Data commodity recommendation method and system based on deep learning

Abstract

The invention relates to a data commodity recommendation method and system based on deep learning, and the method comprises the steps: carrying out the word embedding coding of a user text and a commodity text of a data commodity through a first deep learning model, and obtaining a user text matrix representation and a commodity text matrix representation; respectively carrying out feature extraction and feature compression on the user text matrix representation and the commodity text matrix representation by utilizing a second deep learning model to obtain a user text vector representation and a commodity text vector representation; fusing the user text vector representation and the commodity text vector representation to obtain a fusion vector; and calculating a prediction score of the data commodity according to the fusion vector. According to the method, the data commodity recommendation effect can be improved, the recommended data commodity can meet the purchase demand of the user, and the user experience and the transaction efficiency are improved.

Description

Technical Field

[0001] This application mainly relates to the field of deep learning technology, and specifically relates to a method, system and computer-readable medium for data commodity recommendation based on deep learning. Background Art

[0002] With the rapid development of computer science and technology, as well as fields such as big data, cloud computing, and artificial intelligence, the growth rate of various types of data has increased significantly. At the same time, the increasing maturity of related technologies such as data analysis and data mining has promoted the rapid development of data-related industries, and data has gradually become a unique resource with important commercial and research value. Therefore, the demand for trading valuable data elements has been increasing, and many data trading platforms have emerged.

[0003] Data trading platforms have been in existence for a short time and have limited business scope, and they are different from traditional e-commerce platforms. For example, data trading scenarios are scarce in data sets, and data commodity information is basically represented by text, lacking other types of multimodal data. While traditional e-commerce platforms usually have mature recommendation systems that can effectively help users select products using multimodal data. In contrast, due to the poor performance of the recommendation algorithm in data trading platforms, users face great difficulties in selecting data commodities, which affects the user experience and transaction efficiency of the platform. Summary of the Invention

[0004] The technical problem to be solved by this application is to provide a method, system and computer-readable medium for data commodity recommendation based on deep learning, which can improve the effect of data commodity recommendation, enable the recommended data commodities to meet the purchase needs of users, improve the user experience and transaction efficiency, and the data commodity recommendation method has good scalability and can be extended to other e-commerce recommendation systems for users and commodities represented only by text.

[0005] The technical solution adopted by this application to solve the above technical problem is a method for data commodity recommendation based on deep learning, including: using a first deep learning model to perform word embedding encoding on the user text and commodity text of data commodities respectively to obtain a user text matrix representation and a commodity text matrix representation; using a second deep learning model to perform feature extraction and feature compression on the user text matrix representation and the commodity text matrix representation respectively to obtain a user text vector representation and a commodity text vector representation; fusing the user text vector representation and the commodity text vector representation to obtain a fused vector; calculating the predicted score of the data commodity according to the fused vector.

[0006] In one embodiment of the present application, calculating the predicted score of a data product based on the fusion vector includes: performing a linear transformation on the fusion vector to obtain a first score; calculating a first interaction score and a second interaction score based on the fusion vector; the first interaction score and the second interaction score are used to represent the interaction degree between the user and the product; calculating the predicted score based on the first score, the first interaction score, and the second interaction score.

[0007] In one embodiment of the present application, calculating the first interaction score and the second interaction score based on the fusion vector includes: calculating the first interaction score using the following formula:

[0008]

[0009] Calculating the second interaction score using the following formula:

[0010]

[0011] Where, represents the first interaction score; represents the second interaction score; ui vector represents the fusion vector; fm V represents the factor matrix, fm V is obtained by initialization with a uniform distribution or a normal distribution.

[0012] In one embodiment of the present application, calculating the predicted score based on the first score, the first interaction score, and the second interaction score includes: calculating the predicted score using the following formula:

[0013]

[0014] Where, predict rating represents the predicted score; p1 represents the first fusion coefficient, and p1 is a constant; fm linear_part represents the first score.

[0015] In one embodiment of the present application, calculating the predicted score based on the first score, the first interaction score, and the second interaction score includes: performing a bilinear interaction process on the first interaction score and the second interaction score using the following formula to obtain a bilinear interaction score:

[0016]

[0017] Where, bilinear represents the bilinear interaction score; p2 represents the second fusion coefficient, and p2 is a constant;

[0018] Performing a non-linear mapping and a linear transformation on the bilinear interaction score using the following formula to obtain a third interaction score:

[0019] fminteractions = h(drop_out(relu(mlp(bilinear))))

[0020] Among them, fm interactions represents the third interaction score; mlp(·) represents a multi-layer perceptron, and the multi-layer perceptron is used for non-linear mapping; relu(·) represents an activation function; drop_out(·) represents random dropout processing; h(·) represents an output layer, and the output layer is used for linear transformation;

[0021] The following formula is used to calculate the predicted score:

[0022] predict rating = fm interactions + fm linear_part

[0023] Among them, predict rating represents the predicted score; fm linear_part represents the first score.

[0024] In an embodiment of the present application, calculating the predicted score of the data product according to the fusion vector includes using the following formula to calculate the predicted score:

[0025] predict rating = relu(fc(ui vector ))

[0026] Among them, predict rating represents the predicted score; ui vector represents the fusion vector; fc(·) represents a linear layer; relu(·) represents an activation function.

[0027] In an embodiment of the present application, the first deep learning model includes a BERT model.

[0028] In an embodiment of the present application, the second deep learning model includes a first convolutional neural network and a second convolutional neural network. The first convolutional neural network includes a first convolutional layer, a first pooling layer, and a first fully connected layer. The second convolutional neural network includes a second convolutional layer, a second pooling layer, and a second fully connected layer. Feature extraction and feature compression are respectively performed on the user text matrix representation and the commodity text matrix representation by using the second deep learning model to obtain a user text vector representation and a commodity text vector representation, including: performing feature extraction on the user text matrix representation by using the first convolutional layer to obtain a user text matrix; performing feature extraction on the commodity text matrix representation by using the second convolutional layer to obtain a commodity text matrix; performing feature compression on the user text matrix by using the first pooling layer to obtain a user text compressed matrix; performing feature compression on the commodity text matrix by using the second pooling layer to obtain a commodity text compressed matrix; processing the user text compressed matrix by using the first fully connected layer to obtain a user text vector representation; and processing the commodity text compressed matrix by using the second fully connected layer to obtain a commodity text vector representation.

[0029] In an embodiment of the present application, fusing the user text vector representation and the commodity text vector representation to obtain a fusion vector includes: concatenating the user text vector representation and the commodity text vector representation to obtain a fusion vector; or adding the user text vector representation and the commodity text vector representation to obtain a fusion vector; or taking the inner product of the user text vector representation and the commodity text vector representation to obtain a fusion vector.

[0030] In an embodiment of the present application, the user text includes a user ID and a user name; the commodity text includes one or any combination of a commodity ID, a commodity name, a commodity description, and a user rating of the commodity.

[0031] The present application also proposes a data commodity recommendation system based on deep learning to solve the above technical problems, including: a data commodity display module for sorting and displaying data commodities according to predicted ratings; a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement the above data commodity recommendation method based on deep learning.

[0032] The present application also proposes a computer-readable medium storing computer program code, and the computer program code implements the above data commodity recommendation method based on deep learning when executed by a processor.

