Task-aware data product recommendation method

Abstract

The invention discloses a task-aware data product recommendation method, which comprises the following steps of: 1, constructing a user and data product interaction matrix, and extracting task information at the same time; 2, constructing a heterogeneous graph and updating and embedding graph nodes of the heterogeneous graph; 3, constructing an asymmetric attention network module; 4, constructing a task integration module; and 5, selecting a proper loss function to train and optimize the model for task recommendation. According to the method, the relational graph convolutional network and the attention mechanism are combined, the task information of the user and the data product can be considered, and the user, the data product and the task information can be accurately represented, so that a more accurate data product recommendation effect can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of user preference mining and data product recommendation, and in particular to a task-aware data product recommendation method. Background Art

[0002] The widespread use of the Internet has greatly increased the amount of data generated, and people generate a large amount of data through various online activities in their daily lives. At the same time, the popularity of smartphones and tablets, as well as the emergence of high-speed mobile networks, have made the generation and consumption of mobile data more common, further enriching the content and types of data products. High-quality data products play a key role in the development of machine learning and deep learning technologies, and have irreplaceable value in promoting scientific discoveries, verifying hypotheses, promoting technological progress and interdisciplinary collaboration.

[0003] As the basis for model training and evaluation, data products directly affect the performance and application effect of the model. Existing data product recommendation methods mainly include query-based data product recommendation systems, content-driven data product recommendation systems, and user preference-based data product recommendation systems. However, most of these recommendation methods do not take into account the task information of data products, and most methods do not take into account the potential feature interactions between users and data products, which leads to insufficient modeling of user and data product features and poor recommendation accuracy. Summary of the invention

[0004] The present invention aims to address the deficiencies of the above-mentioned prior art and to provide a task-aware data product recommendation method, so as to not only take into account the task information of users and data products, but also accurately characterize the users, data products and task information, thereby achieving a more accurate data product recommendation effect.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0006] The task-aware data product recommendation method of the present invention is characterized in that it is performed according to the following steps:

[0007] Step 1: Build an interaction matrix between users and data products , and according to the user Description information and data products Description information Extract users separately Task information and data products Task information ;

[0008] Step 1.1: Based on the interaction information between users and data products, obtain the interaction matrix between users and data products , among which, if , indicating that the user Interacted with data products ,like , it means that the user No data product interaction , is the number of users, is the number of data products, and the total number of user and data product pairs is ;

[0009] Get data products The description information is recorded as , the user Descriptive information of the data products that users have interacted with The description information is recorded as , in, For users No. Description information, The number of data products representing user interactions;

[0010] Step 1.2: Use the large language model to learn from users Description information and data products Description information Extracting users from Task information , data products Task information ,user and data products Common task information collection ,in, Indicates user No. tasks, Represents data products No. tasks, Indicates user and data products Common task information collection The tasks, , and Representing users The number of tasks, data products The number of tasks and users and data products Common task information collection The number of tasks in

[0011] The user and data products Common task information collection Middle Tasks The type source is recorded as ;

[0012] Step 2: Build a heterogeneous graph node update embedding module and update the node based on the interaction matrix and task information to build a heterogeneous graph , so that the heterogeneous graph Update the nodes in the embedding to get the representation of the user after the update embedding , representation of data products and task representation ; and update the embedded user representation , representation of data products Convert them respectively and get the converted user representation accordingly and the representation of the transformed data products ;

[0013] Step 3: Build an asymmetric attention network module and Description information and data products Description information Processing to get user Enhanced representation of and data products Enhanced representation of ;

[0014] Step 3.1: For users No. Description information and data products Description information Processing to get user No. Description information Sentence-level representation of and data products Description information Sentence-level representation of ;

[0015] Step 3.2: Use the data product attention mechanism module to Perform sentence-level aggregation to obtain a comprehensive representation of the data product d ;

[0016] Step 3.3: Use the user attention mechanism module to Aggregate the context sentence level representations of all descriptive information of the user Comprehensive representation of ;

[0017] Step 3.4: Use a multi-layer perceptron to analyze the user Comprehensive representation of and data products Comprehensive characterization of Transform and get the user Comprehensive representation after conversion and data products Comprehensive representation after conversion ;

[0018] Step 3.5: Use equation (10) and equation (11) to get the user Enhanced representation of and data products Enhanced representation of :

[0019] (10)

[0020] (11)

[0021] Step 4: Build a task integration module to represent heterogeneous graph-based tasks With task-based representation Processing to get user and data products Task integration representation ;

[0022] Step 5: Build a prediction module and use formula (18) to obtain the user For data products The predicted score :

[0023] (18)

[0024] In formula (18), is the sigmoid function;

[0025] Step 6: Train the data product recommendation network consisting of the heterogeneous graph node update embedding module, the asymmetric attention network module, the task integration module and the prediction module to obtain a trained data product recommendation model, which is used to recommend the input data products and obtain the user's best prediction score for the data product.

