Communication user package recommendation method, device, electronic device and storage medium

Abstract

The present invention provides a communication user package recommendation method, device, electronic device and storage medium, belonging to the field of computer technology, the method comprising: obtaining user sample information; constructing a user connection matrix and a user degree matrix according to the feature information; performing topic extraction on the description information to obtain corresponding topic feature information, and calculating a user similarity matrix according to the topic feature information and the feature information; calculating a second weight matrix according to the user connection matrix, the user degree matrix and the user similarity matrix; constructing a user package model according to the feature information, and recommending packages to target users based on the user similarity matrix, the second weight matrix and the user package model. The present invention can realize accurate recommendation of communication user packages and effectively improve the effect of package recommendation.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device, electronic device and storage medium for recommending packages for communication users. Background Art

[0002] In the prior art, there are three main ways to recommend packages to communication users. The first is random recommendation, when the operator formulates a new package, it recommends it to users without distinction; the second is to judge whether the user is willing to accept the new package based on the user's past package selection, and decide whether to recommend a package to the user based on the result; the third is to decide whether to recommend a new package to the user mainly based on the acceptance of the new package in the user's social circle.

[0003] The first recommendation method in the prior art does not take the actual situation of the user into consideration, so the recommendation effect is poor and resources are seriously wasted. The second recommendation method uses historical consumption habits for recommendation because time will affect consumption habits, and the accuracy is relatively low. The third recommendation method is easily affected by the consumption habits of members of the social circle. If the consumption habits of members of the social circle are inconsistent, it is easy to cause wrong recommendations. Summary of the invention

[0004] The present invention provides a communication user package recommendation method, device, electronic device and storage medium, which are used to solve the problem that the effect of communication user package recommendation in the prior art is not ideal, realize accurate recommendation of communication user packages, and improve the effect of package recommendation.

[0005] The present invention provides a method for recommending a communication user package, comprising:

[0006] Acquire user sample information, where the user sample information includes user description information and user feature information;

[0007] Constructing a user connection matrix and a user degree matrix according to the feature information, wherein the user connection matrix represents the connection between two users, and the user degree matrix represents the number of user node edges;

[0008] Performing topic extraction on the description information to obtain corresponding topic feature information, and calculating a user similarity matrix based on the topic feature information and the feature information, wherein the user similarity matrix represents the similarity between two users;

[0009] Calculating a second weight matrix according to the user connection matrix, the user degree matrix, and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle;

[0010] A user package model is constructed according to the feature information, and a package is recommended to a target user based on the user similarity matrix, the second weight matrix and the user package model.

[0011] According to a communication user package recommendation method provided by the present invention, the subject extraction of the description information to obtain corresponding subject feature information includes:

[0012] Segment the text in the description information, calculate the TF-IDF value of each word in the text, use the TF-IDF value as a feature row, and generate a TF-IDF feature matrix for multiple records;

[0013] The TF-IDF feature matrix is ​​weightedly decomposed using a non-negative matrix factorization method to obtain topic feature information.

[0014] According to a communication user package recommendation method provided by the present invention, the user similarity matrix is ​​calculated according to the topic feature information and the feature information, including:

[0015] Combining the subject feature information and the feature information into a new feature matrix;

[0016] According to the merged feature matrix, the cosine similarity is used to calculate the user similarity matrix.

[0017] According to a communication user package recommendation method provided by the present invention, the second weight matrix is ​​calculated according to the user connection matrix, the user degree matrix and the user similarity matrix, including:

[0018] According to the user connection matrix, a new connection matrix is ​​calculated based on a graphical neural network;

[0019] Calculating a first weight matrix according to the new connection matrix and the user degree matrix;

[0020] Based on the mean of the preset feature information and the user degree matrix, the first weight matrix is ​​optimized to obtain the second weight matrix.

[0021] According to a communication user package recommendation method provided by the present invention, the user package model is constructed according to the characteristic information, comprising:

[0022] Identify the user's consumption category based on the user's historical consumption packages;

[0023] According to the characteristic information of the user and the consumption category identifier, an extreme gradient boosting algorithm is used to construct the user package model;

[0024] The user package model outputs the user's consumption category identifier based on the input user's characteristic information.

[0025] According to a communication user package recommendation method provided by the present invention, the user package model is constructed according to the feature information, and a package is recommended to a target user based on the user similarity matrix, the second weight matrix and the user package model, including:

[0026] Determine the social circle of the target user based on the social circle divided by the graph neural network;

[0027] Determining the similarity between the target user and other members of the social circle according to the user similarity matrix;

[0028] The information of the first preset members with the highest similarity and the information of the target user are respectively input into the user package model, and the user package model outputs the corresponding consumption category identifiers of the preset members and the target user;

[0029] The members with the same consumption category identifier as the target user are retained, and the majority of packages are selected from the members and recommended to the target user.

[0030] According to a communication user package recommendation method provided by the present invention, constructing a user connection matrix and a user degree matrix according to the feature information includes:

[0031] Constructing the user connection matrix according to the interactive information of call duration, call number and location between users;

[0032] If the user connection matrix is ​​represented by a user connection graph, the nodes of the user connection graph represent users, the edges represent the connections between the nodes, and the number of user node edges represents the degree of the user.

