Iot product recommendation method and device based on graph neural network, equipment and medium

Abstract

Embodiments of the application disclose a method, device and equipment for recommending Internet of Things products based on a graph neural network and a medium. The method belongs to the technical field of Internet of Things and artificial intelligence, and comprises the following steps: extracting text information of Internet of Things products from an Internet of Things product library, and constructing a graph structure according to the text information; determining quantified features of the Internet of Things products according to the graph structure and a graph convolution network; if a user is not a new user, mining a dependency relationship between historical behavior sequences of the user in a preset time period to obtain a long-term interest vector feature; activating a short-term interest state of the user according to the quantified features to obtain a short-term interest vector feature; predicting a selection rate of the user for the Internet of Things products through a selection rate prediction model according to attribute information of the user, the Internet of Things products and the short-term interest vector feature; and recommending the Internet of Things products to the user according to a recommendation list generated based on the selection rate. The embodiments of the application improve the recommendation effect and usage rate of the Internet of Things products.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things (IoT) and artificial intelligence (AI) technologies, and in particular to an IoT product recommendation method, apparatus, device, and medium based on graph neural networks. Background Technology

[0002] As a key area for independent innovation in China's next-generation information technology, the Internet of Things (IoT) holds immense potential for innovation. Innovation activities are becoming increasingly active in areas such as chips, sensors, short-range transmission, massive data processing, and comprehensive integration and applications, with innovative elements constantly accumulating. The deepening application of IoT across various industries will spawn numerous new technologies, products, applications, and models. The huge future market demand will bring rare development opportunities and vast growth potential to the IoT.

[0003] Currently, operators primarily rely on account managers to collect target company information online or through in-person visits to understand their needs. These needs are then matched with the operator's IoT products, leading to corresponding product marketing activities or simple product promotion and sales through IoT marketplaces. However, this product recommendation method lacks direction, heavily depends on user experience and manual intervention, and is ill-suited to today's rapidly changing market competition. While some operators use machine learning and deep learning methods for personalized product recommendations, these often simply recommend products to users who have previously ordered similar products, or to users with similar product preferences, resulting in poor recommendation effectiveness and low usage rates for IoT products. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and medium for recommending IoT products based on graph neural networks, aiming to improve the recommendation effect and usage rate of IoT products.

[0005] In a first aspect, embodiments of the present invention provide an IoT product recommendation method based on graph neural networks, comprising:

[0006] Extract text information of IoT products from the IoT product library, and construct a graph structure based on the text information;

[0007] The graph information corresponding to the graph structure is input into a graph convolutional network for iterative learning to obtain the quantitative representation of the keywords, and the quantitative representation of the keywords is fused to obtain the quantitative features of the IoT product.

[0008] If the user is not a new user, then the long-term interest vector features are obtained by mining the dependencies between the user's historical behavior sequences within a preset time period.

[0009] Short-term interest vector features are obtained by activating the short-term interest states of users related to the IoT product based on the quantified features.

[0010] Based on the acquired user attribute information, the IoT product, and the short-term interest vector features, a selection rate prediction model is used to predict the user's selection rate for the IoT product.

[0011] The IoT products are recommended to the user based on the selection rate-generated recommendation list.

[0012] Secondly, embodiments of the present invention also provide an IoT product recommendation device based on a graph neural network, comprising:

[0013] Extraction building units are used to extract text information of IoT products from the IoT product library and construct a graph structure based on the text information;

[0014] The input fusion unit is used to input graph information corresponding to the graph structure into the graph convolutional network for iterative learning to obtain the quantitative representation of the keywords, and to fuse the quantitative representation of the keywords to obtain the quantitative features of the Internet of Things product.

[0015] The mining unit is used to mine the dependencies between the user's historical behavior sequences within a preset time period to obtain long-term interest vector features if the user is not a new user.

[0016] An activation unit is used to activate the short-term interest state of users related to the Internet of Things product based on the quantized features to obtain short-term interest vector features.

[0017] The prediction unit is used to predict the user's selection rate for the IoT product based on the acquired user attribute information, the IoT product, and the short-term interest vector features through a selection rate prediction model.

[0018] The first recommendation unit is used to recommend the IoT products to the user based on the recommendation list generated by the selection rate.

