Metalearning cold start recommendation method based on multi-semantic space fusion

Abstract

The invention relates to a meta-learning cold start recommendation method based on multi-semantic space fusion, and the method comprises the steps: carrying out the feature extraction of the attribute content of a target user, and generating a first content feature vector; generating a corresponding path sequence of the target user under a preset meta path based on the heterogeneous information network, performing feature extraction based on the path sequence, and generating a corresponding first semantic association feature vector of the target user under the meta path; taking the first semantic association feature vector as an initial feature vector, constructing a corresponding hypergraph structure based on an interaction relationship of the target user under the meta-path, and performing multilayer convolution processing on the hypergraph structure to obtain a first high-order interaction feature vector corresponding to the target user; and fusing the first content feature vector, the first semantic association feature vector and the first high-order interaction feature vector, and inputting a fused target feature vector into a meta learning model to determine a target object recommended to a target user. According to the method, the recommendation accuracy in the cold start scene can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a meta-learning cold-start recommendation method based on multi-semantic space fusion. Background Technology

[0002] In the actual operation of intelligent recommendation systems, ensuring recommendation accuracy is of paramount importance.

[0003] Taking movie recommendation scenarios as an example, general recommendations are usually made based on popular movies, ignoring individual differences among users and failing to meet diverse viewing needs. For instance, art-house film enthusiasts may have little interest in popular commercial action movies, resulting in low recommendation accuracy.

[0004] To avoid the problems of general recommendation systems, some solutions have proposed matching recommendation strategies. Specifically, with user authorization, recommendations are made based on user information or movie information. However, in cold start scenarios, new users have limited information, only basic information (such as age and gender) exists, or new movies lack play counts and ratings data, only attributes such as movie title and genre tags are available. This results in limited feature dimensions and shallow data mining, leading to recommendations that often deviate significantly from the user's actual interests. For example, relying solely on the "suspense" tag makes it difficult to distinguish between users who prefer hardcore mystery and those with a light suspense / emotional interest. Furthermore, given the massive and dynamically updated film content and user base of video platforms, traditional solutions struggle to quickly establish accurate associations between new users and new movies during the cold start phase, resulting in low recommendation accuracy.

[0005] Therefore, there is an urgent need for a highly efficient recommendation solution that can deeply mine potential associations and adapt to cold start scenarios, in order to break through the recommendation performance bottleneck of the current recommendation system and improve recommendation accuracy. Summary of the Invention

[0006] Therefore, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product based on meta-learning cold start recommendation with multi-semantic space fusion that can improve recommendation accuracy in the face of the above-mentioned technical problems.

[0007] Firstly, this application provides a meta-learning cold-start recommendation method based on multi-semantic space fusion, the method comprising:

[0008] Extract features from the target user's attribute content to generate a first content feature vector;

[0009] Based on the heterogeneous information network, a path sequence corresponding to the target user under the preset meta-path is generated. Based on the path sequence, feature extraction is performed to generate the first semantic association feature vector corresponding to the target user under the meta-path. The nodes in the heterogeneous information network include user nodes, object nodes, and intermediate association nodes related to user nodes and object nodes. The meta-path is used to capture the multi-hop semantic association pattern between user nodes and object nodes in the heterogeneous information network.

[0010] Using the first semantic association feature vector as the initial feature vector, a corresponding hypergraph structure is constructed based on the interaction relationship of the target user under the meta path. The hypergraph structure is then subjected to multi-layer convolution processing to obtain the first high-order interaction feature vector corresponding to the target user.

[0011] The first content feature vector, the first semantic association feature vector, and the first higher-order interaction feature vector are fused to obtain the target feature vector corresponding to the target user.

[0012] The target feature vector is input into the meta-learning model to determine the target object to recommend to the target user.

[0013] Secondly, this application also provides a meta-learning cold-start recommendation device based on multi-semantic space fusion, the device comprising:

[0014] The content feature extraction module is used to extract features from the attribute content of the target user and generate the first content feature vector;

[0015] The semantic association feature extraction module is used to generate the path sequence corresponding to the target user under the preset meta path based on the heterogeneous information network, and to extract features based on the path sequence to generate the first semantic association feature vector corresponding to the target user under the meta path; the nodes in the heterogeneous information network include user nodes, object nodes, and intermediate association nodes related to user nodes and object nodes.

[0016] The hypergraph convolution module is used to construct a corresponding hypergraph structure based on the interaction relationship of the target user under the metapath, using the first semantic association feature vector as the initial feature vector, and perform multi-layer convolution processing on the hypergraph structure to obtain the first high-order interaction feature vector corresponding to the target user.

[0017] The feature fusion module is used to fuse the first content feature vector, the first semantic association feature vector, and the first higher-order interaction feature vector to obtain the target feature vector corresponding to the target user.

[0018] The recommendation module is used to input the target feature vector into the meta-learning model to determine the target objects to recommend to the target user.

[0019] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the embodiments of this application.

[0020] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps in the embodiments of this application.

[0021] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps in the embodiments of this application.

[0022] The aforementioned meta-learning cold-start recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product based on multi-semantic space fusion extracts features from three dimensions: user content features, semantic association features under meta-paths, and high-order interaction features extracted based on hypergraph depth. Even in cold-start scenarios, this approach can acquire more multi-dimensional user feature information. By deeply aggregating these three dimensions of features and performing object recommendations based on multi-dimensional, deeply aggregated user features (i.e., target feature vectors), it solves the problem of "single feature dimensions and shallow semantic representation" in traditional cold-start recommendation, significantly improving recommendation accuracy. Furthermore, combining the target feature vector with a meta-learning model for recommendation requires only a small amount of user interaction data, reducing the complexity of model training and thus improving recommendation efficiency. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a meta-learning cold-start recommendation method based on multi-semantic space fusion in one embodiment.

[0024] Figures 2 to 5 This is a comparison chart of recommendation performance in four scenarios in one embodiment;

[0025] Figure 6 This is a schematic diagram illustrating the percentage increase in metrics of the recommendation model in this case compared to existing models in one embodiment;

[0026] Figure 7 This is a structural block diagram of a meta-learning cold-start recommendation device based on multi-semantic space fusion in one embodiment;

[0027] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] It should be understood that, unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing specific examples only and are not intended to limit the scope of this application.