[0033] The present invention also provides a communication user package recommendation device, comprising:

[0034] A user sample module is used to obtain user sample information, wherein the user sample information includes user description information and user feature information;

[0035] A feature information module, used to construct a user connection matrix and a user degree matrix according to the feature information, wherein the user connection matrix represents the connection between two users, and the user degree matrix represents the number of user node edges;

[0036] A topic extraction module, used to extract topics from the description information to obtain corresponding topic feature information, and calculate a user similarity matrix based on the topic feature information and the feature information, wherein the user similarity matrix represents the similarity between two users;

[0037] A weight calculation module, used to calculate a second weight matrix according to the user connection matrix, the user degree matrix and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle;

[0038] The package recommendation module is used to build a user package model according to the feature information, and recommend packages to target users based on the user similarity matrix, the second weight matrix and the user package model.

[0039] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-mentioned methods for recommending communication user packages are implemented.

[0040] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described methods for recommending communication user packages.

[0041] The present invention provides a communication user package recommendation method, device, electronic device and storage medium, which construct a user connection matrix and a user degree matrix for user sample information, extract topics from description information, calculate a user similarity matrix, optimize the user weight matrix to mine the user social circle, and realize accurate recommendation of communication user packages through the constructed user package model, which can effectively improve the effect of package recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 It is a flow chart of the communication user package recommendation method provided by the present invention;

[0044] Figure 2 is a schematic diagram of a user connection diagram provided by the present invention;

[0045] Figure 3 It is a flowchart of the subject extraction provided by the present invention;

[0046] Figure 4 It is a schematic diagram of a flow chart of calculating user similarity provided by the present invention;

[0047] Figure 5 It is a schematic diagram of the flow chart of calculating the user weight matrix provided by the present invention;

[0048] Figure 6 It is a schematic diagram of the process of building a user package model provided by the present invention;

[0049] Figure 7 It is a schematic diagram of the process of recommending a meal package provided by the present invention;

[0050] Figure 8 It is a structural schematic diagram of a communication user package recommendation device provided by the present invention;

[0051] Fig. 9 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.

[0054] The technical terms involved in the present invention are described below:

[0055] 1. Theme Analysis (Theme Extraction)

[0056] A topic can be regarded as a probability distribution of terms. An article is represented by a bag of words, which is long. After being mapped to the topic space, since the number of topics K is usually much smaller than the number of terms, the dimension can be reduced through the topic model. Topic analysis mainly consists of five components: input, model assumptions, representation, parameter estimation, and new sample inference.

[0057] (1) Input. The input of the topic model is a document set, which is equivalent to a term-document matrix due to the assumption of exchangeability. Another input is the number of topics K, which is usually determined empirically. The simplest method is to repeat the experiment with different K.

[0058] (2) Model assumptions. The bag of words assumption states that the order of words in a document is irrelevant to the model results. However, in the LDA derived model, some commutativity is broken.

[0059] (3) indicates that they are graphical model and generative model respectively. Note that there are two hyperparameters α and β.

[0060] (4) Parameter estimation: First, we need to select the objective function to be optimized, which is usually the probability value of the entire corpus, and then iteratively optimize the objective function.

[0061] (5) New sample inference: Map the original sample description text to a low-dimensional topic space.

[0062] The main models of topic analysis include singular value decomposition (SVD), latent semantic analysis (LSA), probabilistic latent semantic analysis (PLSA) and LDA (Latent Dirichlet Allocation, a document topic generation model).

[0063] 2. GCN (Graph Neural Network) Algorithm

[0064] The GNN model mainly studies the representation of graph nodes (Graph Embedding), graph edge structure prediction tasks and graph classification problems. The latter two tasks are also based on Graph Embedding. GCN has the following properties: hierarchical structure, nonlinear transformation and end-to-end training, as follows:

[0065] (1) Hierarchical structure. Features are extracted layer by layer, each layer is more abstract and advanced than the previous one.

[0066] (2) Nonlinear transformation: Increase the expressive power of the model.

[0067] (3) End-to-end training. There is no need to define any rules. You only need to label the nodes of the graph and let the model learn by itself, integrating feature information and structural information.

[0068] GCN has four characteristics: GCN is a natural extension of convolutional neural networks in the graph domain; it can simultaneously perform end-to-end learning of node feature information and structural information, and is currently the best choice for graph data learning tasks; graph convolution is extremely applicable and can be used for nodes and graphs of arbitrary topological structures; in tasks such as node classification and edge prediction, the effect is far superior to other methods on public datasets.

[0069] 3. NMF (Nonnegative Matrix Factorization) algorithm.

[0070] The basic idea of ​​NMF can be described as follows: for any given non-negative matrix A, the NMF algorithm can find a non-negative matrix U and a non-negative matrix V, decompose a non-negative matrix into the product of two non-negative matrices on the left and right, reduce the dimension of the matrix, and compress a large amount of data. NMF is an unsupervised learning algorithm, in which the restriction condition is that all elements in W and H must be greater than 0.

