CN111611472A - A method and system for bundle recommendation based on graph convolutional neural network - Google Patents

Abstract

Embodiments of the present invention provide a method and system for bundling recommendation based on a graph convolutional neural network. The method includes: acquiring historical interaction data between users and bundling, historical interaction data between users and items, and affiliation data between bundling and items; and then inputting them into a bundling recommendation model to obtain a recommendation result of user-bundling interaction possibility output by the bundling recommendation model; wherein The bundled recommendation model constructs a unified heterogeneous graph based on the user data set, the bundled data set and the item data set, extracts the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature, and then performs joint prediction and feature connection. It is obtained by training the hard negative sample learning strategy. The embodiment of the present invention reconstructs the interaction and affiliation between users, bundles and items into graphs, and uses the powerful ability of graph neural networks to learn three kinds of associated entity representations from complex topological structures and high-order connectivity, can achieve better recommendation performance.

Description

A method and system for bundle recommendation based on graph convolutional neural network

technical field

The present invention relates to the technical field of bundling recommendation, and in particular, to a method and system for bundling recommendation based on a graph convolutional neural network.

Background technique

The definition of bundled recommendation is to recommend a set of bundled items for the user’s overall consumption. The prevalence of bundled items on e-commerce and content platforms makes bundled recommendation an important task, which can not only avoid users’ monotonous choices and improve users’ experience. experience, but also increase business sales by expanding order size. Since a bundle consists of multiple items, the attractiveness of the bundle depends on the items within the bundle, and the attractiveness of each item within the bundle is affected by the other items displayed together in the bundle. Additionally, users need to be satisfied with most items in the bundle, which means that the interaction between the user and the bundle is more sparse.

While bundling strategies are currently widely used in various places, little work has been done to solve the bundle recommendation problem. Most existing works treat item recommendation and bundle recommendation as two relatively independent tasks and associate them through shared model parameters or a multi-task framework. List Recommender Model (LIRE) and Embedding Factorization Machine (EFM) utilize both user interaction with items and bundled items under the BPR framework. The bundled BPR model (BBPR) leverages the parameters previously learned by the item BPR model. Deep Attention Multi-Task Model (DAM) jointly models user-bundle interaction and user-item interaction in a multi-task manner, transferring the benefits of item recommendation task to bundling recommendation to alleviate the scarcity of user-bundle interaction.

Currently, the following methods are usually used:

Option 1: First determine the corresponding co-selection score of each item pairing among the multiple items, the co-selection score indicates the probability that both items in the item pair are selected by the user, and determine the co-selection map among the multiple items based on the co-selection score. . Each node represents an item, and each edge is associated with a co-selection score, transforming the item bundling problem into a maximum complete N subgraph optimization problem.

Option 2: Calculate the similarity between other items and the current item based on the category or own attributes of the item that the user is currently interested in, and use the most similar item as a supplement to the current item as a bundle, and then supplement other items similar to the bundled items one by one. As a new bundle, multiple item bundles are ranked, and one item bundle is selected from multiple item bundles to recommend to users.

It is not difficult to find that the above schemes have the following limitations: (1) The method of parameter sharing does not explicitly model the relationship between users, items and bundles, and the combination at the loss level in the multi-task framework is difficult to balance the main task and auxiliary weights between tasks; (2) Existing works only consider correlations between items in bundles to learn better item representations to enhance item recommendation tasks. However, when bundling is a recommendation target, the similarity between them is more important and ignored; (3) When users are tied to interactions, their decision-making psychology has not been considered. At the item level, even if the user likes most of the items in the bundle, the user may reject the bundle due to one disliked item. At the bundle level, for two highly similar bundles, the key to influencing the final choice of users is their non-overlapping parts.

SUMMARY OF THE INVENTION

Embodiments of the present invention provide a method and system for bundling recommendation based on a graph convolutional neural network, so as to solve the problem that the weight relationship between users, bundling and items cannot be balanced when performing bundling recommendation in the prior art, and the corresponding decision reference relationship is not enough Comprehensive, resulting in inaccurate and objective recommendation results.