[0033] The technical solution of this application encodes the user text and commodity text through word embedding, converts the text information into a high-dimensional matrix representation, and effectively captures the semantic features of users and commodities; by extracting and compressing the features of the matrix representation, the core feature vectors of users and commodities are further refined; by fusing the feature vectors of users and commodities, the association between user preferences and commodity characteristics is fully considered; based on the fused vector, the predicted score of the data commodity is calculated, and personalized recommendation results can be provided for users. This application can improve the effect of data commodity recommendation, enable the recommended data commodities to meet the user's purchase needs, improve the user experience and transaction efficiency. The data commodity recommendation method has good scalability and can be extended to other e-commerce recommendation systems for users and commodities with only text representations. Description of the Drawings

[0034] To make the above objects, features, and advantages of this application more obvious and understandable, the following provides a detailed description of the specific implementation manners of this application with reference to the accompanying drawings, where:

[0035] Figure 1 is an exemplary flowchart of a data commodity recommendation method based on deep learning according to an embodiment of this application;

[0036] Figure 2 is an exemplary flowchart of a data commodity recommendation method based on deep learning according to another embodiment of this application;

[0037] Figure 3 is a schematic diagram of the model training result of a data commodity recommendation system according to an embodiment of this application;

[0038] Figure 4 is a schematic diagram of the storage order of data commodities in a database according to an embodiment of this application;

[0039] Figure 5 is a schematic diagram of a data commodity recommendation system before using the data commodity recommendation method according to an embodiment of this application;

[0040] Figure 6 is a schematic diagram of a data commodity recommendation system after using the data commodity recommendation method according to an embodiment of this application;

[0041] Figure 7 is a schematic diagram of the recommendation results for users in the same company but different departments according to an embodiment of this application;

[0042] Figure 8 is a schematic diagram of the recommendation results for users in different companies and different departments according to an embodiment of this application;

[0043] Figure 9 is a system block diagram of a data commodity recommendation system based on deep learning according to an embodiment of this application. Detailed Implementation Manner

[0044] To make the above objectives, features, and advantages of the present application more obvious and understandable, the following provides a detailed description of the specific implementation manners of the present application in conjunction with the accompanying drawings.

[0045] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, the present application may be implemented in other ways different from those described herein. Therefore, the present application is not limited by the specific embodiments disclosed below.

[0046] As shown in the present application and the claims, unless the context clearly indicates otherwise, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. A method or device may also include other steps or elements.

[0047] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations before or below are not necessarily executed precisely in sequence. On the contrary, they can be executed in reverse order or simultaneously. At the same time, other operations may be added to these processes, or one or more steps may be removed from these processes.

[0048] First, the analysis of the data product recommendation system (i.e., the data trading platform) in the present application and the inventive concept of the present application are introduced below.

[0049] The main function of the recommendation system is to establish a connection between users and products. On the one hand, it recommends valuable products to users, and on the other hand, it presents products to users who are interested in them. As a special e-commerce trading scenario, the data trading platform has particularities compared with traditional e-commerce platforms, which are mainly reflected in the following aspects: (1) The data trading platform usually only involves data products, while traditional e-commerce platforms cover various product types; (2) The products on traditional e-commerce platforms usually have rich metadata and supplementary information, such as various multimedia forms including text, pictures, videos, etc., and user comments. However, due to the particularity of the products on the data trading platform, it mainly relies on text information to describe products and it is difficult to use other forms of supplementary information for recommendation.

[0050] In recent years, various recommendation system algorithms have emerged for traditional e-commerce scenarios. However, these methods usually require the products themselves to have rich information, such as text introductions, pictures, videos, user comments, and the browsing behaviors of other users, etc., as the input of the recommendation algorithm. For data trading platforms that have existed for a short time, it is often difficult to obtain this information, resulting in the ineffective application of existing recommendation algorithms in data trading scenarios.

[0051] Due to the lack of relevant experience, the recommendation method based on simple rules often has problems such as unreasonable recommendation rules and poor interpretability. Some data trading platforms attempt to use methods such as collaborative filtering and matrix factorization for recommendation, but these methods will reduce the recommendation performance in the case of data sparsity. When the platform scale is large and the number of rated items is relatively small compared to the total number of items, the collaborative filtering method is difficult to provide accurate recommendation results. Therefore, there is an urgent need to design a recommendation algorithm suitable for data trading platforms and with good scalability.

[0052] Based on the actual needs of the data trading scenario, this application proposes a recommendation system solution suitable for data trading platforms to address the problem of users having difficulty selecting items in data trading platforms. The data item recommendation solution of this application models users and items separately, uses a deep learning model (such as a convolutional neural network) to extract the features of users and items, and at the same time considers the interaction relationship between users and items. The data item recommendation method of this application is not only applicable to data trading scenarios, but can also be extended to other e-commerce recommendation systems that only rely on text information to describe users and items, and can be used as a general text data recommendation algorithm. This application provides an efficient, scalable and accurate recommendation solution for data trading platforms, improving the user experience and platform trading efficiency.

[0053] The following will introduce the data item recommendation method based on deep learning of this application.

[0054] This application proposes a data item recommendation method based on deep learning, which can be applied to data trading scenarios. The data item recommendation method based on deep learning of this application can run on a variety of terminal devices, for example, it can run in the controller of a mobile terminal or a server, and can also run in a cloud platform. When the data item recommendation method based on deep learning runs on a cloud platform, the data of the terminal device interacts with the cloud platform data through a wireless network. Exemplarily, the cloud platform can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an interconnected cloud, a multi-cloud, etc. or any combination thereof. This application places no restrictions on the operating environment of the data item recommendation method based on deep learning.

[0055] Figure 1 is an exemplary flowchart of the data item recommendation method based on deep learning in an embodiment of this application. Refer to Figure 1 As shown, the data item recommendation method based on deep learning in this embodiment includes the following steps:

[0056] Step S110: Use a first deep learning model to perform word embedding encoding on the user text and item text of data items respectively to obtain a user text matrix representation and an item text matrix representation.

[0057] Step S120: Use the second deep learning model to perform feature extraction and feature compression on the user text matrix representation and the product text matrix representation respectively, to obtain the user text vector representation and the product text vector representation.

[0058] Step S130: Fuse the user text vector representation and the product text vector representation to obtain a fused vector.

[0059] Step S140: Calculate the predicted score of the data product according to the fused vector.

[0060] The above steps S110 to S140 are described in detail below:

[0061] In step S110, use the first deep learning model to perform word embedding encoding on the user text and product text of the data product respectively, to obtain the user text matrix representation and the product text matrix representation. Exemplarily, by performing word embedding encoding on the text, the text is transformed into a high-dimensional matrix representation, which can effectively capture the semantic information of user preferences and product characteristics.

[0062] In some embodiments, the first deep learning model includes a BERT model. The full name of the BERT model is Bidirectional Encoder Representations from Transformers, that is, bidirectional encoder representations based on Transformer. Exemplarily, in this application, BERT is used to perform word embedding encoding on the user text and product text respectively. The BERT model can capture the context relationship between words in the text through a bidirectional Transformer architecture, generate high-quality word embedding representations, and thus transform the user text and product text into a more accurate and semantic matrix representation. In practical applications, the user text and product text can be obtained from the database, and BERT is used to encode the text information respectively, providing an efficient encoding method for the subsequent feature extraction of the recommendation model.

[0063] Figure 2 is an exemplary flowchart of a data product recommendation method based on deep learning in another embodiment of this application. Refer to Figure 2 As shown, the first deep learning model (such as BERT) of this application can be constructed as Figure 2 the word embedding layer shown. The user and the product are encoded using BERT respectively, and the user text and product text input into the model are in a parallel relationship.

