Multi-interest cross-domain recommendation method, model training method, system, device and medium

Abstract

The application provides a multi-interest cross-domain recommendation method, a model training method, a system, a device and a medium, and relates to the technical field of deep learning. First, target domain and source domain data are obtained, and target domain user and item embeddings and user behavior and source domain interest embeddings are respectively preprocessed. Based on a diffusion model, real and predicted noise is generated according to the target domain user, source domain interest and user behavior embeddings to adjust the model parameters, complete diffusion model training, and then use the trained diffusion model to obtain predicted interest embeddings in combination with the source domain interest and user behavior embeddings. A predicted score is obtained through a fusion network, and then the parameters are adjusted according to the predicted score and a first score to complete fusion network training. Finally, the two are combined to obtain a multi-interest cross-domain recommendation model. The model solves the distribution difference of cross-domain, makes the multi-interest transfer process more suitable for target domain information, can better capture the potential intention of the user, and improves the accuracy of recommendations to cold start users.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a multi-interest cross-domain recommendation method, model training method, system, device and medium. Background Technology

[0002] Recommendation systems help users discover relevant products and services. In recent years, improving the performance of recommendation systems has attracted widespread research attention. However, recommendation systems face challenges in providing accurate recommendations to new users who lack historical interactions. This limitation is known as the cold start problem.

[0003] Cross-domain recommendation improves recommendation performance in the target domain by transferring knowledge from a source domain with historical data to the target domain and leveraging the rich historical behavior in the source domain. Related cross-domain recommendation methods typically transfer user embeddings across domains through personalized preference bridges or finer-grained bridging functions. However, these methods often ignore distributional differences and fail to fully capture users' multiple interests. In addition, cold-start users lack behavioral data in the target domain, and most methods struggle to fully learn the impact of diverse interests on user representations from these interactions, resulting in low accuracy of recommendation results. Summary of the Invention

[0004] The main objective of this disclosure is to propose a multi-interest cross-domain recommendation method, model training method, system, device, and medium that can better capture users' potential intentions and improve the accuracy of recommendations to cold-start users.

[0005] To achieve the above objectives, one aspect of this application proposes a method for training a multi-interest cross-domain recommendation model, comprising the following steps:

[0006] Acquire target domain data and source domain data, wherein the target domain data includes user information, first item information and first rating, and the source domain data includes user information, second item information and user behavior information;

[0007] The target domain data is preprocessed to obtain target domain user embeddings and target domain item embeddings, and the source domain data is preprocessed to obtain user behavior embeddings and source domain interest embeddings.

[0008] Based on the diffusion model, the real noise and predicted noise are obtained according to the target domain user embedding, the source domain interest embedding, and the user behavior embedding;

[0009] The parameters of the diffusion model are adjusted based on the actual noise and the predicted noise to obtain the trained diffusion model;

[0010] Using the trained diffusion model, the predicted interest embedding is obtained based on the source domain interest embedding and the user behavior embedding;

[0011] Based on the fusion network, a predicted score is obtained according to the predicted interest embedding, the user behavior embedding, and the target domain item embedding;

[0012] The loss function of the fusion network is obtained based on the predicted score and the first score;

[0013] The parameters of the fusion network are adjusted according to the loss function to obtain the trained fusion network;

[0014] The trained fusion network and the trained diffusion model are combined to obtain a multi-interest cross-domain recommendation model.

[0015] In some embodiments, the preprocessing of the source domain data to obtain user behavior embeddings and source domain interest embeddings includes the following steps:

[0016] Based on the user information and the second item information, a first interest embedding is obtained;

[0017] Feature extraction is performed on user behavior information to obtain user behavior embeddings;

[0018] The source domain interest embedding is obtained based on the first interest embedding and the user behavior embedding.

[0019] In some embodiments, obtaining the first interest embedding based on the user information and the second item information includes the following steps:

[0020] Feature extraction is performed on the second item information to obtain the attribute information of the second item information;

[0021] Based on the relationship between the user information, the second item information, and the attribute information, multiple meta-paths are constructed.

[0022] Multiple meta-paths are aggregated to obtain the first interest embedding.

[0023] In some embodiments, obtaining the source domain interest embedding based on the first interest embedding and the user behavior embedding includes the following steps:

[0024] The weights of the first interest embedding are obtained through an attention network based on the first interest embedding and the user behavior embedding.

[0025] Using a meta-network, the parameters of the bridging function are obtained based on the weights of the first interest embedding;

[0026] The source domain interest embedding is obtained based on the first interest embedding and the parameters of the bridging function.

[0027] In some embodiments, obtaining real noise and predicted noise based on the diffusion model, according to the target domain user embedding, the source domain interest embedding, and the user behavior embedding, includes the following steps:

[0028] Obtain Gaussian noise;

[0029] The Gaussian noise is sampled to obtain the real noise;

[0030] Add the real noise to the target domain user embedding to obtain a second user embedding;

[0031] Add the real noise to the source domain interest embedding to obtain the second interest embedding;

[0032] The approximator of the diffusion model obtains prediction noise based on the second user embedding, the second interest embedding, and the user behavior embedding, wherein the prediction noise is used to denoise the second interest embedding to obtain the prediction interest embedding.

[0033] In some embodiments, obtaining the predicted score based on the predicted interest embedding, the user behavior embedding, and the target domain item embedding, using a fusion network, includes the following steps:

[0034] Obtain the parameters of the fusion network;

[0035] Based on the predicted interest embedding and the fusion network parameters, the attention score of the predicted interest embedding is obtained;

[0036] Based on the attention score and the predicted interest embedding, a third interest embedding is obtained;

[0037] The weights of the predicted interest embedding are obtained based on the user behavior embedding and the third interest embedding.

[0038] Based on the weights and the predicted interest embedding, the target user embedding is obtained;

[0039] A predicted score is obtained based on the target user embedding and the target domain item embedding.

[0040] On the other hand, embodiments of the present invention propose a multi-interest cross-domain recommendation method, including the following steps:

[0041] Retrieve the user's source domain data and target domain items;

[0042] The source domain data is preprocessed to obtain source domain interest embedding and user behavior embedding;

[0043] Feature extraction is performed on the target domain items to obtain the target domain item embeddings;

[0044] The source domain interest embedding, the user behavior embedding, and the target domain item embedding are input into the multi-interest cross-domain recommendation model to obtain the predicted rating of the user's target domain item. The multi-interest cross-domain recommendation model is obtained by the multi-interest cross-domain recommendation model training method described in any of the preceding embodiments.