In a first aspect, an embodiment of the present invention provides a method for bundling recommendation based on a graph convolutional neural network, including:

Obtain the historical interaction data between the user and the bundle, the historical interaction data between the user and the item, and the affiliation data between the bundle and the item;

Input the user and bundling historical interaction data, the user and item historical interaction data, and the bundling and item affiliation data into the pre-trained bundling recommendation model to obtain the user-bundling interaction output by the bundling recommendation model Likelihood recommendation results; wherein the bundled recommendation model constructs a unified heterogeneous graph based on the user-bundling interaction data set, the bundle-item interaction data set, and the user-item interaction data set, and extracts the convolution propagation features of the item-level graph and the bundle-level graph After convolutional propagation of features, joint prediction and feature connection are performed, and they are obtained by training a learning strategy based on hard negative samples.

Preferably, the bundled recommendation model is obtained through the following steps:

Obtain the user-bundle interaction data set, the bundle-item interaction data set, and the user-item interaction data set, based on the user-bundle interaction data set, the bundle-item interaction data set, and the user constructing the unified heterogeneous graph with the item interaction data set;

extracting the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature based on the unified heterogeneous graph;

The item-level graph convolutional propagation feature and the bundled-level graph convolutional propagation feature are embedded and connected in all layers to obtain a joint prediction expression of the item propagation perspective and the bundled propagation perspective;

The joint prediction expression is trained by using the hard-negative sample learning strategy based on the bundling scenario to obtain the bundling recommendation model.

Preferably, the acquisition of the user-bundling interaction data set, the binding-item interaction data set, and the user-item interaction data set is based on the user-bundling interaction data set, the binding-item interaction data The unified heterogeneous graph is constructed by the collection and the user-item interaction data collection, which specifically includes:

Acquire a number of user information, a number of bundle information, and a number of item information, and define the interaction data of the several user information and the several bundle information as the user and bundle interaction data set, the several bundle information and the several item information. The affiliation is the bundle and the item interaction data collection and the interaction data between the several user information and the several item information is the user and item interaction data collection;

An undirected graph is used to represent the user-bundle interaction data set, the bundle-item interaction data set, and the user-item interaction data set; wherein the undirected graph includes nodes and edges, and the nodes include user nodes , a bundle node and an item node, the edges include user-bundle interaction edges, user-item interaction edges, and bundle-item dependent edges;

The inputs to the user node, the bundle node, and the item node are encoded using one-hot encoding and compressed into a dense real-valued vector.

Preferably, extracting the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature based on the unified heterogeneous graph specifically includes:

Constructing embedding propagation between users and items based on the dense real-valued vector, obtaining an item-level embedding update rule, and obtaining the item-level graph convolution propagation feature from the item-level embedding update rule;

Embedding propagation between bundles and users is constructed based on the dense real-valued vector, and a bundle-level embedding update rule is obtained, and the bundle-level graph convolution propagation feature is obtained from the bundle-level embedding update rule.

Preferably, the item-level graph convolutional propagation feature and the bundled-level graph convolutional propagation feature are embedded in all layers to obtain a joint prediction expression of the item propagation perspective and the bundle propagation perspective, specifically including:

Perform several graph convolution propagations on the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature to obtain several user embedding vectors and several bundled embedding vectors;

The joint prediction expression is obtained by embedding and combining the several user embedding vectors and all layers of the several bundled embedding vectors according to a preset operation mode.

Preferably, the joint prediction expression is trained by adopting a learning strategy based on difficult and negative samples in a bundling scenario to obtain the bundled recommendation model, which specifically includes:

Define observed user binding interaction data and unobserved user binding interaction data based on the joint prediction expression, and construct paired training data with negative sampling based on the observed user binding interaction data and the unobserved user binding interaction data;

The paired training data is introduced with a preset objective function as a model training target and a preset probability, and the training is performed based on the hard-negative sample learning strategy to obtain the bundled recommendation model.

Preferably, the training of the bundled recommendation model further includes setting several model hyperparameters.

In a second aspect, an embodiment of the present invention provides a bundled recommendation system based on a graph convolutional neural network, including:

The acquisition module is used to acquire the historical interaction data between the user and the bundle, the historical interaction data between the user and the item, and the affiliation data between the bundle and the item;

A processing module, configured to input the user and bundling historical interaction data, the user and item historical interaction data and the bundling and item affiliation data into the pre-trained bundling recommendation model, and obtain the output of the bundling recommendation model The recommendation result of user-bundling interaction possibility based on user-bundling interaction data set, bundle-item interaction data set, and user-item interaction data set constructs a unified heterogeneous graph, extracts item-level graph convolution propagation The features and bundled hierarchical graph convolutions propagate the features after joint prediction and feature connection, and are obtained by training based on the hard negative sample learning strategy.

