Personalized Recommendation Method, Device and Medium Based on High-Order Feature Interaction

Abstract

The present application discloses a personalized recommendation method, device and medium based on high-order feature interaction, which relates to the field of deep learning. The method includes: inputting the obtained user information and the knowledge in the knowledge base into a pre-trained personalized recommendation model to output a classification variable indicating whether the user has obtained knowledge; and performing personalized recommendation of knowledge to the user according to the classification variable. The personalized recommendation model is a deep learning-based model, which, compared with traditional shallow models (such as factorization machines, field factorization machines, etc.), can effectively learn low-order features with strong interactions by using a field factorization machine, and can construct high-order features by combining a cross network module and a deep learning module.

Description

Technical Field

[0001] This application relates to the field of deep learning, and specifically to a personalized recommendation method, device, and medium based on high-order feature interaction. Background Art

[0002] With the rapid development of Internet and big data technologies, recommendation systems have made progress in intelligence and personalization. However, in traditional solutions, recommendation systems mainly rely on simple statistics of rating data and user behavior, making it difficult to meet users' higher demands for personalization, real-time performance, and diversification.

[0003] With the application of deep learning and data mining technologies further promoting the intelligent development of recommendation systems, existing technologies still face the following problems to be solved: data sparsity, dynamic changes in user interests, insufficient content attributes, and poor model interpretability, etc. Summary of the Invention

[0004] To solve the above problems, this application proposes a personalized recommendation method based on high-order feature interaction, including:

[0005] Input the obtained user information and knowledge in the knowledge base into a pre-trained personalized recommendation model, output a classification variable indicating whether the user has obtained the knowledge, and perform personalized recommendation of knowledge for the user according to the classification variable;

[0006] Wherein, the personalized recommendation model includes an input layer, an embedding layer, a feature interaction layer, and a combined output layer;

[0007] The input layer inputs the user information and the knowledge in the form of feature vectors;

[0008] The embedding layer converts high-dimensional sparse feature vectors into low-dimensional dense embedding vectors;

[0009] The feature interaction layer includes a deep neural network module, a cross network module, a field factorization machine module, and an attention mechanism module;

[0010] The deep neural network module and the cross network module construct high-order feature interactions through a cross network and a deep neural network;

[0011] The field factorization machine module defines a feature domain through a field factorization machine, groups features, and constructs a linear relationship and a second-order feature interaction relationship of the features;

[0012] The attention mechanism module learns the weights of second-order feature interactions through an attention network;

[0013] The combined output layer makes predictions based on the output vectors of the deep neural network module, the cross network module, and the attention mechanism module.

[0014] In one example, the model training process of the personalized recommendation model includes:

[0015] The input layer obtains a data set and, through encoding, converts the information features in the data set into feature vectors;

[0016] Among them, the information features include a feature column and a label column. The feature column includes multi-dimensional features corresponding to user information and knowledge, and the label column includes whether the user has obtained this knowledge;

[0017] The feature types of the multi-dimensional features include categorical features and continuous features, which are processed through encoding respectively.

[0018] In one example, for the feature vector of each feature domain, the embedding layer converts it into an embedding vector corresponding to this feature domain through the embedding matrix corresponding to this feature domain;

[0019] The embedding vectors corresponding to each feature domain are combined to obtain the output vector of the embedding layer.

[0020] In one example, the deep neural network module includes a Product layer, a hidden layer, and an output layer;

[0021] The Product layer includes a linear part and a non-linear part;

[0022] The linear part outputs the corresponding output result of the linear part through the corresponding weight matrix and the linear signal vector; among them, the linear signal vector is obtained through the corresponding embedding vector;

[0023] The non-linear part outputs the corresponding output result of the non-linear part through the corresponding weight matrix and the quadratic signal vector; among them, the linear signal vector is obtained through multiple embedding vectors.