[0030] like Figure 1 As shown, in one embodiment, a meta-learning cold-start recommendation method based on multi-semantic space fusion is provided. This method is executed by a computer device and specifically includes the following steps:

[0031] S11, extract features from the target user's attribute content to generate the first content feature vector.

[0032] The target users are those who need to be recommended objects (such as items).

[0033] In some examples, the target user's attributes include at least one of the following: demographic characteristics, historical behavioral data, or preference tags. Demographic characteristics may include at least one of age, gender, or region. Historical behavioral data includes clicks, ratings, and favorites records related to the interactive object. Preference tags include explicitly labeled categories of interest. The interactive object can be an item (such as a movie, music, article, or product) or any other object capable of generating interaction; there are no limitations on this.

[0034] Specifically, text-based attribute content (such as preference tags) can be broken down into independent words and converted into one-hot sparse vectors. Numerical attribute content (such as age and rating) is normalized. In this way, multi-source attribute data is integrated into a unified content feature vector, which serves as the basic feature representation of the target user.

[0035] S12, Generate the path sequence corresponding to the target user under the preset meta-path based on the heterogeneous information network, extract features based on the path sequence, and generate the first semantic association feature vector corresponding to the target user under the meta-path.

[0036] Heterogeneous information networks are relational networks built upon different types of nodes. Nodes in a heterogeneous information network include user nodes, object nodes, and intermediate association nodes related to user nodes and object nodes. It should be understood that user nodes in a heterogeneous information network include target user nodes corresponding to target users.

[0037] A set of meta-paths is pre-defined for heterogeneous information networks, including at least one meta-path. Meta-paths can also be called semantic association paths. Each meta-path represents a one-dimensional semantic relationship, used to capture multi-hop semantic association patterns between user nodes and object nodes in heterogeneous information networks.

[0038] Assume there are 4 metapaths mp in the metapath set. i (i = 1, 2, 3, 4), respectively:

[0039] mp1(UM): The meta-path for direct user-movie interaction, in the form of a meta-path. This indicates that user U has rated movie M; this metapath is used to capture the user's explicit preferences.

[0040] mp2(UMUM): The metapath of user-movie-user-movie, the metapath format is as follows: This indicates that user U1 rated movie M1, movie M1 was also rated by user U2, and user U2 rated other movies M2. This meta-path is used to discover user groups with similar movie preferences.

[0041] mp3(UMAM): The metapath of User-Movie-Actor-Movie, the metapath format is as follows This indicates that user U rated movie M1, and one of the main actors in movie M1 is actor A, who also starred in another movie, M2. This meta-path is used to characterize the user's preference transfer for actors.

[0042] mp4(UMDM): The metapath of User-Movie-Director-Movie, the metapath format is as follows: This indicates that user U rated movie M1, which was directed by director D, who also directed another movie, M2. This meta-path is used to capture user preferences for directorial styles.

[0043] In some embodiments, a computer device can perform a random walk based on each meta-path in a heterogeneous information network to obtain a path sequence corresponding to the target user. Each meta-path corresponds to at least one path sequence; the nodes in each path sequence include the target user node corresponding to the target user and the associated object nodes of the target user node; the associated object nodes include object nodes directly associated with the target user node, and object nodes indirectly associated with the target user node through the meta-path (also called extended object nodes), and the indirectly associated object nodes are indirectly associated with the target user node through intermediate associated nodes.

[0044] The following text uses four meta-paths mp1 to mp4 as examples to illustrate the generation of path sequences based on random walks.

[0045] mp1: Starting from user u, a path sequence [u,m] is generated by sampling the movie m directly rated by user u through random walk. Movie m is the object node directly associated with user u.

[0046] mp2: Starting from user u1, a random walk is used to first sample the movie m1 rated by user u1, then sample other users u1 who rated m1, and then sample the movie m2 rated by other users u1, generating the path sequence [u1, m1, u2, m2]. Movie m1 is the object node directly associated with user u1, and movie m2 is the object node indirectly associated with user u1 through user u2 (an intermediate association node).

[0047] mp3: Starting from user u, a random walk is used to sample the movie m1 rated by user u, then the actor a in m1 is sampled, and finally the other movies m2 in which a starred are sampled, generating the path sequence [u, m1, a, m2]. Movie m1 is the object node directly associated with user u, and movie m2 is the object node indirectly associated with user u through actor a (an intermediate association node).

[0048] mp4: Starting from user u, a random walk is used to sample the movie m1 rated by user u, then the director d of movie m1 is sampled, and finally other movies m2 directed by d are sampled, generating the path sequence [u,m1,d,m2]. Movie m1 is the object node directly associated with user u, and movie m2 is the object node indirectly associated with user u through director d (an intermediate association node).

[0049] In some embodiments, a computer device may obtain a set of object nodes corresponding to a meta-path based on the associated object nodes in each path sequence under the same meta-path; and aggregate the object feature vectors corresponding to each associated object node in the set of object nodes to obtain a first semantic association feature vector.

[0050] It should be understood that a meta-path is an abstract path at the graph schema level. It describes a relational pattern between node types. A concrete path sequence is an actual instance-level path existing in the graph, consisting of concrete node instances and concrete edge instances connecting them. For example, User-Movie-Director-Movie is an abstract meta-path, not involving specific user names, movie names, and director names. Zhang San—"Happy Life"—Li Ming—"A Better Life" is a concrete path sequence, where Zhang San and Li Ming are the user name and director name, respectively, and "Happy Life" and "A Better Life" are the movie names.

[0051] Since the path sequence is generated by random walk based on the meta-path, multiple path sequences can be generated under one meta-path. That is, different specific node instances will be randomly sampled under the abstract node type in the same meta-path to form different path sequences.

[0052] For each user u, the computer device can use the same metapath mp i The following is a summary of the associated object nodes in all path sequences formed for user u, resulting in the set of object nodes corresponding to user u under this meta-path.