[0019] Thirdly, embodiments of the present invention also provide a computer device, which is equipped with an Internet of Things (IoT) product recommendation system based on graph neural networks. The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0020] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0021] This invention provides a method, apparatus, device, and medium for recommending IoT products based on graph neural networks. The method includes: extracting textual information of IoT products from an IoT product library and constructing a graph structure based on the textual information; inputting graph information corresponding to the graph structure into a graph convolutional network for iterative learning to obtain quantified representations of keywords, and fusing the quantified representations of the keywords to obtain quantified features of the IoT products; if the user is not a new user, mining the dependencies between the user's historical behavior sequences within a preset time period to obtain long-term interest vector features; activating the short-term interest states of users related to the IoT products based on the quantified features to obtain short-term interest vector features; predicting the user's selection rate for the IoT products using a selection rate prediction model based on the acquired user attribute information, the IoT products, and the short-term interest vector features; and recommending the IoT products to the user based on a recommendation list generated according to the selection rate. The technical solution of this invention can accurately construct a graph structure through the textual information of IoT products, obtain quantified features through a graph convolutional network based on the graph structure, mine the dependencies between users' historical behavior sequences to obtain long-term interest vector features, obtain short-term interest vector features based on the quantified features, predict users' selection rate for IoT products through a selection rate prediction model based on attribute information, IoT products, and short-term interest vector features, and generate a recommendation list to provide personalized recommendations for IoT products to users based on the selection rate. This helps users select the IoT products they need more quickly and accurately, promotes IoT products more deeply, and improves the recommendation effect and usage rate of IoT products. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an IoT product recommendation method based on graph neural networks, provided in an embodiment of the present invention.

[0024] Figure 2 A schematic diagram illustrating the convergence of the loss value during the training and testing process of the selection rate prediction model provided in this embodiment of the invention;

[0025] Figure 3 A flowchart illustrating an IoT product recommendation method based on graph neural networks, provided for another embodiment of the present invention;

[0026] Figure 4 A schematic diagram of text information provided in an embodiment of the present invention;

[0027] Figure 5 A schematic diagram illustrating keyword information provided in an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of the relevance matrix between keywords provided in an embodiment of the present invention;

[0029] Figure 7 A schematic diagram of the structure of the GCN model provided in an embodiment of the present invention;

[0030] Figure 8 A schematic diagram of historical behavior data provided in an embodiment of the present invention;

[0031] Figure 9 A schematic diagram of the structure of the GRU model provided in an embodiment of the present invention;

[0032] Figure 10 A schematic diagram illustrating the recommendation performance of NCF, LSTUR, PNN, DIN, and the IoT product recommendation method based on graph neural networks provided in the embodiments of the present invention;

[0033] Figure 11 A schematic block diagram of an IoT product recommendation device based on a graph neural network, provided for an embodiment of the present invention;

[0034] Figure 12 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0037] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0038] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0039] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrases "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0040] Please see Figure 1 , Figure 1 This is a flowchart illustrating an IoT product recommendation method based on graph neural networks provided in an embodiment of the present invention. The IoT product recommendation method based on graph neural networks of this invention can be applied to IoT product recommendation systems based on graph neural networks. For example, it can be implemented through software programs configured on the IoT product recommendation system based on graph neural networks. Figure 1 As shown, the method includes the following steps S100-S150.

[0041] S100. Extract text information of IoT products from the IoT product database and construct a graph structure based on the text information.

[0042] In this embodiment of the invention, text information of IoT products is extracted from an IoT product database. This text information includes product name, category, tags, functions, etc. Keyword information is obtained by extracting keywords from the text information using BiLSTM (Bi-directional Long Short-Term Memory). An initial feature matrix corresponding to the keyword information is constructed through an embedding layer. An adjacency matrix is ​​constructed based on whether the keywords in the keyword information belong to the same text. A graph structure is then constructed based on the initial feature matrix and the adjacency matrix. Specifically, the text information of the IoT products is first mapped to a vector space in units of words, z = [z1, z2, ..., z...]. n ] represents a text sequence, z i Let z be the word vector of the i-th word in this text; then, the text z is encoded using the BiLSTM model through formulas (1.1) to (1.4):

[0043] f t =σ(W F·[h t-1 , z t ]+b F (1.1)

[0044] i t =σ(W I ·[h t-1 , z t ]+b I (1.2)

[0045] o t =σ(W o ·[h t-1 , z t ]+b o (1.3)

[0046] h t =o t ·tanh(c t (1.4)

[0047] z t The input at time t represents the text information from the IoT product. σ and tanh are activation functions, and the hidden state h is output at each time step. t BiLSTM computes two different hidden layer representations for each text using both sequential and reverse computation. These representations are then concatenated to obtain the final hidden layer representation, with the label with the highest probability used as the representation for each keyword n. i The tags are ultimately used to output keywords within IoT products.