[0026] The task-aware data product recommendation method of the present invention is also characterized in that step 2 is performed as follows:

[0027] Step 2.1: Treat all users, all data products, and all tasks as user nodes , data product node , task node , thus obtaining the node set , taking the interaction or similarity relationship between nodes as edges, thus obtaining the edge set , and construct a heterogeneous graph ;

[0028] Step 2.2: Use the relational graph convolutional network to classify heterogeneous graphs After learning the node representation, we get the user Node representation , data products Node representation and tasks Node representation ;

[0029] Through multi-layer perceptron and Transform and obtain the converted user accordingly Representation and the transformed data products Representation .

[0030] Further, step 3.1 is performed as follows:

[0031] Step 3.1.1: User No. Description information and data products Description information Each word in is input into the BERT model for word embedding, so that the user Description information The Sentences and data products Description information No. Sentences :

[0032] (1)

[0033] In formula (1), for The word embedding of the last word in , for Middle word embedding; U represents The number of words in

[0034] (2)

[0035] In formula (2), for The word embedding of the last word in , for Middle word embedding; D represents The number of words in

[0036] Step 3.1.2: Follow the process in step 3.1.1 to obtain the user No. Description information The sentence sequence and data products Description information The sentence sequence ,in, For users No. Description information The last sentence of For users No. Description information The number of sentences in For data products Description information The last sentence of For data products Description information The number of sentences in

[0037] Step 3.1.3: Use a bidirectional long short-term memory network to and Processing, the corresponding user No. Description information The context sentence level representation and data products Description information The context sentence level representation ,in, for The context sentence level representation of for The context sentence level representation of .

[0038] Further, step 3.2 is performed as follows:

[0039] Step 3.2.1. Calculate using formula (3) The attention weight , thus After the corresponding attention weights are connected, the data product is obtained The overall weight of :

[0040] (3)

[0041] In formula (3), is the weight vector, and are two linear transformation matrices, is the element-wise product, is the sigmoid function, is the activation function, For data products Description information The jth sentence in Contextual representation of represents transpose;

[0042] Step 3.2.2: Use formula (4) to get the data product Comprehensive representation of :

[0043] (4).

[0044] Further, step 3.3 is performed as follows:

[0045] Step 3.3.1: Use formula (5) to get user No. Description information No. Sentences Affinity Matrix :

[0046] (5)

[0047] In formula (5), is a connection operation, is the sentence parameter to be learned, is the activation function, is the mapping function;

[0048] Step 3.3.2: Use formula (6) to get user No. Description information No. Sentences The sentence-level attention weights :

[0049] (6)

[0050] Step 3.3.3: Use formula (7) to get user No. Description information Characterization :

[0051] (7)

[0052] Step 3.3.4: Use formula (8) to get user No. Description information Weight :

[0053] (8)

[0054] In formula (8), is the user parameter to be learned;

[0055] Step 3.3.5: Use formula (9) to get user Comprehensive characterization of :

[0056] (9)

[0057] In formula (9), Indicates user The number of data products that were interacted with.

[0058] Further, step 4 is performed as follows:

[0059] Step 4.1: Use formula (12) to get Tasks Comprehensive representation of , thus obtaining a comprehensive representation of all tasks ,in, Indicates user and data products Common task information collection The number of tasks in

[0060] (12)

[0061] In formula (12), express The importance weights are:

[0062] (13)

[0063] In formula (13), for In the user and data products Common task information collection The number of

[0064] Step 4.2, through the converter Embed and get the comprehensive representation of the task after embedding ,in, For the Tasks The integrated representation of the task after embedding;

[0065] Step 4.3: Calculate users and data products Connections , and through the Bayesian attention mechanism Processing to get user and data products Task integration representation .

