the project embeddings of related nodes g. j→n This information is expected to be helpful, as the relevant node represents V. j→n Closely related in terms of metadata and interaction. As will be discussed later. Figure 2 As shown, items in the target domain are fed into the proposed privacy-preserving knowledge graph structure to retrieve their corresponding nodes V. j and closely related nodes V from different auxiliary sources j→n1 V j→n2 This will be used for further knowledge integration in the future.

[0058] 203: Knowledge Fusion Based on KG;

[0059] As mentioned above, the retrieved nodes may come from different auxiliary domains. Therefore, since the training data in different auxiliary domains may have different scales and distributions, the auxiliary embeddings may exhibit domain differences. Therefore, a fully connected layer is used to adapt the transmitted embedding, with the following formula:

[0060]

[0061] in, b0 and b0 are learnable parameters of the mapping matrix and bias. Since item embeddings from multiple auxiliary domains may differ in dimension, different fully connected layers are applied to the auxiliary embeddings from different sources. Furthermore, if a bridging item has multiple auxiliary item embeddings, all of these embeddings are used for subsequent fusion.

[0062] 204: Use an attention mechanism to adaptively fuse embeddings by considering the different weights of the embeddings;

[0063] Among these methods, the most intuitive way to fuse relevant nodes in the auxiliary domain is to take the average of the adaptive embeddings. However, because these projects have limitations regarding v... j (v j The influence capabilities of the embedded elements (which may also be available themselves) may vary. This invention uses a widely adopted attention mechanism and proposes to adaptively fuse embedded elements by considering different weights of the embeddings, as shown in the following formula:

[0064] k j =g′ j +∑α j , n g′ j→n ,if v j ∈v∩ (8)

[0065]

[0066] Where, α j,n It is g′ j→n Learnable item weights, g′, are used to model the importance of 20 points across various sections. j→n For the target node v j For example, auxiliary node v n Its characteristics.

[0067] The obtained KG-based embeddings comprehensively integrate implicit knowledge learned from multiple auxiliary domains. Knowledge transfer occurs even if candidate query items are not included in the auxiliary domains. By transferring the KG-based embeddings of query items and related nodes, the modeling of items and users in the target domain can be greatly enriched.

[0068] 205: Information Fusion Modeling and Recommendation Prediction

[0069] To obtain a comprehensive model of user and project information, this embodiment of the invention refers to the offline features of user i and project j as p. i and q j This is for embedding local information because it is not affected by auxiliary information.

[0070] This invention provides an embodiment for finding user u. i The historical interaction behavior, implicitly represented as a multi-heat vector y. i,* , which should be y i,t This indicates that user i and item t have interacted. In this embodiment of the invention, these items are retrieved from a privacy-preserving KG to obtain their transmitted KG-based embeddings, which are represented as follows:

[0071] k (i,t) =KKT(v t ),if y (i,t) =1 (10)

[0072] KKT(·) represents performing a KG-based knowledge transfer operation on the input ·, thereby generating a KG-based embedding.

[0073] Since historical interaction items have different levels of importance to represent the cognitive patterns of a given user, embodiments of this invention apply item-level attention to obtain the shifted user embeddings. i , written as:

[0074]

[0075] Where, α t For the impact factor weights of different projects, |y i,* |for yi,* The l1 norm.

[0076] To effectively leverage user and project transfer embeddings to enrich native embeddings, embodiments of this invention further introduce domain-level concerns to accommodate the varying importance of each component, expressed as follows:

[0077] P i =α p p i +α s s i Q j =α q q j +α k k i (12)

[0078] P i and Q j This refers to the final user and project characteristics. Where p... i and q j For the aforementioned local information embedding, s i For the transferred user embedding, k i The project embedding is obtained after the project has undergone KKT(·) operation, and α is the weight of each part of the learning process.

[0079] Example 3

[0080] The effectiveness of the above scheme is evaluated below with specific experimental data, as detailed in the following description:

[0081] 301: Data Composition;

[0082] The training set consists of the MovieLens movie dataset, the Netflix movie dataset, and the BookCrossing book dataset, which include 3952, 1108, and 400 items respectively.