In a third aspect, an embodiment of the present invention provides an electronic device, including:

A memory, a processor, and a computer program stored on the memory and executable on the processor, when the processor executes the program, the processor implements any one of the steps of the graph convolutional neural network-based bundle recommendation method.

In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements any one of the graph convolutional neural network-based bundle recommendations steps of the method.

The method and system for bundling recommendation based on a graph convolutional neural network provided by the embodiments of the present invention reconstruct the interaction and subordination relationships among users, bundles and items into graphs, and utilize the powerful capabilities of graph neural networks from complex Learning representations of three associated entities in topology and higher-order connectivity can achieve better recommendation performance.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are For some embodiments of the present invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

1 is a flowchart of a method for bundling recommendation based on a graph convolutional neural network provided by an embodiment of the present invention;

FIG. 2 is a flowchart of a bundling recommendation system provided by an embodiment of the present invention;

FIG. 3 is a schematic diagram of the definition of a bundled recommendation problem provided by an embodiment of the present invention;

4 is a schematic diagram of a construction method of a unified heterogeneous graph including users, items, and bundles provided by an embodiment of the present invention;

5 is a schematic diagram of a process of feature extraction based on item-level graph convolution propagation provided by an embodiment of the present invention;

6 is a schematic diagram of a process of feature extraction based on bundled hierarchical graph convolution propagation provided by an embodiment of the present invention;

FIG. 7 is a schematic diagram of high-order connectivity in a bundling recommendation scenario provided by an embodiment of the present invention;

8 is a schematic diagram of a joint prediction process based on two perspectives provided by an embodiment of the present invention;

9 is a schematic diagram of a difficult-to-negative sampling process in a bundling scenario provided by an embodiment of the present invention;

FIG. 10 is a flowchart of an implementation case of a bundled recommendation system provided by an embodiment of the present invention;

11 is a structural diagram of a bundled recommendation system based on a graph convolutional neural network provided by an embodiment of the present invention;

FIG. 12 is a structural block diagram of an electronic device provided by an embodiment of the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In order to overcome the limitations of the prior art, the goal of the embodiments of the present invention is to recommend bundles that the user may interact with based on the user's historical interactions with items/bundles and the composition information of bundles and items. Considering the bundling recommendation problem in scenarios such as e-commerce platforms or content platforms, a bundling recommendation method based on graph convolutional neural networks is proposed. Interactions and affiliations between bundles and commodities. Based on two-level graph convolution propagation on the constructed heterogeneous graph, the collaborative filtering signals between users and items/bundles, the semantics of bundling, and bundling are captured by distinguishing the affiliation between item nodes and bundle nodes. similarity between. Considering the cautious attitude of users in choosing bundles, hard negative sampling further encodes fine-grained representations of users and bundles.

FIG. 1 is a flowchart of a method for bundling recommendation based on a graph convolutional neural network provided by an embodiment of the present invention, as shown in FIG. 1 , including:

S1, obtain the historical interaction data between the user and the bundling, the historical interaction data between the user and the item, and the affiliation data between the bundling and the item;

S2: Input the historical interaction data between the user and the bundling, the historical interaction data between the user and the item, and the affiliation data between the bundling and the item into the pre-trained bundling recommendation model, and obtain the user and the bundling recommendation model output. Binding interaction possibility recommendation result; wherein the bundling recommendation model constructs a unified heterogeneous graph based on the user-bundling interaction data set, the bundling-item interaction data set, and the user-item interaction data set, and extracts the convolution propagation features of the item-level graph and the bundling After the hierarchical graph convolution propagates features, joint prediction and feature connection are performed, and they are obtained by training the learning strategy based on hard negative samples.

B和I表示用户，捆绑和物品的集合，并将用户-捆绑交互矩阵，用户-物品交互矩阵和捆绑-物品从属关系矩阵定义为X_M×N＝{x_ub|u∈U,b∈B}，Y_M×O＝{y_ui|u∈U,i∈I}，以及Z_N×O＝{z_bi|b∈B,i∈I}，每个条目都是一个二进制值。观察到的交互x_ub表示用户u曾经与捆绑b交互，观察到的交互y_ui意味着用户u曾经与物品i交互，如图3所示。类似地，条目z_bi＝1意味着捆绑b包含物品i。M，N和O分别表示用户数，捆绑数和商品数。Specifically, by first obtaining the data stored on the platform, the user's historical interaction records of items and bundling and the composition information of the bundling can be obtained, which are reflected in the user-bundling historical interaction data, the user-item historical interaction data, and the bundling-item affiliation. data, as shown in Figure 2. Here, in order to integrate item-level information to improve the accuracy of bundle recommendation, three important pieces of information need to be modeled, using