[0071] The traditional NMF problem can be described as follows:

[0072] Given a matrix Finding non-negative matrices and a non-negative matrix So that V≈WH.

[0073] Non-negative matrix factorization (NMF) decomposes the non-negative matrix V into two smaller non-negative matrices W and H, which are multiplied by each other. Its mathematical expression is:

[0074] V n*m ≈W n*r H r*m .

[0075] In the above formula, the reason why it is approximately equal is that the current solution is not an exact solution, but only a numerical approximate solution, where r is much smaller than n and m. In general, (n+m)r <nm。

[0076] Since the effects of randomly recommending packages, recommending packages based on the user's past package selections, and recommending packages based only on the user's social circle in the prior art are not ideal, in order to accurately recommend packages to users, the present invention provides a communication user package recommendation method, device, electronic device, and storage medium based on topic analysis and GCN social circle discovery. The present invention constructs user sample information and its related matrix based on business scenarios (such as 4G business scenarios, 5G business scenarios, etc.), optimizes the NMF decomposition algorithm and the GCN algorithm, and combines topic analysis, social circle discovery, and classification models (user package models) to recommend packages to users.

[0077] Combine the following Figure 1-Figure 9 The present invention describes a method, device, electronic device and storage medium for recommending packages for communication users.

[0078] Figure 1 The figure is a flow chart of a method for recommending a communication user package provided by the present invention. A method for recommending a communication user package includes:

[0079] Step 101: Obtain user sample information, where the user sample information includes user description information and user feature information.

[0080] For example, the user sample information of target user B may be obtained, or the user sample information of multiple target users may be obtained.

[0081] Optionally, the user description information is textual description information of the user, such as text or voice recording of the communication between the user and customer service personnel.

[0082] Optionally, the characteristic information of the user is information other than the description information of the user, and the user sample information is shown in Table 1.

[0083] Table 1

[0084]

[0085]

[0086] In the above table, DES feature is the user's descriptive information, and the other features are the user's characteristic information, such as user basic information, process, call and other features.

[0087] Step 102: construct a user connection matrix and a user degree matrix according to the feature information. The user connection matrix represents the connections between two users, and the user degree matrix represents the number of user node edges.

[0088] Optionally, the user connection matrix can be constructed based on the interactive information such as call duration, call number, and location between users. If the user connection matrix is ​​represented by a user connection graph, the nodes of the user connection graph represent users, the edges represent the connections between nodes, and the number of user node edges represents the degree of the user. Figure 2 As shown, it is represented by G; Figure 2 In it, it indicates whether there is a connection between users. The node represents the user and the edge represents the connection. If there is a connection between two users, they are connected by an edge; otherwise, they are not connected.

[0089] The user connection matrix is ​​shown in Table 2, denoted by A. Table 2 indicates whether there is a connection between users, which is called the user connection matrix A. The value 1 indicates that there is a connection; 0 indicates that there is no connection.

[0090] Table 2

[0091] User 1 User 2 User 3 User 4 User 1 0 1 1 1 User 2 1 0 1 0 User 3 1 1 0 0 User 4 1 0 0 0

[0092] The nodes of the user connection graph represent users, and the edges represent the connections between nodes. The connection matrix shows the connections between users. Table 2 shows that user 1 is connected to user 2, user 3, and user 4; user 2 is connected to user 1 and user 3; user 3 is connected to user 1 and user 2; and user 4 is only connected to user 1.

[0093] The degree of a user indicates the number of edges of the user node. The degree matrix of the user is shown in Table 3 and is represented by D. Table 3 represents the degree of the user, that is, how many users the user is connected to, which is called the user degree matrix D. Therefore, it is a symmetric matrix with non-zero main diagonal and zero in other positions.

[0094] Table 3

[0095] User 1 User 2 User 3 User 4 User 1 3 0 0 0 User 2 0 2 0 0 User 3 0 0 2 0 User 4 0 0 0 1

[0096] Table 3 shows that the degree of user 1's node is 3, the degree of user 2's node is 2, the degree of user 3's node is 2, and the degree of user 4's node is 1.

[0097] From this, we can see that according to actual business scenarios (such as 4G business, 5G business, etc.), based on the user sample information obtained by the system, the user connection matrix and user degree matrix of the package recommendation system are constructed. The selection of user features is highly correlated with the package recommendation business. The user's connection and degree are generated according to the business scenarios of package recommendation, which is in line with reality.

[0098] Step 103, extracting topics from the description information to obtain corresponding topic feature information, and calculating a user similarity matrix based on the topic feature information and the feature information, wherein the user similarity matrix represents the similarity between two users.

[0099] Optionally, the present invention adopts a weighted non-negative matrix factorization (NMF) decomposition method to perform topic extraction on the description information. The DES features in the above Table 1 are the user's description information, such as information described in text, and the topic features of the extracted text are used as the user's description features.

[0100] Optionally, by combining the extracted topic features with the user's own feature information, the similarity between two users regarding a certain topic feature can be obtained.