[0066] Further, step 4.3 is performed as follows:

[0067] Step 4.3.1: Use formula (14) to get Attention score :

[0068] (14)

[0069] In formula (14), and are 2 matrix parameters, and are 2 vector parameters, is the scalar parameter to be optimized;

[0070] Step 4.3.2: Use formula (15) to get Updated deterministic attention weights :

[0071] (15)

[0072] In formula (15), Indicates user and data products Common task information collection Medium Foreign Tasks The attention weight of

[0073] Step 4.3.3: Modeling using formula (16) Bayesian attention weights :

[0074] (16)

[0075] In formula (16), represents the scale parameter of the Weibull distribution, and , Weibull represents Weibull distribution, is a hyperparameter, is the gamma function;

[0076] Step 4.3.5: Use formula (17) to get user and data products Task integration representation :

[0077] (17)

[0078] In formula (17), Indicates user and data products Common task information collection Remove tasks Foreign Tasks The Bayesian attention weights of .

[0079] Further, step 6 is performed as follows:

[0080] Step 6.1: Use formula (19) to construct the binary cross entropy loss function :

[0081] (19)

[0082] In formula (19), The total number of user-data product pairs;

[0083] Step 6.2: Use formula (20) to get the nodes in the relational graph convolutional network The probability of belonging to each type :

[0084] (20)

[0085] In formula (20), Respectively represent nodes Probability of belonging to users, data products, and tasks;

[0086] Step 6.3: Use formula (21) to construct the cross entropy loss of node type :

[0087] (twenty one)

[0088] In formula (21), Representation Node One-hot representation of the node type, Heterogeneous Graph The number of nodes in ;

[0089] Step 6.4: Use formula (22) to construct the hybrid loss function :

[0090] (twenty two)

[0091] In formula (22), is a hyperparameter;

[0092] Step 6.5: Use the gradient descent method to train the data product recommendation network and calculate the mixed loss function To update the network parameters until the mixed loss function The training process continues until convergence, thus obtaining the trained data product recommendation model.

[0093] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the data product recommendation method, and the processor is configured to execute the program stored in the memory.

[0094] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the data product recommendation method when executed by a processor.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] 1. The present invention proposes a task-aware data product recommendation method, which uses a large language model to extract the task information of users and data products, thereby integrating the task information into the data product recommendation task, which is conducive to improving the subsequent recommendation accuracy.

[0097] 2. The present invention constructs a heterogeneous graph consisting of users, data products, and tasks, and introduces a relational graph convolutional network to update and embed nodes. Ultimately, it can obtain a node representation that integrates both the structural information of the graph and the edge type, greatly improving the representation ability of the nodes and taking into account the interaction information between users and data products, thereby fully representing the user, data product, and task information and improving the accuracy of recommendations.

[0098] 3. The present invention introduces an asymmetric attention mechanism to target the heterogeneity of user description information and data product description information, that is, user description information is multiple different ones for interaction, and data product description information is its own description information. The data product and user description information are represented separately, thereby improving the accuracy of description information representation.

[0099] 4. The present invention introduces a task integration module, which not only considers the source of different task information, but also integrates the task embedding through the Bayesian attention mechanism, fully considering the task information of data products and users, and improving the accuracy of data product recommendation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 Heterogeneous graph and relational graph convolutional networks constructed for the present invention;

[0101] Figure 2 It is the overall structural diagram of the present invention. DETAILED DESCRIPTION

[0102] In this example, a task-aware data product recommendation method integrates the task information of users and data products into the data product recommendation task. Considering the interactive relationship between users and data products and the heterogeneity of user description information and data product description information, a relational graph convolutional network and an asymmetric attention mechanism are introduced to perform the data product recommendation task, thereby improving the accuracy of data product recommendation. Specifically, Figure 2 As shown, the method is performed in the following steps:

[0103] Step 1: Build an interaction matrix between users and data products , and according to the user Description information and data products Description information Extract users separately Task information and data products Task information ;

[0104] Step 1.1: Based on the interaction information between users and data products, obtain the interaction matrix between users and data products , among which, if , indicating that the user Interacted with data products ,like , it means that the user No data product interaction , is the number of users, is the number of data products, and the total number of user and data product pairs is ;