[0064] Exemplarily, the BERT model generally consists of three parts: an input layer, an encoding layer, and an output layer. In the input layer, BERT uses three types of features for representation, namely Token embeddings, Segment embeddings, and Position embeddings. The features finally input to the encoding layer are obtained by adding these three types of embeddings. Among them, Token embeddings are used to vectorize words. For Chinese, Token embeddings can be performed on individual Chinese characters. Segment embeddings are used to embed sentence-level features to distinguish different sentences. Position embeddings are used to embed the position information of words in the text to capture the sequential relationship between words.

[0065] In the encoding layer, BERT uses Transformer to further extract features from the input word embeddings. The encoder of Transformer contains a self-attention mechanism, which can calculate the dependency relationships between different words based on the context information of the words in the text. In addition, Transformer can calculate the features of all words in the input text in parallel, improving the computational efficiency.

[0066] In the output layer, BERT is jointly trained through two major tasks in the pre-training stage, and the specific output varies according to different tasks. These two major tasks are the sentence coherence task and the masked language model task. In the sentence coherence task, the input of BERT is two sentences, and the model needs to determine whether one sentence is the next sentence of the other. Through this task, the model can learn the context relationship between sentences, thus better understanding the semantic structure of the text.

[0067] In the masked language model task, BERT randomly masks a part of the words in the text and tries to predict the masked words based on other words in the context. To enhance the robustness of the model, BERT introduces a technique in the masking process: for the masked words, a small proportion of the words will be restored to the original words, another part of the words will be replaced with other random words, while most of the words remain masked. This design makes it impossible for the model to determine which words need to be predicted, so it relies more on context information for inference, further improving the model's understanding ability of text context and semantic structure.

[0068] In some embodiments, the user text includes a user ID and a user name; the product text includes one or any combination of a product ID, a product name, a product description, and a user rating of the product. Exemplarily, the user ID and the user name reflect the identity characteristics of the user, while the product ID, the product name, the product description, and the user rating demonstrate the attributes of the product and the user's feedback on it from different perspectives. By converting this information into a high-dimensional matrix representation, the personalized preferences of the user and the characteristics of the product can be modeled more accurately, further enhancing the richness and accuracy of semantic expression.

[0069] In step S120, the second deep learning model is used to perform feature extraction and feature compression on the user text matrix representation and the product text matrix representation respectively, to obtain a user text vector representation and a product text vector representation. Exemplarily, through feature extraction and feature compression, the core features of the user and the product can be further refined, converting the high-dimensional matrix representation into a low-dimensional vector representation, which can extract the most representative information in the user preferences and product characteristics while reducing the computational complexity, and removing redundant and noisy data.

[0070] Exemplarily, the second deep learning model of the present application can be a convolutional neural network CNN (Convolutional Neural Network). Referring to Figure 2 as shown, the CNN can be composed of three parts: a convolutional layer, a pooling layer, and a fully connected layer. A convolutional layer contains m neurons for feature extraction, and each neuron corresponds to a convolutional kernel, which is used to generate a new feature map for the row vectors in the input user matrix and product matrix. For all neurons in the convolutional layer, assuming K j is the current convolutional kernel and t is the size of the convolutional kernel, for a row vector V in the user matrix or product matrix, the feature map vector after passing through the convolutional layer can be expressed as:

[0071] F j = f(V * K j + b j ) (1)

[0072] where f represents the activation function of the convolutional layer, which is used to add non-linearity to the network; b j is the bias term of the convolutional layer. F j represents the feature map vector after passing through the convolutional layer, and F j corresponds to the feature map vector obtained by applying the convolutional kernel K j to perform a sliding convolution operation on the "user text matrix representation" or "product text matrix representation". The "user text matrix" and "product text matrix" are composed of several feature map vectors, which can be expressed as [F1, F2,..., Fm], j = 1, 2,..., m.

[0073] This application uses ReLU as the activation function in the convolutional layer, and its formula is:

[0074]

[0075] The pooling layer is used to further compress the feature map extracted by the convolution layer. The pooling operation of the pooling layer uses max-pooling, which extracts the maximum value of the corresponding sliding window in the feature map as matrix information. The maximum pooling operation can be used to process texts of different lengths. After the pooling layer, the vector matrix can be processed into a vector of fixed size:

[0076] O j =max{F1,F2,…,F (n-t+1)} (3)

[0077] Among them, n is the number of words contained in the input user text or product text, and t is the convolution kernel size of the convolution layer.

[0078] Since multiple convolution kernels are needed to extract features from user text or product text, multiple features can be obtained through the convolution layer. The output vector of the convolution layer is:

[0079] O={O1,O2,…,O k} (4)

[0080] Among them, k represents the number of convolution kernels in the convolution layer.

[0081] After obtaining the output vector O and passing it through the fully connected layer, we can get the one-dimensional vector feature representation of the user or product:

[0082] R=f(W×O+g) (5)

[0083] Among them, R represents the one-dimensional feature vector representation of the user or product, W is the weight matrix of the fully connected layer, and g represents the bias term. After the data is preprocessed, the user text matrix representation and the product text matrix representation are processed by CNN to obtain the new feature representations R of the user and product respectively. u and R i .

[0084] In some embodiments, the second deep learning model includes a first convolutional neural network and a second convolutional neural network, the first convolutional neural network includes a first convolutional layer, a first pooling layer, and a first fully connected layer, and the second convolutional neural network includes a second convolutional layer, a second pooling layer, and a second fully connected layer. Exemplarily, the first convolutional neural network may correspond to Figure 2 The left branch in the second convolutional neural network can correspond to Figure 2 The right branch in the figure. The first convolutional neural network and the second convolutional neural network are in parallel.

[0085] Combine Figure 2 As shown, in step S120, the second deep learning model is used to respectively perform feature extraction and feature compression on the user text matrix representation and the commodity text matrix representation to obtain the user text vector representation and the commodity text vector representation, including:

[0086] Step S1201: Use the first convolutional layer to perform feature extraction on the user text matrix representation to obtain the user text matrix; use the second convolutional layer to perform feature extraction on the commodity text matrix representation to obtain the commodity text matrix;

[0087] Step S1202: Use the first pooling layer to perform feature compression on the user text matrix to obtain the user text compression matrix; use the second pooling layer to perform feature compression on the commodity text matrix to obtain the commodity text compression matrix;

[0088] Step S1203: Use the first fully connected layer to process the user text compression matrix to obtain the user text vector representation; use the second fully connected layer to process the commodity text compression matrix to obtain the commodity text vector representation.

[0089] Exemplarily, by adopting two parallel convolutional neural networks to respectively perform feature extraction and feature compression on the user text matrix representation and the commodity text matrix representation, the present application can efficiently extract the core features of users and commodities, enhance the model's ability to capture local features and global semantics of text data, and the finally generated user and commodity vector representations have stronger semantic relevance and representativeness, providing higher-quality feature inputs for subsequent fusion and recommendation calculations.

[0090] The present application provides a flexible and reliable recommendation system solution for scenarios such as data trading that use text to represent user information and commodity information by combining BERT and CNN. In specific implementation, the convolutional layer of the convolutional neural network (CNN) is responsible for performing feature extraction on the user text matrix and the commodity text matrix, and the pooling layer further performs feature compression on the user representation and the commodity representation. After convolution and pooling processing, the corresponding user text compression matrix and commodity text compression matrix respectively pass through the fully connected layer to generate the final user text vector representation and commodity text vector representation.