[0101] It can be seen that the present invention adopts a weighted NMF decomposition method to extract topics from the user's description information. Topic extraction can adopt a matrix decomposition strategy. Since the volume of mobile users on the network is large, a faster NMF algorithm is used to accelerate matrix decomposition and reduce decomposition time. In addition, the traditional NMF algorithm does not take into account the importance of different user samples. The present invention improves the NMF algorithm by constructing a sample weight matrix and introducing the sample weight matrix into the optimization function. On the one hand, iteration can obtain better weights, and on the other hand, the solution effect of the objective function will be better.

[0102] Furthermore, the present invention combines the extracted topic features with the user's own node features to calculate the user similarity matrix. The user similarity includes both the user's own topic attributes and the user's business attributes, and the features are more complete, the calculated similarity is more realistic, and the information is richer.

[0103] Step 104: Calculate a second weight matrix based on the user connection matrix, the user degree matrix, and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle.

[0104] Optionally, the user's social circle can be mined by optimizing the user weight matrix in a graph neural network (GCN) algorithm, and the user can be classified into the social circle to obtain the similarity between the user and the members of the social circle regarding a certain topic feature.

[0105] It can be seen that the present invention uses the optimized user weight matrix in the GCN algorithm to mine the user's social circle. The GCN algorithm is effective in mining social circles, but the traditional GCN algorithm uses the inverse of the degree as the weight value, which only considers whether there is a connection between users, but does not consider the closeness of the connection between users; the present invention optimizes the traditional GCN algorithm, and uses the modified weight matrix It reflects the closeness of the connection between users and further improves the quality of the generated social circle.

[0106] Step 105: construct a user package model according to the feature information, and recommend packages to target users based on the user similarity matrix, the second weight matrix and the user package model.

[0107] Optionally, an extreme gradient boosting (XGBOOST) algorithm may be used to construct a user package model, wherein the user package model is trained based on the input user's characteristic information and consumption category identifier, and then the consumption category identifier of the new user may be output after learning the user package model based on the input new user's characteristic information.

[0108] It can be seen that the present invention constructs a user's package replacement model based on the user's characteristic information and consumption category identifier. The user's package replacement model is used to evaluate the user's package consumption level and increase the user's package attributes.

[0109] Furthermore, based on the above-mentioned user similarity matrix, the present invention utilizes the social circles divided by the optimized GCN algorithm and the user package model to improve the effect of user package recommendation; and through the fusion of these three methods, the recommended packages are more robust, thereby improving the accuracy of package recommendations.

[0110] A specific embodiment will be provided below to describe the above steps 103 to 105.

[0111] Figure 3 The flowchart of the topic extraction provided by the present invention is shown in the figure. In the above step 103, the topic extraction of the description information to obtain the corresponding topic feature information includes:

[0112] Step 301, segment the text in the description information, calculate the TF-IDF value of each word in the text, use the TF-IDF value as a feature row, and generate a TF-IDF feature matrix for multiple records. ij Indicates that i represents the number of users and j represents the number of words.

[0113] TF-IDF (term frequency-inverse document frequency) is a commonly used weighting technique for information retrieval and data mining. TF stands for term frequency, and IDF stands for inverse document frequency.

[0114] Step 302: Use non-negative matrix factorization to perform weighted decomposition on the TF-IDF feature matrix to obtain topic feature information.

[0115] Specifically, since the user has many vocabulary features, it is necessary to reduce the dimension of the features. The features can be abstracted by extracting the themes of these feature words, thereby achieving both dimensionality reduction and compression of feature information. The present invention uses NMF (non-negative matrix factorization) method for topic extraction, and the formula is shown in (1):

[0116] U ij ≈P ik H kj ; (1)

[0117] Among them, U ij Represents the word vector of the jth word in the i-th text, P ik represents the probability correlation between the i-th text and the k-th topic, H kj represents the probability correlation between the jth word and the kth topic, U ij ≈P ik H kj It means decomposing the text word matrix into the product of the text topic matrix and the topic word matrix.

[0118] In the above, P ik is the desired topic matrix. The objective function of formula (1) is shown in formula (2):

[0119]

[0120] in, is the Frobenius norm, Indicates that the optimization algorithm is used to calculate P and H so that The value is the smallest.

[0121] Although formula (2) can be used for topic extraction, it cannot reflect the more important information of certain samples. Therefore, the present invention needs to optimize formula (2) by giving different weights to different samples.

[0122] The topic features extracted based on the present invention are mainly used to calculate the similarity of users. Therefore, the harmonic mean of the standardized samples, such as monthly call duration (thsc), monthly call times (thcs), monthly traffic (11) and monthly consumption (xf), can be used as the initial value of the sample weight. Therefore, the calculation formula of the sample weight is shown in (3):

[0123]

[0124] Among them, thsc represents the call duration, thcs represents the number of monthly calls, ll represents the monthly traffic, and xf represents the monthly consumption. It is expressed as the square root of the product of the four attribute values ​​of the i-th sample.

