Recommended methods, systems, electronic devices, and storage media

Abstract

This application relates to a recommendation method, system, electronic device, and storage medium. The recommendation method includes: converting multi-feature domain tabular data into a shared feature embedding representation through feature embedding; generating multiple parallel gating matrices based on the shared feature embedding representation; multiplying the shared feature embedding representation with each gating matrix bitwise to generate multiple dynamically decoupled feature representations; inputting the multiple dynamically decoupled feature representations into corresponding parallel feature interaction paths, and performing layer-by-layer feature interaction through a feature interaction encoder to obtain an initial interaction representation; generating corrected interaction representations for each interaction level based on the initial interaction representation and the shared feature embedding representation; fusing the corrected interaction representations of the last layer of the parallel feature interaction path to obtain a fused corrected interaction representation; and obtaining a predicted value of the user's interaction probability with candidate items based on the fused corrected interaction representation. By using decoupled features and gating fusion, the accuracy of the recommendation model's prediction is improved.

Description

Technical Field

[0001] This application relates to the field of recommendation system technology, and in particular to recommendation methods, systems, electronic devices, and storage media. Background Technology

[0002] Currently, recommendation models are widely used in e-commerce, content platforms, and online advertising to predict users' preferences for specific items, thereby generating personalized recommendation lists to improve user experience and platform revenue.

[0003] Currently, parallel recommendation models based on deep learning have become mainstream in the industry. They learn the complex interaction relationships between users and items by deploying multiple sub-networks with different functions. However, in related technologies, all parallel sub-networks in such models are forced to share the same set of static feature embeddings in their architecture. This shared embedding paradigm of static feature embeddings causes a severe mismatch between feature supply and functional requirements, resulting in low overall prediction accuracy of the recommendation model.

[0004] There is currently no effective solution to the problem of low overall prediction accuracy in recommendation models in related technologies. Summary of the Invention

[0005] This embodiment provides a recommendation method, system, electronic device, and storage medium to address the problem of low overall prediction accuracy of recommendation models in related technologies.

[0006] Firstly, this embodiment provides a recommended method, including:

[0007] Multi-feature-domain tabular data is transformed into a shared feature embedding representation through feature embedding; wherein, the multi-feature-domain tabular data includes user historical data, user information, and candidate item attribute information;

[0008] Multiple parallel gated matrices are generated based on the shared feature embedding representation. The shared feature embedding representation is then multiplied bitwise with each gated matrix to generate multiple dynamically decoupled feature representations. The multiple parallel gated matrices represent features in different optimization directions.

[0009] The multiple dynamically decoupled feature representations are respectively input into the corresponding parallel feature interaction paths, and the feature interaction encoder performs layer-by-layer feature interaction to obtain the initial interaction representation.

[0010] In each interaction level of the parallel feature interaction path, a modified interaction representation for each interaction level is generated based on the initial interaction representation and the shared feature embedding representation.

[0011] The modified interaction representations of the last layer of the parallel feature interaction path are fused to obtain the fused modified interaction representation.

[0012] Based on the fusion-corrected interaction representation, the predicted value of the probability of user interaction with candidate items is obtained.

[0013] In some embodiments, the process of converting multi-feature-domain tabular data into a shared feature embedding representation via feature embedding includes:

[0014] The original interaction logs and attribute information between users and candidate projects are collected from the client. After data cleaning and feature engineering, multi-feature domain tabular data is generated.

[0015] In some embodiments, the step of generating multiple parallel gating matrices based on the shared feature embedding representation, and multiplying the shared feature embedding representation with each gating matrix bitwise to generate multiple dynamically decoupled feature representations includes:

[0016] Based on the shared feature embedding representation, the attention relationship between features in different optimization directions is modeled to obtain the attention representation;

[0017] Based on the shared feature embedding representation, the contextual information between features in different optimization directions is modeled to obtain a contextual representation;

[0018] The attention representation and the context representation are fused to obtain multiple parallel gating matrices; wherein the dimension of the gating matrix is ​​consistent with the dimension of the shared feature embedding representation.

[0019] In some embodiments, the step of inputting the plurality of dynamically decoupled feature representations into corresponding parallel feature interaction paths, and performing layer-by-layer feature interaction through a feature interaction encoder to obtain an initial interaction representation includes:

[0020] The gating values ​​of each interaction level of the parallel feature interaction path are obtained based on the initial interaction representation.

[0021] The gate value is combined with the shared feature embedding representation to generate auxiliary interaction representations for each interaction level of the parallel feature interaction path;

[0022] The auxiliary interaction representations of each interaction level of the parallel feature interaction path are fused with the initial interaction representation to obtain the modified interaction representation.

[0023] In some embodiments, fusing the modified interaction representations of the last layer of the parallel feature interaction pathway to obtain a fused modified interaction representation includes:

[0024] The modified interaction representation of the last layer of the parallel feature interaction pathway is fused by the parallel feature interaction encoder.

[0025] In some embodiments, obtaining the predicted probability of user interaction with candidate items based on the fused and corrected interaction representation includes:

[0026] Based on the fused and corrected interaction representation, the predicted value of the user's interaction probability with the candidate item is obtained by combining a feature vector classifier with an activation function; wherein, the feature vector classifier includes a logistic regression function and an artificial neural network.

[0027] In some embodiments, the method further includes:

[0028] Based on the predicted probability of user interaction with candidate items and the recommendation settings, a recommendation list of candidate items that matches the user's preferences is output.

[0029] Secondly, this embodiment provides a recommendation system, including: a data feature embedding module, a decoupled feature generation module, and a gated interaction fusion module; wherein:

[0030] The data feature embedding module is used to transform multi-feature domain table data into a shared feature embedding representation through feature embedding; wherein, the multi-feature domain table data includes user historical data, user information, and candidate item attribute information;

[0031] The decoupled feature generation module is used to generate multiple parallel gated matrices based on the shared feature embedding representation, and to perform bitwise multiplication of the shared feature embedding representation with each gated matrix to generate multiple dynamic decoupled feature representations; wherein, the multiple parallel gated matrices represent features in different optimization directions;

[0032] The gated interaction fusion module is used to input the multiple dynamically decoupled feature representations into the corresponding parallel feature interaction paths, and perform layer-by-layer feature interaction through the feature interaction encoder to obtain an initial interaction representation; in each interaction level of the parallel feature interaction path, a modified interaction representation for each interaction level is generated based on the initial interaction representation and the shared feature embedding representation; the modified interaction representations of the last layer of the parallel feature interaction path are fused to obtain a fused modified interaction representation; and based on the fused modified interaction representation, a predicted value of the probability of user interaction with candidate items is obtained.