[0091] During the model training phase, user information, product information, and rating data can be retrieved in batches from the training set. Among them, the user information and product information are first subjected to feature encoding to generate corresponding word embedding representations. Then these word embedding representations are respectively input into two parallel convolutional neural network layers for feature extraction. The basic structure of the convolutional neural network layer in this application, for example, includes: 1 convolutional layer, 1 pooling layer, and 1 fully connected layer. The respectively extracted user feature vector and product feature vector pass through the fusion interaction layer and the rating prediction layer to generate a predicted rating. During this process, the model continuously learns and adjusts parameters to optimize the recommendation effect.

[0092] After training is completed, the trained model parameters can be saved to the data storage subsystem. By loading the saved recommendation model, the preferences of users in the computing platform are calculated, and the preference results are sorted and persisted to the corresponding database table. When making recommendations, the human-computer interaction subsystem can obtain the identity information of the user and interact with the data storage subsystem to obtain their preferred content, and finally display the recommendation results in the front-end visualization interface. Adopting the technical solution of this application can efficiently process text data and provide accurate and personalized recommendation services for users.

[0093] The core code for implementing the convolutional neural network layer in this application is shown in Algorithm 1 below.

[0094] Algorithm 1: DataExchangeRec: CNN layer of the text data recommendation model.

[0095]

[0096] In step S130, the user text vector representation and the product text vector representation are fused to obtain a fused vector. Exemplarily, this step can correspond to Figure 2 the fusion layer therein. By fusing the user text vector representation and the product text vector representation in this application, the user preferences and product characteristics can be deeply associated to form a comprehensive semantic representation. Through the method of vector fusion, the semantic association between the user and the product is modeled into a unified feature space, so as to more comprehensively reflect the matching relationship between the user and the product.

[0097] In some embodiments, fusing the user text vector representation and the product text vector representation to obtain a fused vector includes: concatenating the user text vector representation and the product text vector representation to obtain a fused vector; or adding the user text vector representation and the product text vector representation to obtain a fused vector; or taking the inner product of the user text vector representation and the product text vector representation to obtain a fused vector.

[0098] Exemplarily, the feature information of users and commodities is directly merged through vector concatenation operations, preserving their respective complete semantics and enhancing the expressive power of the features; through vector addition operations, the features of both are added to generate a comprehensive representation, emphasizing the overall matching relationship between user preferences and commodity characteristics; through vector inner product operations, the similarity between user and commodity features can be directly calculated, highlighting the correlation strength between the two.

[0099] In practical applications, considering that there are various different feature vector fusion methods for user and commodity interactions, the recommendation model can be fine-tuned according to specific scenarios. Different user-commodity fusion schemes can be encapsulated as modules. When fusing user vectors and commodity vectors, by specifying different fusion parameters, the corresponding user vector and commodity vector fusion methods can be called to obtain different user-commodity fusion vectors.

[0100] Vector concatenation and vector addition do not have fixed requirements for the dimensions of user vectors and commodity vectors. The PyTorch toolkit provides a large number of array and matrix operation functions. When concatenating user vectors and commodity vectors, torch.cat(inputs, dimension) can be called to perform the concatenation of input vectors, where inputs are the input vectors, which can be more than one, and dimension indicates the dimension along which the vectors are concatenated. To implement vector addition, the two vectors can be directly added and then output to a new vector.

[0101] The core code for fusing user text vector representations and commodity text vector representations is shown in Algorithm 2 below.

[0102] Algorithm 2: FusionLayer.

[0103]

[0104]

[0105] As shown in the above code, after obtaining the user feature representation u_out and commodity feature representation i_out processed by word embedding, different fusion vectors are output according to different fusion strategies used in the fusion layer. By adopting different fusion strategies, the influence of different user and commodity interaction methods on the final recommendation effect of the recommendation model can be compared. When the fusion strategy is cat, the fusion vector is the concatenation of the user feature vector and the commodity feature vector. When the fusion strategy is add, the fusion vector is the vector addition of the user feature vector and the commodity feature vector, and the dimensions of the user feature vector and the commodity feature vector can be ensured to be the same through the setting of word embedding in advance. When the fusion strategy is dot, the fusion vector is the result of the vector inner product operation of the user feature vector and the commodity feature vector.

[0106] In step S140, the predicted score of the data product is calculated according to the fusion vector. Exemplarily, such a setting can make full use of the deep semantic association between the user and the product to achieve accurate personalized recommendation. Based on the fusion vector of the user and the product, this application designs different score prediction strategies. In the actual application process, different score prediction schemes can be encapsulated into modules. When predicting the score of the fusion vector, by specifying different score prediction parameters, the corresponding score prediction scheme can be called.

[0107] In some embodiments, calculating the predicted score of the data product according to the fusion vector includes:

[0108] Step S1401: Perform a linear transformation on the fusion vector to obtain the first score;

[0109] Step S1402: Calculate the first interaction score and the second interaction score according to the fusion vector; the first interaction score and the second interaction score are used to represent the interaction degree between the user and the product;

[0110] Step S1403: Calculate the predicted score according to the first score, the first interaction score, and the second interaction score.

[0111] Exemplarily, the first score can reflect the basic matching degree between the user and the product; the first interaction score and the second interaction score respectively reflect the interaction degree between the user and the product from different perspectives, capturing more detailed semantic associations and potential preferences. By comprehensively considering the first score, the first interaction score, and the second interaction score, the predicted score can more accurately reflect the user's potential interest in the product, and enhance the recommendation effect through multi-level interaction modeling.

[0112] To explore the influence of different score prediction methods on the performance of the recommendation model, this application designs multiple ways of predicting scores according to the Factorization Machines (FM), Neural Factorization Machines (NFM), and Multilayer Perceptron (MLP).

[0113] In some embodiments, step S1402, calculating the first interaction score and the second interaction score according to the fusion vector, includes:

[0114] The first interaction score is calculated using the following formula (6):

[0115]

[0116] The second interaction score is calculated using the following formula (7):

[0117]

[0118] Among them, represents the first interaction score; represents the second interaction score; ui vector represents the fusion vector; fm V represents the factor matrix, fm V is obtained by initializing with a uniform distribution or a normal distribution.

[0119] Exemplarily, the first interaction score and the second interaction score can reflect the interaction intensity between the user and the commodity from different perspectives, providing richer information input for subsequent predicted scores, and improving the accuracy of the recommendation result through multi-level interaction calculation.

[0120] In some embodiments, in step S1403, calculating the predicted score according to the first score, the first interaction score, and the second interaction score includes: calculating the predicted score using the following formula (8):

[0121]

[0122] Among them, predict rating represents the predicted score; p1 represents the first fusion coefficient, and p1 is a constant; fm linear_part represents the first score. Exemplarily, p1 can be set to 0.5, and the present application does not limit the value of p1.

[0123] Exemplarily, by introducing the first fusion coefficient, the present application can effectively balance the weights between the interaction score and the first score, so that the predicted score can capture both the global matching relationship between the user and the commodity and reflect its fine-grained interaction characteristics.

[0124] The following uses an embodiment to introduce the complete predicted score method designed according to the factorization machine FM of the present application. The present application performs score prediction by combining the linear part and the interaction part. The specific processing steps of the algorithm are as follows:

[0125] (1) Initialize the parameters of the factorization machine model, including: the model feature dimension dim, the weight fc.weight and bias fc.bias of the linear layer, the user bias term b users and the commodity bias term b items and the factor matrix fm V .

[0126] (2) Transform the fusion feature vector ui of the user and the commodity through the linear layer vector to obtain the linear part: fm linear_part = fc(ui vector ).