[0105] Get data products The description information is recorded as , considering the sparsity of user information, Descriptive information of the data products that users have interacted with The description information is recorded as , in, For users No. Description information, The number of data products representing user interactions;

[0106] Step 1.2: Use the large language model to learn from users Description information and data products Description information Extracting users from Task information , data products Task information ,user and data products Common task information collection ,in, Indicates user No. tasks, Represents data products No. tasks, Indicates user and data products Common task information collection The tasks, , and Representing users The number of tasks, data products The number of tasks and users and data products Common task information collection The number of tasks in

[0107] Considering the source information of the task, that is, from users, from data products, and from users and data products, the user and data products Common task information collection Middle Tasks The type source is recorded as ;

[0108] Step 2: Based on the interaction matrix and task information to build a heterogeneous graph , and for heterogeneous graphs Update the nodes in the embedding to get the representation of the user after the update embedding , representation of data products and task representation ; thereby updating the representation of the embedded user , representation of data products Convert them respectively and get the converted user representation accordingly and the representation of the transformed data products ;

[0109] Step 2.1: Treat all users, all data products, and all tasks as user nodes , data product node , task node , thus obtaining the node set , taking the interaction or similarity relationship between nodes as edges, thus obtaining the edge set , and construct a heterogeneous graph ;

[0110] Step 2.2: Figure 1 As shown, the relational graph convolutional network is used to After learning the node representation, we get the user Node representation , data products Node representation and tasks Node representation ;

[0111] Through multi-layer perceptron and Transformation is performed so that the node representation is aligned with other subsequent representations, and the converted user Representation and the transformed data products Representation ;

[0112] Step 3: Construct an asymmetric attention network module, including a sentence encoding unit, an attention mechanism unit, and a sentence aggregation unit, and Description information and data products Description information Processing to get user Enhanced representation of and data products Enhanced representation of ;

[0113] Step 3.1: For users No. Description information and data products Description information Processing to get user No. Description information Sentence-level representation of and data products Description information Sentence-level representation of ;

[0114] Step 3.1.1: User No. Description information and data products Description information Each word in is input into the BERT model for word embedding, and the maximum pooling method is used to extract the representation of the sentence, which can capture the strongest signal in each dimension of the sentence, that is, the information that best represents the characteristics of the sentence. Thus, using formula (1) and formula (2) respectively, we can get the user Description information The Sentences and data products Description information No. Sentences :

[0115] (1)

[0116] In formula (1), for The word embedding of the last word in , for Middle word embedding; U represents The number of words in

[0117] (2)

[0118] In formula (2), for The word embedding of the last word in , for Middle word embedding; D represents The number of words in

[0119] Step 3.1.2: Follow the process in step 3.1.1 to obtain the user No. Description information The sentence sequence and data products Description information The sentence sequence ,in, For users No. Description information The last sentence of For users No. Description information The number of sentences in For data products Description information The last sentence of For data products Description information The number of sentences in

[0120] Step 3.1.3: Each sentence only considers its own semantics and ignores the context in the same description information. To further refine the sentence embedding, a bidirectional long short-term memory network is used to embed and The bidirectional long short-term memory network can simultaneously consider the context before and after each sentence, thereby obtaining more comprehensive information and obtaining the user's No. Description information The context sentence level representation and data products Description information The context sentence level representation ,in, for The context sentence level representation of for The context sentence level representation of

[0121] Step 3.2: User description information is the description information of different data products that the user has interacted with. The data product description information is about the data product itself. Therefore, asymmetric attention representation is performed on the user description information and the data product description information. First, considering that the information of different sentences in the data product description information has different importance to the data product description, the data product attention mechanism module is used to Perform sentence-level aggregation to obtain a comprehensive representation of the data product d ;

[0122] Step 3.2.1. Calculate using formula (3) The attention weight , thus After the corresponding attention weights are connected, the data product is obtained The overall weight of :

[0123] (3)

[0124] In formula (3), is the weight vector, and are two linear transformation matrices, is the element-wise product, is the sigmoid function, is the activation function, For data products Description information The jth sentence in Contextual representation of represents transpose;

[0125] Step 3.2.2: Use formula (4) to get the data product Comprehensive representation of :

[0126] (4)

[0127] Step 3.3: Use the user attention mechanism module to Aggregate the context sentence level representations of all descriptive information of the user Comprehensive representation of ;