[0033] Thirdly, this embodiment provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the recommended method described in the first aspect above.

[0034] Fourthly, this embodiment provides a storage medium on which a computer program is stored, which, when executed by a processor, implements the recommended method described in the first aspect above.

[0035] Compared with related technologies, this embodiment provides a recommendation method, system, electronic device, and storage medium. In the recommendation method, firstly, multi-feature domain table data is transformed into a shared feature embedding representation through feature embedding; wherein, the multi-feature domain table data includes user historical data, user information, and candidate item attribute information; secondly, multiple parallel gating matrices are generated based on the shared feature embedding representation, and the shared feature embedding representation is multiplied bitwise with each gating matrix to generate multiple dynamically decoupled feature representations; wherein, the multiple parallel gating matrices represent features in different optimization directions; subsequently, the multiple dynamically decoupled feature representations are input into corresponding parallel feature interaction paths, and feature interaction is performed layer by layer by a feature interaction encoder to obtain an initial interaction representation; further, in each interaction level of the parallel feature interaction path, a modified interaction representation for each interaction level is generated based on the initial interaction representation and the shared feature embedding representation; then, the modified interaction representations of the last layer of the parallel feature interaction path are fused to obtain a fused modified interaction representation; finally, based on the fused modified interaction representation, the predicted value of the user's interaction probability with the candidate item is obtained. It improves the overall prediction accuracy of the recommendation model by dynamically decoupling feature representation and fusion of interactive representation.

[0036] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0037] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0038] Figure 1 This is a hardware structure block diagram of the terminal for the recommended method in this embodiment;

[0039] Figure 2 This is a flowchart of the recommended method in this embodiment;

[0040] Figure 3 This is a structural diagram of a recommended method according to an embodiment of this application;

[0041] Figure 4 This is a structural diagram of a multilayer perceptron module according to an embodiment of this application;

[0042] Figure 5This is a structural diagram of a self-attention module according to an embodiment of this application;

[0043] Figure 6 This is a flowchart illustrating the layer-by-layer interaction between two different sub-networks in one embodiment of this application;

[0044] Figure 7 This is an overall architecture diagram of a conventional parallel recommendation method according to an embodiment of this application;

[0045] Figure 8 This is an embodiment of the present application of the decoupled representation generator and gated fusion module in a traditional parallel recommendation method.

[0046] Figure 9 This is a schematic diagram of the functional modules of a recommendation system according to an embodiment of this application;

[0047] Figure 10 This is a structural block diagram of a recommendation system according to an embodiment of this application. Detailed Implementation

[0048] To better understand the purpose, technical solution, and advantages of this application, the application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0049] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0050] The method embodiments provided in this example can be executed on a terminal, computer, or similar electronic device with a certain computing power. For example, it can run on a terminal. Figure 1 This is a hardware structure block diagram of the terminal for the recommended method in this embodiment. For example... Figure 1 As shown, a terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 and a memory 104 for storing data are also included. The processor 102 may be, but is not limited to, a microprocessor (MCU) or a programmable logic device (FPGA). The terminal may also include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that… Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, the terminal may also include components that are larger than... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown are illustrated.

[0051] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the recommended method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0052] The transmission device 106 is used to receive or send data via a network. This network includes a wireless network provided by the terminal's communication provider. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 can be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0053] This embodiment provides a recommended method. Figure 2 This is a flowchart of the recommended method in this embodiment, such as... Figure 2 As shown, the process includes the following steps:

[0054] Step S210: The multi-feature domain table data is transformed into a shared feature embedding representation through feature embedding; wherein, the multi-feature domain table data includes user historical data, user information and candidate item attribute information.

[0055] Specifically, the multi-feature domain tabular data originates from different logical dimensions, including user historical data, user information, and candidate item attribute information. User historical data includes explicit records such as clicks, purchases, browsing, and ratings, recorded in sequence, representing the user's dynamic interests and preferences during past usage. User information includes user demographic attributes such as age, gender, region, and device information, which can be used to represent the user's static profile. Candidate item attribute information includes data such as the title, click count, and view count of candidate items (e.g., products, videos, movies, and books), which can be used to represent the core characteristics of the candidate items themselves. In particular, all of the above data collection actions were carried out with the user's consent.

[0056] However, different feature domains have different data types and physical meanings, making direct computation impossible. Feature embedding can unify discrete or continuous numerical values ​​into vectors in the same vector space, eliminating the barriers between heterogeneous data types. In the embedding space, features with similar semantics have closer vector distances, enabling deep neural networks to learn deep relationships between features rather than simply memorizing independent features. Therefore, feature embedding can map the aforementioned heterogeneous, high-dimensional, sparse tabular data from different feature domains into a unified low-dimensional, dense, semantic vector space. Subsequent deep neural network models require inputs of fixed-dimensional, numerical, dense vectors. A shared feature embedding layer can transform all input features into vector representations of the same size. For example, in one embodiment, the input form of multi-feature domain tabular data is a one-hot vector. Subsequently, the multi-feature domain tabular data is searched and transformed through an embedding feature table to convert it into a dense vector representation, which is the shared feature embedding representation. All parameters of the embedding feature table can be learned and optimized through backpropagation during network training.

[0057] Step S220: Generate multiple parallel gated matrices based on the shared feature embedding representation, and multiply the shared feature embedding representation with each gated matrix bitwise to generate multiple dynamically decoupled feature representations; wherein, the multiple parallel gated matrices represent features in different optimization directions.

[0058] Specifically, in existing parallel recommendation methods, multiple parallel feature interaction pathways—i.e., sub-networks of deep neural networks—are forced to share the same set of static feature embeddings, leading to feature supply mismatch and optimization conflicts. Based on shared feature embedding representations, decoupled representation generation generates a dedicated gating matrix for each parallel sub-network. The gating matrix can be viewed as a set of "customized instructions" to guide how to extract information from the shared features for the corresponding sub-network's optimization direction. It is precisely through decoupled representation generation that customized feature inputs can be generated for different optimization directions.