[0127] (3) Multiply the fused feature vector and the factor matrix to obtain the interaction part to represent the interaction between user and item features:

[0128] (4) Combine the linear part and the interaction part to obtain the predicted score:

[0129] The above processing steps can be represented by Algorithm 3 as follows.

[0130] Algorithm 3: FactorizationMachine: The factorization machine makes score predictions.

[0131]

[0132]

[0133] In some embodiments, in step S1403, calculating the predicted score according to the first score, the first interaction score, and the second interaction score includes:

[0134] Step S1403a: Perform bilinear interaction processing on the first interaction score and the second interaction score using the following formula (9) to obtain the bilinear interaction score:

[0135]

[0136] where bilinear represents the bilinear interaction score; p2 represents the second fusion coefficient, and p2 is a constant; for example, p2 can be set to 0.5, and the present application does not limit the value of p2.

[0137] Step S1403b: Perform non-linear mapping and linear transformation on the bilinear interaction score using the following formula (10) to obtain the third interaction score:

[0138] fm interactions =h(drop_out(relu(mlp(bilinear)))) (10)

[0139] where fm interactions represents the third interaction score; mlp(·) represents a multi-layer perceptron, and the multi-layer perceptron is used for non-linear mapping; relu(·) represents an activation function; drop_out(·) represents random dropout processing; h(·) represents the output layer, and the output layer is used for linear transformation.

[0140] Step S1403c: Calculate the predicted score using the following formula (11):

[0141] predict rating= fm interactions + fm linear_part (11)

[0142] Among them, predict rating represents the predicted score; fm linear_part represents the first score.

[0143] Exemplarily, the bilinear interaction score can dynamically balance the weights between the first interaction score and the second interaction score, capturing the multi-dimensional interaction characteristics between the user and the commodity. The third interaction score reflects higher-order interaction characteristics. By combining the third interaction score with the first score, the final predicted score is generated, improving the accuracy and comprehensiveness of score calculation and optimizing the modeling ability of the complex relationship between the user and the commodity, thereby providing more relevant and satisfactory data commodity recommendations for the user.

[0144] The following uses an embodiment to introduce the complete predicted score method designed by the present application according to the Neural Factorization Machine (NFM). The specific processing steps of the algorithm are as follows:

[0145] (1) Initialize the model parameters, including: the model feature dimension dim, the weights fc.weight and bias fc.bias of the linear layer (fc), the factor matrix fm V , the weights mlp.weight of the MLP layer, the weights h.weight of the output layer h, and the dropout rate drop_out = 0.5 of the Dropout layer.

[0146] (2) Transform the fused feature vector ui of the user and the commodity through the linear layer vector to obtain the linear part: fm linear_part = fc(ui vector ).

[0147] (3) Calculate the interaction part by combining FM and MLP. First, multiply the fused feature vector ui vector with the factor matrix fm V to obtain the interaction part: and the square term of the interaction part: Then, combine the foregoing interaction part through bilinear interaction: After that, perform a non-linear mapping through the MLP and add a Dropout operation to prevent overfitting; finally, perform a linear transformation through the output layer h to obtain the interaction part: fm interactions = h(drop_out(relu(mlp(bilinear)))).

[0148] (4) Combine the linear part and the interaction part to obtain the final score prediction: predict rating = fm interactions + fmlinear_part 。

[0149] The above processing steps can be represented by the following Algorithm 4.

[0150] Algorithm 4: NeuralFactorizationMachine: The neural factorization machine makes a scoring prediction.

[0151]

[0152]

[0153] In some embodiments, calculating the predicted score of the data product according to the fusion vector includes calculating the predicted score using the following formula (12):

[0154] predict rating =relu(fc(ui vector )) (12)

[0155] where predict rating represents the predicted score; ui vector represents the fusion vector; fc(·) represents the linear layer; relu(·) represents the activation function.

[0156] Exemplarily, calculating the predicted score through formula (12) can reduce the computational complexity, and improve the accuracy and stability of the predicted score by fusing the deep semantic information of the fusion vector and the non-linear characteristics of the ReLU activation function.

[0157] The following uses an embodiment to introduce the complete predicted scoring method designed by this application according to the multi-layer perceptron MLP. The specific processing steps of the algorithm are as follows:

[0158] (1) Initialize the model parameters, including the model feature dimension dim, the weight fc.weight and the bias fc.bias of the linear layer (fc).

[0159] (2) Use the MLP to perform a linear mapping and a non-linear transformation on the fusion feature vector ui vector to obtain the predicted score: predict rating =relu(fc(ui vector ))。

[0160] The above processing steps can be represented by the following Algorithm 5.

[0161] Algorithm 5: MultiLayerPerceptron: The multi-layer perceptron makes a scoring prediction.

[0162]

[0163]

[0164] Exemplarily, in practical applications, after training the model in the manner described above, the trained recommendation model can be saved in the data storage subsystem. Figure 3 It is a schematic diagram of the model training result of the data product recommendation system in an embodiment of the present application. The training result of the data product recommendation model of the present application is as Figure 3 shown.

[0165] The data product recommendation system and experimental results of the present application will be introduced later.

[0166] The data product recommendation system of the present application is equivalent to a data transaction recommendation system based on deep learning. Starting from deep learning theories and methods, the present application proposes a text data recommendation algorithm framework, implements a recommendation algorithm model applicable to the data transaction platform, and applies it to the data transaction scenario. Based on deep learning, this framework encapsulates each process of the recommendation algorithm into modules, can batch process user data and commodity information in the database, and significantly improves the reliability and efficiency of the system.

[0167] The present application solves the problem of the lack of recommendation system methods in the data transaction scenario due to the short existence time and limited business scope of the platform. Through the framework of the present application, the recommendation model architecture can be fine-tuned according to the actual scenario, so as to quickly realize the application. Without relying on more user and commodity auxiliary meta-information, based on the same data set for evaluation, the text data recommendation algorithm model of the present application is superior to the traditional matrix factorization method and probabilistic matrix factorization method in evaluation metrics such as mean squared error, root mean squared error, and mean absolute error.

[0168] To further verify the rationality of the recommendation system model, the present application designs an ablation experiment. By comparing the changes in evaluation metrics under different user feature and commodity feature fusion methods and different rating prediction methods, the scientificity and effectiveness of the recommendation system design are analyzed. In addition, the present application also encapsulates the recommendation algorithm model and the text data enhancement method into modules and applies them to the data transaction platform to verify the functions of the recommendation system. The experimental results show that the recommendation algorithm framework designed by the present application is not only applicable to the data transaction scenario, but also can be extended to other e-commerce scenarios that only use text information representation, and has significant engineering value and practical application significance.

[0169] I. Experimental Environment

[0170] In terms of the experimental environment of the recommendation model, the text data recommendation system algorithm of this application is implemented using the Python 3.6 programming language and the PyTorch 1.8 deep learning framework. In addition, machine learning libraries such as numpy and genism are installed, and the operating system is Ubuntu 18.04. In terms of experimental hardware, the CPU uses a 6-core Xeon E5-2678 v3, the memory is 11GB, the graphics card used is NVIDIA GeForce RTX 2080Ti, and the video memory is 11GB.

[0171] II. Determining the Dataset and Recommendation System Evaluation Metrics

[0172] Due to the support of various e-commerce platforms for the research of recommendation systems, recommendation systems have datasets taken from real scenarios and sufficient in quantity in many scenarios. The research on recommendation systems in scenarios such as movie recommendation, music recommendation, and information flow recommendation has received great attention in recent years. Considering the scarcity of datasets in data trading scenarios and the fact that user and product information in data trading scenarios is represented by text, in terms of dataset selection, this application uses Amazon Review Data (2018) for the training and testing of the text data recommendation algorithm model and the text data generation model. This dataset is provided by Amazon and contains data in multiple categories, including rich text metadata of various products such as fashion, books, music, electrical appliances, and software, and has been widely used in the research of recommendation systems.

