Information recommendation model fine-tuning method and device, and commodity recommendation method and device

Abstract

The application discloses a fine-tuning method and device of an information recommendation model, a commodity recommendation method and device, and the method comprises the following steps: acquiring a first interaction relationship graph and inputting the first interaction relationship graph to an initial information recommendation model; obtaining node representations of an interaction requester and an interaction object through pre-training; clustering the node representations of the interaction object to obtain clustered interaction objects; constructing a second interaction relationship graph for representing the interaction relationship between the interaction requester and the clustered interaction objects according to the clustered interaction objects and the first interaction relationship graph; adjusting the node representations of the interaction requester, the interaction object and the clustered interaction objects according to a cross-layer comparison learning loss obtained by the initial information recommendation model between different layers of nodes on the first interaction relationship graph, a recommendation loss obtained according to the first interaction relationship graph, and a recommendation loss obtained according to the second interaction relationship graph as a joint constraint, so as to obtain a target information recommendation model; and thereby improving the accuracy and comprehensiveness of the model recommendation result.

Description

Technical Field

[0001] This application relates to the field of information recommendation technology, specifically to a method and apparatus for fine-tuning an information recommendation model, a method and apparatus for recommending goods, as well as computer program products and electronic devices. Background Technology

[0002] With the rapid development of internet and information technology, the amount of information available to users on various online platforms (such as e-commerce, social media, video streaming, and news) has grown exponentially. Faced with this massive amount of information resources, users often find it difficult to quickly and accurately find content that matches their personal interests and needs, leading to an increasingly prominent problem of information overload. To address this issue, recommendation systems have emerged and are widely used in personalized information services.

[0003] Recommendation systems analyze users' historical behavioral data (such as browsing history, click behavior, purchase history, rating feedback, etc.) and item attribute information (such as category, tags, content features, etc.) to build user interest models. These models then predict users' preferences for items they haven't yet encountered and recommend potentially interesting items to them. Common recommendation techniques include collaborative filtering-based methods (such as user collaborative filtering and item collaborative filtering), content-based recommendation methods, hybrid recommendation methods, and, in recent years, recommendation methods implemented using deep learning and graph neural network models.

[0004] While existing recommender systems have achieved significant results in improving user experience, they still face many challenges. For example, traditional collaborative filtering methods suffer from data sparsity and cold-start problems; content-based methods may lack recommendation diversity due to insufficient feature extraction; furthermore, existing systems still need improvement in capturing the dynamic evolution of user interests, processing multi-source heterogeneous data, and achieving interpretable recommendations. Summary of the Invention

[0005] This application provides a method for fine-tuning an information recommendation model to solve the problem of inaccurate recommendation information in the prior art.

[0006] This application provides a method for fine-tuning an information recommendation model, including: Obtain a first interaction relationship graph and input it into the initial information recommendation model. The first interaction relationship graph includes a first node representing the interaction requester and a second node representing the interaction object, as well as the connection relationship between the nodes used to represent the interaction relationship between the interaction requester and the interaction object. Through pre-training, the node representations of the interaction requester and the interaction object are obtained respectively; Cluster the node representations of the interactive objects to obtain clustered interactive objects; Based on the clustered interaction objects and the first interaction relationship diagram, a second interaction relationship diagram is constructed to represent the interaction relationship between the interaction requester and the clustered interaction objects; The cross-layer contrastive learning loss obtained from the node representations of different layers in the first interaction graph of the initial information recommendation model, the recommendation loss obtained from the first interaction graph, and the recommendation loss obtained from the second interaction graph are used as joint constraints to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object to obtain the target information recommendation model for information recommendation.

[0007] In some embodiments, clustering the node representations of the interactive objects to obtain clustered interactive objects includes: Project the node representation of the interactive object onto a two-dimensional coordinate system; In the two-dimensional coordinate system, the node representation of the interactive object is divided according to the set division direction and the number of divisions corresponding to the division direction to obtain the clustered interactive object.

[0008] In some embodiments, constructing a second interaction relationship graph representing the interaction relationship between the interaction requester and the clustered interaction object based on the clustered interaction object and the first interaction relationship graph includes: Based on the connection relationship between the interaction requester and the interaction object in the first interaction relationship diagram, if any of the interaction objects in the clustered interaction objects has a connection relationship with the interaction requester, then a connection relationship between the interaction requester and the clustered interaction object is established. Based on the interaction requester, the clustered interaction object, and the connection relationship, construct the second interaction relationship graph.

[0009] In some embodiments, the cross-layer contrastive learning loss obtained from the node representations at different layers of the first interaction graph based on the initial information recommendation model, the recommendation loss obtained from the first interaction graph, and the recommendation loss obtained from the second interaction graph are used as joint constraints to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object to obtain a target information recommendation model for information recommendation, including: Based on the first interaction relationship diagram and the second interaction relationship diagram, the first connection strength between the interaction requester and the interaction object, and the second connection strength between the interaction requester and the clustered interaction object are obtained. Based on the first connection strength, determine the first recommendation loss of the initial information recommendation model; Based on the second connection strength, determine the second recommendation loss of the initial information recommendation model; Based on the first recommendation loss, the second recommendation loss, and the cross-layer contrastive learning loss as joint constraints, the node representations of the interaction requester, the interaction object, and the clustered interaction object are adjusted to obtain a target information recommendation model for information recommendation.

[0010] In some embodiments, obtaining the first connection strength between the interaction requester and the interaction object, and the second connection strength between the interaction requester and the clustered interaction object based on the first interaction relationship graph and the second interaction relationship graph, includes: Based on the graph convolution operation of the first interaction graph and the second interaction graph, the interaction requester representation matrix, the interaction object representation matrix, and the clustered interaction object representation matrix are obtained. The interaction object node vector in the interaction object representation matrix is ​​fused with the cluster interaction object node vector of its corresponding cluster in the cluster interaction object representation matrix to obtain a fused vector. The dot product of the interaction requester node vector in the interaction requester representation matrix and the fused vector is calculated to obtain the first connection strength. The second connection strength is obtained by calculating the dot product between the interaction requester node vector in the interaction requester representation matrix and the clustered interaction object node vector in the clustered interaction object representation matrix.