[0128] Step 3.3.1: Use formula (5) to get user No. Description information No. Sentences Affinity Matrix :

[0129] (5)

[0130] In formula (5), is a connection operation, is the sentence parameter to be learned, is the activation function, is the mapping function;

[0131] Step 3.3.2: Use formula (6) to get user No. Description information No. Sentences The sentence-level attention weights :

[0132] (6)

[0133] Step 3.3.3: Use formula (7) to get user No. Description information Characterization :

[0134] (7)

[0135] Step 3.3.4: Use formula (8) to get user No. Description information Weight :

[0136] (8)

[0137] In formula (8), is the user parameter to be learned;

[0138] Step 3.3.5: Use formula (9) to get user Comprehensive characterization of :

[0139] (9)

[0140] In formula (9), Indicates user The number of data products that are interacted with;

[0141] Step 3.4: Use a multi-layer perceptron to analyze the user Comprehensive characterization of and data products Comprehensive characterization of Transformation is performed so that the user comprehensive representation is aligned with the data product comprehensive representation and other representations, and the user is obtained accordingly. Comprehensive representation after conversion and data products Comprehensive representation after conversion ;

[0142] Step 3.5: Use equation (10) and equation (11) to get the user Enhanced representation of and data products Enhanced representation of :

[0143] (10)

[0144] (11)

[0145] Step 4: Build a task integration module to represent heterogeneous graph-based tasks With task-based representation Processing to get user and data products Task integration representation ;

[0146] Step 4.1: Considering that some tasks may be shared by users and data products, while other tasks belong only to users or data products, shared tasks may have different effects on the interaction modeling of user-data product pairs. Therefore, the source information of the tasks is also considered during task integration. Formula (12) is used to obtain Tasks Comprehensive representation of , thus obtaining a comprehensive representation of all tasks ,in, Indicates user and data products Common task information collection The number of tasks in

[0147] (12)

[0148] In formula (12), express The importance weights are:

[0149] (13)

[0150] In formula (13), for In the user and data products Common task information collection The number of

[0151] Step 4.2, through the converter Embed and get the comprehensive representation of the task after embedding ,in, For the Tasks The integrated representation of the task after embedding;

[0152] Step 4.3: Calculate users and data products Connections , and through the Bayesian attention mechanism Processing to get user and data products Task integration representation ;

[0153] Step 4.3.1: Use formula (14) to get Attention score :

[0154] (14)

[0155] In formula (14), and are 2 matrix parameters, and are 2 vector parameters, is the scalar parameter to be optimized;

[0156] Step 4.3.2: Use formula (15) to get Updated deterministic attention weights :

[0157] (15)

[0158] In formula (15), Indicates user and data products Common task information collection Medium Foreign Tasks The attention weight of

[0159] Step 4.3.3: Modeling using formula (16) Bayesian attention weights :

[0160] (16)

[0161] In formula (16), represents the scale parameter of the Weibull distribution, and , Weibull represents Weibull distribution, is a hyperparameter, is the gamma function;

[0162] Step 4.3.5: Use formula (17) to get user and data products Task integration representation :

[0163] (17)

[0164] In formula (17), Indicates user and data products Common task information collection Remove tasks Foreign Tasks Bayesian attention weights of ;

[0165] Step 5: Build a prediction module and use formula (18) to obtain the user For data products The predicted score :

[0166] (18)

[0167] In formula (18), is the sigmoid function;

[0168] Step 6: Train the data product recommendation network consisting of the heterogeneous graph node update embedding, the asymmetric attention network module, the task integration module and the prediction module to obtain a trained data product recommendation model, which is used to recommend the input data products and obtain the user's best prediction score for the data product;

[0169] Step 6.1: Use formula (19) to construct the binary cross entropy loss function :

[0170] (19)

[0171] In formula (19), The total number of user-data product pairs;

[0172] Step 6.2: Use formula (20) to get the nodes in the relational graph convolutional network The probability of belonging to each type :

[0173] (20)

[0174] In formula (20), Respectively represent nodes Probability of belonging to users, data products, and tasks;

[0175] Step 6.3: Use formula (21) to construct the cross entropy loss of node type :

[0176] (twenty one)

[0177] In formula (21), Representation Node One-hot representation of the node type, Heterogeneous Graph The number of nodes in ;

[0178] Step 6.4: Use formula (22) to construct the hybrid loss function :

[0179] (twenty two)

[0180] In formula (22), is a hyperparameter;

[0181] Step 6.5: Use the gradient descent method to train the data product recommendation network and calculate the mixed loss function To update the network parameters until the mixed loss function The training process continues until convergence, thus obtaining the trained data product recommendation model.