[0125] The sample weight matrix is ​​shown in formula (4):

[0126]

[0127] Where W_samp represents the weight matrix of the sample, the main diagonal is the sample weight, and the rest is 0. The objective function of formula (2) can be optimized by combining the sample weight, as shown in formula (5):

[0128]

[0129] in, The meanings of P and H in formula (1) and (2) are the same, indicating that the optimization algorithm is used to calculate P, H and W_samp so that Minimum.

[0130] The sample weight is used as a parameter and its iteration is as follows:

[0131] W_asmp=W_asmp+λΔW_asmp; (6)

[0132] Among them, ΔW_samp represents the gradient of the parameter weight (i.e., the derivative vector), and W_samp = W_samp-λΔW_samp is used to represent the update of the sample weight. The iteration of the weight helps to achieve better results in the final optimization goal.

[0133] The optimized objective function can quickly perform matrix decomposition and also incorporate the user's importance information into the decomposed topic features.

[0134] It can be seen that the present invention optimizes the NMF algorithm, constructs a weighted matrix, incorporates the importance of samples into the decomposed topics, and the obtained topic matrix is ​​more robust.

[0135] Figure 4 The flowchart of calculating user similarity provided by the present invention is shown in the figure. In the step 103, the user similarity matrix is ​​calculated according to the topic feature information and the feature information, including:

[0136] Step 401: merge the topic feature information and the feature information into a new feature matrix.

[0137] Specifically, the user's feature information is represented by F, the extracted topic feature is P, and the two are combined into a new feature matrix:

[0138] X = (F, P).

[0139] Step 402: Calculate the user similarity matrix using cosine similarity based on the merged feature matrix.

[0140] Specifically, the user similarity measurement uses cosine distance, as shown in formula (7):

[0141]

[0142] Among them, x i represents the feature vector of the i-th user, and the numerator x i x j is the vector product of two eigenvectors, the denominator |x i ||x j | is the product of the lengths of the moduli of the eigenvectors, It represents the vector product of the i-th sample vector and the j-th sample vector, that is, the distance between the two samples.

[0143] The user similarity matrix is ​​calculated based on the user similarity between each two users, as shown in formula (8):

[0144]

[0145] Formula (8) is used to combine the distances between sample vectors into a matrix. Cosine distance, also known as cosine similarity, uses the cosine value of the angle between two vectors in the vector space as a measure of the size of the difference between two individuals.

[0146] Figure 5 The flowchart of calculating the user weight matrix provided by the present invention is shown in the figure. In the step 104, the second weight matrix is ​​calculated according to the user connection matrix, the user degree matrix and the user similarity matrix, including:

[0147] Step 501: Calculate a new connection matrix based on the user connection matrix and the graphical neural network.

[0148] In the above step 102, the user connection matrix A and the user degree matrix D are calculated. In the above step 401, the user feature matrix X is calculated. According to the GCN principle, a new connection matrix is ​​calculated. The calculation is shown in formula (9):

[0149]

[0150] Among them, A represents the user connection matrix, I represents the identity matrix, and λI represents adding 1 to the main diagonal. The new user connection matrix has 1 added to the main diagonal.

[0151] Step 502: Calculate a first weight matrix according to the new connection matrix and the user degree matrix.

[0152] After calculating the new connection matrix, it is necessary to calculate the mean of the adjacent user features H and feature information The calculation formula is shown in (10)(11):

[0153]

[0154]

[0155] Where D represents the degree matrix, The matrix product of the inverse matrix of the degree matrix, the new connection matrix and the user feature matrix is ​​used to calculate the adjacent user features; It means that the degree matrix is ​​added with 1 on the main diagonal, which represents the number of nodes itself and connected nodes; For The main diagonal of is taken inversely, which is used to The features and mean are taken.

[0156] Considering the need to give different weights to different nodes, the traditional GCN algorithm is to multiply To assign different nodes, since there is also left multiplication So by assigning Modify formula (10) for both sides, as shown in formula (12):

[0157]

[0158] in, It means that different weights are assigned to different nodes, which are weighted adjacent user features.

[0159] From this, we can see that formula (12) simply uses the user's degree as the first weight matrix. However, in actual application scenarios, such as takeaway calls and telephone promotions, the degree of users will be abnormally high. Although the degree is relatively high, it is not considered that these people have greater influence. Therefore, it is not appropriate to use only the user's degree as the weight matrix.

[0160] Step 503: Optimize the first weight matrix based on the mean of the preset feature information and the user degree matrix to obtain the second weight matrix.

[0161] Therefore, the present invention needs to recalculate the user node weight matrix. The weight calculation can be based on the preset feature information of a specific user (such as calling features). The calling features, such as the calling call duration, the calling call times, the calling SMS number, etc., are standardized and averaged and represented by the matrix JH. The calculation of JH is shown in formula (13):

[0162]

[0163] Where, x: standardized calling feature, x j Indicates the standardized caller characteristics, such as the call duration, call times, and SMS number of the caller.

[0164] in, Indicates the average of the standardized calling characteristics.

[0165] It is understandable that the above-mentioned calling characteristics, such as calling call duration, calling number, and calling text message quantity, are only examples of the present invention, and the present invention is not limited to these characteristics.