[0059] Subsequently, by multiplying the shared feature embedding representation bitwise with each gating matrix, the gating matrices for different optimization directions of each sub-network are combined with the shared feature embedding representation. Each dimension of the feature vector in the shared feature embedding representation is weighted according to the gating matrix, thereby enabling the generation of feature representations with different optimization directions from the same set of shared feature embedding representations.

[0060] For example, in one embodiment, decoupled representation generation achieves its function through a gating information selection mechanism. Specifically, decoupled representation generation takes a shared feature embedding representation as input, analyzes the feature context of the shared feature embedding representation, and computes it twice in parallel to obtain two gating matrices with different optimization directions. Subsequently, the two gating matrices are multiplied bitwise with the shared feature embedding representation to generate two sets of dynamically decoupled feature representations with different optimization directions, which serve as inputs to two parallel sub-networks.

[0061] Step S230: Input multiple dynamically decoupled feature representations into the corresponding parallel feature interaction paths respectively, and perform layer-by-layer feature interaction through the feature interaction encoder to obtain the initial interaction representation.

[0062] Following the previous step, the dynamically decoupled feature representations, modulated by a dedicated gating matrix, have evolved into feature sets with different emphases. The dynamically decoupled feature representations with different optimization directions are then input into their respective parallel sub-networks. The input method follows the principle of "dedicated lines for dedicated tasks." For example, the first dynamically decoupled feature representation is fixedly connected and input into the first sub-network, while the second dynamically decoupled feature representation is fixedly connected and input into the second sub-network. This design ensures that each sub-network can learn independently based on its own customized dynamically decoupled feature representation input.

[0063] Subsequently, deep neural networks are used to explore the complex relationships between dynamically decoupled features. Each dynamically decoupled feature is input from the first layer of each sub-network, passed through a non-linear transformation to the second layer, and so on. At each layer, the deep neural network learns feature combinations at different levels of abstraction. Shallow layers learn simple features, while deeper layers capture highly non-linear and complex relationships, including long-term user preferences, contextual intent, and item attributes. The fundamental purpose of feature interaction is to go beyond the individual dynamically decoupled features themselves, and to build a network of relationships between these features to gain a deeper understanding of the motivations behind user behavior, achieving a deeper characterization of user interests and thus enabling accurate prediction of recommendation lists.

[0064] Step S240: In each interaction level of the parallel feature interaction path, a modified interaction representation for each interaction level is generated based on the initial interaction representation and the shared feature embedding representation.

[0065] After the layer-by-layer interaction in step S230, the obtained initial interaction representation already contains the dynamic decoupling feature information learned through the sub-network. However, due to the inherent characteristics of deep neural network training, the initial interaction representation may also contain some noise or have insufficient optimization. Therefore, after each layer of initial interaction is completed, a correction process will be executed to adjust the initial interaction representation.

[0066] Specifically, firstly, a bit-level gating value is learned based on the initial interaction representations output from each interaction layer of the sub-network. This gating value can identify which dimensions in the current feature are valid information and which are noise. Next, the gating value is combined with the shared feature embedding representation to generate an auxiliary representation. Finally, the auxiliary representation is fused with the initial interaction representation to obtain a corrected interaction representation with less noise and higher quality information. This layer-by-layer calibration mechanism is equivalent to setting a quality check and enhancement process after each layer, ensuring that the dynamically decoupled features maintain high fidelity throughout the deep propagation process.

[0067] Step S250: The modified interaction representation of the last layer of the parallel feature interaction path is fused to obtain the fused modified interaction representation.

[0068] This step aims to integrate the differentiated deep dynamic decoupling features learned by each sub-network to form a unified representation. Specifically, the final layer, a modified interaction representation, is obtained after layer-by-layer optimization of each sub-network. Each modified interaction representation carries deep interaction information from different optimization directions. Through feature concatenation, weighted summation, and an attention-based fusion strategy, the modified interaction representations are organically combined, resulting in a fused modified interaction representation that possesses the advantages of different optimization directions. This fusion not only achieves information complementarity but also produces excellent results through the synergistic effect between features. It provides high-quality feature input for subsequent prediction of user-candidate item interaction probabilities, thereby improving the accuracy of recommended candidate items and the generalization ability of deep neural networks.

[0069] For example, the final layer modified interaction representations of the first and second sub-networks, obtained after modification by each sub-network, carry different types of deep interaction information: the final layer modified interaction representation of the first sub-network focuses more on explicit feature cross-patterns, while the final layer modified interaction representation of the second sub-network is better at capturing implicit nonlinear relationships. By using a fusion strategy, these two sets of modified interaction representations are combined, so that the obtained fused modified interaction representation has the advantages of both explicit and implicit feature interactions.

[0070] Step S260: Based on the fusion and correction of the interaction representation, obtain the predicted value of the probability of user interaction with candidate items.

[0071] Specifically, predictions are made by fusing and correcting interaction representations, typically using a single-layer neural network with activation functions or a logistic regression model, mapping high-dimensional feature vectors to probability values ​​between 0 and 1. This probability value quantifies the likelihood that a user will engage in a specific interaction (such as clicking or purchasing) with a particular candidate item in a given context. This approach enables the prediction of recommended candidate items from multi-feature domain tabular data. More importantly, because the input features are high-quality representations that have undergone dynamic decoupling, hierarchical gating correction, and cross-sub-network integration, the final predicted output accurately reflects the user's true interests, improving the accuracy of the predicted candidate items and providing a reliable basis for generating high-quality recommendation lists.

[0072] Compared to other related technologies, parallel recommendation models based on deep learning have become the mainstream architecture for improving prediction performance. They typically employ dual-path or multi-path designs, capturing feature interactions of different patterns through parallel sub-networks. However, parallel recommendation models suffer from a common and fundamental architectural flaw: all parallel sub-networks are forced to share the same set of static feature embeddings. This seemingly efficient parameter-sharing mechanism actually causes serious functional limitations. Specifically, during model training, the embedding parameters need to simultaneously respond to the competing optimization objectives of multiple sub-networks. This multi-task competition for the same set of parameters leads to inherent gradient conflicts. This conflict turns the learning process of embedding parameters into a compromise between the demands of different sub-networks, rather than their respective optimal solutions. Ultimately, this causes the model to converge to a performance-constrained suboptimal state, resulting in a decrease in the overall prediction accuracy of the recommendation model.