[0048] The initial feature matrix X = [x1, x2, x3, ..., x4] of the product keywords is obtained through the embedding layer. Nall(n) Then, based on whether the keywords co-occur in the same text segment, an adjacency matrix A is constructed. Based on the feature matrix X and the adjacency matrix A, a graph structure G0 with keywords as nodes is obtained.<X,A> .

[0049] S110. Input the graph information corresponding to the graph structure into the graph convolutional network for iterative learning to obtain the quantitative representation of the keywords, and fuse the quantitative representation of the keywords to obtain the quantitative features of the Internet of Things product.

[0050] In this embodiment of the invention, the graph information includes the keyword information, the initial feature matrix, and the adjacency matrix. A relevance matrix is ​​constructed by measuring the relevance between keywords in the keyword information using Pointwise Mutual Information (PMI), and this relevance matrix is ​​used as the learning label for the graph convolutional network (GCN) to quantize the graph structure. The initial feature matrix and the adjacency matrix are input into the GCN, and the graph structure is iteratively learned according to a preset loss function to obtain the quantized representation of the keywords. Specifically, PMI is used to measure the relevance between keywords ni to obtain a relevance matrix F0, which is used as the label η for the graph result G0 quantized by the GCN. The initial feature matrix X and the adjacency matrix A are used as inputs to the GCN, and a preset loss function L is defined. gcn L represents the mean squared error of the similarity between each pair of keywords. gcn As shown in formula (1.5):

[0051]

[0052] Among them, the keyword n i With n j The label between them is η ij n i The vectorized representation of v i n j The vectorized representation of v j Then, according to the propagation rules of GCN, the characteristics of the keywords are obtained as shown in formula (1.6):

[0053] V = f(X, A) = σ(A) * σ(A * XW (0) W ( l ) (1.6)

[0054] After processing by GCN, graph G0 yields a graph structure G containing key information about IoT products and their interrelationships. G is composed of keywords and their features, as shown in formula (1.7).

[0055] G =<V,A> (1.7)

[0056] The quantitative representations of the keywords are fused to obtain the quantitative characteristics of the IoT product. Specifically, when determining keyword n... i The vectorized representation of v i Then, according to n i m times in the same product iThe vectorized representation d of the IoT product is shown in formula (1.8), where d is also a quantitative characteristic of the IoT product:

[0057]

[0058] S120. If the user is not a new user, then the long-term interest vector features are obtained by mining the dependencies between the user's historical behavior sequences within a preset time period.

[0059] In this embodiment of the invention, if a user is not a new user (indicating they are an existing user), then the dependencies between the user's historical behavior sequences within a preset time period are mined to obtain long-term interest vector features. Specifically, a Gate Recurrent Unit (GRU) is used to mine the user's historical behavior sequences [d1, d2, ..., dn] for IoT products. T Based on the dependencies between them, we obtain the vector representation of the user's long-term interests [h1, h2, ..., h]. T Understandably, historical behavior sequences are, for example, sequences of browsing / ordering / using behavior.

[0060] S130. Based on the quantization features, activate the short-term interest state of users related to the Internet of Things product to obtain short-term interest vector features.

[0061] In this embodiment of the invention, an attention mechanism is used to calculate the user's interest state and the attention score of the IoT product; the attention score is embedded in a GRU structure to activate the user's short-term interest state related to the IoT product, thus obtaining short-term interest vector features. Specifically, based on the representation d of the IoT product in graph G, the attention mechanism and GRU are used to analyze the short-term interest state h′ related to the candidate product d0 in the user's long-term interest. t This is represented. Specifically, the user's interest state h can be obtained from the attention mechanism. t Attention score a of IoT product d0 t As shown in formula (1.9):

[0062]

[0063] Among them, a t This represents the correlation between IoT products and users' short-term interests, and will be a t Specifically, embedding the attention score a into the GRU structure involves... t With the original update gate u′ of GRU t Multiply to obtain a new update gate The new update gate, as shown in Equation (1.10), not only incorporates the dimensional features of the original update gate but also the correlation features between the short-term interests of IoT products and users. This makes the short-term interest state h′ related to the candidate IoT products in the long-term interest more relevant. t As shown in Equation (1.11), the short-term interest vector features are obtained:

[0064]

[0065]

[0066] S140. Based on the acquired user attribute information, the IoT product, and the short-term interest vector features, the user's selection rate for the IoT product is predicted using a selection rate prediction model.