[0166] The product of the obtained mean and degree pair is used as the user's final weight Therefore, the optimized second weight matrix is ​​shown in formula (14):

[0167]

[0168] in, Express The correction is done by multiplying and adding feature information.

[0169] Therefore, the present invention optimizes formula (12) to formula (15):

[0170]

[0171] in, Indicates that after the correction Modify the weighted adjacent user features.

[0172] The optimized H reflects the real importance of the user in the social circle. The node information and edge information of the graph data structure can be converted into features that can be used by the neural network. Then, a two-layer GCN is used to calculate the social circle of each user node. The calculation formula is shown in (16):

[0173]

[0174] in, Indicates that the sofmax method is used to calculate the probability distribution of the user's community.

[0175] It can be seen that the present invention optimizes the user weight matrix in the GCN algorithm, that is, optimizes the first weight matrix into the second weight matrix, and the second weight matrix reflects the closeness of the connection between users. By mining social circles through the optimized second weight matrix and summarizing users into social circles, the quality of the generated social circles can be further improved.

[0176] Figure 6 The flowchart of constructing a user package model provided by the present invention is shown in the figure. In the above-mentioned step 105, the user package model is constructed according to the feature information, including:

[0177] Step 601, identify the consumption category of the user according to the user's historical consumption packages.

[0178] For example, based on the user's historical consumption packages, the user is labeled to identify the user's consumption category. For example, the consumption category identification can be divided into low-consumption users, medium-consumption users, and high-consumption users.

[0179] Step 602: construct the user package model using the extreme gradient boosting (eXtremeGradient Boosting, XGBoost) algorithm according to the characteristic information of the user and the consumption category identifier.

[0180] Optionally, characteristic information of the user, such as historical call charges, traffic flow, recharge and other characteristic information of the user.

[0181] XGBoost is a gradient boosting algorithm and residual decision tree. Its basic idea is: one tree, one tree is gradually added to the model. Every time a CRAT decision tree is added, the overall effect (the objective function decreases) should be improved. Multiple decision trees (multiple single weak classifiers) are used to form a combined classifier, and a certain weight is assigned to each leaf node. The decision tree is a classification and regression tree CRAT (classification and regression tree, CART). The CART decision tree is a binary tree. The values ​​of the internal node features are "yes" and "no". The left branch is the branch with the value of "yes", and the right branch is the branch with the value of "no". CART will distribute the input to each leaf node according to the attributes, and each leaf node will correspond to a score value (weight).

[0182] Step 603: the user package model outputs the user's consumption category identifier based on the input user's characteristic information.

[0183] For example, the characteristic information of user A (such as the user's historical call charges, traffic, recharge, etc.) is input, and the user package model outputs a consumption category identifier of a low-consumption user, a medium-consumption user, or a high-consumption user.

[0184] It can be seen that a user package model is constructed based on the user's characteristic information and consumption category identifier. The user package model is used to evaluate the user's package consumption level and increase the user's package attributes.

[0185] Figure 7 The flowchart of the package recommendation provided by the present invention is shown in the figure. In the above step 105, the user package model is constructed according to the feature information, and the package recommendation is made to the target user based on the user similarity matrix, the second weight matrix and the user package model, including:

[0186] Step 701, determining the social circle of the target user based on the social circle divided by the graph neural network.

[0187] Specifically, the social circle divided by the GCN algorithm optimized in step 503 is used to find the social circle where the target user B is located. Assume that the social circle where the target user B is located is D1, and the members of D1 include user_m1, user_m2, ... user_mn, n members.

[0188] Step 702: Determine the similarity between the target user and other members of the social circle based on the user similarity matrix.

[0189] Specifically, according to the user similarity matrix, the similarity between the target user B and the n members is found, and the result is: cor_1, cor_2, ... cor_n.

[0190] Step 703, taking the information of the first preset members with the highest similarity and the information of the target user and inputting them into the user package model respectively, and the user package model outputs the corresponding consumption category identifiers of the preset members and the target user.

[0191] Specifically, the first k members with the highest similarity are selected, namely user_mk1, user_mk2, ... user_mkk, and the features of the target user B and the k members are input into the user package model, and the user's consumption level is determined by the user package model.

[0192] Step 704: retain members with the same consumption category identifier as the target user, and select the majority of packages from among the members to recommend to the target user.

[0193] Specifically, members with the same consumption level as target user B are retained, and then the majority of packages are selected from these members and recommended to target user B.

[0194] It can be seen that the present invention improves the effect of user package recommendation based on the above-mentioned user similarity matrix, the social circle divided by the optimized GCN algorithm and the user package model; and through the fusion of these three methods, the recommended packages are more robust, which improves the accuracy of package recommendations.

[0195] The communication user package recommendation device provided by the present invention is described below. The communication user package recommendation device described below and the communication user package recommendation method described above can be referenced to each other.

[0196] Figure 8 The structure diagram of the communication user package recommendation device provided by the present invention is shown in the figure. A communication user package recommendation device 800 includes a user sample module 810, a feature information module 820, a theme extraction module 830, a weight calculation module 840 and a package recommendation module 850. Among them,

[0197] The user sample module 810 is used to obtain user sample information, where the user sample information includes user description information and user feature information.