[0053] Taking user u as Zhang San and the metapath mp3 as an example, suppose we can generate two path sequences: [Zhang San, "Today", Jack, "Future"] and [Zhang San, "Weather", Lily, "Universe"]. Here, "Today" and "Weather" are the movies m1 directly associated with user "Zhang San". "Jack" and "Lily" are actors in the movies "Today" and "Weather" respectively (i.e., 'a' in the path sequence [u,m1,a,m2]). "Future" and "Universe" are other movies starring actors "Jack" and "Lily" respectively (i.e., 'm2' in the path sequence [u,m1,a,m2]). The movies "Today", "Weather", "Future", and "Universe" are all associated object nodes of Zhang San, forming a set of object nodes.

[0054] For each associated object node in the object node set, a corresponding feature vector can be generated, denoted as the object feature vector. For example, for each associated object node, a corresponding object feature vector can be generated based on the corresponding object attribute content. Object attribute content can include basic attribute information such as the object name. Taking a movie as an example, object attribute content can include at least one of the following: movie name, release date, or movie genre. For the specific processing of generating object feature vectors, please refer to the specific processing of generating user content feature vectors based on content attribute content; it will not be repeated here.

[0055] Computer devices can aggregate the object feature vectors corresponding to each associated object node (for example, by averaging the object feature vectors corresponding to each associated object node) to obtain the first semantic association feature vector (also known as the meta-path feature vector or meta-path embedding vector) corresponding to the user under the meta-path.

[0056] Taking the associated object as a movie as an example, the first semantic association feature vector can be calculated using the following formula:

[0057]

[0058] in, This indicates that user u is in the i-th metapath mp i The corresponding first semantic association feature vector, e m Let m be the feature vector (i.e., the object feature vector) of the movie m (i.e., the associated object node). This indicates that user u is in the metapath mp i The corresponding set of object nodes, Represents a collection of object nodes The number of associated object nodes.

[0059] Taking four meta-paths as an example, Where d is the embedding dimension, for example, d=32, and each dimension corresponds to the semantic features of a meta-path. Let be the first semantic association feature vectors corresponding to user u under the metapaths mp1 to mp4 respectively.

[0060] In this scheme, by leveraging multi-hop association mining in the meta-path semantic space, the object nodes of indirect interactions can be expanded based on the meta-path, thereby capturing deep semantic chains. Compared with the limitations of traditional methods that rely solely on direct interactions, this expands the content coverage dimension of the recommendation system and extracts deeper semantic features of associations.

[0061] S13, using the first semantic association feature vector as the initial feature vector, construct the corresponding hypergraph structure based on the interaction relationship of the target user under the metapath, perform multi-layer convolution processing on the hypergraph structure, and obtain the first high-order interaction feature vector corresponding to the target user.

[0062] It should be understood that a corresponding hypergraph structure is constructed for each meta-path. Assuming there are p meta-paths, p first semantic association feature vectors will be generated. The computer device uses these p first semantic association feature vectors as initial feature vectors to construct p hypergraph structures. Nodes in each hypergraph structure include user nodes and object nodes; user nodes in the hypergraph structure include the target user node corresponding to the target user. The hyperedge connections in the hypergraph structure are dynamically generated based on the semantic association rules of the corresponding meta-paths, used to capture complex relationships between nodes across different dimensions.

[0063] Assuming there are 4 meta-paths, the first semantic association feature vector... Using the initial features, construct four independent hypergraph structures. in, This represents the i-th metapath mp i The constructed hypergraph structure, v i Represents hypergraph structure The node in ε i Represents hypergraph structure The hyperedges in the graph. The four hypergraph structures are as follows:

[0064] (Corresponding to metapath mp1:UM): Hypergraph structure The superedge connects user u and its rated movies, i.e., superedge ε = R(u), where R(u) is the set of rated movies of user u, used to capture the direct interaction between user and movies;

[0065] (Corresponding to the metapath mp2:UMUM): The hyperedge connects user groups with shared movie ratings, i.e., the hyperedge. in, This is a set of movies that users u1 and u2 both rated, used to identify user groups with similar preferences;

[0066] (Corresponding to the metapath mp3:UMAM): The hyperedge connects users who have a preference for the same actor and their associated movies, i.e., the hyperedge. Where 'a' represents the lead actor in the movie rated by the user. For all the movies that actor A has starred in, this is used to capture semantic associations along the actor dimension;

[0067] (Corresponding to the metapath mp4:UMDM): The hyperedge connects users who have a preference for the same director and their associated movies, i.e., the hyperedge ε = D(d), where d is the director of the movie rated by the user, and D(d) is all the movies directed by director d, which is used to capture the semantic associations in the director dimension.

[0068] For each hypergraph structure, multiple convolutional layers are performed (e.g., configuring the number of convolutional layers to L). In each current convolutional layer, based on the feature vector of the target user node at the current convolutional layer, the hypergraph structure is convolved to obtain the feature vector of the target user node at the next convolutional layer. This next convolutional layer can then be used as the new current convolutional layer for further convolutional processing. It should be understood that the feature vector of the target user node at the first convolutional layer is the first semantic association feature vector. In this way, multiple convolutional layers are performed on the hypergraph structure.

[0069] Specifically, in each current convolutional layer l (i.e., the l-th layer), for each hyperedge corresponding to the target user node (i.e., each hyperedge to which it belongs), the feature vectors of each node belonging to the hyperedge are aggregated and normalized to obtain the hyperedge aggregated features. The target user node is included among the nodes belonging to the hyperedge. For example, the transpose of the hyperedge matrix can be multiplied by the node feature matrix corresponding to the hypergraph structure (i.e., the feature vectors of each node in the hypergraph structure) to achieve hyperedge feature aggregation, obtaining the initial hyperedge aggregated features of each hyperedge. Then, the initial hyperedge aggregated features of each hyperedge are multiplied by the inverse of the hyperedge degree matrix to complete the normalization process, obtaining the hyperedge aggregated features corresponding to each hyperedge.

[0070] Then, the computer device can aggregate and normalize the hyperedge aggregation features of each hyperedge to which the target user node belongs, to obtain the feature vector of the target user node in the next convolutional layer (i.e., the l+1th layer).