[0173] At the same time, considering that due to the limitations of the data trading e-commerce platform scenario, it is difficult for the scale of users and products to reach the scale of ordinary e-commerce platforms, and training a recommendation algorithm model requires a relatively sufficient dataset. This application considers using the Generative Adversarial Network (GAN) to perform data augmentation on user and product text data.

[0174] For user text data, in addition to the user ID, user name, and user's rating of the product, it also includes information such as the user's avatar and the user's comments on other products. For product text data, in addition to the product ID and product description, it also includes other information such as the product price, picture, brand, and category. Since the model of this application is applicable to recommendation systems for text data, it is hoped to make recommendations for such scenarios with as little information as possible. In addition to the user ID, user name, product ID, product name, product description, and user's rating of the product, this application does not require other information.

[0175] This application has respectively conducted performance tests on the text data recommendation algorithm and the text data enhancement method on this data set. Subsequently, when conducting functional tests on the text data recommendation algorithm, this application selected a real data transaction scenario to apply the recommendation algorithm model of this application.

[0176] The text data recommendation algorithm model of this application aims to fit the user's rating of goods with the model. Its essence is a regression task. Therefore, this application adopts the evaluation indicators commonly used for the rating prediction problem of the recommendation system. The evaluation indicators adopted by this application are:

[0177] 1. Mean Squared Error (MSE). Its calculation formula is:

[0178]

[0179] By calculating the predicted ratings of all users in the data set for goods, the predicted ratings are summed after squaring the differences from the true ratings r u,i , and then divided by the number of data in the data set D. One piece of data can be considered as a user's rating of a good. Since the mean squared error represents the error between the predicted rating and the true rating, the smaller the mean squared error indicator, the better the performance of the model. Since the mean squared error calculates the square of the error, if the ratings calculated for some samples deviate greatly from the true ratings, the overall mean squared error calculated will be very large.

[0180] 2. Root Mean Squared Error (RMSE). Its calculation formula is:

[0181]

[0182] After calculating the predicted ratings of all users for goods, first calculate the differences from the true ratings and square the differences. After summing all the squared difference results, divide by the number of data in the data set, and finally take the square root of the obtained result, which is the final RMSE result. Similar to the mean squared error, the root mean squared error represents to a certain extent the error between the calculated rating and the true rating. Therefore, the smaller this indicator, the better the effect of the recommendation model. However, the root mean squared error is less sensitive to outliers and less punishing for errors compared to the mean squared error.

[0183] 3. Mean Absolute Error (MAE). Its calculation formula is:

[0184]

[0185] The mean absolute error calculates the average of the absolute values of the errors between the true ratings and the predicted ratings. The smaller the value of the mean absolute error, the better the performance of the model can be considered.

[0186] where u and i respectively represent the user dataset and the item dataset, and r u,i is the true rating of user u for item i recorded in the dataset, while is the predicted rating of the model for the current user and item. During the process of testing the recommendation algorithm model in this application, the above three metrics were calculated simultaneously.

[0187] III. Performance Testing of Recommendation Algorithm

[0188] To verify the recommendation performance of the text data recommendation algorithm based on deep learning, this application uses data of multiple categories provided by Amazon Review Data. Amazon provides datasets of multiple sub - categories. To verify that the recommendation algorithm model of this application is applicable to the recommendation scenario of text data, this application selects 5 datasets with different categories and different data scales: Apps for Android (containing 752937 pieces of data), Clothing, shoes and Jewerly (containing 278677 pieces of data), Digital Music (containing 64706 pieces of data), Health and Personal Care (containing 346355 pieces of data), Video Games (containing 231780 pieces of data). For each sample of the dataset of each category, this application only extracts 5 features as experimental samples: reviewerID (user number ID), reviewerName (user name), asin (item number ID), title (item name), overall (user's rating of the item). The user number ID and the user name are concatenated as the text feature of the user, and the item number ID and the item name are concatenated as the text feature of the item, thereby obtaining the user representation and item representation in text form and the corresponding ratings. Additionally, due to the different data scales of the selected different datasets, when testing the recommendation algorithm, the evaluation metrics of the 5 - category datasets are averaged as the final evaluation result of the recommendation algorithm.

[0189] This application divides the dataset in a ratio of 8:1:1, and the three resulting subsets are the training set, the validation set, and the test set. Among them, the training set is used for parameter training within the model. The validation set is used to evaluate the model's performance and, to a certain extent, for users to adjust the hyperparameters of the recommendation model. The test set is used to evaluate the performance of the recommendation model after the model training is completed according to the evaluation metrics of the recommendation system. This application tests the performance of the recommendation algorithm model using the test set in terms of three metrics: mean squared error, root mean squared error, and mean absolute error.

[0190] Meanwhile, for performance comparison, this application selects two recommendation algorithms that are also based on text data for comparison, namely the Matrix Factorization (MF) method and the Probabilistic Matrix Factorization (PMF) method. Matrix factorization is one of the common collaborative filtering algorithms, which takes the rating matrix of users and items as input and uses the alternating least squares method to minimize the objective function. The probabilistic matrix factorization method models the latent representations of users and items by introducing a Gaussian distribution.

[0191] For the MF method and the PMF method, the dimensions of the user latent vector and the item latent vector are uniformly set to 50, which are reasonable parameters summarized by this application through experiments. To verify the rationality of the design of the text data recommendation algorithm model based on deep learning in this application and its superiority compared to other recommendation algorithms, this application compares the experimental results of the recommendation evaluation metrics of the three recommendation algorithms when using different rating prediction methods.

[0192] To explore the impact of different rating prediction methods on the performance of different recommendation algorithms, under the condition that the data encoding, the structure of the recommendation algorithm model, and the fusion method of the item feature vector and the user feature vector are the same, this application compares the three evaluation metrics of MSE, RMSE, and MAE for different rating prediction methods. The presented results are the averages of the results of different models on 5 categories of datasets.

[0193] When the rating prediction layer uses the prediction method constructed according to the Factorization Machine (FM), the performance of different methods is shown in Table 1 below. It can be seen that the recommendation algorithm (Our) in this application outperforms the Matrix Factorization (MF) algorithm and the Probabilistic Matrix Factorization (PMF) algorithm in all three recommendation metrics. This is because the MF algorithm and the PMF algorithm are affected by data sparsity, while the recommendation algorithm in this application uses deep learning technology to reasonably encode and extract features from the data, better modeling the latent representations of users and items. Therefore, the predicted ratings have lower errors than the MF algorithm and the PMF algorithm.

[0194] When using mean square error as an evaluation indicator, the value obtained is greater than the value obtained using root mean square error and mean absolute error. This is because mean square error is more sensitive to errors. The difference between the true score and the predicted score of some data will cause the overall MSE to be higher. The mean absolute error only calculates the difference between the average true score and the predicted score, so it is relatively insensitive to incorrect predictions and will have a lower error value.

[0195] Table 1 Performance comparison of various recommendation methods based on FM as the rating prediction layer

[0196]

[0197] When the rating prediction layer is built based on a neural factorization machine, the performance of different methods is shown in Table 2 below.