[0045] On the other hand, embodiments of the present invention propose a multi-interest cross-domain recommendation model training system, comprising:

[0046] The first module is used to acquire target domain data and source domain data, wherein the target domain data includes user information, first item information and first rating, and the source domain data includes user information, second item information and user behavior information;

[0047] The second module is used to preprocess the target domain data to obtain target domain user embeddings and target domain item embeddings, and to preprocess the source domain data to obtain user behavior embeddings and source domain interest embeddings.

[0048] The third module is used to obtain real noise and predicted noise based on the diffusion model, according to the target domain user embedding, the source domain interest embedding, and the user behavior embedding;

[0049] The fourth module is used to adjust the parameters of the diffusion model based on the real noise and the predicted noise to obtain the trained diffusion model;

[0050] The fifth module is used to obtain the predicted interest embedding based on the source domain interest embedding and the user behavior embedding using the trained diffusion model;

[0051] The sixth module is used to obtain a predicted score based on the fusion network, according to the predicted interest embedding, the user behavior embedding, and the target domain item embedding;

[0052] The seventh module is used to obtain the loss function of the fusion network based on the predicted score and the first score;

[0053] The eighth module is used to adjust the parameters of the fusion network according to the loss function to obtain the trained fusion network;

[0054] The ninth module is used to combine the trained fusion network and the trained diffusion model to obtain a multi-interest cross-domain recommendation model.

[0055] On the other hand, embodiments of the present invention provide an electronic device comprising:

[0056] At least one processor;

[0057] At least one memory for storing at least one program;

[0058] When the at least one program is executed by the at least one processor, the at least one processor implements the multi-interest cross-domain recommendation model training method or the multi-interest cross-domain recommendation method as described in the preceding embodiments.

[0059] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions, which are used to cause a computer to execute the multi-interest cross-domain recommendation model training method or the multi-interest cross-domain recommendation method as described in the preceding embodiments.

[0060] The technical solution described above has at least one of the following advantages or beneficial effects: By modeling the distribution of user interests using a diffusion model, the distribution characteristics of interests are gradually transferred from the source domain to the target domain, ensuring distribution alignment. This solves the problem of cross-domain distribution differences and enables more effective transfer of multiple user interests. Furthermore, user interests are reconstructed in the target domain, thereby effectively capturing the user's potential intent. The fusion network aggregates different interests, making the multi-interest transfer process more adaptable to target domain information, thus further improving the performance of the fusion process and enhancing the accuracy of recommendations to users. Attached Figure Description

[0061] Figure 1 This is a flowchart of the multi-interest cross-domain recommendation model training method provided in the embodiments of this application;

[0062] Figure 2 This is a schematic diagram of a heterogeneous network structure provided in an embodiment of this application;

[0063] Figure 3 This is a schematic diagram illustrating the impact of multiple interests on cross-domain recommendations, provided in an embodiment of this application.

[0064] Figure 4 This is a schematic diagram of the multi-interest cross-domain recommendation model structure provided in the embodiments of this application;

[0065] Figure 5 This is a schematic diagram of meta-path aggregation provided in an embodiment of this application;

[0066] Figure 6 This is a flowchart illustrating the operation of the diffusion model provided in an embodiment of this application;

[0067] Figure 7 This is a flowchart of the multi-interest cross-domain recommendation method provided in the embodiments of this application;

[0068] Figure 8 This is a flowchart illustrating the operation of the multi-interest cross-domain recommendation model provided in this application embodiment;

[0069] Figure 9These are experimental results of DICDR generalization based on different neural networks provided in the embodiments of this application;

[0070] Figure 10 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0071] 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.

[0072] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0073] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0074] First, let's analyze some of the terms used in this application:

[0075] Diffusion models are a class of deep learning methods based on probabilistic generative models, which have achieved significant results in recent years in fields such as image generation, speech synthesis, and text generation. Diffusion models simulate a process of gradually transforming a data distribution into a Gaussian noise distribution (forward diffusion), and then reconstruct high-quality data samples from the noise by learning the inverse process (backward diffusion).

[0076] Cross-Domain Recommendation (CDR) is a technique that utilizes transfer learning in recommender systems to address the cold start and data sparsity problems, thereby improving the performance and diversity of recommender systems. CDR improves recommendation performance by leveraging source domain information from rich data to assist the target domain of sparse data.

[0077] In recent years, improving the performance of recommender systems has attracted widespread research attention. However, recommender systems face challenges in providing accurate recommendations to new users who lack historical interactions; this limitation is known as the cold start problem. Cross-domain recommendation offers a feasible solution to this problem by transferring knowledge from a source domain with historical data to a target domain.

[0078] Mapping functions are the most common method in cross-domain recommendation, establishing a paradigm for transferring user embeddings using bridging functions. However, these mapping-based methods often overlook the diversity of user preferences. Therefore, increasing research focuses on improving performance by mining diverse user interests. Among these efforts, to more intuitively represent the relationship between users and items, many studies attempt to capture diverse user interests by constructing Heterogeneous Information Networks (HINs). Figure 2 As shown, Figure 2 This study demonstrates a heterogeneous network structure. By aggregating multidimensional interactions using meta-paths in HIN, researchers are able to reveal complex relationships and gain richer insights into user preferences. Interactions involving users, items, and brands can be represented as meta-paths, which reflect different impacts on user interests.

[0079] Existing methods often fail to consider the inherent differences in the distribution of user interests across domains, which limits their ability to effectively generalize to the target domain. For example, ... Figure 3 As shown, Figure 3 This illustrates the impact of different interests on target outcomes. Existing methods, when recommending movies to NBA enthusiasts, often emphasize basketball themes, resulting in primarily basketball-related music recommendations while neglecting users' potential preferences for other vibrant music genres. Secondly, due to the diversity of user interests, these methods struggle to accurately capture user intent within the target domain, especially for new users. Consequently, the items recommended by these methods often lack novelty and diversity.

[0080] Based on this, the embodiments of this disclosure provide a multi-interest cross-domain recommendation method, model training method, system, device and medium, which can provide more accurate recommendations for users with cold start in the target domain, capture users' potential intentions and make recommendations more in line with actual interests.

[0081] Reference Figure 1 As shown, Figure 1 This is an optional flowchart of a multi-interest cross-domain recommendation model training method provided in some embodiments of this application. A multi-interest cross-domain recommendation model training method of this invention includes, but is not limited to, steps S100 to S900.