[0182] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0183] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A task-aware data product recommendation method, characterized in that: The steps are as follows: Step 1: Build an interaction matrix between users and data products , and according to the user Description information and data products Description information Extract users separately Task information and data products Task information ; Step 1.1: Based on the interaction information between users and data products, obtain the interaction matrix between users and data products , among which, if , indicating that the user Interacted with data products ,like , it means that the user No data product interaction , is the number of users, is the number of data products, and the total number of user and data product pairs is ; Get data products The description information is recorded as , the user Descriptive information of the data products that users have interacted with The description information is recorded as , in, For users No. Description information, The number of data products representing user interactions; Step 1.2: Use the large language model to learn from users Description information and data products Description information Extracting users from Task information , data products Task information ,user and data products Common task information collection ,in, Indicates user No. tasks, Represents data products No. tasks, Indicates user and data products Common task information collection The tasks, , and Representing users The number of tasks, data products The number of tasks and users and data products Common task information collection The number of tasks in The user and data products Common task information collection Middle Tasks The type source is recorded as ; Step 2: Build a heterogeneous graph node update embedding module and update the node based on the interaction matrix and task information to build a heterogeneous graph , so that the heterogeneous graph Update the nodes in the embedding to get the representation of the user after the update embedding , representation of data products and task representation ; and update the embedded user representation , representation of data products Convert them respectively and get the converted user representation accordingly and the representation of the transformed data products ; Step 3: Build an asymmetric attention network module and Description information and data products Description information Processing to get user Enhanced representation of and data products Enhanced representation of ; Step 3.1: For users No. Description information and data products Description information Processing to get user No. Description information Sentence-level representation of and data products Description information Sentence-level representation of ; Step 3.2: Use the data product attention mechanism module to Perform sentence-level aggregation to obtain a comprehensive representation of the data product d ; Step 3.3: Use the user attention mechanism module to Aggregate the context sentence level representations of all descriptive information of the user Comprehensive characterization of ; Step 3.4: Use a multi-layer perceptron to analyze the user Comprehensive characterization of and data products Comprehensive characterization of Transform and get the user Comprehensive representation after conversion and data products Comprehensive representation after conversion ; Step 3.5: Use equation (10) and equation (11) to get the user Enhanced representation of and data products Enhanced representation of : (10) (11) Step 4: Build a task integration module to represent heterogeneous graph-based tasks With task-based representation Processing to get user and data products Task integration representation ; Step 5: Build a prediction module and use formula (18) to obtain the user For data products The predicted score : (18) In formula (18), is the sigmoid function; Step 6: Train the data product recommendation network consisting of the heterogeneous graph node update embedding module, the asymmetric attention network module, the task integration module and the prediction module to obtain a trained data product recommendation model, which is used to recommend the input data products and obtain the user's best prediction score for the data product.

2. The task-aware data product recommendation method according to claim 1, characterized in that: Described step 2 is carried out as follows: Step 2.1: Treat all users, all data products, and all tasks as user nodes , data product node , task node , thus obtaining the node set , taking the interaction or similarity relationship between nodes as edges, thus obtaining the edge set , and construct a heterogeneous graph ; Step 2.2: Use the relational graph convolutional network to classify heterogeneous graphs After learning the node representation, we get the user Node representation , data products Node representation and tasks Node representation ; Through multi-layer perceptron and Transform and obtain the converted user accordingly Representation and the transformed data products Representation .

3. The task-aware data product recommendation method according to claim 2, characterized in that: Step 3.1 is performed as follows: Step 3.1.1: User No. Description information and data products Description information Each word in is input into the BERT model for word embedding, so that the user Description information The Sentences and data products Description information No. Sentences : (1) In formula (1), for The word embedding of the last word in , for Middle word embedding; U represents The number of words in (2) In formula (2), for The word embedding of the last word in , for Middle word embedding; D represents The number of words in Step 3.1.2: Follow the process in step 3.1.1 to obtain the user No. Description information The sentence sequence and data products Description information The sentence sequence ,in, For users No. Description information The last sentence of For users No. Description information The number of sentences in For data products Description information The last sentence of For data products Description information The number of sentences in Step 3.1.3: Use a bidirectional long short-term memory network to and Processing, the corresponding user No. Description information The context sentence level representation and data products Description information The context sentence level representation ,in, for The context sentence level representation of for The context sentence level representation of .