[0024] In one example, the cross network module includes multiple cross layers;

[0025] Each cross layer performs a cross calculation on the feature vector input in the current layer and the output vector of the previous layer, adjusts the contribution of each layer in the cross calculation through the weight vector of the current layer, and obtains the output of the current layer through the adjusted result, the original input vector, and the bias term.

[0026] In one example, the field factorization machine module defines the features corresponding to a single class for each feature field, and each feature field includes multiple features;

[0027] For each feature, assign multiple hidden vectors to the feature;

[0028] For each feature in feature interaction, calculate through the hidden vector of the feature in the feature domain where the other feature is located, and based on the calculation results of all feature interactions, as well as the feature weights and global bias terms corresponding to all features, output the corresponding linear relationship and second-order feature interaction relationship.

[0029] In one example, the attention mechanism module processes the weight matrix corresponding to the second-order feature interaction, the calculation result of the hidden vector, and the corresponding bias term through an activation function, and obtains the corresponding attention score based on the processing result and the attention network parameters, and normalizes the attention score.

[0030] In one example, the combined output layer concatenates the output vectors of the deep neural network module, the cross network module, and the attention mechanism module, as well as the bias term, and makes a prediction on the concatenated vector through an activation function.

[0031] On the other hand, the present application also proposes a personalized recommendation device based on high-order feature interaction, including:

[0032] At least one processor; and,

[0033] A memory communicatively connected to the at least one processor; wherein,

[0034] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: perform the personalized recommendation method based on high-order feature interaction described in any of the above examples.

[0035] On the other hand, the present application also proposes a non-volatile computer storage medium storing computer-executable instructions, characterized in that the computer-executable instructions are set to: perform the personalized recommendation method based on high-order feature interaction described in any of the above examples.

[0036] The personalized recommendation method based on high-order feature interaction proposed by the present application can bring the following beneficial effects:

[0037] The personalized recommendation model in the present application is a deep learning-based model, which, compared with traditional shallow models (such as factorization machines, field factorization machines, etc.), can effectively learn low-order features with strong interactions by using field factorization machines, and can combine a cross network module and a deep learning module to construct high-order features.

[0038] Compared with traditional models such as DeepFM, which only implicitly process feature interactions through deep neural networks, the parallel architecture of the personalized recommendation model in this application enables it to explicitly construct high-order feature interactions of any order through the cross network.

[0039] The personalized recommendation model introduces a Product layer before the hidden layer of the deep learning module to capture the non-linear relationship between features, further enhancing the ability of high-order feature interactions.

[0040] Compared with traditional models such as xDeepFM and EDCN, although they also combine explicit and implicit high-order feature construction, they do not, like the personalized recommendation model in this application, use a shallow module to construct second-order feature interactions, and they pay even less attention to the weights of each second-order feature interaction, causing the traditional model to lose the ability to capture low-order key information.

[0041] The personalized recommendation model can effectively capture the complex relationship between users and knowledge, thereby effectively improving the pertinence and accuracy of knowledge recommendation for users. It solves problems such as data sparsity, dynamic changes in user interests, insufficient content attributes, and poor model interpretability in traditional solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of this application and form a part of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0043] Figure 1 It is a schematic flowchart of the personalized recommendation method based on high-order feature interaction in an embodiment of this application;

[0044] Figure 2 It is a schematic diagram of the architecture of the personalized recommendation model in one case of an embodiment of this application;

[0045] Figure 3 It is a schematic diagram of the conversion process of the embedding layer in one case of an embodiment of this application;

[0046] Figure 4 It is a schematic diagram of the structure of the deep neural network module in one case of an embodiment of this application;

[0047] Figure 5 It is a schematic diagram of the computational visualization of the cross network module in one case of an embodiment of this application;

[0048] Figure 6 It is a schematic diagram of the personalized recommendation device based on high-order feature interaction in an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following will clearly and completely describe the technical solutions of this application in combination with specific embodiments of this application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

[0050] The following will, in combination with the drawings, elaborate on the technical solutions provided by each embodiment of this application.