[0067] In this embodiment of the invention, user attribute information is obtained, including gender and age; the short-term interest vector features, gender, age, and the IoT product are concatenated to obtain concatenated input features; the concatenated input features are input into a selection rate prediction model, and the concatenated input features are processed by the Dice activation function, Softmax function, and a preset selection loss function in the selection rate prediction model to generate the user's selection rate for the IoT product. It should be noted that the selection rate prediction model is a multilayer perceptron, and the convergence of the loss value during training and testing is as follows... Figure 2 As shown. Specifically, the concatenated input features are as shown in formula (1.12):

[0068] γ=Concat(h′ t (E0, S0, d0) (1.12)

[0069] Where, h′ t Gender (S0) and age (E0) represent the user's short-term interest characteristics, while d0 represents the IoT product. γ is input into a multi-layer perceptron (MLP) for selection rate prediction. The Dice activation function is used between fully connected layers to learn non-linear relationships, as shown in equation (1.13).

[0070] R i =Dice(W i R i-1 +b i ), i=1, 2, 3,..., c (1.13)

[0071] Among them, W i and b i These are the trainable parameter matrix and bias in the hidden layer, respectively, R.c The hidden units output from the c-th layer are ultimately used to perform selection rate prediction on the last hidden layer using the Softmax function, as shown in formula (1.14):

[0072]

[0073] in, The selection rate predicted by the model is represented by the preset selection loss function as shown in Equation (1.15):

[0074]

[0075] Where y∈{0,1} represents the actual click situation. If y=0, it means that the user has not clicked, purchased or interacted with this IoT product. Otherwise, it means that the user has interacted with the IoT product. N represents the size of the training batch. Finally, the result is converted into the selection rate of candidate product d0 to realize personalized recommendation of IoT products.

[0076] S150. Recommend the IoT products to the user based on the recommendation list generated by the selection rate.

[0077] In this embodiment of the invention, the selection rates are sorted from high to low to generate a recommendation list, and IoT products that are ranked higher in the recommendation list are recommended to users to improve the recommendation effect and usage rate of IoT products.

[0078] Figure 3 This is a flowchart illustrating an IoT product recommendation method based on a graph neural network, provided in another embodiment of the present invention. Figure 3 As shown, the IoT product recommendation method based on graph neural networks in this embodiment includes steps S200-S290. Steps S200-S250 are similar to steps S100-S150 in the previous embodiment and will not be described again here. The following details the additional steps S260-S290 in this embodiment.

[0079] S260. If the user is the new user, then recommend keywords related to the IoT product to the user;

[0080] S270. If the keyword selected by the user is received, the selected keyword is taken as the keyword of interest.

[0081] S280. Calculate the importance of the interest keywords for the IoT products using a preset calculation formula, and recommend important IoT products to the user based on the importance of the IoT products.

[0082] S290: Receive user behavior data generated based on the important IoT products, and mine the dependencies between the behavior data to obtain the long-term interest vector features, then execute step S230.

[0083] In this embodiment of the invention, when faced with new users who have no information or behavior, it is impossible to obtain their interaction history with IoT products. This leads to the cold start problem in the recommendation system, making it impossible to mine the interest characteristics of new customers or users for IoT product recommendations. In this embodiment, IoT product-related keywords are first pushed to new users. When a user selects k (10≤k≤20) keywords n1, n2, n3, ..., n..., ... k Then, based on the keywords they selected, the importance of each keyword to the corresponding IoT product is calculated using TF-IDF. Specifically, the frequency and inverse document frequency of the selected keywords in the IoT product are statistically analyzed using the TF and IDF calculation formulas. Based on the frequency and inverse document frequency, the TF-IDF value is calculated using the TF-IDF calculation formula to determine the importance of the selected keywords to the corresponding IoT product. More specifically, the TF-IDF calculation steps are as follows:

[0084] Statistical keywords n i In IoT products j The frequency of occurrence in the text is calculated using the formula shown in formula (1.16):

[0085]

[0086] Wherein, N(d) j n i ) represents the keyword n i In IoT products j The number of times it appears in N(d) j * represents IoT product d j The total number of keywords in the document is calculated, and then the inverse document frequency is counted. The specific calculation formula is shown in Formula (1.17):

[0087]

[0088] Where N represents the number of IoT products, N(n i ) represents the keyword n i The more IoT products a keyword appears in, the higher its TF-IDF value indicates, indicating that the keyword is more important to that IoT product. The formula for calculating TF-IDF is shown in formula (1.18) below:

[0089] TF-IDF(d j n i ) = TF(dj n i )·IDF(n i (1.18)

[0090] The keyword n ​​was obtained through calculation. i (i = 1, 2, 3, ..., k) represent the k most important IoT products (10 ≤ k ≤ 20). These IoT products are then pushed to users, and behavioral data [d1, d2, ..., dk] is generated based on their choices. T This allows us to mine short-term and long-term interests based on behavioral data, thereby determining users' different selection rates for various IoT products and providing personalized recommendation services for IoT products.

[0091] CTR=f3(d0,[d1,d2,...,d T ], profile) (1.19)

[0092] In summary, whether for new or existing users, the embodiments of this invention can provide them with reasonable personalized recommendation services for IoT products.

[0093] For ease of understanding, a specific embodiment will be used for illustration below.

[0094] Extract multi-dimensional textual information about products from the IoT product database. This textual information includes data such as product name, category, tags, and functions. Specifically, for example... Figure 4 As shown, Figure 4 For the extracted text information, in Figure 4 Only a portion of the text information is displayed. This is achieved through BiLSTM... Figure 4 Keyword information obtained by extracting keywords from text information, such as Figure 5 As shown, by Figure 5 It can be seen that corresponding keywords have been extracted for each product. After obtaining the keywords contained in each IoT product, the keywords are embedded into the vector space to obtain the initial feature matrix. Next, construct an adjacency matrix based on whether the keywords appear in the same product, as shown in formula (1.20). Thus, we obtain a graph structure with keywords as nodes and 0 / 1 as edges.

[0095]

[0096] Using PMI to measure the relevance between keywords yields the following results: Figure 6 The relevance matrix between the keywords is shown.

[0097] After obtaining the relevance matrix, it is used as the label for the GCN model to learn. The graph structure described above is used as the input to the GCN model to construct a GCN model containing two convolutional layers. The GCN model structure diagram is as follows. Figure 7 As shown, by Figure 7 As can be seen, the GCN model ultimately outputs a 64-dimensional keyword vector, which is equivalent to obtaining the 64-dimensional quantitative features d of IoT products.

[0098] After obtaining the quantitative characteristics of IoT products, historical user behaviors such as ordering / browsing of IoT products are collected to obtain the final dataset:

[0099] The user's historical behavior sequence [d1, d2, ..., d] T The input is used to model the GRU system and obtain the user's long-term interest features [h1, h2, ..., h]. T Historical behavior data corresponding to historical behavior sequences, such as Figure 8 As shown, the structure diagram of GRU is as follows: Figure 9 As shown.

[0100] Then, an attention mechanism combined with GRU is used to activate interest states with high relevance to the candidate product in long-term interest features. Finally, combined with the user's basic attributes, a fully connected layer γ = Concat(h′) is used. t (E0, S0, d0) is used to obtain the predicted selection rate of users for products, and a recommendation list is generated based on the selection rate.

[0101] Figure 10 The recommended effects of NCF, LSTUR, PNN, DIN, and the recommended method of this invention are demonstrated. Figure 10 It can be seen that the recommended effect of the present invention is better.