[0082] Step S100: Obtain target domain data and source domain data, wherein the target domain data includes user information, first item information and first rating, and the source domain data includes user information, second item information and user behavior information;

[0083] Step S200: Preprocess the target domain data to obtain target domain user embedding and target domain item embedding, and preprocess the source domain data to obtain user behavior embedding and source domain interest embedding.

[0084] Step S300: Based on the diffusion model, obtain the real noise and predicted noise according to the target domain user embedding, source domain interest embedding, and user behavior embedding;

[0085] Step S400: Adjust the parameters of the diffusion model based on the actual noise and the predicted noise to obtain the trained diffusion model;

[0086] Step S500: Using the trained diffusion model, the predicted interest embedding is obtained based on the source domain interest embedding and the user behavior embedding.

[0087] Step S600: Based on the fusion network, a predicted score is obtained according to the predicted interest embedding, user behavior embedding, and target domain item embedding;

[0088] Step S700: Based on the predicted score and the first score, obtain the loss function of the fusion network;

[0089] Step S800: Adjust the parameters of the fusion network according to the loss function to obtain the trained fusion network;

[0090] Step S900: Combine the trained fusion network and the trained diffusion model to obtain a multi-interest cross-domain recommendation model.

[0091] In step S100 of some embodiments, the data is divided into a source domain and a target domain, and the data includes a set of users u∈U d A set of items i∈I d and a set of ratings r ui ∈R d Where d∈{s,t} represent the source domain(s) and the target domain(t), respectively, and the target domain can be represented as U. t The source domain user can be represented as U. s For overlapping users, it is defined as u o ∈U s ∩U t User information can be represented as overlapping user u o Items in the target domain and the source domain typically do not overlap; items in the target domain are represented as I. t This refers to the first item information; the source domain item is represented as I. s This refers to the second item information. The rating can represent a user's level of liking for an item; the rating of a target domain user for a target domain item can be represented as R. t This is the first rating; the rating of source domain users for source domain items can be represented as R. sFor each user, some historical behaviors can be found in the source domain, represented as... This historical behavior can be represented as user behavior information, where n represents the number of interacted items in the source domain. In the pre-trained model trained in the target domain, and Let represent the embeddings of the i-th user and the j-th item, respectively, where n represents the dimension of all embeddings.

[0092] In step S200 of some embodiments, the target domain data is preprocessed to obtain target domain user embeddings and target domain item embeddings. The target domain user embedding is a vector that can be obtained by feature extraction from users in the target domain and their interaction behavior with items (such as ratings) using an encoder or similar method. It can represent the user's overall preferences in the target domain, summarizing the user's interests and behaviors. The target domain item embedding is a vector representation of the target domain items, which can be feature extracted using an encoder or similar method. For example, the target domain item embedding can represent the item's category, brand, tags, etc. The source domain data is preprocessed to obtain user behavior embeddings and source domain interest embeddings. The user behavior embedding is a vector that can represent the user's historical behavioral characteristics in the source domain. Historical behavior can be records of the user's purchases, searches, browsing, etc. The source domain interest embedding is a vector that can represent the user's specific preferences for a particular item in the source domain.

[0093] In some embodiments, step S200 may include, but is not limited to, steps S210 to S230:

[0094] Step S210: Obtain the first interest embedding based on user information and second item information;

[0095] Step S220: Extract features from user behavior information to obtain user behavior embedding;

[0096] Step S230: Obtain the source domain interest embedding based on the first interest embedding and the user behavior embedding.

[0097] Please refer to Figure 4 , Figure 4This diagram illustrates the overall architecture of a multi-interest cross-domain recommendation model. The source domain user interest is represented by the first interest embedding, obtained through meta-path aggregation. User history in the source domain contains rich information, such as actions on different items (purchases, browsing, ratings, etc.) and interactions with other users. The behavior encoder processes and extracts features from this complex and diverse historical behavior data, transforming it into a specific vector representation (user behavior embedding). Through encoding, it extracts key features and patterns, allowing this information to exist in a form more suitable for subsequent computation. The bridging function, building upon this foundation, uses the output of the behavior encoder to influence the calculation of attention scores for different items, thereby affecting the final personalized interest encoding (source domain interest embedding). This better reflects the user's actual interests and preferences, improving the personalization of the recommendation.

[0098] In some embodiments, step S210 may include, but is not limited to, steps S211 to S213:

[0099] Step S211: Extract features from the second item information to obtain the attribute information of the second item information;

[0100] Step S212: Construct multiple meta-paths based on the relationship between user information, second item information, and attribute information;

[0101] Step S213: Aggregate multiple meta-paths to obtain the first interest embedding.

[0102] In some embodiments, steps S211 to S213 include a description of the appearance and content of the source domain item. Key information can be extracted from this information to obtain the source domain item's related attribute information, including but not limited to the item's category and brand. Please refer to [reference needed]. Figure 5 It defines multiple meta-paths to represent the influence of factors such as item category, brand, and other users, and obtains user interest representations from multiple dimensions by aggregating information at the node and path levels.

[0103] In some embodiments, four types of nodes are defined: User U s Item I s Category C s And Brand B s Each item has a category c i ∈C s And a brand b i ∈B s A heterogeneous network was constructed using these nodes. By combining these nodes, various meta-paths can be built. Specifically, relationships exist between users and items, items and categories, and items and brands, constructing a set of meta-paths {ρ1, ρ2, ..., ρ...}.k For each user, different meta-paths are aggregated from this set, where each meta-path generates a unique embedding that captures specific aspects of the user's interests.

[0104] To improve efficiency, selecting neighbors based on meta-paths for aggregation is crucial. For example... Figure 5 As shown, from the user Initially, selecting n meta-paths can generate at most n meta-path-based neighbors, each neighbor... Belongs to N ρ The aggregation process includes two types of nodes: the user's own nodes and neighboring user nodes. The attention score for each meta-path is defined as shown in Equation (1).

[0105]

[0106] Among them, f ρ (·,·) is a linear layer. Let N represent the end user of each meta-path. Starting from a given user, a set of neighboring nodes will be determined based on different meta-paths. ρ .

[0107] The embedding of each metapath is obtained by assigning weights to each neighbor node and summing them, as shown in Equation (2).

[0108]

[0109] Where σ represents the activation function.

[0110] By evaluating the importance of each meta-path, all meta-paths related to the current user are integrated to generate the final representation of the user's interests, as shown in Equation (3).

[0111]

[0112] in, The importance of each path is indicated by equations (4) and (5).

[0113]

[0114] Among them, W p W is the weight matrix, b is the bias matrix, and in the linear layer, W... p Both b and W are essentially learnable parameter matrices. p Used to perform linear transformations and add offsets to the input features; specifically, W p With input Multiply, input Map it to a new space, and add the appropriate offset to it using b.