[0051] As Figure 1 shown, the embodiments of this application provide a personalized recommendation method based on high-order feature interaction, including:

[0052] S101: Input the obtained user information and the knowledge in the knowledge base into a pre-trained personalized recommendation model, and output a classification variable indicating whether the user has obtained the knowledge.

[0053] The initiation of personalized recommendation can be active or passive. For example, after a user issues a question on a search platform, corresponding personalized recommendation is made passively according to the user's needs. Or, on an e-commerce platform, corresponding products are actively recommended to the user according to the user's purchase records.

[0054] The user information and the knowledge in the knowledge base are both preset and obtained. Based on different actual scenarios, the knowledge and the knowledge base can also be set to corresponding types. For example, the knowledge can include: web addresses, product knowledge, skill knowledge, business knowledge, etc.

[0055] When personalized recommendation needs to be made to the user, according to the current recommendation scenario (such as, including recommending products, recommending books, recommending services, etc.), select the corresponding knowledge base, and use the classification variable output by the personalized recommendation model to indicate whether the user has obtained the corresponding knowledge.

[0056] The classification variable output by the personalized recommendation model (Higher-Order Feature Interaction Model, HFIM) is a binary classification variable 0 or 1, indicating whether the user has obtained the knowledge.

[0057] S102: Perform personalized recommendation of knowledge to the user according to the classification variable.

[0058] For different recommendation scenarios, the final personalized recommendation results may also vary. For example, for user search queries, knowledge that the user has not yet obtained can be preferentially recommended. For product recommendations, based on the products that the user has already purchased (corresponding to the knowledge that has been obtained), the same or similar products can be recommended. For book recommendations, based on the books that the user has already read (corresponding to the knowledge that has been obtained), similar or unread books can be recommended. For business recommendations, based on the business operations that the user has already carried out (corresponding to the knowledge that has been obtained), and referring to the business operation process, the next business operations to be carried out can be recommended.

[0059] The HIFM model realizes the linear relationship and second-order feature interaction between features through a field factorization machine, and introduces an attention network to learn the weight information of second-order interaction features; constructs high-order feature interactions explicitly through a cross network, and implicitly constructs high-order feature interactions using a deep neural network; finally, concatenates the outputs of the deep neural network, the cross network, and the field factorization machine passing through the attention network as the final prediction result of the model.

[0060] As Figure 2 shown, the personalized recommendation model includes an input layer, an embedding layer, a feature interaction layer, and a combined output layer. The input layer inputs the processed user information and knowledge in the dataset in the form of feature vectors. The embedding layer converts the high-dimensional sparse feature vectors into low-dimensional dense embedding vectors. The feature interaction layer includes a deep neural network module, a cross network module, a field factorization machine module, and an attention mechanism module; the deep neural network module and the cross network module construct high-order feature interactions through the cross network and the deep neural network; the field factorization machine module defines the feature fields through the field factorization machine, groups the features, and constructs the linear relationship and second-order feature interaction relationship of the features; the attention mechanism module learns the weights of second-order feature interactions through the attention network. The combined output layer makes predictions based on the output vectors of the deep neural network module, the cross network module, and the attention mechanism module.

[0061] The input layer is the first step of the HIFM model, and its purpose is to input the information features of the user and the information features of the knowledge into the model in the form of vectors. During the model training process of the personalized recommendation model, the original dataset is set as the training sample (after training, this dataset corresponds to the actual data content to be predicted).

[0062] During the training process, the dataset is obtained, and through encoding, the information features in the dataset are converted into feature vectors.

[0063] Among them, the information features include a feature column and a label column. The feature column includes multi-dimensional features corresponding to user information and knowledge, such as various features like the user's gender, age, hobbies, knowledge type, etc. The label column includes whether the user has obtained the knowledge (for example, by determining whether the user has clicked on the knowledge on the web page or platform, it is determined whether the user has obtained the knowledge).