[0011] In some embodiments, obtaining the node representations of the interaction requester and the interaction object through pre-training includes: Based on the graph convolution operation of the first interaction graph, the node representations of the interaction requester and the interaction object corresponding to the graph convolution layer are obtained. Random noise is introduced into the node representation of the interaction requester and / or the node representation of the interaction object to obtain the perturbed node representation of the interaction requester and / or the perturbed node representation of the interaction object. By performing cross-layer comparative learning on the perturbation node representations of the interaction requester and / or the interaction object corresponding to different graph convolutional layers, the initial node representations of the interaction requester and the interaction object are obtained.

[0012] This application also provides a fine-tuning device for an information recommendation model, comprising: The acquisition unit is used to acquire a first interaction relationship graph and input it into the initial information recommendation model. The first interaction relationship graph includes a first node representing the interaction requester and a second node representing the interaction object, as well as the connection relationship between the nodes representing the interaction relationship between the interaction requester and the interaction object. The first obtaining unit is used to obtain the node representations of the interaction requester and the interaction object respectively through pre-training; The second obtaining unit is used to cluster the node representations of the interactive objects to obtain clustered interactive objects. The construction unit is configured to construct a second interaction relationship diagram representing the interaction relationship between the interaction requester and the clustering interaction object based on the clustering interaction object and the first interaction relationship diagram; The fine-tuning unit is used to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object based on the cross-layer contrastive learning loss obtained by the node representations in different layers of the first interaction graph, the recommendation loss obtained by the first interaction graph, and the recommendation loss obtained by the second interaction graph as joint constraints, so as to obtain a target information recommendation model for information recommendation.

[0013] This application also provides a product recommendation method, including: In response to a user's access request to the application platform, the system obtains the user's historical data and the application platform's product data. The historical data and the product data are input into the target information recommendation model trained by the fine-tuning method of the above information recommendation model to obtain the predicted relationship value of the connection relationship between the user and the product data; Based on the predicted relationship value, the target recommended products are obtained for recommendation to the user.

[0014] This application also provides a product recommendation device, including: The acquisition unit is used to acquire the user's historical data and the product data of the application platform in response to the user's access request; The obtaining unit is used to input the historical data and the product data into the target information recommendation model trained by the fine-tuning method of the above information recommendation model, and obtain the predicted relationship value of the connection relationship between the user and the product data; The recommendation unit is used to obtain target recommended products to recommend to the user based on the predicted relationship value.

[0015] This application also provides a computer program product, comprising: a computer program, which, when executed by a processor, implements the fine-tuning method of the information recommendation model described above, or implements the product recommendation method described above.

[0016] Compared with the prior art, this application has the following advantages: This application provides a method for fine-tuning an information recommendation model. It involves constructing a first interaction relationship graph to mine horizontal interaction relationship data between the requester and the interaction object. Then, by clustering the interaction object nodes in the interaction relationship data of the first interaction relationship graph, clustered interaction objects are obtained. Next, based on the clustered interaction objects and the first interaction relationship graph, a second interaction relationship graph is constructed to mine vertical hierarchical interaction relationship data between the requester and the clustered interaction objects. The cross-layer contrastive learning loss of the node representations of the requester and interaction object across different layers of the obtained initial information recommendation model, along with the recommendation losses obtained from the first and second interaction relationship graphs respectively, are used as joint constraints to adjust the node representations of the requester, interaction object, and the clustered interaction objects, thereby obtaining a target information recommendation model for information recommendation. When recommending information, the target information recommendation model can take the interaction relationship between the requester and the interaction object in the horizontal direction as the basis, and further explore the interaction relationship between the requester and the clustered interaction object in the vertical direction by clustering the interaction object and the second interaction relationship graph. This makes the adjusted target information recommendation model more accurate, comprehensive and insightful in recommending information and more in line with the deep potential needs of the requester. Attached Figure Description

[0017] Figure 1 This is a flowchart of a fine-tuning method for an information recommendation model provided in this application.

[0018] Figure 2 This is a schematic diagram of the structure of a fine-tuning device for an information recommendation model provided in this application.

[0019] Figure 3 This is a flowchart of a product recommendation method provided in this application.

[0020] Figure 4 This is a schematic diagram of the structure of a product recommendation device provided in this application.

[0021] Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of this application, the application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. However, this application can be implemented in many other ways different from those described below. Therefore, based on the embodiments provided in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0023] It should be noted that the terms "first," "second," "third," etc., in the claims, specification, and drawings of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. Such data are interchangeable where appropriate so that the embodiments of this application described herein can be implemented in a sequence other than that shown or described in this application. Furthermore, the terms "comprising," "having," and their variations are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0024] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0025] It should be understood that in the embodiments of this application, "B corresponding to A", "B corresponding to A", "A corresponds to B" or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0026] The fine-tuning method for the information recommendation model provided in this application can be applied to product recommendation scenarios, such as product recommendations in lifestyle shopping application service platforms and video recommendations in lifestyle entertainment application service platforms. Based on the aforementioned background technology, with the continuous development of artificial intelligence and computer technology, using artificial intelligence methods such as deep learning neural networks to achieve information recommendation has become a widely used recommendation method in recommendation systems in recent years. However, existing deep learning methods model information such as products and videos in a flat manner during the information recommendation process, resulting in recommendation information being analyzed from a horizontal or parallel perspective, exploring the relationships between information. This approach clearly lacks vertical or deep-level relationship mining, leading to unsatisfactory recommendation information.

[0027] Based on this, this application provides a method for fine-tuning an information recommendation model, so as to improve the recommendation effect in information recommendation scenarios. For example... Figure 1 As shown, the method includes: Step S101: Obtain the first interaction relationship graph and input it into the initial information recommendation model, wherein the first interaction relationship graph includes a first node representing the interaction requester and a second node representing the interaction object, as well as the connection relationship between the nodes used to represent the interaction relationship between the interaction requester and the interaction object; Step S102: Obtain the node representations of the interaction requester and the interaction object respectively through pre-training; Step S103: Cluster the node representations of the interactive objects to obtain clustered interactive objects; Step S104: Based on the clustering interaction object and the first interaction relationship diagram, construct a second interaction relationship diagram to represent the interaction relationship between the interaction requester and the clustering interaction object; Step S105: Using the cross-layer contrastive learning loss obtained from the node representations of different layers in the first interaction relationship graph of the initial information recommendation model, the recommendation loss obtained from the first interaction relationship graph, and the recommendation loss obtained from the second interaction relationship graph as joint constraints, adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object to obtain the target information recommendation model for information recommendation.