[0073] Steps S210 to S260 above involve: First, transforming multi-feature-domain tabular data into a shared feature embedding representation through feature embedding; where the multi-feature-domain tabular data includes user historical data, user information, and candidate item attribute information; second, generating multiple parallel gating matrices based on the shared feature embedding representation, and multiplying the shared feature embedding representation with each gating matrix bitwise to generate multiple dynamically decoupled feature representations; where the multiple parallel gating matrices represent features in different optimization directions; next, inputting the multiple dynamically decoupled feature representations into corresponding parallel feature interaction paths, and performing layer-by-layer feature interaction through a feature interaction encoder to obtain initial interaction representations; then, in each interaction level of the parallel feature interaction path, generating corrected interaction representations for each interaction level based on the initial interaction representations and the shared feature embedding representations; further, fusing the corrected interaction representations of the last layer of the parallel feature interaction path to obtain fused corrected interaction representations; finally, based on the fused corrected interaction representations, obtaining the predicted value of the user's interaction probability with the candidate item. This method improves the overall prediction accuracy of the recommendation model by fusing dynamically decoupled feature representations and interaction representations.

[0074] Optionally, in one embodiment, the multi-feature domain table data is transformed into a shared feature embedding representation through feature embedding, including: collecting the original interaction logs and attribute information between the user and the candidate items from the client, and generating multi-feature domain table data after data cleaning and feature engineering.

[0075] First, raw interaction data between users and candidate items is collected from the client-side distributed log system. This data includes implicit feedback such as user click sequences, dwell time, search query records, and purchase history, as well as explicit feedback such as user ratings, favorites, and shares. Simultaneously, user attribute information, including demographic characteristics, device information, and geographic location, and candidate item attribute information, including product category, content tags, price range, and creator information, are collected. Notably, all data collection processes adhere to privacy protection guidelines, and sensitive information is anonymized.

[0076] Subsequently, the collected raw interaction data underwent preprocessing. Outliers and missing values ​​were removed through data cleaning. Furthermore, short-term user behavior characteristics were statistically analyzed based on time windows, and a sliding window algorithm was used to construct long-term user interest profiles. Categorical features were subjected to one-hot encoding or binning, while numerical features were standardized and normalized. Finally, structured multi-feature-domain tabular data was generated, where each feature domain represents a specific semantic dimension.

[0077] Furthermore, in one embodiment, multiple parallel gated matrices are generated based on the shared feature embedding representation, and the shared feature embedding representation is multiplied bitwise with each gated matrix to generate multiple dynamically decoupled feature representations, including:

[0078] Attention representation is obtained by modeling the attention relationship between features in different optimization directions based on shared feature embedding representation; contextual information between features in different optimization directions is obtained by modeling the contextual information based on shared feature embedding representation; attention representation and contextual representation are fused to obtain multiple parallel gating matrices; the dimension of the gating matrix is ​​consistent with the dimension of the shared feature embedding representation.

[0079] The preprocessed multi-feature domain table data is input into the feature embedding layer of a deep neural network. For high-dimensional, sparse categorical features in the multi-feature domain table, they are mapped to low-dimensional, dense vectors by looking up the embedding table. For numerical features in the multi-feature domain table, they are projected into the same vector space as the categorical features through linear transformation or a fully connected layer. For multi-valued categorical features in the multi-feature domain table, pooling operations are used to aggregate their embedding representations. Finally, the embedding vectors of all feature domains are concatenated to form a unified, shared feature embedding representation. This process ensures that the original heterogeneous data can be effectively transformed into a unified representation suitable for deep learning networks.

[0080] Based on the obtained shared feature embedding representation, a gating matrix is ​​generated through two parallel feature analysis modules: on the one hand, a self-attention mechanism is used to model the correlation between features in different optimization directions, and an attention representation is obtained by calculating the attention weights between features. This representation can highlight feature combinations that are important to a specific optimization objective. On the other hand, a multilayer perceptron network is used to extract the contextual representation of features. This network captures the global contextual information of features at the instance level through nonlinear transformation.

[0081] Subsequently, the obtained attention representation and context representation are fused and concatenated, then subjected to dimensionality transformation through a fully connected layer, and normalized by an activation function (e.g., an S-shaped growth curve), finally generating a gating matrix with dimensions completely consistent with the shared feature embedding representation. This fusion process ensures that the gating matrix considers both the relative importance of features and incorporates instance-specific contextual information.

[0082] Finally, dynamic decoupling of features is achieved through element-wise multiplication: each generated gate matrix is ​​multiplied element-wise with the original shared feature embedding representation. This operation is equivalent to reweighting the different dimensions of the original features, thereby generating multiple dynamically decoupled feature representations with different optimization directions. Each decoupled feature representation undergoes feature selection enhancement for a specific optimization direction, providing customized input features for different subsequent sub-networks.

[0083] In one embodiment, multiple dynamically decoupled feature representations are input into corresponding parallel feature interaction paths, and feature interactions are performed layer by layer through a feature interaction encoder to obtain an initial interaction representation, including:

[0084] The gating values ​​of each interaction level of the parallel feature interaction path are obtained based on the initial interaction representation; the gating values ​​are combined with the shared feature embedding representation to generate the auxiliary interaction representations of each interaction level of the parallel feature interaction path; the auxiliary interaction representations of each interaction level of the parallel feature interaction path are fused with the initial interaction representation to obtain the modified interaction representation.

[0085] Multiple dynamically decoupled feature representations are input into corresponding parallel sub-networks. Each sub-network performs feature transformation and interaction learning through layer-by-layer forward propagation of a feature interaction encoder. In the first layer, the dynamically decoupled feature representation is used as input, and after linear transformation and nonlinear activation function processing, the initial interaction representation of the first layer is obtained. In subsequent layers, the output of the previous layer (the corrected interaction representation obtained after modifying the initial interaction representation) is used as the input of the current layer, and feature interaction calculation is repeated to obtain the initial interaction representation of each layer.

[0086] Next, after the initial interaction of features at each layer is completed, a gating fusion operation is performed. Based on the initial interaction representation of the current layer, bit-level gating values ​​are calculated through a lightweight gating network. This gating network can consist of fully connected layers and activation functions, used to evaluate the importance of each dimension in the initial interaction representation. These gating values ​​are then multiplied bitwise with the shared feature embedding representation to generate an auxiliary interaction representation. Essentially, this step utilizes the original feature information to supplement and correct the feature representation after multiple transformations.