[0064] The feature types of the multi-dimensional features include categorical features and continuous features, which are processed by encoding respectively. For the categorical features, the feature processing method adopted is to convert the feature data into the form of One-Hot encoding. For example, the gender feature is encoded as 10 or 01. If the first digit is 1, it represents male, and if the second digit is 1, it represents female. For continuous features, a discretization method is selected to process the continuous feature data.

[0065] Embedding layer: For the feature vectors of each feature domain, through the embedding matrix corresponding to the feature domain, they are converted into the embedding vectors corresponding to the feature domain. The embedding vectors corresponding to each feature domain are combined to obtain the output vector of the embedding layer.

[0066] Although One-Hot encoding standardizes the model input to a certain extent, when dealing with features with a large number of categories such as knowledge type, it will cause a sharp increase in the data dimension and generate high-dimensional sparse data. In this case, the effective expression of the features may be affected. If directly calculated, it will lead to a very large computational amount of the model and cause losses to the model performance. Therefore, the HFIM model introduces an embedding layer after the input layer to convert high-dimensional sparse features into low-dimensional dense vectors. The conversion process of the embedding is shown in Formula 1:

[0067] Formula 1

[0068] Among them, represents the feature vector of the th feature domain, represents the embedding matrix of the th feature domain, represents the embedding vector of the th feature domain.

[0069] As Figure 3 shown, feature vectors of different dimensions , after being processed by the embedding layer, a group of embedding vectors with the same dimension can be obtained. During the conversion process, each parameter in each embedding matrix is first initialized to a certain floating-point value, and finally the optimal parameter results are learned through the backpropagation of the model. The definition of the output vector E of the embedding layer is shown in Formula 2:

[0070] Formula 2;

[0071] Among them, represents the output vector of the embedding layer, represents the th embedding vector of the feature domain, represents the total number of feature domains, represents the dimensionality size of the embedding vector.

[0072] As Figure 4 shown, the deep neural network module includes a Product layer, a hidden layer, and an output layer. HIFM implicitly constructs high-order feature combinations through a deep neural network. Its essence is a feedforward neural network, but it is difficult to fully capture the relationships between features by only privately learning high-order features in a fully connected manner. Therefore, a Product layer is added in front of the first hidden layer to further capture the non-linear relationships between features and achieve the cross of features.

[0073] The Product layer includes a linear part and a non-linear part. The linear part outputs the corresponding linear part output result through the corresponding weight matrix and the linear signal vector; among them, the linear signal vector is obtained through the corresponding embedding vector; the non-linear part outputs the corresponding non-linear part output result through the corresponding weight matrix and the quadratic signal vector; among them, the linear signal vector is obtained through multiple embedding vectors.

[0074] The Product layer mainly focuses on the original information of the embedding vectors and the interaction information of each embedding vector, making the information richer and thus improving the model effect. This layer is composed of a linear part and a non-linear part The specific calculation formulas are shown in Formula 3 and Formula 4:

[0075] Formula 3;

[0076] Formula 4;

[0077] Among them, represents the linear signal vector, represents the quadratic signal vector, represents the Hadamard product operation, represents the embedding vector, represents and the inner product of the vectors, , represent the weight matrix.

[0078] As Figure 4As shown in the figure, on the left side of the Product layer, the embedded vector is directly copied, and a constant signal "1" is introduced in the embedding layer to multiply with it. On the right side, the inner product cross calculation is performed pairwise on the embedded vectors. Then, the fully connected layer converts the dimensions of the two parts into the same dimension and inputs them into the subsequent hidden layer. By introducing the Product layer, the interaction between features is further enhanced.

[0079] The cross network module includes multiple cross layers; for each cross layer, the feature vector input by the current layer is used to perform cross calculation with the output vector of the previous layer, and the contribution of each layer in the cross calculation is adjusted by the weight vector of the current layer. The output of the current layer is obtained by adding the adjusted result to the original input vector and the bias term.