B and I represent the set of users, bundles and items, and define the user-bundle interaction matrix, user-item interaction matrix and bundle-item affiliation matrix as X _M×N = {x _ub |u∈U,b∈B }, Y _M×O = {y _ui |u∈U,i∈I}, and Z _N×O ={z _bi |b∈B,i∈I}, each entry is a binary value. The observed interaction x _ub means that user u has ever interacted with bundle b, and the observed interaction y _ui means that user u has ever interacted with item i, as shown in Figure 3. Similarly, the entry _zbi = 1 means that bundle b contains item i. M, N and O represent the number of users, the number of bundles and the number of items, respectively.

Therefore, the problem of bundling recommendation for users in the platform can be expressed as the following form, input: user-bundling history interaction X _M×N , user-item historical interaction Y _M×O and bundling-item affiliation Z _N×O ; output : bundling recommendation model, which is used to estimate the possibility of user u interacting with bundle b, i.e. the user-bundling interaction possibility recommendation result.

The embodiment of the present invention reconstructs the interaction and affiliation between users, bundles and items into graphs, and utilizes the powerful ability of graph neural networks to learn the representations of three associated entities from complex topological structures and higher-order connectivity , can achieve better recommendation performance.

Based on the above embodiment, the bundled recommendation model is obtained through the following steps:

Obtain the user-bundle interaction data set, the bundle-item interaction data set, and the user-item interaction data set, based on the user-bundle interaction data set, the bundle-item interaction data set, and the user constructing the unified heterogeneous graph with the item interaction data set;

extracting the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature based on the unified heterogeneous graph;

The item-level graph convolutional propagation feature and the bundled-level graph convolutional propagation feature are embedded and connected in all layers to obtain a joint prediction expression of the item propagation perspective and the bundled propagation perspective;

The joint prediction expression is trained by using the hard-negative sample learning strategy based on the bundling scenario to obtain the bundling recommendation model.

Specifically, to establish a bundled recommendation model, the first step is to uniformly model users, items, and complex relationships of bundles to construct a unified heterogeneous graph, then to extract features based on the item level graph convolution propagation, and to perform graph convolution propagation based on the bundle level. Extract features, then make predictions based on the two propagation perspectives of items and bundles, and finally perform model training based on difficult negative sampling in bundled scenarios.

Based on any of the foregoing embodiments, the acquiring the user-bundling interaction data set, the binding-item interaction data set, and the user-item interaction data set, based on the user-bundling interaction data set, the binding The unified heterogeneous graph is constructed by the interaction data set with the item and the user interaction data set with the item, which specifically includes:

Acquire a number of user information, a number of bundle information, and a number of item information, and define the interaction data of the several user information and the several bundle information as the user and bundle interaction data set, the several bundle information and the several item information. The affiliation is the bundle and the item interaction data collection and the interaction data between the several user information and the several item information is the user and item interaction data collection;

An undirected graph is used to represent the user-bundle interaction data set, the bundle-item interaction data set, and the user-item interaction data set; wherein the undirected graph includes nodes and edges, and the nodes include user nodes , a bundle node and an item node, the edges include user-bundle interaction edges, user-item interaction edges, and bundle-item dependent edges;

The inputs to the user node, the bundle node, and the item node are encoded using one-hot encoding and compressed into a dense real-valued vector.

由用户节点

捆绑节点b∈B和物品节点i∈I组成，边E是由对应x_ub＝1的用户-捆绑交互边缘(u,b)，对应y_ui＝1的用户-物品交互边缘(u,i)和对应z_bi＝1的捆绑-物品从属边缘(b,i)组成。用于统一建模用户，物品和捆绑复杂关系的异构图构造过程如图4所示。Specifically, to explicitly model the relationship between users, bundles and items, a unified heterogeneous graph is first constructed. Interaction and affiliation data can be represented by an undirected graph G=(V,E), where nodes

by user node

The bundle node b∈B and the item node i∈I are composed, and the edge E is composed of the user-bundle interaction edge (u,b) corresponding to x _ub =1, and the user-item interaction edge (u,i) corresponding to y _ui =1 and the bundle-item dependent edge (b, i) corresponding to z _bi = 1. The heterogeneous graph construction process for unified modeling of complex relationships of users, items, and bundles is shown in Figure 4.