[0102] It should be noted that, in this embodiment, the product modeling method integrates both product feature information and inter-product association information into a product-node topology, and uses a graph convolutional network to model and learn this graph. Regarding the user interest modeling method, it considers not only the user's long-term interests over a period of time but also short-term interests highly relevant to IoT products within those long-term interests, thus reproducing a more realistic picture of user interests. In modeling short-term interests, the attention score is added to the update gate of the GRU structure; specifically, the attention score is multiplied by the original update gate to obtain a new update gate, preserving the original dimensional information of the update gate. This approach allows for the acquisition of both product-specific features and inter-product association features, enabling more comprehensive quantification of IoT products and improving the efficiency and accuracy of personalized recommendations in subsequent stages. Therefore, the recommendation method in this embodiment of the invention fully utilizes the characteristics of the product and takes into full account people's long-term and short-term interests, effectively improving the product recommendation effect, that is, improving recommendation efficiency and accuracy, thereby increasing the usage rate of IoT products. Moreover, it can also realize personalized recommendations for IoT products to address the cold start problem that may be caused by new users.

[0103] Figure 11 This is a schematic block diagram of an IoT product recommendation device 200 based on a graph neural network provided in an embodiment of the present invention. Figure 11 As shown, corresponding to the above-described IoT product recommendation method based on graph neural networks, the present invention also provides an IoT product recommendation apparatus 200 based on graph neural networks. This IoT product recommendation apparatus 200 based on graph neural networks includes units for executing the above-described IoT product recommendation method based on graph neural networks. Specifically, please refer to... Figure 11 The IoT product recommendation device 200 based on graph neural networks includes an extraction and construction unit 201, an input fusion unit 202, a mining unit 203, an activation unit 204, a prediction unit 205, and a first recommendation unit 206.

[0104] The extraction and construction unit 201 is used to extract text information of IoT products from the IoT product library and construct a graph structure based on the text information; the input fusion unit 202 is used to input the graph information corresponding to the graph structure into a graph convolutional network for iterative learning to obtain the quantitative representation of keywords, and fuse the quantitative representation of keywords to obtain the quantitative features of the IoT products; the mining unit 203 is used to mine the dependency relationship between the user's historical behavior sequence within a preset time period to obtain long-term interest vector features if the user is not a new user; the activation unit 204 is used to activate the short-term interest state of users related to the IoT products according to the quantitative features to obtain short-term interest vector features; the prediction unit 205 is used to predict the user's selection rate for the IoT products through a selection rate prediction model based on the obtained user attribute information, the IoT products, and the short-term interest vector features; and the first recommendation unit 206 is used to recommend the IoT products to the user based on the recommendation list generated by the selection rate.

[0105] In some embodiments, such as this one, the extraction building unit 201 includes an extraction unit, a first building unit, and a second building unit.

[0106] The extraction unit is used to extract text information of IoT products from the IoT product library and extract keywords from the text information using BiLSTM to obtain keyword information; the first construction unit is used to construct an initial feature matrix corresponding to the keyword information through an embedding layer; the second construction unit is used to construct an adjacency matrix based on whether the keywords in the keyword information belong to the same text, and construct a graph structure based on the initial feature matrix and the adjacency matrix.

[0107] In some embodiments, such as this one, the input fusion unit 202 includes a third construction unit and an iteration unit.

[0108] The third construction unit is used to measure the relevance between keywords in the keyword information through PMI to construct a relevance matrix, and uses the relevance matrix as a learning label for the graph convolutional network to perform quantization learning on the graph structure; the iterative unit is used to input the initial feature matrix and the adjacency matrix into the graph convolutional network, and perform iterative learning on the graph structure according to a preset loss function to obtain the quantization representation of the keywords.

[0109] In some embodiments, such as this one, the activation unit 204 includes a first computing unit and an activation subunit.

[0110] The first calculation unit is used to calculate the user's interest state and the attention score of the IoT product through an attention mechanism; the activation unit is used to embed the attention score into the GRU structure to activate the short-term interest state of the user related to the IoT product to obtain short-term interest vector features.

[0111] In some embodiments, such as this one, the prediction unit 205 includes an acquisition unit, a splicing unit, and a prediction subunit.

[0112] The acquisition unit is used to acquire user attribute information, including gender and age; the splicing unit is used to splice the short-term interest vector features, gender, age, and IoT products to obtain spliced ​​input features; the prediction subunit is used to input the spliced ​​input features into the selection rate prediction model, and process the spliced ​​input features through the Dice activation function, Softmax function, and preset selection loss function in the selection rate prediction model to generate the user's selection rate for the IoT products.

[0113] In some embodiments, such as this one, the IoT product recommendation device 200 based on graph neural networks further includes a second recommendation unit, a receiving unit, a third recommendation unit, and a receiving mining unit.