[0080] The cross between features can provide richer information than individual features. Although deep neural networks can implicitly construct high-order features through multiple non-linear transformations, they have problems of insufficient interpretability and low efficiency. Therefore, through the Cross Network, feature crosses are applied at each layer to explicitly construct high-order feature combinations. The cross network is stacked by multiple cross layers, and each layer captures higher-order crosses of the input features. Specifically, each layer calculates the outer product of the input feature vector with itself based on the output of the previous layer, then adjusts the contribution of these crosses by taking the dot product with the weight vector of this layer, and finally adds the result to the original input vector and a bias term to form the output of this layer. The calculation formula is shown in Formula Five:

[0081] Formula Five;

[0082] Among them, and respectively represent the output of the cross layer of the th layer and the th layer of the cross network, represents the learning parameter of the th layer, represents the offset of the th layer. As Figure 5 shown, based on the residual mechanism, the output of each cross layer needs to be added to the output of the previous layer. For example, the output of the th layer includes the constructed cross feature and the output of the th layer.

[0083] The above is the macroscopic calculation method of the cross network. The following details the calculation process of each layer through the specific calculation of the first layer. Let be the embedded vector output by the embedding layer, and the embedding dimension is , the bias term is 0, and the output of the first cross layer is shown in Equation (6):

[0084] Equation (6);

[0085] As can be seen from Equation (6), in the first layer, first is multiplied by its transpose to perform an outer product operation to form a matrix, where each element in the matrix is the product of two elements in the embedding vector. Then, the matrix is multiplied by the parameter vector to generate second-order cross terms. Finally, based on the residual mechanism, the second-order cross terms are added to the output of the previous layer . By repeating this process and stacking multiple cross layers, high-order features can be effectively constructed. In addition, the time and space complexity of the cross network are linear in the input dimension. Therefore, the complexity introduced by introducing the cross network is negligible compared to the deep neural network module.

[0086] The field-aware factorization machine module defines the features corresponding to each single class for each feature field, and each feature field includes multiple features; for each feature, multiple latent vectors are assigned to this feature; for each feature in the feature interaction, calculations are performed through the latent vectors of this feature in the feature fields where the other features are located, and based on the calculation results of all feature interactions, as well as the feature weights and global bias terms corresponding to all features, the corresponding linear relationship and second-order feature interaction relationship are output.

[0087] The field-aware factorization machine (FFM) is an improvement based on the factorization machine (FM), aiming to improve the accuracy of the prediction model in dealing with complex interactions between features. The field-aware factorization machine groups features by introducing the concept of fields, which fully improves the interaction ability between features. The relationship between the defined fields and features is . Each category of features of knowledge can be called a feature field, and each feature field contains multiple features. For example, the knowledge type can be used as a feature field, and the technology category, as a type of knowledge, can be considered as a feature. In the FM model, each feature uses the same latent vector when interacting with other features, while the FFM model assigns multiple latent vectors to each feature, enabling features from different fields to use different latent vectors to respond when interacting. The expression of FFM is shown in Equation (7):

[0088] Equation (7);

[0089] where, represents the global bias term, represents the total number of features, Represents the weight of the th feature, represents the vector representation of the th feature, and respectively represent the latent vectors of feature and feature in the other domain. Represents the dot product of the latent vectors. The difference from the FM model lies in that the latent vector changes from to , that is, each feature changes from the original unique latent vector to a group of latent vectors, which enables the FFM model to capture the interactions between features more meticulously.

[0090] The attention mechanism module processes the weight matrix corresponding to the second-order feature interaction, the calculation result of the latent vector, and the corresponding bias term through an activation function, and obtains the corresponding attention score according to the processing result and the attention network parameters, and normalizes the attention score.