[0087] Finally, the auxiliary interaction representation is fused with the initial interaction representation to obtain the revised interaction representation. The fusion method can employ weighted summation, feature concatenation, or attention-based fusion strategies. This fusion operation preserves the complex patterns obtained through deep interaction learning while incorporating fundamental information from the original features, effectively avoiding information loss and feature bias in deep networks. The resulting revised interaction representation contains deep semantic information while maintaining feature accuracy, providing a more reliable feature representation for the next layer of feature interaction or the final prediction task.

[0088] In one embodiment, fusing the modified interaction representation of the last layer of the parallel feature interaction path to obtain a fused modified interaction representation includes: fusing the modified interaction representation of the last layer of the parallel feature interaction path through a parallel feature interaction encoder.

[0089] We collect the corrected interaction representations of the outputs of the last layer of each parallel sub-network. These corrected interaction representations carry differentiated feature information learned by different sub-networks, such as explicit feature crossover patterns and implicit nonlinear relationships. Each corrected interaction representation is a high-quality feature representation optimized through multi-layer feature interactions and gated fusion within each sub-network.

[0090] Next, feature fusion is employed to deeply integrate the modified interaction representations of the last layer. First, each modified interaction representation is standardized to ensure consistency across different feature scales. Then, an attention mechanism is used to calculate the importance weights of the modified interaction table of the last layer. Specifically, this involves inputting the modified interaction table of the last layer into the fully connected layer for feature transformation, calculating the attention distribution using a normalized exponential function, and then weighting and summing the modified interaction representations of the last layer according to their weights. Furthermore, a combined strategy of feature concatenation and dimensionality reduction can be used: first, all representations are concatenated along the feature dimensions, and then dimensionality compression and information fusion are performed through the fully connected layer.

[0091] During the fusion process, the complementarity between features of different sub-networks was specifically considered. A cross-feature calculation module was used to compute the outer product relationship between the representations of different sub-networks, capturing the synergistic effect between features. Simultaneously, a residual connection mechanism was introduced to preserve important information from the original features and prevent information loss during the fusion process. The final fused and corrected interactive representation integrates the advantages of each sub-network, including both local feature interaction information and global semantic associations.

[0092] In one embodiment, obtaining a predicted value of the user's interaction probability with candidate items based on the fused and corrected interaction representation includes: obtaining a predicted value of the user's interaction probability with candidate items by combining a feature vector classifier with an activation function based on the fused and corrected interaction representation; wherein the feature vector classifier includes a logistic regression function and an artificial neural network.

[0093] The fusion-corrected interactive representation is input to the prediction layer, which includes a feature vector classifier and a corresponding activation function. The feature vector classifier can be a linear classifier based on logistic regression, which performs a linear transformation on the input features through a weight matrix and a bias term; or it can be a nonlinear classifier based on an artificial neural network, which performs a nonlinear feature transformation through a multilayer perceptron containing at least one hidden layer.

[0094] In the specific calculation process, the selected feature vector classifier is first fused and modified with interactive representations. If a logistic regression function is used, the linear transformation result is directly mapped to a probability value between 0 and 1 using the sigmoid growth curve activation function. If an artificial neural network is used, a nonlinear transformation is first performed in the hidden layer using activation functions such as modified linear units or hyperbolic tangent functions, and finally, the predicted probability is obtained by using the sigmoid growth curve activation function in the output layer.

[0095] To ensure prediction accuracy, cross-entropy loss function can be used during the training phase, and regularization can be introduced to prevent overfitting. During the deployment phase of the recommendation model, a calibration layer can be used to post-process the predicted probabilities to ensure consistency between the predicted probabilities and the actual results.

[0096] In addition, in one embodiment, the above recommendation method further includes: outputting a recommendation list of candidate items that match the user's preferences based on the predicted value of the probability of user interaction with candidate items and recommendation settings.

[0097] Based on the predicted probability values ​​of user interaction with candidate items, and combined with pre-set business rules and diversity requirements, the candidate items are evaluated and ranked in multiple dimensions. Specifically, the candidate items are first initially ranked according to the predicted scores. Then, the results are adjusted by comprehensively considering business factors such as content freshness, the balance of candidate item category distribution, and regional adaptability, ultimately generating a recommendation list that meets both user personalized preferences and operational needs. This recommendation list is displayed through a user interface and is continuously optimized and updated based on real-time feedback data to ensure that the recommendation list always remains synchronized with the dynamic changes in user interests.

[0098] Figure 3 This is a structural diagram of a recommended method according to an embodiment of this application. (As shown...) Figure 3 As shown, the recommendation method is bottom-up and includes: a feature embedding layer, a feature interaction layer, and a prediction layer.

[0099] The feature embedding layer transforms the raw feature data into a shared feature embedding representation. The original feature data is sparse, typically represented as a high-dimensional binary sparse vector. The feature embedding layer transforms this into a dense vector form of the shared feature embedding representation. The representation is given, where d represents the dimension of the embedding representation, and F is the number of features in each input instance, which will be used subsequently. This represents the concatenated feature embedding, where .

[0100] The feature interaction layer includes a decoupled representation generator and a gated fusion module (GateF). The decoupled representation generator takes a shared feature embedding representation as input and, in this embodiment, outputs a first and a second dynamically decoupled feature representation, respectively. These are applied after the feature embedding layer to facilitate learning the dynamically decoupled feature representation in conjunction with the subsequent parallel interaction structure. The decoupled representation generator enhances the flexibility of feature representation through dynamic feature representation; furthermore, it allows adaptation to the interaction structures of different depth interaction encoders, such as... Figure 3The paper presents two interaction modules: a deep neural network (DNN) and an interaction structure (CNv2). To obtain dynamic feature representations, the decoupled representation generator first uses a lightweight structure to generate a feature-wise weight matrix. This lightweight structure contains a multi-layer perceptron (MLP) module for capturing instance-level contextual information and a self-attention (SelfAtt) module for modeling the relationships between decoupled features. Since feature combinations vary with instances, their contextual relationships also need to be dynamically adapted. The joint output of the MLP and SelfAtt modules allows each feature embedding to adaptively adjust based on features in different instances. Furthermore, these decoupled feature representations are dynamically optimized to align with the specific interaction patterns of their respective subnetworks.