[0071] For example, the hyperedge matrix can be multiplied by the hyperedge aggregation features corresponding to each hyperedge to achieve node feature aggregation, thus obtaining the initial aggregated features corresponding to the target user node (i.e., the features after aggregating other related node features through hyperedges). Then, the initial aggregated features corresponding to the target user node can be normalized using the inverse matrix of the node degree matrix to obtain the normalized aggregated features of the target user node, completing the standardization of node features. Further, the normalized aggregated features of the target user node can be linearly transformed. For layers other than the last one, the linearly transformed aggregated features are sequentially processed by activation functions, layer normalization, and Dropout operations (random deactivation) to obtain the feature vector of the target user node in the next convolutional layer (i.e., layer l+1). That is, the feature vector output after these operations is used as the initial feature vector of the target user node in the next convolutional layer (i.e., layer l+1), and the next convolutional layer is then processed. The activation function can be a ReLU nonlinear mapping, layer normalization is used to stabilize the training process, and the Dropout operation is used to suppress model overfitting.

[0072] In some examples, convolution processing for each layer can be achieved using the following formula:

[0073]

[0074] Among them, X (l+1) X is the feature vector of the target user node at layer l; (l) H is the feature vector of the target user node at layer l; H is the hyperedge matrix, used to characterize the relationship between nodes and hyperedges; D v For the node degree matrix, D e W is the hyperedge degree matrix, which describes the association strength of nodes and hyperedges, respectively; (l) With b (l) Here, represents the weight matrix and bias vector for the linear transformation of the l-th layer; ReLU(·) is the non-linear activation function, LayerNorm(·) is the layer normalization operation, and Dropout(·,q) is the random deactivation operation with a dropout rate q. For example, the dropout rate q can be 0.5.

[0075] Furthermore, the computer device can aggregate the feature vectors of the target user node across multiple convolutional layers to obtain the first high-order interaction feature vector of the target user in the hypergraph structure. This mechanism can balance feature extraction depth and computational complexity by fusing semantic information from different layers.

[0076] For example, an average pooling strategy can be used to average the feature vectors of the target user node across multiple convolutional layers to obtain the first high-order interaction feature vector. For instance, the first high-order interaction feature vector can be calculated using the following formula: in, For target user u in SuperGraph The first high-order interactive feature vector in the matrix.

[0077] It should be understood that by using hypergraph convolution, higher-order relationships such as the common preferences of user groups for similar content can be captured. Compared with traditional graph models, this enables recommendation systems to better explore the potential common preferences of user groups and optimize recommendation performance in cold start scenarios.

[0078] S14, the first content feature vector, the first semantic association feature vector, and the first higher-order interaction feature vector are fused to obtain the target feature vector corresponding to the target user.

[0079] In some embodiments, when there are multiple meta-paths, the first semantic association feature vector under each meta-path and the first higher-order interaction feature vector generated based on the corresponding hypergraph structure correspond to the same semantic space. The computer device can concatenate the first semantic association feature vector and the first higher-order interaction feature vector under the same semantic space to obtain the first fused feature vector under each semantic space; and perform semantic weighted aggregation on the first fused feature vectors under multiple semantic spaces to obtain the first multi-semantic space feature vector corresponding to the target user.

[0080] In some embodiments, the target user corresponds to a first dataset; the first dataset includes a first support set and a first query set; the first support set includes a set of first objects that the target user has rated; the first query set is a set of second objects that the target user has not yet rated.

[0081] Taking four meta-paths as an example, there are four semantic spaces. This is achieved through an adaptive weight matrix. The first fusion feature vectors of the four semantic spaces are weighted and summed to generate the first multi-semantic space feature vector of the target user u. Wherein, weight w i It is the weight corresponding to the i-th semantic space, and the weight w i It is obtained by normalizing the adaptation loss of each semantic space using softmax. For example, Among them, L i For the i-th metapath mp i The adaptation loss for the corresponding semantic space. The adaptation loss for each semantic space is the rating loss of the first fused feature vector of each semantic space to the first object in the first support set. Specifically, based on the correlation between the first fused feature vector of each semantic space and the object feature vectors of each first object in the first support set, the target user's rating for each first object is predicted. The adaptation loss for each semantic space is obtained based on the difference between the actual rating of each first object and the corresponding predicted rating.

[0082] Furthermore, the first multi-semantic space feature vector can be concatenated with the first content feature vector corresponding to the target user to obtain the final user feature vector of the target user, which can be denoted as the target feature vector.

[0083] It should be understood that performing hypergraph convolution operations on hypergraph structures extracts high-order interaction features between users and items, addressing the problem of insufficient association mining caused by data sparsity in cold-start scenarios. Hypergraph convolution, through multi-layer nonlinear transformations and feature aggregation, can capture high-order relationships such as shared preferences of user groups for similar content and semantic associations across meta-paths, making it more suitable for modeling complex semantic structures in cold-start recommendation compared to traditional graph models.

[0084] S15, input the target feature vector into the meta-learning model to determine the target object to recommend to the target user.

[0085] The computer device can input the target feature vector into the meta-learning model, and combine it with the first support set and the first query set corresponding to the target user to determine the target object to recommend to the target user.

[0086] Specifically, the computer device can input the target feature vector into a meta-learning model. Based on the correlation between the target feature vector and the object feature vectors of each first object in the first support set, the meta-learning model predicts the target user's first rating for each first object. A first loss is calculated based on the difference between the target user's true rating and the first rating for the same first object. For example, a sub-loss can be calculated for each first object based on the difference between the target user's true rating and the first rating. Furthermore, the sub-losses for multiple first objects can be averaged to obtain the first loss. The computer device can then update the local parameters of the meta-learning model based on the first loss.

[0087] Furthermore, the computer device can use a meta-learning model with locally updated parameters to calculate the correlation between the target feature vector and the object feature vectors of each second object in the first query set, in order to predict the target user's second rating for each second object; based on each second rating, the target object recommended by the target user is determined from each second object in the first query set. For example, the target object recommended by the target user can be determined from each second object in the first query set in descending order of the second rating. For instance, the second objects with the highest preset ranking can be selected as the target objects recommended to the target user.