[0091] Generally speaking, not all feature interactions play a positive role in the performance of the model. Some unimportant feature interactions may even affect the performance of the model. Therefore, the HIF-CSR model assigns different weights to different second-order feature interactions by introducing an attention network, enabling the model to automatically learn the importance of feature interactions. The weights are calculated through an attention network, and the calculation process is shown in Formulas Eight and Nine:

[0092] Formula Eight;

[0093] Formula Nine;

[0094] Among them, are the attention network parameters, is the activation function, represents the weight matrix of the feature interaction, represents the dot product of the latent vectors, represents the bias term, is the attention score, represents the normalized attention score.

[0095] The combined output layer concatenates the output vectors of the deep neural network module, the cross network module, the attention mechanism module, and the bias term, and makes a prediction on the concatenated vector through an activation function.

[0096] The combined output layer concatenates the outputs of the deep neural network module, the cross network module, and the output of the field factorization machine passing through the attention network module, and predicts whether the user obtains and browses the knowledge through an activation function. The deep neural network captures complex patterns by learning the non-linear combination of input features, while the cross network focuses on learning the cross combinations between features, which is very valuable for understanding the interactions between features. The field factorization machine further enhances the model's ability to handle interactions between different feature domains, allowing the model to learn the interactions between features at a finer granularity and learn the weights of different interaction features. Information is merged by concatenating the output vectors of each layer end to end, thus forming a more comprehensive feature representation. This fusion strategy not only enhances the model's ability to capture complex patterns in the data, but also provides a rich information basis for the final computational output, thereby improving the prediction accuracy and generalization ability of the entire personalized recommendation model.

[0097] The combined output layer uses an activation function to calculate the concatenated output vector to obtain the final output of the model. The output of the final model is shown in Equation Ten:

[0098] Equation Ten;

[0099] where and represent the outputs of the deep neural network and the cross network respectively, represents the output of the field factorization machine passing through the attention network, represents the bias term.

[0100] The personalized recommendation model in this application is a deep learning-based model. Compared with traditional shallow models (such as factorization machines, field factorization machines, etc.), it can effectively learn low-order features with strong interactions by using a field factorization machine, and can combine a cross network module and a deep learning module to construct high-order features.

[0101] Compared with traditional models such as DeepFM, which only implicitly processes feature interactions through a deep neural network, the parallel architecture of the personalized recommendation model in this application enables it to explicitly construct high-order feature interactions of any order through a cross network.

[0102] The personalized recommendation model introduces a Product layer before the hidden layer of the deep learning module to capture the non-linear relationship between features, further strengthening the ability of high-order feature interactions.

[0103] Compared with traditional models such as xDeepFM and EDCN, although they also combine the construction of explicit and implicit high-order features, they do not use a shallow module to construct second-order feature interactions like the personalized recommendation model in this application, and they do not pay attention to the weights of each second-order feature interaction, which makes the traditional model lose the ability to capture low-order key information.

[0104] The personalized recommendation model can effectively capture the complex relationship between users and knowledge, thereby effectively improving the pertinence and accuracy of recommending knowledge to users. It solves problems such as data sparsity, dynamic changes in user interests, insufficient content attributes, and poor model interpretability in traditional solutions.

[0105] As Figure 6 shown, the embodiment of the present application also provides a personalized recommendation device based on high-order feature interaction, including:

[0106] At least one processor; and,

[0107] A memory communicatively connected to the at least one processor; wherein,

[0108] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: perform the personalized recommendation method based on high-order feature interaction described in any of the above embodiments.

[0109] The embodiment of the present application also provides a non-volatile computer storage medium storing computer-executable instructions, and the computer-executable instructions are set to: perform the personalized recommendation method based on high-order feature interaction described in any of the above embodiments.

[0110] Each embodiment in the present application is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0111] The device and medium provided by the embodiment of the present application correspond one-to-one with the method. Therefore, the device and medium also have beneficial technical effects similar to those of the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the device and medium are not described herein again.