[0114] The second recommendation unit is used to recommend keywords related to the IoT product to the user if the user is a new user; the receiving unit is used to use the selected keyword as an interest keyword if it receives the keyword selected by the user; the receiving unit is used to calculate the importance of the interest keyword to the IoT product using a preset calculation formula, and recommend important IoT products to the user based on the importance of the IoT product; the receiving mining unit is used to receive behavioral data generated by the user based on the important IoT products, mine the dependency relationship between the behavioral data to obtain the long-term interest vector feature, and execute the step of activating the short-term interest state of the user related to the IoT product based on the quantitative feature to obtain the short-term interest vector feature.

[0115] In some embodiments, such as this one, the third recommendation unit includes a second calculation unit and a third calculation unit.

[0116] The second calculation unit is used to calculate the frequency and inverse document frequency of the interest keywords in the IoT product using the TF calculation formula and the IDF calculation formula; the third calculation unit is used to calculate the TF-IDF value based on the frequency and the inverse document frequency using the TF-IDF calculation formula to obtain the importance of the interest keywords for the IoT product.

[0117] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned IoT product recommendation device 200 based on graph neural networks and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0118] The aforementioned IoT product recommendation device based on graph neural networks can be implemented as a computer program, which can, for example... Figure 12 It runs on the computer device shown.

[0119] Please see Figure 12 , Figure 12 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 900 is a device for building IoT product recommendations based on graph neural networks.

[0120] See Figure 12 The computer device 900 includes a processor 902, a memory, and an interface 907 connected via a system bus 901, wherein the memory may include a storage medium 903 and internal memory 904.

[0121] The storage medium 903 can store an operating system 9031 and a computer program 9032. When the computer program 9032 is executed, it enables the processor 902 to execute the aforementioned IoT product recommendation method based on graph neural networks.

[0122] The processor 902 provides computing and control capabilities to support the operation of the entire computer device 900.

[0123] The internal memory 904 provides an environment for the execution of the computer program 9032 in the storage medium 903. When the computer program 9032 is executed by the processor 902, the processor 902 can execute an Internet of Things product recommendation method based on a graph neural network.

[0124] This interface 905 is used for communication with other devices. Those skilled in the art will understand that... Figure 12The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 900 to which the present application is applied. The specific computer device 900 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0125] The processor 902 is used to run a computer program 9032 stored in a memory to implement any embodiment of the above-described IoT product recommendation method based on graph neural networks.

[0126] It should be understood that in the embodiments of this application, the processor 902 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0127] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the wireless communication system to implement the process steps of the embodiments of the above methods.

[0128] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform any of the embodiments of the above-described IoT product recommendation method based on graph neural networks.

[0129] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0130] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, wireless communication software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0131] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0132] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0133] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This wireless communication software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a user wireless communication device, terminal, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0134] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0135] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A graph neural network-based Internet of Things product recommendation method, characterized in that, include: Extract text information of IoT products from the IoT product library, and construct a graph structure based on the text information; The graph information corresponding to the graph structure is input into a graph convolutional network for iterative learning to obtain the quantitative representation of the keywords, and the quantitative representation of the keywords is fused to obtain the quantitative features of the IoT product. If the user is not a new user, then the long-term interest vector features are obtained by mining the dependencies between the user's historical behavior sequences within a preset time period. Short-term interest vector features are obtained by activating the short-term interest states of users related to the IoT product based on the quantified features. Based on the acquired user attribute information, the IoT product, and the short-term interest vector features, a selection rate prediction model is used to predict the user's selection rate for the IoT product. The IoT products are recommended to the user based on the recommendation list generated from the selection rate. If the user is a new user, then keywords related to the IoT product will be recommended to the user; If the user selects the keyword, then the selected keyword will be used as the keyword of interest. The importance of the interest keywords to the IoT products is calculated using a preset calculation formula, and important IoT products are recommended to the user based on the importance of the IoT products. Receive user behavior data generated based on the important IoT products, mine the dependencies between the behavior data to obtain the long-term interest vector features, and execute the step of activating the short-term interest state of users related to the IoT products based on the quantified features to obtain short-term interest vector features.