[0088] In the aforementioned method, a multi-semantic space fusion mechanism is used to deeply aggregate features from three dimensions: user content features, semantic association features under meta-paths, and high-order interaction features extracted based on hypergraph depth. This allows for object recommendation based on multi-dimensional, deeply aggregated target feature vectors, solving the problem of "single feature dimensions and shallow semantic representation" in traditional cold-start recommendation and significantly improving recommendation accuracy. Secondly, combining the target feature vectors with a meta-learning model for recommendation requires only a small amount of user interaction data, reducing the complexity of model training and thus improving recommendation efficiency.

[0089] Furthermore, a decoupled architecture of multi-semantic space features and meta-learning is proposed, which supports flexible migration to cold start scenarios in multiple fields such as e-commerce, social networking, and information. For example, the meta-learning model in movie recommendation can be applied to product recommendation scenarios, which has broad technical scalability and commercial application value.

[0090] In some embodiments, the meta-learning model is obtained through iterative training based on a sample set; the sample set includes multiple sample users and sample objects rated by each sample user; the heterogeneous information network is constructed based on the sample set.

[0091] Taking movies as an example of user interaction, the sample set can contain 1,000,209 rating records of 3,881 movies from 6,040 users, with a rating range of 1-5 points and a sparsity of 95.73%.

[0092] In each training round, a meta-learning task group is determined, consisting of t tasks. Each task is a recommendation task for a target user in the sample set, where t ≥ 1 and is a positive integer. Therefore, in each training round, 16 target users can be sampled to form the meta-learning task group. Taking t = 16 as an example... For example, stratified random sampling can be performed for multiple cold start types (WU / UC / IC / UIC) to determine the meta-learning task group corresponding to each cold start type, with each group containing 16 target sample users.

[0093] Each target sample user has a corresponding second support set and second query set. For example, the second support set and second query set are constructed according to the following rules: only users with a first number (e.g., 13-100) rating records (including rated objects) are retained. In each training round, target sample users are randomly sampled from the retained users to form a meta-learning task group. For each target sample user, a second number (e.g., 10) of rating records (including rated objects) can be randomly selected as the second query set Q for that target sample user; that is, the second query set includes second sample objects rated by the target sample user, and the number of second sample objects can be the second number. The second support set S is determined from the remaining rating records, ensuring... That is, the second support set includes the first sample objects rated by the target sample users, and the number of the first sample objects can be a third number. The sum of the second and third numbers is less than or equal to the first number.

[0094] For each target sample user, the computer device can obtain a second content feature vector corresponding to that target sample user. This second content feature vector is generated based on the target sample user's attribute content. It should be understood that the second content feature vectors corresponding to multiple target sample users can form a content feature matrix, with each target sample user's second content feature vector being a row of data in the content feature matrix. The principle for generating the second content feature vector is the same as the principle for generating the first content feature matrix; the only difference is the target user (i.e., the target sample user and the target user are different), which will not be elaborated further here.

[0095] For each target user sample, the computer device can generate a sequence of sample paths for the target user under each meta-path based on a heterogeneous information network. Based on this sequence, a second semantic association feature vector corresponding to the target user under each meta-path is generated. Using this second semantic association feature vector as the initial feature vector, the computer device can construct a corresponding sample hypergraph structure based on the interaction relationships of the target user under each meta-path. Multi-layer convolution processing is then performed on the sample hypergraph structure to generate a second higher-order interaction feature vector corresponding to the target user under each meta-path.

[0096] It should be understood that the principle for generating the second semantic association feature vector and the second higher-order interaction feature vector is the same as the principle for generating the first semantic association feature vector and the first higher-order interaction feature vector, and will not be repeated here. The second semantic association feature vectors corresponding to multiple target sample users form a semantic association feature matrix (also known as a meta-path feature matrix), and the first higher-order interaction feature vectors corresponding to multiple target sample users form a higher-order interaction feature matrix.

[0097] The computer device can fuse the second content feature vector, the second semantic association feature vector, and the second higher-order interaction feature vector corresponding to the same target sample user to obtain a user feature matrix; the user feature matrix includes the final feature vectors corresponding to each of the t target sample users.

[0098] Specifically, for the metapath mp i , will each target sample user u t The second semantic association feature vector under this meta-path With the second higher-order interaction feature vector By splicing and merging, the metapath mp is obtained. i The corresponding semantic space fusion feature matrix Fusion feature matrix F' i This includes multiple target sample users in the metapath mp i The second fusion feature vectors corresponding to each of the following are respectively.

[0099] Thus, with multiple meta-paths, multiple semantic space fusion feature matrices can be obtained. Furthermore, the fusion feature matrices of multiple semantic spaces can be weighted and summed (i.e., the second fusion feature vectors corresponding to the same target sample user in the fusion feature matrices of multiple semantic spaces are weighted and summed) to generate a multi-semantic space feature matrix. The multi-semantic space feature matrix includes the second multi-semantic space feature vectors corresponding to each of the multiple (t) target sample users. The calculation method for the weights used in calculating the second multi-semantic space feature vectors can refer to the relevant processing for calculating the first multi-semantic space feature vectors; simply replace the first support set with the second support set, which will not be elaborated here.

[0100] Furthermore, the computer device can train the initial meta-learning model to be trained in this round based on the user feature matrix and the second support set and second query set corresponding to each of the t target sample users, and enter the next round of training based on the meta-learning model trained in this round, until the training stopping condition is met.

[0101] In some embodiments, the computer device may, for each target sample user, predict the third rating of the target sample user for each first sample object based on the correlation between the final feature vector corresponding to the target sample user in the user feature matrix and the object feature vector of each first sample object; and calculate the support set loss based on the difference between the target sample user's true rating and the third rating for the same first sample object.

[0102] Specifically, the sub-loss corresponding to each first sample object can be obtained based on the difference between the true rating and the third rating of the target sample users for the same first sample object. The support set loss is obtained by averaging the sub-losses corresponding to multiple first sample objects.

[0103] For example, the calculation of each target sample user u is performed according to the following formula. t The corresponding support set loss L s (u t ):

[0104]

[0105] in, For the t-th target sample user u t The corresponding second support set is the target sample user u. t The set of first sample objects with which interaction behavior has occurred; This represents the total number of objects in the first sample. For target sample user u t For the first sample object m t1 Real ratings; Is the target sample user u t The corresponding final feature vector, The first sample object m t1 The object's feature vector;

[0106] The rating prediction function f() predicts the target sample user u based on the current global prior parameter θ. t For each first sample object m t1 The third rating.