[0028] The steps S101 to S105 described above will be described in detail below.

[0029] Regarding step S101: Obtain the first interaction relationship graph and apply it to the initial information recommendation model, wherein the first interaction relationship graph includes a first node representing the interaction requester and a second node representing the interaction object, as well as the connection relationship between the nodes used to represent the interaction relationship between the interaction requester and the interaction object.

[0030] The purpose of step 101 is to construct a first interaction relationship graph based on data interaction record information. This first interaction relationship graph represents the connection relationships between nodes in an interaction requester and an interaction object. It can be a structured data graph, such as a bipartite graph. In the context of recommendation applications, it can be understood as a user-item graph, including: a first node of the interaction requester (also called an interaction requester node), such as a user node (representing a user in the application service platform); and a second node of the interaction object (also called an interaction object node), such as an item node, representing recommended objects such as goods, videos, and text in the application service platform. The data interaction record information includes information about the interaction requester, information about the interaction object, and interaction behavior information. The interaction behavior information can include actions such as purchasing, browsing, searching, collecting, and rating. The connection relationships between nodes between the interaction requester and the interaction object can be established through edges based on interaction relationships (such as interaction behaviors). In this embodiment, the first interaction relationship graph can be a user-item graph. Users represent the interaction requester, Items represent the interaction object, and the edges between them constitute the graph data structure. In this embodiment, the first interaction graph (such as a user-item graph) can be used as input data into the initial recommendation model. The initial information recommendation model in this embodiment can be a neural network model.

[0031] Regarding step S102: Obtain the node representations of the interaction requester and the interaction object respectively through pre-training.

[0032] The purpose of step S102 is to obtain node representations (also known as node embeddings or node vectors) by pre-training the first interaction graph. The specific implementation process may include: Step S102-1: Based on the graph convolution operation of the first interaction graph, obtain the node representation of the interaction requester and the node representation of the interaction object corresponding to the graph convolution layer.

[0033] Step S102-2: Introduce random noise into the node representation of the interaction requester and / or the node representation of the interaction object to obtain the perturbed node representation of the interaction requester and / or the perturbed node representation of the interaction object.

[0034] Step S102-3: Based on cross-layer comparative learning of the perturbation node representations of the interaction requester and / or the interaction object corresponding to different graph convolutional layers, the initial node representations of the interaction requester and the interaction object are obtained.

[0035] In the embodiment of step S102-1, the graph convolution operation refers to the process of processing graph data (the first interaction graph) using graph convolutional layers in a graph neural network (GNN). This process aims to learn node representations, i.e., embedding representations. By aggregating information about each node and its neighbors to capture graph structure and node features, the purpose is to: transmit node representations through the graph convolution operation to capture collaborative filtering signals between the interaction requester and the interaction object. The interaction requester node is represented as i (i.e., the user node), and the node representation can be represented as... The interactive object node is represented as j (project node), and the node representation can be expressed as... The convolutional layer in the diagram is the convolutional layer of the neural network, denoted by k. The transmission mechanism can be represented by the following formula (1): (1) Here, A is the adjacency matrix representing the first interaction graph (which can be a bipartite graph), indicating the interaction relationships between users and items. If user i interacts with item j, then Aij = 1; otherwise, it is 0. D is the degree matrix of the first interaction graph, a diagonal matrix where the diagonal element Di represents the degree of node i (i.e., the number of edges connected to it). For the first interaction graph, D can be divided into the interaction requester part and the interaction object part. E is a matrix of size (m+n)×d, where m is the number of interaction requesters (users), n is the number of interaction objects (items, items, or goods, etc.), and d is the representation dimension. Let represent the representation matrix of the k-th layer, and represent the node representation after k layers of graph convolution. Initially, It can be a randomly initialized representation or a pre-trained representation based on certain features. K represents the number of layers in the GNN. With each additional layer, the node's embedding further aggregates information from its neighbors, thereby capturing a wider range of contextual information. This indicates that the adjacency matrix A is normalized to eliminate the impact of differences in degree between different nodes, so that the contribution of each node to its neighbors is more balanced.

[0036] This formula represents the update process, or propagation process, of node representations from layer k to layer (k+1). Specifically, it describes the embedding process of layer (k+1). It is achieved by embedding the k-th layer With the normalized adjacency matrix The result is obtained by multiplication. This allows the aggregation of a node embedding with its neighboring node embeddings, thereby gradually constructing a node embedding that contains broader contextual information.

[0037] After obtaining the node representation based on step S102-1, noise is introduced into the node representation in step S102-2 to enhance the uniformity of the node representation, thereby improving the recommendation performance of the information recommendation model.

[0038] In the embodiment of step S102-2, the node representation is passed based on the graph convolution operation in step S102-1 to capture the collaborative filtering signal between the interaction requester and the interaction object, thus obtaining the node representation. Introducing random noise into the node representation can be expressed by the following formula (2): (2) in, This represents the perturbation node characterization after the introduction of random noise. Let be the introduced random noise vector. This formula represents the node representation of node i in the k-th layer. Introducing random noise Similarly, for node j in the kth layer... Introducing random noise Obtain the perturbation node characterization after introducing random noise. ,pass This represents the perturbation node characterization matrix.

[0039] After obtaining the perturbation node representation based on step S102-2, it is necessary to obtain the initial node representation through cross-layer contrastive learning, or to obtain the cross-layer contrastive learning loss, i.e., to execute step S102-3. In this process, the cross-layer contrastive learning loss can be obtained by the following formula (3): (3) in, Denotes the cross-layer contrastive learning loss, 0≤ ≤ , is a hyperparameter representing the layer used for comparative learning with the Kth layer. express The similarity between representations, exp(.) denotes the natural exponential function. These represent the positions of node i in the context of the network. Layers and in Layer node representation. Similarly, express Similarity between representations These represent the positions of node i in the context of the network. Layer node representation and node j in The layer represents the node representation. B indicates that i is taken from a mini-batch dataset. T represents the temperature parameter, used to adjust the distribution of similarity scores, making the model focus more on highly similar sample pairs. The formula aims to maximize the similarity between positive sample pairs (i.e., the representation of the same node in different views) (e.g., the numerator) while minimizing the similarity between negative sample pairs (i.e., the representations of different nodes i and j in different views) (e.g., the denominator). In other words, it compares a given node i with all nodes i. This generates discriminative node representations through contrastive learning, ultimately yielding the target node representation. The above explanation uses node i (the requester node) as an example; the same method can be used to obtain the target node representation for node j (the object node), thus obtaining the target node representation for the requester. and the target node representation of the interactive object .