[0115] q is a semantic-level attention vector, a learnable parameter vector. Specifically, q is dot-producted with the hidden representation after a non-linear transformation tanh to compute an attention score for each input, indicating the importance of each input in the final representation.

[0116] Ultimately, a group of interests were obtained. ,in, Indicates user u i The embedding vector of the j-th interest. Each interest embedding In space The expression in the diagram represents the number of interests, where k represents the number of interests and n represents the dimension of the embedding. The first interest embedding is obtained from the interest embeddings of multiple users, meaning that the first interest embedding contains multiple E... i The first interest embedding is represented as Figure 4 Source domain user interests.

[0117] In some embodiments, step S230 may include, but is not limited to, steps S231 to S233:

[0118] Step S231: Obtain the weight of the first interest embedding based on the first interest embedding and the user behavior embedding through the attention network;

[0119] Step S232: Using a meta-network, the parameters of the bridging function are obtained based on the weights of the first interest embedding;

[0120] Step S233: Obtain the source domain interest embedding based on the parameters of the first interest embedding and the bridging function.

[0121] In step S210, the interest set of the source user is obtained through meta-path aggregation. In some embodiments, steps S231 to S233 are performed via an interest-personalized bridging function. It is encoded, where j represents the j-th bridging function. Specifically, since different items contribute differently to interest, different items are weighted through an attention mechanism, as shown in equation (6).

[0122]

[0123] in, Indicates user u i The personalized weight of the j-th interest, where the weight of the first interest embedding represents the weight of multiple interests of each user embedded in the first interest, including multiple interests of different users. a l It is an item v l Attention score. Attention score a l The result is obtained from equation (7).

[0124] a l =Softmax(h(v) j ;φ h )), (7)

[0125] Where h(·) represents the attention network, φ h The h(·) parameter represents the parameter.

[0126] Will get The input is fed into a meta-network to obtain the bridging function. parameters As shown in equation (8).

[0127]

[0128] Where f e (·) is a meta-network, defined by parameter φ. e Parameterization. The meta-network is a two-layer feedforward network. It is a vector whose size depends on the structure of the bridging function.

[0129] To accommodate the size of the bridging parameters, the vector Remodeling into a matrix Finally, the personalized interest embedding is obtained, as shown in Equation (9).

[0130]

[0131] in This represents the interest embedding through personalized bridging encoding. Through the aforementioned multiple personalized bridging mechanisms, the personalized interests of the i-th user can be obtained. Used for interest transfer in input diffusion models.

[0132] In step S300 of some embodiments, please refer to Figure 4 The part within the dashed box represents the diffusion model, while the rest belongs to the fusion network.

[0133] During the diffusion phase, the diffusion model gradually transforms the data into pure noise by progressively adding noise. For input x0, at each step t of the diffusion process, a new data point is generated by adding noise, which can be represented by equation (10).

[0134]

[0135] in, The mean is Gaussian distribution, β tI represents the variance of the noise, and Ii represents the identity matrix, indicating that the added noise is isotropic. The noise injection at each step is scheduled by a predefined β. t Control, where t represents the current diffusion step.

[0136] Since the diffusion process follows a Markov chain, each state x t Depends only on the previous state x t-1 Thus, by utilizing the properties of Markov chains, x can be directly derived from the initial input x0. t This relationship is represented as shown in equation (11).

[0137]

[0138] in And α t =1-β t .

[0139] In the reverse process, by gradually removing noise, from pure noise x t The original data x0 is recovered. Specifically, for the current denoising step x... t Next step x t-1 The reverse process can be represented as shown in equations (12) and (13).

[0140]

[0141] Here, x0 is unknown and needs to be predicted using a neural network.

[0142] The reverse process, which is approximately the forward process, is shown in equation (14).

[0143]

[0144] Where, μ θ (x t ,t) is the predicted mean, ∑ θ (x t ,t) is the variance of the prediction, both of which are learned by a neural network with parameter θ. The predicted noise is used to denoise the noisy data during the reverse process.

[0145] In some embodiments, step S300 may include, but is not limited to, steps S310 to S350:

[0146] Step S310: Obtain Gaussian noise;

[0147] Step S320: Sample the Gaussian noise to obtain the real noise;

[0148] Step S330: Add real noise to the target domain user embedding to obtain the second user embedding;

[0149] Step S340: Add real noise to the source domain interest embedding to obtain the second interest embedding;

[0150] Step S350: Using the approximator of the diffusion model, prediction noise is obtained based on the second user embedding, the second interest embedding, and the user behavior embedding. The prediction noise is used to denoise the second interest embedding to obtain the prediction interest embedding.

[0151] In some embodiments, please refer to steps S310 to S350. Figure 6 Gaussian noise refers to a type of noise whose probability density function follows a Gaussian distribution (i.e., a normal distribution). Gaussian noise sequences are generated using tools such as the `numpy.random.normal` function in Python, according to a set mean and standard deviation. The generated Gaussian noise is a continuous distribution, but the model needs specific samples as real noise at each diffusion time step. Therefore, random sampling is performed from the generated Gaussian noise. The samples obtained at each sampling time step may differ, and this randomness ensures the diversity and uncertainty of the diffusion process. For example, at a certain time step, a sample is drawn from the generated Gaussian noise; this is the real noise for that time step, and it will be used to add noise to the target domain user embedding and the source domain interest embedding in subsequent steps.

[0152] The acquisition of prediction noise primarily relies on the approximator of the diffusion model. The approximator typically consists of a complex neural network structure, such as a multilayer perceptron (MLP). It receives information such as a second user embedding (the target domain user embedding after adding real noise), a second interest embedding (the source domain interest embedding processed with real noise), and user behavior embeddings. Internally, the approximator processes these inputs through a series of linear transformations and nonlinear activation functions, ultimately outputting prediction noise. Utilizing this prediction noise effectively removes the noise added to the source domain interest embedding, recovering user interest information that more closely approximates reality, providing a reliable basis for subsequent recommendation tasks.

[0153] During the forward pass, Gaussian noise is gradually added to the user embedding in the target domain. Let... Represents multiple user embeddings of the i-th user in the target domain. The splicing process is as follows: According to formula (11), the process of adding noise at each step is shown in formula (15).

[0154]

[0155] Where t represents the time step of the diffusion process. To ensure that the reverse process does not start from complete noise, a smaller α is used. t This is used to control the noise level. Through this process, a user representation U consisting of random noise is ultimately obtained.i .