[0107] Furthermore, the local parameters in the initial meta-learning model of this round can be updated and adjusted based on the support set loss. For example, local parameters can be updated using gradient descent: Where θ is the global prior parameter and α is the learning rate.

[0108] Furthermore, the computer device can, based on the updated local parameters of the meta-learning model, predict the fourth rating of the target sample user for each second sample object based on the correlation between the final feature vector corresponding to the target sample user and the object feature vector of each second sample object; calculate the second loss corresponding to the target sample user based on the difference between the target sample user's true rating and the fourth rating for the same second sample object; average the second losses corresponding to t target sample users to obtain the third loss; update and adjust the global parameters of the initial meta-learning model based on the third loss, and use the meta-learning model with updated global parameters as the initial meta-learning model in the next round of training.

[0109] In some examples, each target sample user u t Group-level loss L on the query set group(i.e., the third loss) can be calculated using the following formula:

[0110]

[0111] in, For target sample user u t The corresponding second query set is the target sample user u. t A collection of second sample objects with which interaction behavior has occurred; The total number of objects in the second sample; For target sample user u t For the second sample object m t2 Real ratings; Is the target sample user u t The corresponding final feature vector, It is the second sample object m t2 The object's feature vector; The rating prediction function f() is based on the updated parameters. Predicted target sample user u t For each second sample object m t2 The fourth rating; Is the target sample user u t The corresponding second loss.

[0112] For example, the Adam optimizer can be used to update θ as a global prior parameter with a learning rate η, balancing personalized needs with cross-group generalization ability.

[0113] To verify the effectiveness of the cold start scenario proposed in this application, the inventors conducted a quantitative comparison and evaluation using three types of indicators: mean absolute error (MAE, used to measure the absolute deviation between the predicted score and the actual score), root mean square error (RMSE, used to amplify the impact of extreme deviations to verify robustness), and normalized depreciation cumulative gain (NDCG@k, used to evaluate the quality of the recommendation ranking).

[0114] Taking the object of user interaction as an example, and combining the state of "whether the user / item appears" during the meta-training phase, we divide the scenarios into three cold start scenarios and one baseline scenario (warm start), and clarify the validation boundaries:

[0115] User cold start (UC): The task is for new users who did not appear in the meta-training phase, and whose interactive items (support sets, query sets) already existed in the meta-training phase.

[0116] Item Cold Start (IC): This task targets existing users who were already in the meta-training phase, and whose interacted items (support sets, query sets) did not appear in the meta-training phase.

[0117] User-Item Cold Start (UIC): This task targets new users and items that did not appear during the meta-training phase.

[0118] Warm-up scenario: The task targets users and items that have already appeared during the meta-training phase.

[0119] Based on the above metrics and scenario definitions, the recommendation performance of the proposed model (MSFML) is compared and validated with existing models (DropoutNet, MeteEmb, MeLU, MetaHIN, etc.) in different scenarios. The experimental results are shown in Table 1 (↓ indicates that the smaller the metric value, the better the performance; ↑ indicates that the larger the metric value, the better the performance).

[0120] Table 1

[0121]

[0122]

[0123] The comparison of metrics across the four scenarios in Table 1 shows that the recommendation model in this case establishes a performance baseline in the warm-start scenario. In the cold-start scenario (UC / IC / UIC), thanks to the collaborative mechanism of "multi-semantic fusion, hypergraph convolution, and meta-learning dynamic adaptation", it achieves dual optimization of "rating prediction accuracy (MAE / RMSE)" and "recommendation ranking quality (NDCG@5)", verifying the targeted improvement effect of this case on cold-start recommendation and demonstrating significant performance advantages compared to existing models.

[0124] Furthermore, the inventors of this application also repeatedly observed the changing trends of MAE, RMSE, and NDCG@5 in four scenarios—warm start, user cold start (UC), item cold start (IC), and user-item cold start (UIC)—by controlling the number of hypergraph convolutional layers (1-9 layers). (Specifically, as shown in the figure below) Figures 2 to 5 As shown), the conclusion is as follows:

[0125] When the number of hypergraph convolutional layers is precisely configured to 7, the three core evaluation metrics—Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Normalized Diminished Cumulative Gain (NDCG@5)—show highly synergistic optimal performance. Specifically, MAE and RMSE decrease to optimal values ​​at this layer count, while NDCG@5 simultaneously climbs to its peak value across all scenarios. This demonstrates that the recommendation model in this case achieves optimal prediction accuracy for scores, reduces the deviation between predicted and actual scores, and optimizes the ranking quality of the recommended list, significantly improving both accuracy and relevance.

[0126] This conclusion verifies that the innovative architecture of "multi-semantic fusion + hypergraph convolution + meta-learning" constructed in this invention has solid design value. The 7-layer hypergraph convolution, with its deep mining capabilities of "user-item-multi-dimensional semantics (such as director, actor associations, etc.)," ​​constructs a rich and accurate high-order semantic association network. On the one hand, it effectively reduces MAE and RMSE; on the other hand, it significantly improves NDCG@5, achieving a dual breakthrough and optimization in "rating prediction accuracy" and "recommendation ranking quality."

[0127] in addition, Figure 6 This illustrates the percentage increase in metrics achieved by the recommendation model in this case compared to the existing MetaHIN model. From... Figure 6 It can be seen that the average NDCG@5 growth rate in the three cold start scenarios (User-cold, Item-cold, and User-Item-cold) is about 2.81%, which is significantly higher than the 2.57% growth rate in the warm start scenario. In particular, in the item-cold scenario, the NDCG@5 growth rate is 4.63%, which verifies that the proposed solution, through hypergraph convolution and meta-learning, significantly optimizes the performance of cold start recommendation and highlights the advantages of the present invention in cold start scenarios.

[0128] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0129] Based on the same inventive concept, this application also provides a device for implementing the above-mentioned meta-learning cold-start recommendation method based on multi-semantic space fusion. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the meta-learning cold-start recommendation device based on multi-semantic space fusion provided below can be found in the limitations of the meta-learning cold-start recommendation method based on multi-semantic space fusion described above, and will not be repeated here.