[0101] Specifically, the structure of the multilayer perceptron module in this embodiment is as follows: Figure 4 As shown, the concatenated decoupled feature representation embedding is used as input. The MLP module adopts a multi-layer structure, and the calculation method of each layer is as follows:

[0102] ;

[0103] in , , and Let these represent the output, weight matrix, bias vector, and activation function of the i-th layer, respectively. It is the concatenated feature embedding. The final output dimension of the MLP module structure. The embedding size is fixed at d, which represents the compressed context-aware information of the input instance.

[0104] Regarding the self-attention module structure in this embodiment, such as Figure 5 As shown. Its main function is to model the interaction relationships between different decoupled features within the same instance. The SelfAtt module structure takes the decoupled feature representation embedding as input and projects it onto three independent matrices, including the query matrix ( ), key matrix ( ) and value matrix ( Then, the attention weights on the value matrix are calculated using scaled dot product attention and a normalized exponential function, as shown in the following formula:

[0105] ;

[0106] Where E is the shared feature embedding representation. , and These are the projection matrices of the query matrix, key matrix, and value matrix, respectively. Represents the attention dimension. Output. It can capture the contextual dependencies between features.

[0107] Next, the output of the SelfAtt module structure is multiplied element by element. The final output of the MLP module structure Fusion allows each decoupled feature to incorporate contextual information. Subsequently, a sigmoid growth curve activation function is used to calculate the bit-level weight matrix. The calculation process is as follows:

[0108] ;

[0109] in, It is a commonly used symbol for the activation function of the S-shaped growth curve. It is a digit-wise multiplication.

[0110] The decoupled representation generator uses two independent SelfAtt modules and an MLP module to generate two gated matrices: and These correspond to the two deep interactive encoder (DNN) and CNv2 sub-networks used in this example, respectively. Using these gating matrices, dynamic decoupling feature representations specific to each sub-network are generated. and The calculation process is as follows:

[0111] ;

[0112] Two sets of dynamic decoupling feature representations and This refers to the first and second dynamic decoupling feature representations mentioned in this embodiment, which are subsequently input into two parallel DNN and CNv2 sub-networks. Since the downstream sub-networks have different characteristics, the learned bitwise weights will differ accordingly, ensuring that the decoupled embeddings can be customized for each sub-network. Based on the dynamic decoupling feature representations of different sub-networks, two key issues can be effectively addressed: firstly, it overcomes the limitation of fixed decoupling feature representations in traditional methods, where the same feature represents the same thing in different instances, failing to reflect the dynamic differences between instances; secondly, it solves the problem of single feature representation, where traditional methods still use a unified decoupling feature representation when facing different interaction structures, making it difficult to meet the diverse feature representation needs of each sub-network.

[0113] The feature interaction layer also contains a gated fusion module, which is embedded in the deep feature interaction layer. The overall first and second dynamically decoupled feature representations and the shared feature embedding representation are taken as input, and the outputs are the corrected interaction representations of the last layer of the first sub-network and the corrected interaction representations of the last layer of the second sub-network. The deep interaction layer is responsible for modeling deep feature interaction information, and typically uses a complex and efficient deep interaction encoder. In this embodiment, an improved interaction network and a deep neural network are used for explicit high-order feature interaction and implicit interaction modeling, respectively. The two deep interaction encoders adopt a layer-by-layer interaction form, and their structures are as follows: Figure 6 As shown. Specifically, the computation process of the l-th interaction layer of CNv2 and the l-th layer of the DNN is defined as follows:

[0114] ;

[0115] in, , , , as well as Let represent the interaction representation of the i-th layer, the weight matrix of CNv2, the weight matrix of the DNN, the bias vector of CNv2, and the bias vector of the DNN, respectively. The input feature representations of the CNv2 and DNN subnetworks are denoted as […]. and ,in , . and This refers to uncorrected, layer-by-layer interaction information. As the interaction depth increases, the performance of deep neural networks significantly decreases due to the accumulation and amplification of noise in higher-order uncorrected interactions. Therefore, to address this issue, a gated fusion module is designed in this application.

[0116] In this embodiment, the gated fusion module is a lightweight interaction representation correction module, applied to each layer of the aforementioned deep interaction encoder. The gated fusion module adaptively corrects the interaction information vector at each layer. In this embodiment, the gated fusion module utilizes a gating mechanism, using the initial interaction information layer by layer as initial information, to dynamically extract correction information from the shared feature representation. Specifically, the gated fusion module uses the shared feature representation... The initial interaction representation is used as input. Taking CNv2 in the sub-network as an example, based on the interaction representation of the l-th layer... The corresponding correction information is calculated as follows:

[0117] ;

[0118] in is the projection matrix of the l-th interactive layer in the CNv2 deep interactive encoder. and These represent the corresponding hierarchical gating vector and hierarchical correction information vector, respectively.

[0119] Subsequently, the corrected information will be displayed. This is superimposed on the corresponding initial interaction information. Specifically, for the CNv2 and DNN sub-networks used in this embodiment, the recursive calculation formula after applying the gated fusion module to each layer is as follows:

[0120] ;

[0121] at the same time, Let l be the projection matrix of the l-th interactive layer in the CNv2 deep interactive encoder. is the projection matrix of the l-th interactive layer in the DNN deep interactive encoder.

[0122] In the final layer of the two parallel deep interaction representations of CNv2 and DNN, the final layer of the sub-network's corrected interaction representation is obtained. At this point, through layer-by-layer optimization by the gated fusion module, the interaction noise in the final layer of the corrected interaction representation of the two parallel sub-networks is significantly reduced. In one embodiment, subsequent experiments have verified that the recommended method at this point can model more than 20 layers of interaction information while maintaining improved performance.

[0123] In this embodiment, the prediction layer uses the last layer of the first sub-network to correct the interaction representation. The last layer of the second sub-network corrects the interaction representation The input is the predicted score of the candidate item, and the output is the predicted score of the candidate item. First, the modified interaction representations of the last layer of the first sub-network and the modified interaction representations of the last layer of the second sub-network are fused. In this embodiment, a direct connection method is used to obtain the fused modified representation. ,Right now Based on the fusion-corrected representation, a multilayer perceptron or logistic regression model is used to calculate the predicted score of the candidate items. Subsequently, based on the predicted score and the actual scores of the data, the loss function of the recommendation model is calculated. In this embodiment, the cross-entropy loss function is used:

[0124] ;

[0125] in, For the true score of the data, This is for predicting scores. The cross-entropy loss function described above is used to train the network, with the goal of minimizing the cross-entropy loss.