[0040] Therefore, the specific implementation process of step S102-3 may include: Step S102-31: Perform cross-layer contrastive learning based on the perturbation node representations of the interactive requester and / or the perturbation node representations of the interactive object corresponding to different graph convolutional layers to obtain the cross-layer contrastive learning loss.

[0041] Step S102-32: Determine the initial node representation of the interaction requester and the initial node representation of the interaction object based on the cross-layer contrastive learning loss.

[0042] The initial node representation of the interactive requester is obtained based on the above pre-training. and the initial node representation of the interactive object Afterwards, it is necessary to reduce the dimensionality of the initial node representation and modify the initial node representation. Modeling. For example, in this embodiment, the initial node representation corresponding to the interactive object node can be represented using an 8-dimensional representation (this is just an example and not limited to 8 dimensions). The node representation of the interactive object is projected onto a two-dimensional space for data compression. This projection process can be performed according to step S103.

[0043] Regarding step S103: Clustering the node representations of the interactive objects to obtain clustered interactive objects. Based on the content of step S102 above, the node representations in step S103 can be the initial node representations.

[0044] The purpose of step S103 is to cluster the node representations of the interactive objects. Specifically, the interactive objects can be clustered according to the node representations of the interactive objects in a two-dimensional coordinate system to obtain clustered interactive objects.

[0045] Step S103-1: Project the node representation of the interactive object into a two-dimensional coordinate system; Step S103-2: In the two-dimensional coordinate system, the node representation of the interactive object is divided according to the set division direction and the number of divisions corresponding to the division direction to obtain the clustered interactive object.

[0046] In this embodiment, the two-dimensional coordinate system may be a coordinate system with hyperparameters, such as radial hyperparameters and / or angular hyperparameters. That is, the division direction in step S103-2 may include radial division direction and / or angular division direction. Therefore, the specific implementation process of step S103-2 may include: Step S103-21: Set the radial partitioning direction and / or the angular partitioning direction, and set the number of partitions corresponding to the radial partitioning direction and / or the number of partitions corresponding to the angular partitioning direction; for example, in an item node clustering graph, the hyperparameter of the radial partitioning direction can be represented by rho, i.e., ρ, and the hyperparameter of the angular partitioning direction can be represented by theta, i.e., θ. The radial partitioning direction refers to the direction of the coordinate axis radius, i.e., the direction from the origin of the coordinate axis to the node. The angular partitioning direction refers to the direction of counterclockwise rotation along the positive x-axis of the coordinate axis.

[0047] Step S103-22: According to the radial division direction and the corresponding division number, the node representation of the interactive object is divided according to similarity; and / or, according to the angular division direction and the corresponding division number, the node representation of the interactive object is divided; Specifically, this process can be implemented using the t-SNE algorithm, first calculating the similarity between nodes i and j in the high-dimensional space, then adjusting the similarity between nodes i and j in the low-dimensional space based on the similarity in the high-dimensional space, to obtain the probability (also called joint probability) that nodes i and j are neighbors in the low-dimensional space. Based on the t-SNE algorithm, the following Student t-distribution formula (4) can be used to calculate: (4) in, The node representation of two interacting objects in a high-dimensional space, such as node i and node j above, is: Nodes projected into a lower-dimensional space can be used This means, that is: The projection process needs to satisfy and the probability of nodes in low-dimensional space. The probability of approaching a node in a higher-dimensional space . Represents a node The probability of being neighbors. KL divergence (Kullback-Leibler divergence) is the method to adjust the position of nodes in the low-dimensional space by minimizing the Kullback-Leibler divergence of the similarity distribution between the high-dimensional and low-dimensional spaces. Located within the neighborhood of x and y. The number of neighborhoods can be measured by hyperparameters (e.g., perplexity). distance(·, ·) represents the distance metric, such as squared Euclidean distance. By optimizing x' and y' in two-dimensional space, interactive object nodes in high-dimensional space can be represented. ( The t-SNE algorithm is used to reduce the dimensionality of the representation and convert it into a two-dimensional visual representation. ( This allows them to be divided into different regions in a two-dimensional coordinate system according to radial and angular directions.

[0048] Step S103-23: Node representations of interactive objects belonging to the same partitioned region are clustered into a group of clustered interactive objects. Based on step S103-23... The node representations of interactive objects are clustered in two-dimensional coordinates. As mentioned earlier, this embodiment introduces radial partitioning hyperparameter ρ and angular partitioning hyperparameter θ. In this embodiment, item node clustering is performed using ρ ≥ 1 and θ ≥ 1. The two hyperparameters represent the number of partitions in different directions. This embodiment provides a reference example, where the number of radial partitions ρ = 1 and the number of angular partitions θ = 4, resulting in a total of 1 × 4 = 4 groups. The node representations of interactive objects located within a unified partition region are clustered into one group, such as... and As a group , As a group , and As a group , As a group .

[0049] The above describes the clustering process in step S103. It's understood that clustering algorithms can also include K-Means, DBSCAN, etc. The advantages of the clustering method provided in this embodiment are: first, it can determine which sector an item falls into and which category it belongs to, which is simpler, more intuitive, and easier to explain compared to the DiffPool soft clustering method (which assigns a probability to each category); second, it only requires the representation of the item itself and does not need to introduce external information such as item category labels, thus having strong versatility; and third, the partitioning method can be flexibly adjusted according to the characteristics of the dataset (in the experiment, different partitioning granularities resulted in different optimal parameters, and the optimal parameters vary based on the dataset). Therefore, flexibly adjusting the partitioning granularity according to the characteristics of the dataset can more accurately capture user preferences.

[0050] Based on the execution process of step S103 above, the clustering of interaction node representations in high-dimensional space to node representations in low-dimensional space can be obtained, and clustered interaction objects in different partitioning directions can be obtained. Based on this, a second interaction relationship graph is constructed to represent the data structure of the interaction relationship between the interaction requester and the clustered interaction object, i.e., step S104 is executed.