2. The graph neural network-based Internet of Things product recommendation method according to claim 1, characterized in that, The step of extracting text information of IoT products from the IoT product database and constructing a graph structure based on the text information includes: Text information of IoT products is extracted from the IoT product database, and keyword information is obtained by extracting keywords from the text information using BiLSTM. An initial feature matrix corresponding to the keyword information is constructed through the embedding layer; An adjacency matrix is ​​constructed based on whether the keywords in the keyword information belong to the same text, and a graph structure is constructed based on the initial feature matrix and the adjacency matrix.

3. The graph neural network-based Internet of Things product recommendation method according to claim 2, characterized in that, The graph information includes the keyword information, the initial feature matrix, and the adjacency matrix. The step of inputting the graph information corresponding to the graph structure into a graph convolutional network for iterative learning to obtain the quantized representation of the keywords includes: The relevance matrix is ​​constructed by measuring the relevance between keywords in the keyword information using PMI, and the relevance matrix is ​​used as the learning label for the graph convolutional network to perform quantitative learning on the graph structure. The initial feature matrix and the adjacency matrix are input into the graph convolutional network, and the graph structure is iteratively learned according to a preset loss function to obtain the quantized representation of the keywords.

4. The graph neural network-based Internet of Things product recommendation method according to claim 1, characterized in that, The step of activating the short-term interest state of users related to the IoT product based on the quantified features to obtain short-term interest vector features includes: The user's interest state and the attention score of the IoT product are calculated using an attention mechanism. The attention score is embedded into the GRU structure to activate the short-term interest state of users associated with the IoT product, thereby obtaining short-term interest vector features.

5. The graph neural network-based Internet of Things product recommendation method according to claim 1, characterized in that, The step of predicting the user's selection rate for the IoT product using a selection rate prediction model based on the acquired user attribute information, the IoT product, and the short-term interest vector features includes: Obtain user attribute information, wherein the attribute information includes gender and age; The short-term interest vector features, the gender, the age, and the IoT product are concatenated to obtain the concatenated input features; The spliced ​​input features are input into the selection rate prediction model, and the spliced ​​input features are processed by the Dice activation function, Softmax function and preset selection loss function in the selection rate prediction model to generate the user's selection rate for the Internet of Things product.

6. The graph neural network-based Internet of Things product recommendation method according to claim 1, characterized in that, The calculation of the importance of the interest keywords for the IoT product using a preset calculation formula includes: The frequency and inverse document frequency of the interest keywords in the IoT products are statistically analyzed using the TF and IDF calculation formulas. Based on the frequency and inverse document frequency, the TF-IDF value is calculated using the TF-IDF calculation formula to determine the importance of the interest keywords to the IoT products.

7. An Internet of Things product recommendation apparatus based on a graph neural network, characterized by, include: Extraction building units are used to extract text information of IoT products from the IoT product library and construct a graph structure based on the text information; The input fusion unit is used to input graph information corresponding to the graph structure into the graph convolutional network for iterative learning to obtain the quantitative representation of the keywords, and to fuse the quantitative representation of the keywords to obtain the quantitative features of the Internet of Things product. The mining unit is used to mine the dependencies between the user's historical behavior sequences within a preset time period to obtain long-term interest vector features if the user is not a new user. An activation unit is used to activate the short-term interest state of users related to the Internet of Things product based on the quantized features to obtain short-term interest vector features. The prediction unit is used to predict the user's selection rate for the IoT product based on the acquired user attribute information, the IoT product, and the short-term interest vector features through a selection rate prediction model. The first recommendation unit is used to recommend the IoT products to the user based on the recommendation list generated by the selection rate. The second recommendation unit is used to recommend keywords related to the IoT product to the user if the user is a new user. The receiving unit is configured to, if it receives the keyword selected by the user, use the selected keyword as the keyword of interest. The third recommendation unit is used to calculate the importance of the interest keywords to the IoT products using a preset calculation formula, and recommend important IoT products to the user based on the importance of the IoT products. A receiving and mining unit is used to receive behavioral data generated by users based on the important IoT products, mine the dependencies between the behavioral data to obtain the long-term interest vector features, and execute the step of activating the short-term interest state of users related to the IoT products based on the quantified features to obtain short-term interest vector features.

8. A computer device, comprising: The computer device is equipped with an Internet of Things (IoT) product recommendation system based on graph neural networks. The computer device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-6.

Images