[0198] Table 2 Performance comparison of each recommendation method based on NFM as the rating prediction layer

[0199]

[0200] When the rating prediction layer is built based on a multi-layer perceptron, the performance of different methods is shown in Table 3 below.

[0201] Table 3 Performance comparison of each recommendation method based on MLP as the rating prediction layer

[0202]

[0203] It can be seen that even if a variety of different rating prediction methods are used, the text data recommendation system built by this application based on deep learning, which considers the integration of user text feature representation and product text feature representation, is superior to the traditional matrix decomposition method and the probability matrix decomposition method in terms of mean square error, root mean square error, and mean absolute error. This shows that the use of deep learning methods to extract features from text and the use of a fusion method to consider the interaction between users and products help improve the accuracy of the recommendation system.

[0204] At the same time, it can be seen that for the same data encoding method and recommendation model, the factor decomposition machine or multi-layer perceptron used in the score prediction layer of this application is better than the neural factor decomposition machine. When the factor decomposition machine is used, the best recommendation effect is achieved in all three recommendation algorithms. Therefore, the factor decomposition machine can be used as the score prediction layer in the recommendation algorithm of the data trading platform.

[0205] To explore the impact of different fusion methods of user feature vectors and commodity feature vectors in the fusion layer of the text data recommendation framework of this application on the recommendation performance, under the same data encoding method and recommendation algorithm model structure, the scoring prediction method in the scoring prediction layer was tried to be changed, and the dataset with the sub-category of Digital Music was selected as the experimental dataset to observe the results of the recommendation model of this application in three evaluation metrics.

[0206] When the fusion method of user and commodity features is to concatenate the user feature vector and the commodity feature vector, the evaluation metric results corresponding to different scoring prediction methods are shown in Table 4.

[0207] Table 4 Performance comparison of each scoring prediction method using concatenation as the fusion layer

[0208]

[0209] When the fusion method of user and commodity features is to perform vector addition on the user feature vector and the commodity feature vector, the evaluation metric results corresponding to different scoring prediction methods are shown in Table 5.

[0210] Table 5 Performance comparison of each scoring prediction method using vector addition as the fusion layer

[0211]

[0212] When the fusion method of user and commodity features is to perform vector inner product on the user feature vector and the commodity feature vector, the evaluation metrics corresponding to different scoring prediction methods are shown in Table 6.

[0213] Table 6 Performance comparison of each scoring prediction method using vector inner product as the fusion layer

[0214]

[0215] The experimental comparison results show that using vector addition as the fusion method is relatively close to using concatenation in some metrics, and both methods are better than using vector inner product as the fusion method. This may be because using vector inner product as the fusion method seriously changes the dimensions of user and commodity feature vectors, thus affecting the fitting effect of the predicted score.

[0216] Overall, when using the vector concatenation method as the fusion method of user feature vector and commodity feature vector, the recommendation model can achieve good results in the three evaluation metrics. Therefore, when comparing with other recommendation models, the recommendation algorithm model of this application also uses concatenation in the fusion layer to fuse user feature vector and commodity feature vector. In the recommendation algorithm finally applied to the data trading platform, concatenation can be used as the fusion method of user feature vector and commodity feature vector in the fusion layer of the recommendation framework.

[0217] IV. Function Verification of Recommendation Algorithm

[0218] The recommendation algorithm model of this application is trained based on a publicly available dataset. The dataset on which the text data recommendation algorithm model of this application depends contains only user text information, product text information, and the rating records of users for products. Therefore, the recommendation model of this application is applicable to scenarios such as data trading, that is, scenarios where the information of users and products is represented in text, and there is a lack of other forms of auxiliary information such as pictures, videos, and review texts. This application deploys the data product recommendation method model of this application on a data trading platform, obtains all user and product text information from the mysql database, and makes personalized recommendations for users based on the user preferences calculated by the recommendation algorithm.

[0219] After applying the recommendation algorithm model to the data trading platform, personalized recommendations can be made according to the ratings between users and products calculated by the model and the sorting result of the predicted ratings.

[0220] Figure 4 It is a schematic diagram of the storage order of data products in the database in an embodiment of this application. Figure 5 It is a schematic diagram of the data product recommendation system before using the data product recommendation method in an embodiment of this application. Figure 5 The recommended results shown in Figure 4 are basically displayed according to the storage order of data products in the database in

[0221] Figure 6 It is a schematic diagram of the data product recommendation system after using the data product recommendation method in an embodiment of this application. Referring to Figure 6 shown, after applying the data product recommendation method of this application, the predicted ratings are calculated based on different users and products in the database, and the recommended results are obtained after sorting.

[0222] Figure 7 It is a schematic diagram of the recommended results of users in the same company but different departments in an embodiment of this application. Figure 8 It is a schematic diagram of the recommended results of users in different companies and different departments in an embodiment of this application. Figure 7 and Figure 8 respectively show the differences in the recommended results obtained for users who are in the same company but different departments as the users in Figure 6 , and users in different companies and different departments. It can be seen that there is a certain similarity in the recommended results among users in the same company but different departments because there is a certain similarity in the encoding of user text information among such users. The significant difference between the recommended results obtained for them and those of users in different companies and different departments also proves this point.

[0223] By encapsulating the data product recommendation method of the present application as a module into the data product recommendation system, personalized recommendation services can be quickly provided to the users of the data trading platform.

[0224] The present application further includes a data product recommendation system based on deep learning, including: a data product display module, a memory, and a processor. Among them, the data product display module is used to sort and display data products according to the predicted scores; the memory is used to store instructions executable by the processor; the processor is used to execute the instructions to implement the data product recommendation method based on deep learning described above.

[0225] Figure 9 It is a system block diagram of the data product recommendation system based on deep learning according to an embodiment of the present application. Refer to Figure 9 As shown, the data product recommendation system 900 based on deep learning may include an internal communication bus 901, a processor 902, a read-only memory (ROM) 903, a random access memory (RAM) 904, and a communication port 905. The data product recommendation system 900 based on deep learning may further include a hard disk 906. The internal communication bus 901 can realize data communication between the components of the data product recommendation system 900 based on deep learning. The processor 902 can make judgments and issue prompts. In some embodiments, the processor 902 may be composed of one or more processors. The communication port 905 can realize the data communication between the data product recommendation system 900 based on deep learning and the outside. In some embodiments, the data product recommendation system 900 based on deep learning can send and receive information and data from the network through the communication port 905. The data product recommendation system 900 based on deep learning may further include different forms of program storage units and data storage units, such as a hard disk 906, a read-only memory (ROM) 903, and a random access memory (RAM) 904, which can store various data files used for computer processing and / or communication, as well as possible program instructions executed by the processor 902. The processor executes these instructions to implement the main part of the method. The results processed by the processor are transmitted to the user device through the communication port and displayed on the user interface.

[0226] The above-mentioned data product recommendation method based on deep learning can be implemented as a computer program, stored in the hard disk 906, and loaded into the processor 902 for execution to implement the data product recommendation method based on deep learning of the present application.

[0227] The present application further includes a computer-readable medium storing computer program code, and the computer program code implements the data product recommendation method based on deep learning described above when executed by the processor.

[0228] When the deep learning-based data product recommendation method is implemented as a computer program, it can also be stored in a computer-readable storage medium as an article of manufacture. For example, the computer-readable storage medium may include, but is not limited to, magnetic storage devices (such as hard disks, floppy disks, magnetic strips), optical disks (such as compact discs (CDs), digital versatile discs (DVDs)), smart cards, and flash memory devices (such as electrically erasable programmable read-only memories (EPROMs), cards, sticks, key drives). In addition, the various storage media described in this application can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" may include, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain, and / or carry code and / or instructions and / or data).