[0112] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A personalized recommendation method based on high-order feature interaction, characterized in that: include: Input the acquired user information and the knowledge in the knowledge base into the pre-trained personalized recommendation model, and output a classification variable indicating whether the user has acquired the knowledge; Providing personalized knowledge recommendations to users based on the classification variables; The personalized recommendation model includes an input layer, an embedding layer, a feature interaction layer, and a combined output layer; The input layer inputs the user information and the knowledge in the form of feature vectors; The embedding layer converts the high-dimensional sparse feature vector into a low-dimensional dense embedding vector; The feature interaction layer includes a deep neural network module, a cross network module, a domain factorization machine module, and an attention mechanism module; The deep neural network module and the cross network module construct high-order feature interactions through the cross network and the deep neural network; The domain factorization machine module defines a feature domain through a domain factorization machine, groups features, and constructs linear relationships of features and second-order feature interaction relationships; The attention mechanism module learns the weights of second-order feature interactions through an attention network; The combined output layer makes predictions based on the output vectors of the deep neural network module, the cross network module, and the attention mechanism module; The personalized recommendation model includes the following steps in the model training process: The input layer obtains a data set and converts the information features in the data set into a feature vector through encoding; The information features include a feature column and a label column. The feature column includes user information and multi-dimensional features corresponding to knowledge. The label column includes whether the user has acquired the knowledge. The feature types of the multi-dimensional features include categorical features and continuous features, which are processed by encoding respectively; The deep neural network module includes a Product layer, a hidden layer, and an output layer; The Product layer includes a linear part and a nonlinear part; The linear part outputs the corresponding linear part output result through the corresponding weight matrix and the linear signal vector; wherein the linear signal vector is obtained through the corresponding embedding vector; The nonlinear part outputs the corresponding nonlinear part output result through the corresponding weight matrix and the secondary signal vector; wherein the linear signal vector is obtained through multiple embedding vectors.

2. The personalized recommendation method based on high-order feature interaction according to claim 1, characterized in that: The embedding layer converts the feature vector of each feature domain into an embedding vector corresponding to the feature domain through the embedding matrix corresponding to the feature domain; The embedding vectors corresponding to the feature fields are combined to obtain the output vector of the embedding layer.

3. The personalized recommendation method based on high-order feature interaction according to claim 1, characterized in that: The cross network module includes a plurality of cross layers; Each cross layer performs cross calculations with the feature vector of the current layer input and the output vector of the previous layer, and adjusts the contribution of each layer in the cross calculation through the weight vector of the current layer, and obtains the output of the current layer through the adjusted result, the original input vector and the bias term.

4. The personalized recommendation method based on high-order feature interaction according to claim 1, characterized in that: The domain factorization machine module defines the features of each feature domain corresponding to a single class, and each feature domain includes multiple features; For each feature, multiple latent vectors are assigned to the feature; For each feature in the feature interaction, the latent vector of the feature in the feature domain of the other feature is calculated, and based on the calculation results of all feature interactions, as well as the feature weights and global bias terms corresponding to all traits, the corresponding linear relationship and second-order feature interaction relationship are output.

5. The personalized recommendation method based on high-order feature interaction according to claim 4, characterized in that: The attention mechanism module processes the weight matrix corresponding to the second-order feature interaction, the calculation result of the latent vector, and the corresponding bias term through an activation function, obtains the corresponding attention score according to the processing result and the attention network parameter, and normalizes the attention score.

6. The personalized recommendation method based on high-order feature interaction according to claim 1, characterized in that: The combined output layer concatenates the output vectors of the deep neural network module, the cross network module, the attention mechanism module, and the bias term, and predicts the concatenated vector through an activation function.

7. A personalized recommendation device based on high-order feature interaction, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: provide a personalized recommendation method based on high-order feature interaction as described in any one of claims 1 to 6.

8. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are configured as: a personalized recommendation method based on high-order feature interaction as described in any one of claims 1 to 6 above.

Images