[0083] 302: Parameter settings and evaluation;

[0084] In the model of this embodiment, the latent embedding dimension of items and users is set to 128. The optimizer is the Adam optimizer, the model learning rate is 0.001, and it is trained for 100 epochs. The nearest neighbor nodes mentioned in the text are set to the top-10. The evaluation used includes two main metrics: hit rate (HR@N) and normalized discounted cumulative gain (NDCG@N). In the experiment, recommendation results were tested for top-3, top-5, and top-10.

[0085] Hit Rate (HR@N): Emphasizes the accuracy of the model's recommendations, i.e., whether the user's desired items are included in the model's recommendations.

[0086]

[0087] Normalized Discount Cumulative Gain (NDCG@N): This emphasizes the position of the user's desired item in the model's recommendation list; the earlier it appears, the better.

[0088]

[0089] 303: Comparison algorithm;

[0090] In the experiment, our method was compared with six other methods, three of which were single-domain methods and three of which were cross-domain methods.

[0091] Single-domain methods include: Neural collaborative filtering (NCF) algorithm. [9] Based on the Bayesian personalized ranking (BPR) algorithm

[10] And Deep Matrix Factorization (DMF) algorithms

[11] .

[0092] Cross-domain methods include: Dual-target cross-domain recommendation (DTCDR) algorithms.

[12] Cross-domain recommendation without sharing user-relevant data (NATR)

[13] And the Toward Equivalent Transformation of User Preferences in Cross Domain Recommendation (ETLCDR) algorithm

[14] .

[0093] Tables 1 and 2 show the bit augmentation results of our method and other methods on the MovieLens and Netflix datasets for 4-16 bits (optimal values ​​of the evaluation metrics are indicated in bold). As can be seen from Table 1, the cross-domain method significantly outperforms the single-domain method in both HR and NDCG. The method proposed in this embodiment achieves the highest PSNR and SSIM scores among all algorithms on the test set.

[0094] Table 1 compares the recommendation performance of Movielens as the target domain.

[0095]

[0096]

[0097] Table 2 Comparison of Recommendation Performance for Netflix as the Target Domain

[0098]

[0099] Example 4

[0100] Training and loss function

[0101] In order to train the proposed model and avoid overfitting positive samples, embodiments of the present invention randomly select some unobserved interactions as negative samples.

[0102] The final prediction objective function can be obtained as follows:

[0103]

[0104] Among them, Y + ,Y - This represents observed and unobserved user item interactions. Ω(θ) is the regularization, and λ is a hyperparameter controlling the regularization.

[0105] l is used to calculate the binary cross-entropy loss:

[0106]

[0107] References

[0108] [1]Adomavicius G,Tuzhilin A.Toward the next generation of recommendersystems:Asurvey of the state-of-the-art and possible extensions[J].IEEE transactions on knowledge and data engineering, 2005,17(6):734-749.

[0109] [2]Zhang F,Yuan N J,Lian D,et al.Collaborative knowledge baseembedding for recommender systems[C] / / Proceedings of the 22nd ACM SIGKDDinternational conference on knowledge discovery and data mining.2016:353-362.

[0110] [3]Mnih A,Salakhutdinov R R.Probabilistic matrix factorization[J].Advances in neural information processing systems,2007,20.

[0111] [4]Berkovsky S,Kuflik T,Ricci F.Cross-domain mediation incollaborative filtering[C] / / User Modeling 2007:11th International Conference,UM 2007,Corfu,Greece,July 25-29,2007.Proceedings 11.Springer BerlinHeidelberg,2007:355-359.

[0112] [5]Gao S,Luo H,Chen D,et al.Cross-domain recommendation via cluster-level latent factor model[C] / / Machine Learning and Knowledge Discovery inDatabases:European Conference,ECML PKDD 2013,Prague,Czech Republic,September23-27,2013,Proceedings,Part II 13.Springer Berlin Heidelberg,2013:161-176.

[0113] [6]Perera D,Zimmermann R.Exploring the use of time-dependent cross-network information for personalized recommendations[C] / / Proceedings of the25th ACM international conference on Multimedia.2017:1780-1788.