[0156] In the reverse process, an approximator is trained to progressively remove noise and recover the original embedding representation. Noise is added to the previously extracted user interest at time step t to simulate the interest reconstruction process.

[0157] This embodiment employs a classifier-free guided method to generate samples aligned with user embeddings in the target domain. This method calculates the difference between the conditional approximator and the unconditional approximator, i.e., ∈ θ (x t |y) and ∈ θ (x t This ensures that the final embedding more accurately captures user interests. The score function without classifier guidance is defined as shown in Equation (16).

[0158]

[0159] in It is a conditional embedded system, which is composed of user behavior information in the source domain controlled by an attention mechanism.

[0160] Furthermore, using This indicates that according to formula (11) to The noise signal obtained after adding noise, where The overlapping multiple interests of the i-th user were encoded through personalized bridging. This was derived from the noise embedding representation. To begin, noise is gradually removed to generate interest embeddings, as shown in Equation (17).

[0161]

[0162] Among them, z t The noise is sampled from a standard normal distribution. In each denoising step, noise predictions are adjusted using a classifier-free approach to ensure the generated embeddings are aligned with the target user's interests. To accelerate the reverse process, a DPM-solver is used as the sampler, thereby improving generation efficiency and reducing model complexity while eliminating the need for an external classifier.

[0163] In this embodiment, a multilayer perceptron (MLP) is used as the backbone network of the approximator, with two zero-linear layers at the outermost layer. The temporal encoder generates a sinusoidal position embedding for each time step, while the conditional coding layer generates the conditional transition input for the diffusion model. Specifically, this is based on user behavior information in the source domain. As conditional information, both the time layer and the conditional layer are implemented as linear layers to simplify the model. This combination can accurately predict the noise added during the diffusion process.

[0164] In the diffusion model, after obtaining personalized codes, noise is added and sampling is performed to obtain interests that include the distribution of the target domain. The fusion network then computes a weighted fusion of multiple interests to generate the final user representation for prediction.

[0165] In step S400 of some embodiments, a simplified loss function can be defined to train the diffusion model. This is achieved by minimizing the difference between the actual noise and the model's predicted noise at each step; the loss function... As shown in equation (18).

[0166]

[0167] in, Represents real noise, ∈ θ (x t ,t) represents the noise predicted by the model. The goal is to enable the model to learn an accurate reverse process to denoise the data.

[0168] In this embodiment, θ represents all parameters of the diffusion model. A time step t is randomly selected, where t∈[0,T], and the noise scheduling β is incorporated into the target domain user embedding U. i Then an approximator is used to predict the noise within it. The difference between the actual noise and the predicted noise is represented by the definition shown in equation (19).

[0169]

[0170] By adjusting all parameters of the diffusion model, the difference between real noise and predicted noise is reduced, enabling the interest embedding generated by the diffusion model to align with the interests of the target user, thus achieving consistency in interests.

[0171] In step S500 of some embodiments, the personalized interests of the i-th user are input into the trained diffusion model. and corresponding user behavior It can be obtained Predicted interest embeddings are obtained from H of multiple users.

[0172] In step S600 of some embodiments, for cold-start users, the lack of sufficient information in the target domain means that the generation of the target user embedding still heavily depends on the regulation of the source domain user interests. Therefore, this problem is solved by encoding the target domain interests and guiding the source domain interactions.

[0173] In some embodiments, step S600 may include, but is not limited to, steps S610 to S630:

[0174] Step S610: Obtain the parameters of the fusion network;

[0175] Step S620: Obtain the attention score of the predicted interest embedding based on the predicted interest embedding and the fusion network parameters;

[0176] Step S630: Obtain the third interest embedding based on the attention score and the predicted interest embedding;

[0177] Step S640: Obtain the weights of the predicted interest embedding based on the user behavior embedding and the third interest embedding;

[0178] Step S650: Obtain the target user embedding based on the weights and the predicted interest embedding.

[0179] Step S660: Obtain the predicted score based on the target user embedding and the target domain item embedding.

[0180] In steps S610 to S650 of some embodiments, the diffusion model is obtained from the trained model. Each interest embedding Based on interest embedding, attention scores can be calculated to determine the weight of each interest.

[0181] pass This represents the different forms of interest expressed in user embeddings, by Forming a third interest embedding, in which, The definition is shown in equation (20).

[0182]

[0183] in, It is a parameter that needs to be learned; it is a parameter of the fusion network. This is represented as an attention score.

[0184] To compute the user embedding in the target domain, guidance generated by source domain signal alignment is also required; therefore, f is used. h (·;·;φ h To calculate the source domain Interaction history for target domain interest The effect is shown in equation (21).

[0185]

[0186] in, This represents the weight of each interest under the guidance of the source domain, for f h Two linear layers are used, with parameter φ. h To learn about the differences in interests among different users. Composed of multiple projects This can be represented as user behavior information, and the weights of the predicted interest embedding can be represented as multiple [weights] from different users. composition.

[0187] Finally, for all Perform a weighted summation to obtain the target user embedding, which contains multiple... As shown in equation (22).

[0188]

[0189] in, β j These are the weights of the interest embedding after each transformation, which can be learned through training.

[0190] In step S660 of some embodiments, when calculating the predicted score, specific mathematical operations are typically used to combine the two embeddings. The feature information contained in the user embedding and the item embedding, after specific operations, can reflect the user's potential rating tendency for the item. The predicted score can provide a decision basis for the recommendation system, ranking items in the target domain according to their scores and recommending higher-scoring items to the user, thereby achieving the purpose of cross-domain recommendation, improving the efficiency of users discovering items of interest, and improving the performance of the recommendation system.

[0191] In step S700 of some embodiments, in order to improve training performance, a mapping-based and task-based optimization method is adopted, directly using the performance of the final recommendation task as the optimization objective. φ represents the parameters of the fusion network. The mapping-based optimization loss is defined as shown in equation (23).

[0192]

[0193] in, It is the target user embedding generated by the fusion network. This method ensures It is embedded closer to the target user.

[0194] Furthermore, this embodiment focuses on the scoring task, and the loss function based on the task is defined as shown in equation (24).

[0195]

[0196] in, r represents the interactions of overlapping users in the target domain. ui This is the actual score, u t (i t ) T It is a predicted score.

[0197] because and Loss taskAll of these factors will affect the optimization of the fusion network, and the final loss function is expressed as shown in equation (25).

[0198]

[0199] Where λ is a hyperparameter.