[0051] Regarding step S104: Based on the clustering interaction object and the first interaction relationship graph, construct a second interaction relationship graph to represent the interaction relationship between the interaction requester and the clustering interaction object.

[0052] The purpose of step S104 is to construct a second interaction relationship graph, such as a user-clustered item graph. This second interaction relationship graph represents a data structure diagram showing the interaction relationships between the interaction requester and the clustered interaction objects. Specifically, the clustered interaction object is C, which can be a set including clustered interaction object nodes, and can include one or more interaction objects. The original node set V of the first interaction graph is... Expand to Draw an edge (i.e., a connection) between nodes U and C, such as the requesting node (e.g., the user node). Clustering interaction object nodes (e.g., item clustering interaction nodes) There are connections between nodes u and c, represented by edges. Node u1 is connected to c2 because u1 is connected to i3 in the first interaction graph, and i3 belongs to c2 in the clustering interaction object. Node u2 is connected to c1 because u2 is connected to i2 in the first interaction graph, and i2 belongs to c1 in the clustering interaction object. Node u3 is connected to c3 and c4 because u3 is connected to i5 and i6 in the first interaction graph, and i5 belongs to c3 and i6 belongs to c4 in the clustering interaction object. Node u4 is connected to c1 and c3 because u4 is connected to i1 and i4 in the first interaction graph, and i1 belongs to c1 and i4 belongs to c3 in the clustering interaction object.

[0053] In this embodiment, if the interactive object is a product, it could be something like a toothbrush, toothpaste, or facial cleanser; clustered interactive objects could be like toiletries. If the interactive object is a meal, it could be something like rice, stir-fried pork with chili peppers, or boiled fish; clustered interactive objects could be like lunch, dinner, or Chinese food. Therefore, the interactive object can be a specific interactive object in different application scenarios, for example, an object with usage and / or consumption attributes. Clustered interactive objects are a higher-level category after classifying two or more interactive objects according to their common attributes, functions, and / or usage scenarios.

[0054] Therefore, the specific implementation process of step S104 includes: Step S104-1: Based on the connection relationship between the interaction requester and the interaction object in the first interaction relationship diagram, if any of the interaction objects in the clustered interaction objects has a connection relationship with the interaction requester, then establish the connection relationship between the interaction requester and the clustered interaction object.

[0055] Step S104-3: Construct the second interaction relationship graph based on the interaction requester, the clustered interaction object, and the connection relationship.

[0056] Based on the above, the first interaction graph and the second interaction graph have been obtained in this embodiment. Next, the node representations need to be adjusted according to the cross-layer contrastive learning loss, recommendation loss, etc., as joint constraints to obtain the target information recommendation model for the information recommendation side. That is, step S105 is executed.

[0057] Regarding step S105: Based on the cross-layer contrastive learning loss obtained from the node representations of different layers in the first interaction relationship graph of the initial information recommendation model, the recommendation loss obtained from the first interaction relationship graph, and the recommendation loss obtained from the second interaction relationship graph as joint constraints, the node representations of the interaction requester, the interaction object, and the clustered interaction object are adjusted to obtain the target information recommendation model for information recommendation.

[0058] The purpose of step S105 is to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object using cross-layer contrastive learning loss and recommendation loss as joint constraints, thereby obtaining a fine-tuned model for use in downstream scenarios. The specific implementation process may include: Step S105-1: Based on the first interaction relationship diagram and the second interaction relationship diagram, obtain the first connection strength between the interaction requester and the interaction object, and the second connection strength between the interaction requester and the clustered interaction object; Step S105-2: Determine the first recommendation loss of the initial information recommendation model based on the first connection strength; Step S105-3: Determine the second recommendation loss of the initial information recommendation model based on the second connection strength; Step S105-4: Based on the first recommendation loss, the second recommendation loss, and the cross-layer contrastive learning loss as joint constraints, adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object to obtain the target information recommendation model for information recommendation.

[0059] The specific implementation process of step S105-1 includes: Step S105-11: Based on the graph convolution operation of the first interaction relationship graph and the second interaction relationship graph, obtain the interaction requester representation matrix, the interaction object representation matrix, and the clustered interaction object representation matrix.

[0060] Step S105-12: Perform feature fusion between the interaction object node vector in the interaction object representation matrix and the cluster interaction object node vector of its corresponding cluster in the clustering interaction object representation matrix to obtain a fusion vector, and calculate the dot product between the interaction requester node vector in the interaction requester representation matrix and the fusion vector to obtain the first connection strength.

[0061] Step S105-13: Calculate the dot product of the interaction requester node vector in the interaction requester representation matrix and the clustering interaction object node vector in the clustering interaction object representation matrix to obtain the second connection strength.

[0062] In step S105-11, the graph convolution operation on the first and second interaction graphs can be performed in the same way as in Formula 1 above to obtain the interaction requester representation matrix. Interactive object representation matrix Clustering interaction object representation matrix U represents the set of interaction requester nodes, I represents the set of interaction object nodes, and C represents the set of clustered interaction objects.

[0063] The first connection strength in step S105-12 can be obtained using the following formula (5): (5) in, Indicates the first connection strength. This represents the node vector of the interactive requester corresponding to the i-th row in the interactive requester representation matrix. Represents node vectors The transpose of . This represents the vector of the interactive object node corresponding to the j-th row in the interactive object representation matrix. This represents the cluster interaction object node vector corresponding to the k-th row in the cluster interaction object representation matrix. This represents the cluster membership degree, that is, the degree to which the interactive object corresponding to row j belongs to the cluster of row k (or it can be a weight). If the interactive object node j belongs to cluster k, then... ,otherwise . The superscript ft (Fine-tuned) indicates the fine-tuned representation vector of a node, representing the state of the model parameters. It's important to note that fine-tuning is performed through gradient descent based on the loss; representation fine-tuning is not performed when calculating connection strength.

[0064] The second connection strength in step S105-13 can be obtained using the following formula (6): Or expressed as (6) in, or Let represent the transpose of the representation vector of the i-th interactive requester node in the interactive requester representation matrix. or This represents the representation vector of the k-th cluster interaction object node in the cluster interaction object representation matrix, i.e., the representation vector of the class node. If we use... Indicate, then This refers to the clustered interaction object node representation vector. The formula represents the connection strength obtained by taking the dot product of the interaction requester node representation vector selected from the interaction requester representation matrix and the clustered interaction object node representation vector selected from the clustered interaction object representation matrix. The representation update direction involved in model training can be to increase the connection strength of "user-item pairs" with interaction records and decrease the connection strength of "user-item pairs" without interaction records, thereby recommending items of interest to users based on the connection strength during the inference phase.