[0126] Figure 7 This is an overall architecture diagram of a conventional parallel recommendation method according to an embodiment of this application. Figure 7As shown, the architecture includes a feature embedding layer, a feature interaction layer, and a prediction layer. The core architecture of the recommender system, the feature embedding layer, is responsible for learning fixed feature representations, i.e., shared feature embedding representations; the feature interaction layer contains two parallel deep feature interaction encoders, which focus on modeling explicit and implicit feature interaction patterns, respectively; the prediction layer is responsible for fusing the two sets of interaction information and predicting the recommendation score of candidate items through a logistic regression function or a multilayer perceptron. This score will be used to generate the final recommendation list. Figure 8 This is an embodiment of the overall architecture of this application, which applies a decoupled representation generator and a gated fusion module to a traditional parallel recommendation method. This overall architecture demonstrates the ingenuity of this embodiment in placing the decoupled representation generator between the feature embedding layer and the feature interaction layer. Taking the shared feature embedding representation as input, it dynamically models the interaction relationships and contextual information between features to generate a first and a second dynamically decoupled feature representation with differentiated semantic tendencies. The gated fusion module is integrated into each interaction level of the feature interaction layer. By combining the initial interaction representation of the current layer with the shared feature embedding representation, the gated fusion module generates a corrected information vector, resulting in a more accurate recommendation score. Furthermore, the layer-by-layer calibration mechanism can alleviate the noise accumulation problem in deep neural networks.

[0127] It is worth mentioning that both the decoupled representation generator and the gated fusion module adopt a plug-and-play design concept, which can be independently embedded into existing parallel recommendation models to improve performance, or work together to optimize recommendation effects to the greatest extent, demonstrating good versatility and scalability.

[0128] Figure 9 This is a schematic diagram of the functional modules of a recommendation system according to an embodiment of this application. Figure 9 As shown, it includes a data preprocessing module, a feature embedding module, a decoupled representation generator module, a parallel interaction and gating fusion module, a prediction layer, and an item recommendation module.

[0129] The data preprocessing module is responsible for collecting, cleaning, and storing data, generating multi-feature domain tabular data that can be used for network training. This data covers multiple dimensions, including user information, user historical behavior sequences, item attributes, and contextual information, forming the system's first set of raw feature data.

[0130] The feature embedding module transforms the raw feature data into a unified feature representation. This module receives sparse vectors with one-hot encoding, maps them to dense vector representations through an embedding table lookup operation, and outputs a shared feature embedding representation.

[0131] The decoupled representation generator module, as a core component, is responsible for generating differentiated feature representations from a unified feature representation. Based on a shared feature embedding representation, this module generates two sets of dynamically decoupled feature representations through parallel computation, providing customized input features for different subsequent sub-networks.

[0132] The core computing unit comprises a parallel interaction and gating fusion module. It includes a parallel multi-layer feature interaction encoder that receives decoupled feature representations and performs deep feature interactions. A gating fusion submodule embedded in each interaction layer dynamically calibrates the interaction process by combining the current layer's interaction representation with the original shared feature representation. After layer-by-layer optimization, the final output consists of two sets of high-quality, corrected interaction representations for the last layer of the sub-networks.

[0133] The prediction layer fuses the final layer modified interaction representations of the two sub-networks and calculates the predicted score of the candidate item through the prediction network.

[0134] The project recommendation module sorts and filters candidate projects based on predicted scores and business rules, generates a final personalized recommendation list, and displays it to the user.

[0135] Figure 10 This is a structural block diagram of the recommendation system 10 in this embodiment, as follows: Figure 10 As shown, the recommendation system 10 includes: a data feature embedding module 12, a decoupled feature generation module 14, and a gated interaction fusion module 16; wherein: the data feature embedding module 12 is used to transform multi-feature domain table data into a shared feature embedding representation through feature embedding; wherein, the multi-feature domain table data includes user historical data, user information, and candidate item attribute information; the decoupled feature generation module 14 is used to generate multiple parallel gate matrices based on the shared feature embedding representation, and multiply the shared feature embedding representation with each gate matrix bitwise to generate multiple dynamically decoupled feature representations; wherein, multiple parallel The gating matrix represents the features of different optimization directions; the gating interaction fusion module 16 is used to input multiple dynamically decoupled feature representations into the corresponding parallel feature interaction paths, and perform layer-by-layer feature interaction through the feature interaction encoder to obtain the initial interaction representation; in each interaction level of the parallel feature interaction path, the modified interaction representation of each interaction level is generated according to the initial interaction representation and the shared feature embedding representation; the modified interaction representation of the last layer of the parallel feature interaction path is fused to obtain the fused modified interaction representation; based on the fused modified interaction representation, the predicted value of the user's interaction probability with the candidate item is obtained.

[0136] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0137] This embodiment also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0138] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0139] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0140] S1, transform the multi-feature-domain tabular data into a shared feature embedding representation through feature embedding; wherein, the multi-feature-domain tabular data includes user historical data, user information, and candidate item attribute information;

[0141] S2 generates multiple parallel gated matrices based on the shared feature embedding representation, and multiplies the shared feature embedding representation with each gated matrix bitwise to generate multiple dynamically decoupled feature representations; wherein, the multiple parallel gated matrices represent features in different optimization directions;

[0142] S3, input multiple dynamically decoupled feature representations into the corresponding parallel feature interaction paths respectively, and perform layer-by-layer feature interaction through the feature interaction encoder to obtain the initial interaction representation;

[0143] S4, in each interaction level of the parallel feature interaction path, generate the modified interaction representation of each interaction level based on the initial interaction representation and the shared feature embedding representation;

[0144] S5, fuse the modified interaction representation of the last layer of the parallel feature interaction path to obtain the fused modified interaction representation;

[0145] S6. Based on the fusion and correction of the interaction representation, the predicted value of the probability of user interaction with candidate items is obtained.

[0146] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated in this embodiment.

[0147] Furthermore, in conjunction with the recommended methods provided in the above embodiments, this embodiment can also provide a storage medium for implementation. This storage medium stores a computer program; when executed by a processor, the computer program implements any of the recommended methods in the above embodiments.