[0200] In step S800 of some embodiments, an optimization algorithm is typically used to adjust the parameters of the fusion network according to the loss function. The optimization algorithm updates the parameters based on the gradient information of the loss function.

[0201] After multiple rounds of training and parameter tuning, the loss function is minimized, enabling the fusion network to better learn the relationship between information in the source and target domains. It can more accurately generate predicted scores by utilizing predicted interest embeddings, user behavior embeddings, and target domain item embeddings, thereby improving the performance of the entire multi-interest cross-domain recommendation model and enabling it to provide better recommendation results even when facing complex scenarios such as cold-start users.

[0202] In step S900 of some embodiments, combining the trained fusion network and diffusion model can give full play to the advantages of both and achieve more accurate cross-domain recommendations.

[0203] The trained diffusion model plays a crucial role in handling the distribution transfer of user interests. It receives interest embeddings encoded by a personalized bridging function, adds Gaussian noise to the target domain user embeddings during the forward pass, and progressively removes noise using source domain interaction information during the backward pass. This achieves effective transfer and distribution alignment of multiple user interests, while simultaneously outputting the denoised interest embeddings to the fusion network. For example, in cross-domain recommendation from the movie domain to the music domain, the diffusion model can capture users' interests in different movie genres (such as action movies, comedies, etc.) within the movie domain and transfer this interest distribution information to the music domain, providing a richer interest foundation for subsequent recommendations.

[0204] The trained fusion network focuses on dynamically generating accurate user embeddings and using source domain interactions to guide the calculation of interest weights. It encodes different interests through a personalized meta-network, fully considering the differences between various user interests, and inputs the interests encoded by the personalized bridging function into the diffusion model. Before generating the target user embedding, the fusion network receives the interest embeddings output by the diffusion model and can accurately assign weights to different interests based on historical behavioral information from the source domain. For example, in a music recommendation scenario, if a user's preference for a specific type of movie in the source domain has a potential correlation with certain music styles, the fusion network can use this information to adjust the calculation of the user's interest weights for music, thereby generating a target user embedding that better reflects the user's actual interests.

[0205] The processed and transferred interest embedding information provided by the diffusion model is input into the fusion network, providing a more comprehensive and accurate foundation for the fusion network to calculate user interest weights and generate target user embeddings. The fusion network, based on the output of the diffusion model and information such as target domain item embeddings, further optimizes the user embeddings and ultimately generates predicted scores, achieving cross-domain recommendation.

[0206] Please refer to Figure 7 , Figure 7 The multi-interest cross-domain recommendation method provided in this embodiment of the invention includes, but is not limited to, steps S910 to S940.

[0207] Step S910: Obtain the user's source domain data and target domain items;

[0208] Step S920: Preprocess the source domain data to obtain source domain interest embedding and user behavior embedding;

[0209] Step S930: Extract features from the target domain items to obtain the target domain item embeddings;

[0210] Step S940: Input the source domain interest embedding, user behavior embedding, and target domain item embedding into the multi-interest cross-domain recommendation model to obtain the predicted rating of the user's target domain item. The multi-interest cross-domain recommendation model is obtained by the multi-interest cross-domain recommendation model training method as described in the previous embodiment.

[0211] In steps S910 to S940 of some embodiments, by integrating source domain data and target domain item information, a precise predicted score is generated using multi-interest cross-domain recommendation. The predicted score can accurately reflect the user's potential preference for items in the target domain, thereby providing a decision-making basis for the recommendation system, realizing efficient multi-interest cross-domain recommendation, effectively alleviating the cold start problem, and improving the universality and accuracy of the recommendation system in different domains.

[0212] In some embodiments, such as Figure 8 As shown, a music album is rated. In the source film domain, users are interested in genres such as action, family, and romance, which can be categorized into different film genres or actors. Different source path information can be interpreted as the influence of film actors, users, and actor types on users. After migrating through different source path information, it is clear that film actors do not effectively guide the generation of user information in the target music domain.

[0213] In some embodiments, a series of experiments were conducted to evaluate the performance and robustness of the multi-interest cross-domain recommendation model.

[0214] First, we obtain the dataset for the experiment. The Amazon review dataset is one of the most widely used public datasets in e-commerce recommendation systems, suitable for evaluating various recommendation algorithms. This experiment uses the Amazon-5 Core dataset, which ensures that each user and item has at least five rating records. This characteristic guarantees the richness and diversity of the data, which helps improve the generalization ability of the recommendation model.

[0215] Table 1 shows the dataset and scenario details. For the cross-domain recommendation experiment, three popular categories were selected: movies and TV, CDs and Vinyl, and books. Based on these categories, three specific cross-domain recommendation scenarios were defined: Scenario 1: Movies → Music, Scenario 2: Books → Movies, and Scenario 3: Books → Music. To challenge the cold start problem from different perspectives, different cold start ratios were set for test users in the three scenarios. Specifically, the proportion of test users β was set to 20%, 50%, and 80%, with the remaining users used for training. As β increases, the recommendation task becomes more challenging.

[0216] Table 1. Statistics on cross-domain recommendation scenarios

[0217]

[0218] Here, overlap represents the number of overlapping users, and proportion represents the percentage of overlapping users in the total number of users.

[0219] The Amazon reviews dataset contains rating data (0-5 stars), and the mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate performance on the test set, as shown in Equations (26) and (27).

[0220]

[0221] Where Φ represents the test set, r i，j and These represent the actual rating and the predicted rating, respectively.

[0222] The best results are shown in bold, and * indicates the paired t-test result of DICDR and the best baseline at the 0.05 level.

[0223] TGT is a simple target model specifically trained on target domain data;

[0224] CMF is trained using all overlapping users in the source and target domains;

[0225] EMCDR uses matrix factorization to learn embeddings, and then transfers the user embeddings from the source domain to the target domain through a network;

[0226] SSCDR proposes a CDR framework based on semi-supervised mapping with a shared bridging function that is trained using overlapping users or items.

[0227] DCDCSR is a bridging-based approach that calculates differences between domains from the perspective of a user or item.

[0228] LACDR employs an encoder-decoder structure and aligns data through a low-dimensional latent space.

[0229] PTUPCDR uses a meta-network, taking user feature embeddings as input, to generate a personalized bridging function for each user.

[0230] In this experiment, the multi-interest cross-domain recommendation model is the diffusion multi-interest cross-domain recommendation framework (DICDR).

[0231] To ensure fairness in the experiment, the number of pre-training and formal training rounds for all models in each scenario was set to 10. The initial learning rate of the Adam optimizer was adjusted in the range {0.001, 0.005, 0.01, 0.02, 0.1} via grid search.