[0065] The recommendation loss involved in steps S105-2 to S105-4 can be obtained in this embodiment using the following formula (7): (7) Among them, the Represents a sigmoid function, with positive edges ( ) belongs to the edge set, negative edges ( It does not belong to the edge set.

[0066] By substituting the connection strengths in equations (5) and (6) into equation (7), we can obtain the first recommended loss for the first connection strength. The second recommended loss regarding the second connection strength .

[0067] For example: First recommendation loss It can be represented as: in, The connection strength score between the user and the positive sample items. . The connection strength score for users to negative sample items. .

[0068] For example: second recommendation loss It can be represented as: in, This represents the connection strength score between a user and positive samples of a clustered interaction object. . This represents the connection strength score between a user and negative samples of a clustered interaction object. .

[0069] Based on this, the first recommended loss can also be... Second recommended loss Cross-layer contrastive learning loss It is determined to be a mixed loss, that is: The initial information recommendation model is adjusted and iterated through a hybrid loss to obtain the target information recommendation model. Among these, cross-layer contrastive learning loss is used. It can be obtained according to the above formula (3).

[0070] This can also be understood as: determining the first recommendation loss, the second recommendation loss, and the cross-layer comparative learning loss as a hybrid recommendation loss; and determining whether the hybrid recommendation loss can meet the model convergence requirements.

[0071] Therefore, step S105-5 can be based on adjusting the node representations of the interaction requester, the interaction object, and the clustered interaction object by using the recommendation loss and cross-layer contrastive learning loss as joint constraints. After the recommendation loss of the adjusted initial information recommendation model converges, the initial information recommendation model can be determined as the target information recommendation model.

[0072] Through the first recommended loss Second recommendation loss This enables the initial information recommendation model to perform fine-grained learning not only horizontally between the requester and the interaction object, but also vertically between the requester and the clustered interaction object during the deep learning process. Furthermore, this is achieved by introducing cross-layer contrastive learning loss. This approach strongly aligns and brings node representations from different structural levels closer together, avoiding semantic biases caused by horizontal and vertical learning. Therefore, the collaborative optimization of this hybrid loss enables the final target information recommendation model to mine the deep relationships between the requester and the object of interaction at a deep level, effectively mitigating data sparsity during the inference stage and significantly improving the accuracy and comprehensiveness of information recommendation.

[0073] In step S105-5, when the recommendation loss causes the model to fail to meet the convergence requirements, the node representations of the interaction requester, the interaction object, and the clustered interaction object are adjusted by using cross-layer comparative learning loss and recommendation loss as joint constraints. This adjusts the initial information recommendation model and ultimately obtains a target information recommendation model that meets the convergence requirements, which is then used for information recommendation on the interactive service application platform.

[0074] This application provides an embodiment of a fine-tuning method for an information recommendation model. It involves constructing a first interaction relationship graph to mine horizontal interaction relationship data between the requester and the interaction object. Then, by clustering the interaction object nodes in the interaction relationship data of the first interaction relationship graph, clustered interaction objects are obtained. Next, based on the clustered interaction objects and the first interaction relationship graph, a second interaction relationship graph is constructed to mine vertical hierarchical interaction relationship data between the requester and the clustered interaction objects. The cross-layer contrastive learning loss of the node representations of the requester and interaction object across different layers of the obtained initial information recommendation model, along with the recommendation losses obtained from the first and second interaction relationship graphs respectively, are used as joint constraints to adjust the node representations of the requester, interaction object, and the clustered interaction objects, thereby obtaining a target information recommendation model for information recommendation. When recommending information, the target information recommendation model, based on the horizontal interaction relationship between the requester and the interaction object, further mines the vertical interaction relationship between the requester and the clustered interaction objects by clustering interaction objects and a second interaction relationship graph. This allows the adjusted target information recommendation model to recommend information more accurately, comprehensively, and insightfully, and to better meet the deeper potential needs of the requester. This overcomes the poor recommendation performance caused by the limited data relationship mining in existing information recommendation methods that rely on flattened modeling.

[0075] The above is a detailed description of an embodiment of a fine-tuning method for an information recommendation model provided in this application. Corresponding to the aforementioned embodiment of a fine-tuning method for an information recommendation model, this application also discloses an embodiment of a fine-tuning device for an information recommendation model. Please refer to [link / reference]. Figure 2 Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0076] like Figure 2 As shown, Figure 2 This is a schematic diagram of the structure of a fine-tuning device for an information recommendation model provided in this application. The device includes: The input data preparation unit 201 is used to obtain a first interaction relationship graph and input it into the initial information recommendation model. The first interaction relationship graph includes a first node representing the interaction requester and a second node representing the interaction object, as well as the connection relationship between the nodes representing the interaction relationship between the interaction requester and the interaction object. The first obtaining unit 202 is used to obtain the node representations of the interaction requester and the interaction object respectively through pre-training; The second obtaining unit 203 is used to cluster the node representations of the interactive objects to obtain clustered interactive objects; Construction unit 204 is configured to construct a second interaction relationship diagram representing the interaction relationship between the interaction requester and the clustering interaction object based on the clustering interaction object and the first interaction relationship diagram; The fine-tuning unit 205 is used to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object based on the cross-layer contrastive learning loss obtained by the node representations of different layers on the first interaction relationship graph of the initial information recommendation model, the recommendation loss obtained by the first interaction relationship graph, and the recommendation loss obtained by the second interaction relationship graph as joint constraints, so as to obtain a target information recommendation model for information recommendation.

[0077] In this embodiment, the first obtaining unit 202 includes: a convolution sub-unit, a noise sub-unit, and an obtaining sub-unit; The convolutional subunit is used to obtain the node representation of the interaction requester and the node representation of the interaction object corresponding to the graph convolutional layer based on the graph convolution operation of the first interaction graph. The noise subunit is used to introduce random noise into the node representation of the interaction requester and / or the node representation of the interaction object obtained in the convolution subunit, so as to obtain the perturbed node representation of the interaction requester and / or the perturbed node representation of the interaction object. The obtaining subunit is used to perform cross-layer comparative learning based on the perturbation node representations of the interaction requester and / or the perturbation node representations of the interaction object corresponding to different graph convolutional layers, so as to obtain the initial node representations of the interaction requester and the initial node representations of the interaction object.