[0229] It should be understood that the embodiments described above are merely illustrative. The embodiments described in this application can be implemented in hardware, software, firmware, middleware, microcode, or any combination thereof. For hardware implementation, the processor can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, and / or other electronic units designed to perform the functions described in this application, or a combination thereof.

[0230] Some aspects of this application can be executed entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software can all be referred to as "data blocks", "modules", "engines", "units", "components", or "systems". The processor can be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processor devices (DAPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, or a combination thereof. In addition, aspects of this application may be embodied as a computer product located in one or more computer-readable media, the product including computer-readable program code. For example, the computer-readable medium may include, but is not limited to, magnetic storage devices (such as hard disks, floppy disks, magnetic tapes...), optical disks (such as compact discs CD, digital versatile discs DVD...), smart cards, and flash memory devices (such as cards, sticks, key drives...).

[0231] A computer-readable medium may include a propagated data signal having computer program code embodied therein, for example, on a baseband or as part of a carrier wave. The propagated signal may take a variety of forms, including electromagnetic, optical, and the like, or any suitable combination thereof. A computer-readable medium may be any computer-readable medium other than a computer-readable storage medium that can be connected to an instruction execution system, apparatus, or device to effect communication, propagation, or transmission for use by a program. The program code located on the computer-readable medium may be propagated via any appropriate medium, including radio, cable, fiber optic cable, RF signals, or similar media, or any combination of the foregoing media.

[0232] The basic concepts have been described above. Obviously, for those skilled in the art, the above application disclosure is only an example and does not constitute a limitation to this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are proposed in this application, so such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0233] Meanwhile, this application uses specific terms to describe the embodiments of this application. Such as "an embodiment", "one embodiment", and / or "some embodiments" mean a certain feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that the "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more at different positions in this specification is not necessarily referring to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application may be combined appropriately.

[0234] In some embodiments, numbers are used to describe components and attribute quantities. It should be understood that such numbers used for the description of embodiments are, in some examples, modified by the modifiers "about", "approximately", or "substantially". Unless otherwise stated, "about", "approximately", or "substantially" indicate that the said numbers allow a variation of ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may vary according to the characteristics required by individual embodiments. In some embodiments, the numerical parameters should consider the specified significant digits and adopt the method of retaining the general number of digits. Although the numerical ranges and parameters used in some embodiments of this application to confirm the breadth of their scope are approximate values, in specific embodiments, such numerical settings are as precise as possible within the feasible range.

Claims

1. A data product recommendation method based on deep learning, characterized in that, Including: Performing word embedding encoding on the user text and the product text of the data product by using a first deep learning model to obtain a user text matrix representation and a product text matrix representation; Performing feature extraction and feature compression on the user text matrix representation and the product text matrix representation respectively by using a second deep learning model to obtain a user text vector representation and a product text vector representation; Fusing the user text vector representation and the product text vector representation to obtain a fused vector; Calculating a predicted score of the data product according to the fused vector.

2. The data product recommendation method according to claim 1, wherein Calculating a predicted score of the data product according to the fused vector includes: Performing a linear transformation on the fused vector to obtain a first score; Calculating a first interaction score and a second interaction score according to the fused vector; the first interaction score and the second interaction score are used to represent the interaction degree between the user and the product; Calculating the predicted score according to the first score, the first interaction score, and the second interaction score.

3. The data product recommendation method according to claim 2, wherein Calculating a first interaction score and a second interaction score according to the fused vector includes: Calculating the first interaction score by using the following formula: Calculating the second interaction score by using the following formula: Among them, represents the first interaction score; represents the second interaction score; ui vector represents the fusion vector; fm V represents the factor matrix, fm V is obtained by initializing with a uniform distribution or a normal distribution.

4. The data product recommendation method according to claim 3, wherein Calculating the predicted score according to the first score, the first interaction score, and the second interaction score includes: calculating the predicted score by using the following formula: Among them, predict rating represents the predicted score; p1 represents the first fusion coefficient, and p1 is a constant; fm linear_part represents the first score.

5. The data product recommendation method according to claim 3, wherein Calculating the predicted score according to the first score, the first interaction score, and the second interaction score includes: Performing a bilinear interaction process on the first interaction score and the second interaction score by using the following formula to obtain a bilinear interaction score: where bilinear represents the bilinear interaction score; p2 represents a second fusion coefficient, and p2 is a constant; Performing a non-linear mapping and a linear transformation on the bilinear interaction score by using the following formula to obtain a third interaction score: fm interactions = h(drop_out(relu(mlp(bilinear)))) wherein, fm interactions represents the third interaction score; mlp(·) represents a multi-layer perceptron, which is used for the non-linear mapping; relu(·) represents an activation function; drop_out(·) represents a random dropout process; h(·) represents an output layer, which is used for the linear transformation; Calculating the predicted score by using the following formula: predict rating = fm interactions + fm linear_part Among them, predict rating represents the predicted score; fm linear_part represents the first score.

6. The data product recommendation method according to claim 1, wherein Calculating a predicted score of the data product according to the fused vector includes calculating the predicted score by using the following formula: predict rating = relu(fc(ui vector )) Among them, predict rating represents the predicted score; ui vector represents the fusion vector; fc(·) represents a linear layer; relu(·) represents an activation function.

7. The data product recommendation method according to any one of claims 1-6, characterized in that The first deep learning model includes a BERT model.

8. The data product recommendation method according to any one of claims 1-6, characterized in that, The second deep learning model includes a first convolutional neural network and a second convolutional neural network. The first convolutional neural network includes a first convolutional layer, a first pooling layer, and a first fully-connected layer. The second convolutional neural network includes a second convolutional layer, a second pooling layer, and a second fully-connected layer; Performing feature extraction and feature compression on the user text matrix representation and the product text matrix representation respectively by using a second deep learning model to obtain a user text vector representation and a product text vector representation includes: Performing feature extraction on the user text matrix representation by using the first convolutional layer to obtain a user text matrix; performing feature extraction on the product text matrix representation by using the second convolutional layer to obtain a product text matrix; Performing feature compression on the user text matrix by using the first pooling layer to obtain a user text compressed matrix; performing feature compression on the product text matrix by using the second pooling layer to obtain a product text compressed matrix; Process the user text compression matrix using the first fully connected layer to obtain the user text vector representation; process the product text compression matrix using the second fully connected layer to obtain the product text vector representation.

9. The data product recommendation method according to any one of claims 1-6, characterized in that Fuse the user text vector representation and the product text vector representation to obtain a fused vector, including: Perform vector concatenation on the user text vector representation and the product text vector representation to obtain the fused vector; or Perform vector addition on the user text vector representation and the product text vector representation to obtain the fused vector; or Perform vector inner product on the user text vector representation and the product text vector representation to obtain the fused vector.

10. The data product recommendation method according to any one of claims 1-6, characterized in that The user text includes a user ID and a user name; the product text includes one or any combination of a product ID, a product name, a product description, and a user rating for the product.

11. A data product recommendation system based on deep learning, characterized in that, Including: A data product display module for sorting and displaying data products according to the predicted rating; A memory for storing instructions executable by a processor; A processor for executing the instructions to implement the deep learning-based data product recommendation method according to any one of claims 1-10.

12. A computer-readable medium storing computer program code, characterized in that, The computer program code, when executed by a processor, implements the deep learning-based data product recommendation method according to any one of claims 1-10.

Images