[0232] During the interest extraction process, the metapath set is set as {uiu, uibcu, uiciu}, where u represents the user, i represents the item, b represents the item's brand, and c represents the item's category. λ is set to 0.1. The personalized bridging function in PTUPCDR and DICDR uses the same metanetwork structure, consisting of two linear layers and an activation function with 2×k hidden units, where k represents the embedding dimension. The embedding dimensions of users, items, and metapath neighbor nodes are set to k. In the diffusion network, the input and output dimensions of the approximator are set to k, and the MLP consists of three layers with a linear layer and a hidden unit size of 256.

[0233] Table 2 compares DICDR with the five baseline models mentioned above to verify its effectiveness. The experiments show the results for three cross-domain recommendation (CDR) scenarios under different β settings, where β represents the proportion of cold-start users. The experimental results clearly show that DICDR achieves significant results in all three scenarios. The TGT model, trained solely on target domain data, performs the worst. In contrast, CMF utilizes auxiliary source domain information, thus improving performance. However, CMF fails to distinguish between different domains and ignores the domain transfer problem. EMCDR constructs a cross-domain embedding vector model, utilizing data more effectively and further improving performance. PTUPCDR achieves personalized transfer by constructing bridging functions independently for different users, but the transfer of single-user embeddings fails to capture users' multiple interests, and aligning inter-domain information through bridging functions is challenging. Compared to the best baseline PTUPCDR, DICDR improved MAE by 13.99% and RMSE by 13.29% in Scenario 1, by 10.96% and 9.76% in Scenario 2, and by 16.84% and 15.47% in Scenario 3.

[0234] This is attributed to the DICDR method's better utilization of the rich and diverse information between users or items in the source and target domains, particularly considering the influence of multiple item attributes and other users. The diffusion model, by transferring source domain information and achieving distribution alignment, alleviates the burden on the bridging function to some extent. The fusion network then aggregates different interests, making the multi-interest transfer process more adaptable to target domain information, thereby further improving the performance of the fusion process.

[0235] In the experiments, matrix factorization (MF) was initially used for evaluation. However, MF is a non-neural network model and may be too simplistic to effectively handle the complexity of large-scale real-world recommendation data. Although matrix factorization algorithms perform well in recommendation systems, they do not fully demonstrate the robustness and compatibility of DICDR.

[0236] Therefore, EMCDR, PTUPCDR, and MAFCDR were further applied to two more complex neural network models: GMF and YouTube DNN. GMF improves prediction performance by assigning different weights in the dot product prediction function through a deeper neural network. YouTube DNN is a dual-tower model. To ensure the reliability of the experiments, all settings were kept consistent. To demonstrate the stronger performance of generalized DICDR in cold-start scenarios, tests were conducted at β = 80%.

[0237] like Figure 9As shown, generalization experiments were conducted on three base models: (a) MF, (b) GMF, and (c) YouTube DNN. The figure presents the average results of five runs. EMCDR, PTUPCDR, and DICDR all improved cold-start recommendation performance. With improvements to the base models, GMF and YouTube DNN achieved significant improvements over MF. The MAE and RMSE results in the figure demonstrate that the generalized DICDR performs well in most cold-start scenarios.

[0238] Table 2. Performance Comparison Data of Cross-Domain Recommendation Models in Different Scenarios

[0239]

[0240] Table 3 Ablation Experiment Data

[0241]

[0242] As shown in Table 3, ablation experiments were conducted to further explore the impact of each component in the DICDR model on performance. Specifically, the following models were evaluated:

[0243] I-CDR: Removes the diffusion model, retains the optimization method of DICDR, and uses only the fusion network to generate target domain user embeddings.

[0244] D-CDR: Remove the fusion network and use only the average weighting method to generate target domain user embeddings.

[0245] In I-CDR, it can be viewed as a variant of PTUPCDR. If the interest is replaced with user embeddings, it degenerates into PTUPCDR. Therefore, it can be concluded that introducing more user information improves the model's performance to some extent. In D-CDR, relying solely on the diffusion model to generate user embeddings still outperforms PTUPCDR, indicating that the introduction of the diffusion model effectively addresses the problem of insufficient distribution information in bridging methods. When the fusion network and the diffusion model are combined, DICDR significantly outperforms other algorithms, demonstrating the crucial role of each module.

[0246] This invention also provides a multi-interest cross-domain recommendation model training system, comprising:

[0247] The first module is used to acquire target domain data and source domain data. The target domain data includes user information, first item information and first rating, while the source domain data includes user information, second item information and user behavior information.

[0248] The second module is used to preprocess the target domain data to obtain the target domain user embedding and the target domain item embedding, and to preprocess the source domain data to obtain the user behavior embedding and the source domain interest embedding.

[0249] The third module is used to obtain real noise and predicted noise based on the diffusion model, according to the target domain user embedding, the source domain interest embedding, and the user behavior embedding.

[0250] The fourth module is used to adjust the parameters of the diffusion model based on the actual noise and the predicted noise to obtain the trained diffusion model.

[0251] The fifth module is used to obtain the predicted interest embedding based on the source domain interest embedding and the user behavior embedding using the trained diffusion model.

[0252] The sixth module is used to obtain a predicted score based on the fusion network, according to the predicted interest embedding, user behavior embedding, and target domain item embedding.

[0253] The seventh module is used to obtain the loss function of the fusion network based on the predicted score and the first score;

[0254] The eighth module is used to adjust the parameters of the fusion network according to the loss function to obtain the trained fusion network;

[0255] The ninth module is used to combine the trained fusion network and the trained diffusion model to obtain a multi-interest cross-domain recommendation model.

[0256] It is understood that the content of the above-described multi-interest cross-domain recommendation model training method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above-described multi-interest cross-domain recommendation model training method embodiments, and the beneficial effects achieved are also the same as those achieved in the above-described multi-interest cross-domain recommendation model training method embodiments.

[0257] The following is combined Figure 10 The electronic devices described in the embodiments of this application will be described in detail.

[0258] like Figure 10 , Figure 10 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0259] The processor 1100 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0260] The memory 1200 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1200 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1200 and is called and executed by the processor 1100 to execute the multi-interest cross-domain recommendation model training method or the multi-interest cross-domain recommendation method of the embodiments of this disclosure.

[0261] The input / output interface 1300 is used to realize information input and output;

[0262] The communication interface 1400 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0263] Bus 1500 transmits information between various components of the device (e.g., processor 1100, memory 1200, input / output interface 1300, and communication interface 1400);

[0264] The processor 1100, memory 1200, input / output interface 1300 and communication interface 1400 are connected to each other within the device via bus 1500.