[0078] In this embodiment, the second obtaining unit 203 is specifically used to cluster the interactive object according to the node representation of the interactive object in a two-dimensional coordinate system to obtain clustered interactive objects, including: a projection subunit and an obtaining subunit; The projection subunit is used to project the node representation of the interactive object onto the two-dimensional coordinate system; The obtaining sub-unit is used to divide the node representation of the interactive object in the two-dimensional coordinate system according to the set division direction and the number of divisions corresponding to the division direction, so as to obtain the clustered interactive object.

[0079] The obtained sub-units include: setting sub-units, dividing sub-units, and clustering sub-units; The setting subunit is used to set the radial division direction and / or the angular division direction, and to set the number of divisions corresponding to the radial division direction and / or the number of divisions corresponding to the angular division direction; The segmentation subunit is used to segment the node representation of the interactive object according to similarity based on the radial segmentation direction and the corresponding segmentation number; and / or, to segment the node representation of the interactive object according to similarity based on the angular segmentation direction and the corresponding segmentation number. The clustering subunit is used to represent nodes of the interactive objects belonging to the same partitioned region as a group of clustered interactive objects.

[0080] The construction unit 204 includes: an establishment subunit and a construction subunit; The establishment subunit is used to establish a connection relationship between the interaction requester and the interaction object based on the connection relationship between the interaction requester and the interaction object in the first interaction relationship graph. If any interaction object in the clustered interaction objects has a connection relationship with the interaction requester, then the connection relationship between the interaction requester and the clustered interaction object is established. The construction subunit is used to construct the second interaction relationship graph based on the interaction requester, the clustered interaction object, and the connection relationship.

[0081] The fine-tuning unit 205 includes: an intensity calculation subunit, a first determination subunit, a second determination subunit, and an adjustment subunit; The strength calculation subunit is used to obtain, based on the first interaction relationship graph and the second interaction relationship graph, the first connection strength between the interaction requester and the interaction object, and the second connection strength between the interaction requester and the clustered interaction object. The first determining subunit is used to determine the first recommendation loss of the initial information recommendation model based on the first connection strength; The second determining subunit is used to determine the second recommendation loss of the initial information recommendation model based on the second connection strength; The adjustment subunit is used to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object based on the first recommendation loss, the second recommendation loss, and the cross-layer contrastive learning loss as joint constraints, so as to obtain a target information recommendation model for information recommendation.

[0082] The strength calculation subunit further includes: a convolution subunit, a first calculation subunit, and a second calculation subunit. The convolution subunit is used to obtain an interaction requester representation matrix, an interaction object representation matrix, and a clustered interaction object representation matrix based on graph convolution operations on the first interaction graph and the second interaction graph. The first calculation subunit is used to perform feature fusion of the interaction object node vectors in the interaction object representation matrix with the corresponding clustered interaction object node vectors of their respective clusters in the clustered interaction object representation matrix to obtain a fused vector, and to calculate the dot product of the interaction requester node vectors in the interaction requester representation matrix and the fused vector to obtain the first connection strength. The second calculation subunit is used to calculate the dot product of the interaction requester node vectors in the interaction requester representation matrix and the clustered interaction object node vectors in the clustered interaction object representation matrix to obtain the second connection strength.

[0083] The above is a description of an embodiment of a fine-tuning device for an information recommendation model provided in this application. For details of the device embodiment, please refer to the relevant content of the above method embodiment, which will not be described in detail here.

[0084] Based on the above, this application also provides a product recommendation method, such as... Figure 3 As shown, Figure 3 This is a flowchart of a product recommendation method provided in this application, the method including: Step S301: In response to the user's access request to the application platform, obtain the user's historical data and the application platform's product data; wherein, the product data may be the application platform's full product data or the full product data within a range defined by a certain division, such as the full product data within the POI range.

[0085] Step S302: Input the historical data and the product data into the target information recommendation model trained by the fine-tuning method of the above information recommendation model to obtain the predicted relationship value of the connection relationship between the user and the product data; Step S303: Based on the predicted relationship value, obtain the target recommended product to be recommended to the user.

[0086] In this embodiment, the interactive service application platform can be a shopping service application platform, an audio-visual service application platform, a technical forum service platform, a catering service application platform, etc. The target information recommendation model can recommend products based on the ranking results of the predicted relationship values, for example, by sorting in descending or ascending order. When sorted in descending order, products ranked higher are recommended; when sorted in ascending order, products ranked lower are recommended. The predicted relationship values ​​can be the connection strength value, matching score, interaction probability value, recommendation score, preference value, or interest value between the user and the product data obtained by the target information recommendation model.

[0087] For details regarding steps S301 to S303, please refer to the descriptions of steps S101 to S105 above; they will not be repeated here.

[0088] Accordingly, this application also provides a product recommendation device, such as... Figure 4 As shown, Figure 4 This is a schematic diagram of a product recommendation device provided in this application. The device includes: The acquisition unit 401 is used to acquire the user's historical data and the product data of the application platform in response to the user's access request; The obtaining unit 402 is used to input the historical data and the product data into the target information recommendation model trained by the fine-tuning method of the above information recommendation model, and obtain the predicted relationship value of the connection relationship between the user and the product data; The recommendation unit 403 is used to obtain target recommended products to recommend to the user based on the predicted relationship value.

[0089] The description of the product recommendation device can be found in the above-mentioned steps S101 to S105 and steps S301 to S303, which will not be repeated here.

[0090] Based on the above, this application also provides a computer program product, including: a computer program, which, when executed by a processor, implements the content of the fine-tuning method of the information recommendation model, or implements the content of the product recommendation method.

[0091] Based on the above, this application also provides an electronic device, such as... Figure 5 As shown, it includes: Processor 501; The memory 502 is used to store a program for processing data generated by the electronic device. When the program is read and executed by the processor, it executes the content of the fine-tuning method of the information recommendation model, or implements the content of the product recommendation method.