[0130] In one embodiment, such as Figure 7 As shown, a meta-learning cold-start recommendation device based on multi-semantic space fusion is provided. The device includes:

[0131] The content feature extraction module 702 is used to extract features from the attribute content of the target user and generate a first content feature vector;

[0132] The semantic association feature extraction module 704 is used to generate a path sequence corresponding to the target user under a preset meta-path based on a heterogeneous information network, and to perform feature extraction based on the path sequence to generate a first semantic association feature vector corresponding to the target user under the meta-path; the nodes in the heterogeneous information network include user nodes, object nodes, and intermediate association nodes related to the user nodes and object nodes.

[0133] The hypergraph convolution module 706 is used to construct a corresponding hypergraph structure based on the interaction relationship of the target user under the metapath, using the first semantic association feature vector as the initial feature vector, and to perform multi-layer convolution processing on the hypergraph structure to obtain the first high-order interaction feature vector corresponding to the target user.

[0134] The feature fusion module 708 is used to fuse the first content feature vector, the first semantic association feature vector and the first high-order interaction feature vector to obtain the target feature vector corresponding to the target user.

[0135] The recommendation module 710 is used to input the target feature vector into the meta-learning model to determine the target object to be recommended to the target user.

[0136] In some embodiments, the meta-learning model is obtained through iterative training based on a sample set; the sample set includes multiple sample users and sample objects rated by each sample user; the heterogeneous information network is constructed based on the sample set. The apparatus further includes:

[0137] The model training module (not shown in the figure) is used to determine the meta-learning task group in each training round. The meta-learning task group includes t tasks, each task being a recommendation task for a target sample user in the sample set, where t≥1 and is a positive integer. Each target sample user has a corresponding second support set and a second query set. For each target sample user, the module obtains the second content feature vector corresponding to the target sample user, which is generated based on the attribute content of the target sample user. The module generates a sample path sequence of the target sample user under the meta-path based on the heterogeneous information network, and generates a second semantic association feature vector corresponding to the target sample user under the meta-path based on the sample path sequence. The module uses the second semantic association feature vector as the initial feature vector. The vector is constructed based on the interaction relationship of the target sample users under the meta-path to build the corresponding sample hypergraph structure. The sample hypergraph structure is subjected to multi-layer convolution processing to generate the second high-order interaction feature vector corresponding to the target sample user. The second content feature vector, the second semantic association feature vector, and the second high-order interaction feature vector corresponding to the same target sample user are fused to obtain the user feature matrix. The user feature matrix includes the final feature vectors corresponding to each of the t target sample users. The initial meta-learning model to be trained in this round is trained based on the user feature matrix and the second support set and the second query set corresponding to each of the t target sample users. The next round of training is carried out based on the meta-learning model trained in this round until the training stopping condition is met.

[0138] The modules in the aforementioned meta-learning cold-start recommendation device based on multi-semantic space fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0139] In one embodiment, a computer device is provided, which may be a terminal or a server. The internal structure diagram of the computer device may be as follows: Figure 8As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a meta-learning cold-start recommendation method based on multi-semantic space fusion.

[0140] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0141] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the embodiments of this application.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the embodiments of this application.

[0143] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the embodiments of this application.

[0144] It should be noted that the user information (including but not limited to user device information, user / object attribute content, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0145] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. The processors involved in the embodiments provided in this application can be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited thereto.

[0146] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0147] The above embodiments are merely illustrative of several implementation methods of this application and should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A meta-learning cold-start recommendation method based on multi-semantic space fusion, characterized in that, The method includes: Extract features from the target user's attribute content to generate a first content feature vector; Based on a heterogeneous information network, a path sequence corresponding to the target user under a preset meta-path is generated. Based on the path sequence, feature extraction is performed to generate a first semantic association feature vector corresponding to the target user under the meta-path. The nodes in the heterogeneous information network include user nodes, object nodes, and intermediate association nodes related to the user nodes and object nodes. The meta-path is used to capture the multi-hop semantic association pattern between the user nodes and the object nodes in the heterogeneous information network. Using the first semantic association feature vector as the initial feature vector, a corresponding hypergraph structure is constructed based on the interaction relationship of the target user under the metapath. The hypergraph structure is then subjected to multi-layer convolution processing to obtain the first high-order interaction feature vector corresponding to the target user. The first content feature vector, the first semantic association feature vector, and the first higher-order interaction feature vector are fused to obtain the target feature vector corresponding to the target user. The target feature vector is input into the meta-learning model to determine the target object to recommend to the target user.

2. The method according to claim 1, characterized in that, The step of generating the path sequence corresponding to the target user under the preset meta-path based on the heterogeneous information network includes: In the heterogeneous information network, a random walk is performed based on the meta-path to obtain the path sequence corresponding to the target user; Wherein, there is at least one meta-path, and each meta-path corresponds to at least one path sequence; the nodes in each path sequence include the target user node corresponding to the target user and the associated object nodes of the target user node; the associated object nodes include object nodes directly associated with the target user node, and object nodes indirectly associated with the target user node extended through the meta-path, and the indirectly associated object nodes are indirectly associated with the target user node through the intermediate associated nodes.

3. The method according to claim 2, characterized in that, The step of extracting features based on the path sequence to generate a first semantic association feature vector corresponding to the target user under the meta-path includes: Based on the associated object nodes in each path sequence under the meta-path, the set of object nodes corresponding to the meta-path is obtained; The first semantic association feature vector is obtained by aggregating the object feature vectors corresponding to each associated object node in the object node set.