4. The task-aware data product recommendation method according to claim 3, characterized in that: Step 3.2 is performed as follows: Step 3.2.

1. Calculate using formula (3) The attention weight , thus After the corresponding attention weights are connected, the data product is obtained The overall weight of : (3) In formula (3), is the weight vector, and are two linear transformation matrices, is the element-wise product, is the sigmoid function, is the activation function, For data products Description information The jth sentence in Contextual representation of represents transpose; Step 3.2.2: Use formula (4) to get the data product Comprehensive characterization of : (4)。 5. The task-aware data product recommendation method according to claim 4, characterized in that: Step 3.3 is performed as follows: Step 3.3.1: Use formula (5) to get user No. Description information No. Sentences Affinity Matrix : (5) In formula (5), is a connection operation, is the sentence parameter to be learned, is the activation function, is the mapping function; Step 3.3.2: Use formula (6) to get user No. Description information No. Sentences The sentence-level attention weights : (6) Step 3.3.3: Use formula (7) to get user No. Description information Characterization : (7) Step 3.3.4: Use formula (8) to get user No. Description information Weight : (8) In formula (8), is the user parameter to be learned; Step 3.3.5: Use formula (9) to get user Comprehensive characterization of : (9) In formula (9), Indicates user The number of data products that were interacted with.

6. The task-aware data product recommendation method according to claim 5, characterized in that: Step 4 is performed as follows: Step 4.1: Use formula (12) to get Tasks Comprehensive representation of , thus obtaining a comprehensive representation of all tasks ,in, Indicates user and data products Common task information collection The number of tasks in (12) In formula (12), express The importance weights are: (13) In formula (13), for In the user and data products Common task information collection The number of Step 4.2, through the converter Embed and get the comprehensive representation of the task after embedding ,in, For the Tasks The integrated representation of the task after embedding; Step 4.3: Calculate users and data products Connections , and through the Bayesian attention mechanism Processing to get user and data products Task integration representation .

7. The task-aware data product recommendation method according to claim 6, characterized in that: Step 4.3 is performed as follows: Step 4.3.1: Use formula (14) to get Attention score : (14) In formula (14), and are 2 matrix parameters, and are 2 vector parameters, is the scalar parameter to be optimized; Step 4.3.2: Use formula (15) to get Updated deterministic attention weights : (15) In formula (15), Indicates user and data products Common task information collection Medium Foreign Tasks The attention weight of Step 4.3.3: Modeling using formula (16) Bayesian attention weights : (16) In formula (16), represents the scale parameter of the Weibull distribution, and , Weibull represents Weibull distribution, is a hyperparameter, is the gamma function; Step 4.3.5: Use formula (17) to get user and data products Task integration representation : (17) In formula (17), Indicates user and data products Common task information collection Remove tasks Foreign Tasks The Bayesian attention weights of .

8. The task-aware data product recommendation method according to claim 7, characterized in that: Step 6 is performed as follows: Step 6.1: Use formula (19) to construct the binary cross entropy loss function : (19) In formula (19), The total number of user-data product pairs; Step 6.2: Use formula (20) to get the nodes in the relational graph convolutional network The probability of belonging to each type : (20) In formula (20), Respectively represent nodes Probability of belonging to users, data products, and tasks; Step 6.3: Use formula (21) to construct the cross entropy loss of node type : (21) In formula (21), Representation Node One-hot representation of the node type, Heterogeneous Graph The number of nodes in ; Step 6.4: Use formula (22) to construct the hybrid loss function : (22) In formula (22), is a hyperparameter; Step 6.5: Use the gradient descent method to train the data product recommendation network and calculate the mixed loss function To update the network parameters until the mixed loss function The training process continues until convergence, thus obtaining the trained data product recommendation model.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the data product recommendation method according to any one of claims 1 to 8, and the processor is configured to execute the program stored in the memory.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the data product recommendation method according to any one of claims 1 to 8 are executed.

Images