[0092] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0093] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0094] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0095] 1. Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0096] 2. Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A method for fine-tuning an information recommendation model, characterized in that, include: Obtain a first interaction relationship graph and input it into the initial information recommendation model. The first interaction relationship graph includes a first node representing the interaction requester and a second node representing the interaction object, as well as the connection relationship between the nodes used to represent the interaction relationship between the interaction requester and the interaction object. Through pre-training, the node representations of the interaction requester and the interaction object are obtained respectively; Cluster the node representations of the interactive objects to obtain clustered interactive objects; Based on the clustered interaction objects and the first interaction relationship diagram, a second interaction relationship diagram is constructed to represent the interaction relationship between the interaction requester and the clustered interaction objects; The cross-layer contrastive learning loss obtained from the node representations of different layers in the first interaction graph of the initial information recommendation model, the recommendation loss obtained from the first interaction graph, and the recommendation loss obtained from the second interaction graph are used as joint constraints to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object to obtain the target information recommendation model for information recommendation.

2. The method according to claim 1, characterized in that, The process of clustering the node representations of the interactive objects to obtain clustered interactive objects includes: Project the node representation of the interactive object onto a two-dimensional coordinate system; In the two-dimensional coordinate system, the node representation of the interactive object is divided according to the set division direction and the number of divisions corresponding to the division direction to obtain the clustered interactive object.

3. The method according to claim 1, characterized in that, The step of constructing a second interaction relationship graph representing the interaction relationship between the interaction requester and the clustered interaction object based on the clustered interaction object and the first interaction relationship graph includes: Based on the connection relationship between the interaction requester and the interaction object in the first interaction relationship diagram, if any of the interaction objects in the clustered interaction objects has a connection relationship with the interaction requester, then a connection relationship between the interaction requester and the clustered interaction object is established. Based on the interaction requester, the clustered interaction object, and the connection relationship, construct the second interaction relationship graph.

4. The method according to claim 1, characterized in that, The method of adjusting the node representations of the interaction requester, the interaction object, and the clustered interaction object based on the cross-layer contrastive learning loss obtained from the node representations of different layers in the first interaction graph according to the initial information recommendation model, the recommendation loss obtained from the first interaction graph, and the recommendation loss obtained from the second interaction graph as joint constraints to obtain a target information recommendation model for information recommendation, includes: Based on the first interaction relationship diagram and the second interaction relationship diagram, the first connection strength between the interaction requester and the interaction object, and the second connection strength between the interaction requester and the clustered interaction object are obtained. Based on the first connection strength, determine the first recommendation loss of the initial information recommendation model; Based on the second connection strength, determine the second recommendation loss of the initial information recommendation model; Based on the first recommendation loss, the second recommendation loss, and the cross-layer contrastive learning loss as joint constraints, the node representations of the interaction requester, the interaction object, and the clustered interaction object are adjusted to obtain a target information recommendation model for information recommendation.

5. The method according to claim 4, characterized in that, The step of obtaining the first connection strength between the interaction requester and the interaction object, and the second connection strength between the interaction requester and the clustered interaction object based on the first interaction relationship graph and the second interaction relationship graph, includes: Based on the graph convolution operation of the first interaction graph and the second interaction graph, the interaction requester representation matrix, the interaction object representation matrix, and the clustered interaction object representation matrix are obtained. The interaction object node vector in the interaction object representation matrix is ​​fused with the cluster interaction object node vector of its corresponding cluster in the cluster interaction object representation matrix to obtain a fused vector. The dot product of the interaction requester node vector in the interaction requester representation matrix and the fused vector is calculated to obtain the first connection strength. The second connection strength is obtained by calculating the dot product between the interaction requester node vector in the interaction requester representation matrix and the clustered interaction object node vector in the clustered interaction object representation matrix.

6. The method according to claim 1, characterized in that, The step of obtaining the node representations of the interaction requester and the interaction object through pre-training includes: Based on the graph convolution operation of the first interaction graph, the node representations of the interaction requester and the interaction object corresponding to the graph convolution layer are obtained. Random noise is introduced into the node representation of the interaction requester and / or the node representation of the interaction object to obtain the perturbed node representation of the interaction requester and / or the perturbed node representation of the interaction object. By performing cross-layer comparative learning on the perturbation node representations of the interaction requester and / or the interaction object corresponding to different graph convolutional layers, the initial node representations of the interaction requester and the interaction object are obtained.

7. A fine-tuning device for an information recommendation model, characterized in that, include: The acquisition unit is used to acquire a first interaction relationship graph and input it into the initial information recommendation model. The first interaction relationship graph includes a first node representing the interaction requester and a second node representing the interaction object, as well as the connection relationship between the nodes representing the interaction relationship between the interaction requester and the interaction object. The first obtaining unit is used to obtain the node representations of the interaction requester and the interaction object respectively through pre-training; The second obtaining unit is used to cluster the node representations of the interactive objects to obtain clustered interactive objects. The construction unit is configured to construct a second interaction relationship diagram representing the interaction relationship between the interaction requester and the clustering interaction object based on the clustering interaction object and the first interaction relationship diagram; The fine-tuning unit is used to adjust the node representations of the interaction requester, the interaction object, and the clustered interaction object based on the cross-layer contrastive learning loss obtained by the node representations in different layers of the first interaction graph, the recommendation loss obtained by the first interaction graph, and the recommendation loss obtained by the second interaction graph as joint constraints, so as to obtain a target information recommendation model for information recommendation.

8. A product recommendation method, characterized in that, include: In response to a user's access request to the application platform, the system obtains the user's historical data and the application platform's product data. The historical data and the product data are input into the target information recommendation model trained by the fine-tuning method of the information recommendation model according to any one of claims 1 to 6, and the predicted relationship value of the connection relationship between the user and the product data is obtained. Based on the predicted relationship value, the target recommended products are obtained for recommendation to the user.

9. A product recommendation device, characterized in that, include: The acquisition unit is used to acquire the user's historical data and the product data of the application platform in response to the user's access request; The obtaining unit is used to input the historical data and the product data into the target information recommendation model trained by the fine-tuning method of the information recommendation model according to any one of claims 1 to 6, and obtain the predicted relationship value of the connection relationship between the user and the product data; The recommendation unit is used to obtain target recommended products to recommend to the user based on the predicted relationship value.

10. A computer program product, characterized in that, include: A computer program, when executed by a processor, implements the fine-tuning method of the information recommendation model as described in any one of claims 1-6, or the product recommendation method as described in claim 8.

Images