[0198] The feature information module 820 is used to construct a user connection matrix and a user degree matrix according to the feature information. The user connection matrix represents the connection between two users, and the user degree matrix represents the number of user node edges.

[0199] The topic extraction module 830 is used to extract topics from the description information to obtain corresponding topic feature information, and calculate a user similarity matrix based on the topic feature information and the feature information. The user similarity matrix represents the similarity between two users.

[0200] The weight calculation module 840 is used to calculate a second weight matrix according to the user connection matrix, the user degree matrix and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle.

[0201] The package recommendation module 850 is used to construct a user package model according to the feature information, and recommend packages to target users based on the user similarity matrix, the second weight matrix and the user package model.

[0202] Optionally, the feature information module 820 is further configured to perform the following steps:

[0203] Constructing the user connection matrix according to the interactive information of call duration, call number and location between users;

[0204] If the user connection matrix is ​​represented by a user connection graph, the nodes of the user connection graph represent users, the edges represent the connections between the nodes, and the number of user node edges represents the degree of the user.

[0205] The topic extraction module 830 is further configured to perform the following steps:

[0206] Segment the text in the description information, calculate the TF-IDF value of each word in the text, use the TF-IDF value as a feature row, and generate a TF-IDF feature matrix for multiple records;

[0207] The TF-IDF feature matrix is ​​weightedly decomposed using a non-negative matrix factorization method to obtain topic feature information.

[0208] The topic extraction module 830 is further configured to perform the following steps:

[0209] Combining the subject feature information and the feature information into a new feature matrix;

[0210] According to the merged feature matrix, the cosine similarity is used to calculate the user similarity matrix.

[0211] The weight calculation module 840 is further configured to perform the following steps:

[0212] According to the user connection matrix, a new connection matrix is ​​calculated based on a graphical neural network;

[0213] Calculating a first weight matrix according to the new connection matrix and the user degree matrix;

[0214] Based on the mean of the preset feature information and the user degree matrix, the first weight matrix is ​​optimized to obtain the second weight matrix.

[0215] The package recommendation module 850 is further configured to perform the following steps:

[0216] Identify the user's consumption category based on the user's historical consumption packages;

[0217] According to the characteristic information of the user and the consumption category identifier, an extreme gradient boosting algorithm is used to construct the user package model;

[0218] The user package model outputs the user's consumption category identifier based on the input user's characteristic information.

[0219] The package recommendation module 850 is further configured to perform the following steps:

[0220] Determine the social circle of the target user based on the social circle divided by the graph neural network;

[0221] Determining the similarity between the target user and other members of the social circle according to the user similarity matrix;

[0222] The information of the first preset members with the highest similarity and the information of the target user are respectively input into the user package model, and the user package model outputs the corresponding consumption category identifiers of the preset members and the target user;

[0223] The members with the same consumption category identifier as the target user are retained, and the majority of packages are selected from the members and recommended to the target user.

[0224] Fig. 9 An example of a physical structure diagram of an electronic device is shown in FIG. Fig. 9 As shown, the electronic device may include: a processor 910, a communication interface 920, a memory 930 and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other through the communication bus 940. The processor 910 may call the logic instructions in the memory 930 to execute the above-mentioned communication user package recommendation method, and the method includes:

[0225] Acquire user sample information, where the user sample information includes user description information and user feature information;

[0226] Constructing a user connection matrix and a user degree matrix according to the feature information, wherein the user connection matrix represents the connection between two users, and the user degree matrix represents the number of user node edges;

[0227] Performing topic extraction on the description information to obtain corresponding topic feature information, and calculating a user similarity matrix based on the topic feature information and the feature information, wherein the user similarity matrix represents the similarity between two users;

[0228] Calculating a second weight matrix according to the user connection matrix, the user degree matrix, and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle;

[0229] A user package model is constructed according to the feature information, and a package is recommended to a target user based on the user similarity matrix, the second weight matrix and the user package model.

[0230] In addition, the logic instructions in the above-mentioned memory 930 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0231] On the other hand, the present invention further provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, when the program instructions are executed by a computer, the computer can execute the communication user package recommendation method provided by the above methods, the method comprising:

[0232] Acquire user sample information, where the user sample information includes user description information and user feature information;

[0233] Constructing a user connection matrix and a user degree matrix according to the feature information, wherein the user connection matrix represents the connection between two users, and the user degree matrix represents the number of user node edges;

[0234] Performing topic extraction on the description information to obtain corresponding topic feature information, and calculating a user similarity matrix based on the topic feature information and the feature information, wherein the user similarity matrix represents the similarity between two users;

[0235] Calculating a second weight matrix according to the user connection matrix, the user degree matrix, and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle;

[0236] A user package model is constructed according to the feature information, and a package is recommended to a target user based on the user similarity matrix, the second weight matrix and the user package model.