[0265] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the above-described multi-interest cross-domain recommendation model training method or multi-interest cross-domain recommendation method.

[0266] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0267] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0268] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0269] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0270] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0271] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises 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 such processes, methods, products, or apparatus.

[0272] It should be understood that in this application, "at least one" means one or more, and "more" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can mean: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0273] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0274] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0275] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0276] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0277] The preferred embodiments of the present disclosure have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present disclosure. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present disclosure shall be within the scope of the claims of the present disclosure.

Claims

1. A training method for a multi-interest cross-domain recommendation model, characterized in that, Includes the following steps: Acquire target domain data and source domain data, wherein the target domain data includes user information, first item information and first rating, and the source domain data includes user information, second item information and user behavior information; The target domain data is preprocessed to obtain target domain user embeddings and target domain item embeddings, and the source domain data is preprocessed to obtain user behavior embeddings and source domain interest embeddings. Based on the diffusion model, the real noise and predicted noise are obtained according to the target domain user embedding, the source domain interest embedding, and the user behavior embedding; The parameters of the diffusion model are adjusted based on the actual noise and the predicted noise to obtain the trained diffusion model; Using the trained diffusion model, the predicted interest embedding is obtained based on the source domain interest embedding and the user behavior embedding; Based on the fusion network, a predicted score is obtained according to the predicted interest embedding, the user behavior embedding, and the target domain item embedding; The loss function of the fusion network is obtained based on the predicted score and the first score; The parameters of the fusion network are adjusted according to the loss function to obtain the trained fusion network; The trained fusion network and the trained diffusion model are combined to obtain a multi-interest cross-domain recommendation model.

2. The multi-interest cross-domain recommendation model training method according to claim 1, characterized in that, The process of preprocessing the source domain data to obtain user behavior embeddings and source domain interest embeddings includes the following steps: Based on the user information and the second item information, a first interest embedding is obtained; Feature extraction is performed on user behavior information to obtain user behavior embeddings; The source domain interest embedding is obtained based on the first interest embedding and the user behavior embedding.

3. The multi-interest cross-domain recommendation model training method according to claim 2, characterized in that, The step of obtaining the first interest embedding based on the user information and the second item information includes the following steps: Feature extraction is performed on the second item information to obtain the attribute information of the second item information; Based on the relationship between the user information, the second item information, and the attribute information, multiple meta-paths are constructed. Multiple meta-paths are aggregated to obtain the first interest embedding.

4. The multi-interest cross-domain recommendation model training method according to claim 2, characterized in that, The step of obtaining the source domain interest embedding based on the first interest embedding and the user behavior embedding includes the following steps: The weights of the first interest embedding are obtained through an attention network based on the first interest embedding and the user behavior embedding. Using a meta-network, the parameters of the bridging function are obtained based on the weights of the first interest embedding; The source domain interest embedding is obtained based on the first interest embedding and the parameters of the bridging function.

5. The multi-interest cross-domain recommendation model training method according to claim 1, characterized in that, The method for obtaining real noise and predicted noise based on the diffusion model, according to the target domain user embedding, the source domain interest embedding, and the user behavior embedding, includes the following steps: Obtain Gaussian noise; The Gaussian noise is sampled to obtain the real noise; Add the real noise to the target domain user embedding to obtain a second user embedding; Add the real noise to the source domain interest embedding to obtain the second interest embedding; The approximator of the diffusion model obtains prediction noise based on the second user embedding, the second interest embedding, and the user behavior embedding, wherein the prediction noise is used to denoise the second interest embedding to obtain the prediction interest embedding.

6. The multi-interest cross-domain recommendation model training method according to claim 1, characterized in that, The method of obtaining a predicted score based on a fusion network, according to the predicted interest embedding, the user behavior embedding, and the target domain item embedding, includes the following steps: Obtain the parameters of the fusion network; Based on the predicted interest embedding and the fusion network parameters, the attention score of the predicted interest embedding is obtained; Based on the attention score and the predicted interest embedding, a third interest embedding is obtained; The weights of the predicted interest embedding are obtained based on the user behavior embedding and the third interest embedding. Based on the weights and the predicted interest embedding, the target user embedding is obtained; A predicted score is obtained based on the target user embedding and the target domain item embedding.

7. A multi-interest cross-domain recommendation method, characterized in that, Includes the following steps: Retrieve the user's source domain data and target domain items; The source domain data is preprocessed to obtain source domain interest embedding and user behavior embedding; Feature extraction is performed on the target domain items to obtain the target domain item embeddings; The source domain interest embedding, the user behavior embedding, and the target domain item embedding are input into the multi-interest cross-domain recommendation model to obtain the predicted rating of the user's target domain item. The multi-interest cross-domain recommendation model is obtained by the multi-interest cross-domain recommendation model training method as described in any one of claims 1-6.

8. A training system for a multi-interest cross-domain recommendation model, characterized in that, include: The first module is used to acquire target domain data and source domain data, wherein the target domain data includes user information, first item information and first rating, and the source domain data includes user information, second item information and user behavior information; The second module is used to preprocess the target domain data to obtain target domain user embeddings and target domain item embeddings, and to preprocess the source domain data to obtain user behavior embeddings and source domain interest embeddings. The third module is used to obtain real noise and predicted noise based on the diffusion model, according to the target domain user embedding, the source domain interest embedding, and the user behavior embedding; The fourth module is used to adjust the parameters of the diffusion model based on the real noise and the predicted noise to obtain the trained diffusion model; The fifth module is used to obtain the predicted interest embedding based on the source domain interest embedding and the user behavior embedding using the trained diffusion model; The sixth module is used to obtain a predicted score based on the fusion network, according to the predicted interest embedding, the user behavior embedding, and the target domain item embedding; The seventh module is used to obtain the loss function of the fusion network based on the predicted score and the first score; The eighth module is used to adjust the parameters of the fusion network according to the loss function to obtain the trained fusion network; The ninth module is used to combine the trained fusion network and the trained diffusion model to obtain a multi-interest cross-domain recommendation model.

9. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the multi-interest cross-domain recommendation model training method as described in any one of claims 1-6 or the multi-interest cross-domain recommendation method as described in claim 7.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, When the processor executes the program, it is used to implement the multi-interest cross-domain recommendation model training method as described in any one of claims 1-6 or the multi-interest cross-domain recommendation method as described in claim 7.

Images