[0148] 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, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0149] It should be understood that the specific embodiments described herein are merely illustrative of the application and not intended to limit it. All other embodiments derived by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0150] Obviously, the accompanying drawings are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, for those skilled in the art, certain design, manufacturing, or production modifications made based on the technical content disclosed in this application are merely conventional technical means and should not be considered as insufficient disclosure of this application.

[0151] The term "embodiment" in this application refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily imply the same embodiment, nor does it imply that it is mutually exclusive with or independent of other embodiments. It will be clearly or implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0152] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of patent protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A recommendation method, characterized in that, The recommendation method includes: Multi-feature-domain tabular data is transformed into a shared feature embedding representation through feature embedding; wherein, the multi-feature-domain tabular data includes user historical data, user information, and candidate item attribute information; Multiple parallel gated matrices are generated based on the shared feature embedding representation. The shared feature embedding representation is then multiplied bitwise with each gated matrix to generate multiple dynamically decoupled feature representations. The generation of multiple parallel gated matrices based on the shared feature embedding representation includes: modeling attention relationships between features with different optimization directions based on the shared feature embedding representation to obtain attention representations; modeling contextual information between features with different optimization directions based on the shared feature embedding representation to obtain contextual representations; and fusing the attention representations and contextual representations to obtain multiple parallel gated matrices. The dimension of each gated matrix is ​​consistent with the dimension of the shared feature embedding representation. The multiple parallel gated matrices represent the features with different optimization directions. The multiple dynamically decoupled feature representations are respectively input into the corresponding parallel feature interaction paths, and the feature interaction encoder performs layer-by-layer feature interaction to obtain the initial interaction representation. In each interaction level of the parallel feature interaction path, a modified interaction representation for each interaction level is generated based on the initial interaction representation and the shared feature embedding representation. The feature interaction path includes a corresponding feature interaction encoder. The feature interaction encoder is used to perform linear transformation and nonlinear activation function processing on the input of each interaction level in the corresponding feature interaction path to generate the initial interaction representation of the current level. In each interaction level, the input of the first layer is the dynamic decoupling feature representation, and the input of the interaction layers other than the first layer is the corrected interaction representation generated by the previous layer. Based on the initial interaction representation and the shared feature embedding representation, generating modified interaction representations for each interaction level includes: obtaining a gating value for each interaction level of the parallel feature interaction path based on the initial interaction representation; combining the gating value with the shared feature embedding representation to generate auxiliary interaction representations for each interaction level of the parallel feature interaction path; and fusing the auxiliary interaction representations for each interaction level of the parallel feature interaction path with the initial interaction representation to obtain the modified interaction representation. The modified interaction representations of the last layer of the parallel feature interaction path are fused to obtain the fused modified interaction representation. Based on the fusion-corrected interaction representation, the predicted value of the probability of user interaction with candidate items is obtained.

2. The recommended method according to claim 1, characterized in that, The process of transforming multi-feature-domain tabular data into a shared feature embedding representation through feature embedding includes: The original interaction logs and attribute information between users and candidate projects are collected from the client. After data cleaning and feature engineering, multi-feature domain tabular data is generated.

3. The recommended method according to claim 1, characterized in that, The step of fusing the modified interaction representations of the last layer of the parallel feature interaction path to obtain the fused modified interaction representation includes: The modified interaction representation of the last layer of the parallel feature interaction pathway is fused by the parallel feature interaction encoder.

4. The recommended method according to claim 1, characterized in that, The step of obtaining the predicted probability of user interaction with candidate items based on the fused and corrected interaction representation includes: Based on the fused and corrected interaction representation, the predicted value of the user's interaction probability with the candidate item is obtained by combining a feature vector classifier with an activation function; wherein, the feature vector classifier includes a logistic regression function and an artificial neural network.

5. The recommended method according to claim 1, characterized in that, The method further includes: Based on the predicted probability of user interaction with candidate items and the recommendation settings, a recommendation list of candidate items that matches the user's preferences is output.

6. A recommendation system, characterized in that, include: The module comprises a data feature embedding module, a decoupled feature generation module, and a gated interaction fusion module; among which: The data feature embedding module is used to transform multi-feature domain table data into a shared feature embedding representation through feature embedding; wherein, the multi-feature domain table data includes user historical data, user information, and candidate item attribute information; The decoupled feature generation module is used to generate multiple parallel gating matrices based on the shared feature embedding representation, and to perform positional multiplication of the shared feature embedding representation with each gating matrix to generate multiple dynamically decoupled feature representations. The generation of multiple parallel gating matrices based on the shared feature embedding representation includes: modeling the attention relationship between features with different optimization directions based on the shared feature embedding representation to obtain an attention representation; modeling the context information between features with different optimization directions based on the shared feature embedding representation to obtain a context representation; and fusing the attention representation and the context representation to obtain multiple parallel gating matrices. The dimension of the gating matrices is consistent with the dimension of the shared feature embedding representation. The multiple parallel gating matrices represent the features with different optimization directions. The gated interaction fusion module is used to input the multiple dynamically decoupled feature representations into corresponding parallel feature interaction paths, and perform layer-by-layer feature interaction through a feature interaction encoder to obtain an initial interaction representation. In each interaction level of the parallel feature interaction path, a revised interaction representation for each interaction level is generated based on the initial interaction representation and the shared feature embedding representation. Each feature interaction path includes a corresponding feature interaction encoder. The feature interaction encoder is used to perform linear transformation and nonlinear activation function processing on the input of each interaction level in the corresponding feature interaction path to generate the initial interaction representation for the current level. In each interaction level, the input of the first layer is the dynamically decoupled feature representation, and the inputs of the interaction layers other than the first layer are... The input is the corrected interaction representation generated in the previous layer; based on the initial interaction representation and the shared feature embedding representation, the corrected interaction representations for each interaction level are generated, including: obtaining the gate value of each interaction level of the parallel feature interaction path based on the initial interaction representation; combining the gate value with the shared feature embedding representation to generate auxiliary interaction representations for each interaction level of the parallel feature interaction path; fusing the auxiliary interaction representations of each interaction level of the parallel feature interaction path with the initial interaction representation to obtain the corrected interaction representation; fusing the corrected interaction representations of the last layer of the parallel feature interaction path to obtain the fused corrected interaction representation; and based on the fused corrected interaction representation, obtaining the predicted value of the user's interaction probability with the candidate item.

7. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the recommended method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the recommended method according to any one of claims 1 to 5.

Images