[0237] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform the communication user package recommendation method provided above, the method comprising:

[0238] Acquire user sample information, where the user sample information includes user description information and user feature information;

[0239] Constructing a user connection matrix and a user degree matrix according to the feature information, wherein the user connection matrix represents the connection between two users, and the user degree matrix represents the number of user node edges;

[0240] Performing topic extraction on the description information to obtain corresponding topic feature information, and calculating a user similarity matrix based on the topic feature information and the feature information, wherein the user similarity matrix represents the similarity between two users;

[0241] Calculating a second weight matrix according to the user connection matrix, the user degree matrix, and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle;

[0242] A user package model is constructed according to the feature information, and a package is recommended to a target user based on the user similarity matrix, the second weight matrix and the user package model.

[0243] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0244] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0245] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for recommending packages for communication users, characterized in that: include: Acquire user sample information, where the user sample information includes user description information and user feature information; Constructing a user connection matrix and a user degree matrix according to the characteristic information of the user, wherein the user connection matrix represents the connection between two users, and the user degree matrix represents the number of users connected with the user; Performing topic extraction on the description information of the user to obtain corresponding topic feature information, and calculating a user similarity matrix based on the topic feature information and the feature information of the user, wherein the user similarity matrix represents the similarity between two users; Calculating a second weight matrix according to the user connection matrix, the user degree matrix, and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle; Constructing a user package model according to the characteristic information of the user, wherein the user package model is used to output a consumption category identifier of the user according to the characteristic information of the input user; Using the second weight matrix as a user weight matrix in a graph neural network, and determining the social circle of the target user based on the social circles divided by the graph neural network; Determining the similarity between the target user and other members of the social circle according to the user similarity matrix; The information of the first preset members with the highest similarity and the information of the target user are respectively input into the user package model, and the user package model outputs the corresponding consumption category identifiers of the preset members and the target user; The members with the same consumption category identifier as the target user are retained, and the majority of packages are selected from the members and recommended to the target user.

2. The method for recommending packages for communication users according to claim 1, characterized in that: The subject extraction of the user's description information to obtain corresponding subject feature information includes: Segment the text in the user's description information, calculate the TF-IDF value of each word in the text, use the TF-IDF value as a feature row, and generate a TF-IDF feature matrix for multiple records; The TF-IDF feature matrix is ​​weightedly decomposed using a non-negative matrix factorization method to obtain topic feature information.

3. The method for recommending packages for communication users according to claim 1, characterized in that: The calculating the user similarity matrix according to the topic feature information and the user feature information includes: Combining the subject feature information and the user feature information into a new feature matrix; According to the merged feature matrix, the cosine similarity is used to calculate the user similarity matrix.

4. The method for recommending packages for communication users according to claim 1, characterized in that: The calculating a second weight matrix according to the user connection matrix, the user degree matrix and the user similarity matrix comprises: According to the user connection matrix, a new connection matrix is ​​calculated based on a graphical neural network; Calculating a first weight matrix according to the new connection matrix and the user degree matrix; Based on the mean of the preset feature information and the user degree matrix, the first weight matrix is ​​optimized to obtain the second weight matrix.

5. The method for recommending packages for communication users according to claim 1, characterized in that: The step of constructing a user package model according to the characteristic information of the user includes: Identify the user's consumption category based on the user's historical consumption packages; According to the characteristic information of the user and the consumption category identifier, the user package model is constructed using an extreme gradient boosting algorithm.

6. The method for recommending packages for communication users according to claim 1, characterized in that: The step of constructing a user connection matrix and a user degree matrix according to the characteristic information of the user comprises: Constructing the user connection matrix according to the interactive information of call duration, call number and location between users; If the user connection matrix is ​​represented by a user connection graph, the nodes of the user connection graph represent users, the edges represent the connections between the nodes, and the number of user node edges represents the degree of the user.

7. A communication user package recommendation device, characterized in that: include: A user sample module is used to obtain user sample information, wherein the user sample information includes user description information and user feature information; A feature information module, used to construct a user connection matrix and a user degree matrix according to the feature information of the user, wherein the user connection matrix represents the connection between two users, and the user degree matrix represents the number of users connected with the user; A topic extraction module is used to extract topics from the description information of the user to obtain corresponding topic feature information, and calculate a user similarity matrix based on the topic feature information and the feature information of the user, wherein the user similarity matrix represents the similarity between two users; A weight calculation module, used to calculate a second weight matrix according to the user connection matrix, the user degree matrix and the user similarity matrix, wherein the second weight matrix represents the importance of the user in the social circle; A package recommendation module is used to build a user package model according to the characteristic information of the user, and the user package model is used to output the user's consumption category identifier according to the input user's characteristic information; the second weight matrix is ​​used as the user weight matrix in the graph neural network, and the social circle where the target user is located is determined based on the social circle divided by the graph neural network; the similarity between the target user and other members of the social circle is determined according to the user similarity matrix; the information of the first preset members with the highest similarity and the information of the target user are respectively input into the user package model, and the user package model outputs the corresponding consumption category identifiers of the preset members and the target user; members with the same consumption category identifier as the target user are retained, and the majority of packages are selected from the members to be recommended to the target user.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the communication user package recommendation method as described in any one of claims 1 to 6 are implemented.

9. A non-transitory 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 communication user package recommendation method as claimed in any one of claims 1 to 6 are implemented.

Images