4. The method according to claim 1, characterized in that, The target user corresponds to a target user node in the hypergraph structure; The step of performing multi-layer convolution processing on the hypergraph structure to obtain the first high-order interaction feature vector corresponding to the target user includes: Based on the feature vector of the target user node in the current convolutional layer, the hypergraph structure is convolved to obtain the feature vector of the target user node in the next convolutional layer, and the next convolutional layer is used as the new current convolutional layer to continue convolution processing; the feature vectors of the target user node in multiple convolutional layers are aggregated across layers to obtain the first high-order interaction feature vector of the target user in the hypergraph structure. The step of performing convolution processing on the hypergraph structure based on the feature vector of the target user node in the current convolutional layer to obtain the feature vector of the target user node in the next convolutional layer includes: In the current convolutional layer, for each hyperedge corresponding to the target user node, the feature vectors of each node belonging to the hyperedge are aggregated and normalized to obtain the hyperedge aggregated feature; the target user node is included among the nodes belonging to the hyperedge, and the feature vector of the target user node under the first convolutional layer is the first semantic association feature vector. The hyperedge aggregation features of each hyperedge to which the target user node belongs are aggregated and normalized to obtain the feature vector of the target user node in the next convolutional layer.

5. The method according to claim 1, characterized in that, When there are multiple meta-paths, the first semantic association feature vector under each meta-path and the first higher-order interaction feature vector generated based on the corresponding hypergraph structure correspond to the same semantic space. The step of fusing the first content feature vector, the first semantic association feature vector, and the first higher-order interaction feature vector to obtain the target feature vector corresponding to the target user includes: The first semantic association feature vector and the first higher-order interaction feature vector in the same semantic space are concatenated to obtain the first fusion feature vector in each semantic space. The first fused feature vectors under multiple semantic spaces are semantically weighted and aggregated to obtain the first multi-semantic space feature vector corresponding to the target user. The first multi-semantic space feature vector is concatenated with the first content feature vector to obtain the target feature vector.

6. The method according to any one of claims 1 to 5, characterized in that, The target user corresponds to a first dataset; the first dataset includes a first support set and a first query set; the first support set includes a set of first objects that the target user has already rated; The first query set is a set of second objects that the target user has not yet rated; The step of inputting the target feature vector into the meta-learning model to determine the target object to recommend to the target user includes: The target feature vector is input into the meta-learning model, and the meta-learning model predicts the first rating of the target user for each of the first objects based on the correlation between the target feature vector and the object feature vectors of each of the first objects in the first support set. A first loss is calculated based on the difference between the target user's actual rating and the first rating for the same first object, and the local parameters of the meta-learning model are updated based on the first loss. The correlation between the target feature vector and the object feature vector of each second object in the first query set is calculated using the meta-learning model after local parameter updates, so as to predict the second rating of the target user for each second object; Based on each of the second scores, the target object recommended by the target user is determined from each of the second objects in the first query set.

7. The method according to claim 6, characterized in that, The meta-learning model is obtained through iterative training based on a sample set; the sample set includes multiple sample users and sample objects rated by each sample user; The heterogeneous information network is constructed based on the sample set; The method further includes: In each round of training, a meta-learning task group is determined, which includes t tasks. Each task is a recommendation task for a target sample user in the sample set, where t ≥ 1 and is a positive integer. Each target sample user has a corresponding second support set and a second query set. For each target sample user, a second content feature vector corresponding to the target sample user is obtained, wherein the second content feature vector is generated based on the attribute content of the target sample user; Based on the heterogeneous information network, a sample path sequence of the target sample user under the meta-path is generated, and a second semantic association feature vector corresponding to the target sample user under the meta-path is generated based on the sample path sequence. Using the second semantic association feature vector as the initial feature vector, a corresponding sample hypergraph structure is constructed based on the interaction relationship of the target sample user under the metapath. The sample hypergraph structure is subjected to multi-layer convolution processing to generate the second high-order interaction feature vector corresponding to the target sample user. The second content feature vector, the second semantic association feature vector, and the second higher-order interaction feature vector corresponding to the same target sample user are fused to obtain a user feature matrix; the user feature matrix includes t final feature vectors corresponding to each of the target sample users. The initial meta-learning model to be trained in this round is trained based on the user feature matrix and the second support set and second query set corresponding to each of the t target sample users. The next round of training is then carried out based on the meta-learning model trained in this round until the training stopping condition is met.

8. The method according to claim 7, characterized in that, The second support set includes a first sample object rated by the target sample user, and the second query set includes a second sample object rated by the target sample user. The process of training the initial meta-learning model for this round of training based on the user feature matrix, the second support set corresponding to each of the t target sample users, and the second query set, and then proceeding to the next round of training based on the meta-learning model trained in this round, includes: For each target sample user, based on the correlation between the final feature vector corresponding to the target sample user in the user feature matrix and the object feature vector of each first sample object, the third rating of the target sample user for each first sample object is predicted; The support set loss is calculated based on the difference between the actual rating and the third rating of the target sample user for the same first sample object, and the corresponding local parameters in the initial meta-learning model of this round are updated and adjusted based on the support set loss. Based on the meta-learning model with updated local parameters, for each target sample user, the fourth rating of the target sample user for each second sample object is predicted according to the correlation between the final feature vector corresponding to the target sample user and the object feature vector of each second sample object. Based on the difference between the target sample user’s actual rating and the fourth rating for the same second sample object, calculate the second loss corresponding to the target sample user; The third loss is obtained by averaging the second loss corresponding to t target sample users. The global parameters of the initial meta-learning model are updated and adjusted according to the third loss, and the meta-learning model with updated global parameters is used as the initial meta-learning model in the next round of training.

9. A meta-learning cold-start recommendation device based on multi-semantic space fusion, characterized in that, The device includes: The content feature extraction module is used to extract features from the attribute content of the target user and generate the first content feature vector; The semantic association feature extraction module is used to generate a path sequence corresponding to the target user under a preset meta-path based on a heterogeneous information network, and to perform feature extraction based on the path sequence to generate a first semantic association feature vector corresponding to the target user under the meta-path; the nodes in the heterogeneous information network include user nodes, object nodes, and intermediate association nodes related to the user nodes and object nodes; The hypergraph convolution module is used to construct a corresponding hypergraph structure based on the interaction relationship of the target user under the metapath, using the first semantic association feature vector as the initial feature vector, and to perform multi-layer convolution processing on the hypergraph structure to obtain the first high-order interaction feature vector corresponding to the target user. The feature fusion module is used to fuse the first content feature vector, the first semantic association feature vector and the first high-order interaction feature vector to obtain the target feature vector corresponding to the target user. The recommendation module is used to input the target feature vector into the meta-learning model to determine the target object to recommend to the target user.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images