[0114] [7]Yan M,Sang J,Xu C,et al.Aunified video recommendation by cross-network user modeling[J].ACM Transactions on Multimedia Computing,Communications,and Applications(TOMM),2016,12(4):1-24.

[0115] [8]Song Y,Yang X,Xu C.Self-supervised calorie-aware heterogeneousgraph networks for food recommendation[J].ACM Transactions on MultimediaComputing,Communications,and Applications(TOMM),2022.

[0116] [9]He X,Liao L,Zhang H,et al.Neural collaborative filtering[C] / / Proceedings of the 26th international conference on world wide web.2017:173-182.

[0117]

[10] Rendle S,Freudenthaler C,Gantner Z,et al.BPR:Bayesianpersonalized ranking from implicit feedback[J].arXiv preprint arXiv:1205.2618,2012.

[0118]

[11] Xue HJ, Dai X, Zhang J, et al.Deep matrix factorization models forrecommender systems[C] / / IJCAI.2017,17:3203-3209.

[0119]

[12] Zhu F,Chen C,Wang Y,et al.Dtcdr:Aframework for dual-target cross-domain recommendation[C] / / Proceedings of the 28th ACM International Conference on Information and Knowledge Management.2019:1533-1542.

[0120]

[13] Gao C, Chen X, Feng F, et al. Cross-domain recommendation without sharing user-relevant data[C] / / The world wide web conference.2019:491-502.

[0121]

[14] Chen X, Zhang Y, Tsang IW, et al. Toward Equivalent Transformation of User Preferences in Cross Domain Recommendation[J]. ACM Transactions on Information Systems, 2023, 41(1):1-31.

[0122] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A collaborative propagation method for multi-platform group cognitive patterns based on privacy protection, characterized in that, The method addresses the issues of data sparsity and cold start, and also protects user privacy by preventing the leakage and sharing of user characteristic information. The method includes: For each current project, retrieve the nearest neighbor project nodes in the project and its auxiliary domain from the knowledge graph, obtain the offline features of the current project and the retrieved nearest neighbor projects, and perform attention fusion on these features to generate knowledge transfer features based on KG. For each user, we focus on historical interaction information, obtain the characteristics of the target user's historical interaction items after knowledge transfer, and then perform attention fusion to obtain the characteristics of historical interactions. By focusing on domains and combining KG-based embeddings with user item embeddings in the target domain, features that incorporate auxiliary information from multiple platforms are obtained and ultimately used for prediction. The knowledge graph is used to collect implicit user item cognition patterns from different domains, and the KG is used for privacy protection; the training set consists of movie data and book datasets, and the method is used in the field of cross-domain collaborative propagation in movie and book recommendation scenarios.

2. The collaborative propagation method for multi-platform group cognitive patterns based on privacy protection according to claim 1, characterized in that, The specific steps for retrieving each current item and its adjacent items in the knowledge graph from the knowledge graph are as follows: For each item in the target domain Retrieve the corresponding node in KG retrieve it The most relevant neighbor node; The retrieved relevant nodes are represented as { }, retrieved nodes The project embedding also collects knowledge learned from the auxiliary domain, which is represented as... .

3. The collaborative propagation method for multi-platform group cognitive patterns based on privacy protection according to claim 1, characterized in that, Items in the target domain are fed into the proposed privacy-preserving knowledge graph structure to retrieve their corresponding nodes. and closely related nodes from different auxiliary sources Used for knowledge fusion.

4. The collaborative propagation method for multi-platform group cognitive patterns based on privacy protection according to claim 3, characterized in that, The knowledge fusion during KG retrieval is as follows: ； ； in, It utilizes the auxiliary nodes obtained through retrieval. For the target node Learnable item weights are used to model the importance of items. For the target node In other words, auxiliary nodes Features; , and These are learnable parameters of the mapping matrix and the bias.

5. The collaborative propagation method for multi-platform group cognitive patterns based on privacy protection according to claim 1, which assigns different importance to historical interaction items to represent the cognitive patterns of a given user, is characterized in that... Transferred user embedding for: ； in, Weighting of impact factors for different projects; for of Norm; , KKT(·) represents a KG-based knowledge transfer operation on the input; This indicates that user i and project t have interacted.

Images