For users, items, and bundle nodes on the construction graph, the input is further encoded with one-hot encoding, which is then compressed into a dense real-valued vector as follows:

表示用户u，物品i和捆绑b的独热特征向量。P，Q和R分别表示用户嵌入，物品嵌入和捆绑嵌入的矩阵。in

One-hot feature vector representing user u, item i and bundle b. P, Q and R denote the matrices of user embedding, item embedding and bundle embedding, respectively.

Based on any of the foregoing embodiments, the extraction of the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature based on the unified heterogeneous graph specifically includes:

Constructing embedding propagation between users and items based on the dense real-valued vector, obtaining an item-level embedding update rule, and obtaining the item-level graph convolution propagation feature from the item-level embedding update rule;

Embedding propagation between bundles and users is constructed based on the dense real-valued vector, and a bundle-level embedding update rule is obtained, and the bundle-level graph convolution propagation feature is obtained from the bundle-level embedding update rule.

Specifically, on the one hand, the user's preference for the items in the bundle can attract the user's attention and interest in the bundle. Since bundled items are often carefully designed, they are often functionally compatible with each other and constitute some semantics to influence the user's choice context. For example, bundles with mattresses and bed frames reflect what bedroom home means, while bundles with suits and ties reflect workplace attire.

To capture the user's preference for items and the features of the item itself, an embedding propagation between users and items is first constructed. Then, information pooling from items to bundles can capture the semantic information of bundles from the item level. The propagation and pooling-based embedding update rules for user u, item i and bundle b can be formulated as:

分别表示用户u，物品i和捆绑b的邻居。聚合函数aggregate可以是诸如简单均值函数，带采样的均值函数，最大池化等函数等。通过这个特殊的传播机制，可以减轻捆绑数据稀疏性的影响，并自然提高模型处理的冷启动能力。基于物品层级图卷积传播的特征提取的过程如图5所示。where W ₁ is a learnable weight, b ₁ is a learnable bias, and σ is the nonlinear activation function LeakyReLU.

are the neighbors of user u, item i, and bundle b, respectively. Aggregate functions can be functions such as simple mean function, mean function with sampling, max pooling, etc. With this special propagation mechanism, the effects of bundled data sparsity can be mitigated and the cold-start capability of model processing can be naturally improved. The process of feature extraction based on item-level graph convolution propagation is shown in Figure 5.

On the other hand, the close association between items in a bundle makes two bundles that share some items very similar. Similarity can be inferred from how many items they share. For example, a computer rig that shares five components is closer in performance than two, and a movie list that shares ten movies is closer thematically than five. For users, bundles that share more items can often be considered at the same time.

A bundle-to-user graph convolutional embedding propagation is first performed to learn user preferences at the bundle level. Then, user-to-bundle embedding propagation is performed to extract the bundle's overall properties. Since highly overlapping bundles exhibit similar patterns in attracting users, the weighted embedding propagation is performed on the bundle-item-bundle meta-path to capture the substitution relationship between bundles, using the degree of bundle overlap as a weight. The embedded update rules at the bundle level can be expressed as:

where W ₂ and b ₂ are the trainable transformation matrix and bias, respectively. M _l represents the neighbors of bundle b on the bundle-item-bundle meta-path. β _bb′ represents the degree of overlap between bundles after normalization. Propagation between attributes manifested by similar b helps bundle learning of better representations and further enhances message passing between u and b. The process of feature extraction based on bundled hierarchical graph convolution propagation is shown in Figure 6.

Based on any of the above embodiments, the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature are embedded and connected in all layers to obtain a joint prediction expression of the item propagation perspective and the bundle propagation perspective, specifically include:

Perform several graph convolution propagations on the item-level graph convolution propagation feature and the bundled-level graph convolution propagation feature to obtain several user embedding vectors and several bundled embedding vectors;

The joint prediction expression is obtained by embedding and combining the several user embedding vectors and all layers of the several bundled embedding vectors according to a preset operation mode.

Specifically, due to the desire to utilize graph neural networks to learn higher-order connectivity of user-bundle interactions, user-item interactions, and item-bundle affiliations. In particular, higher-order connectivity of affiliations that have never been considered before. For example, bundle b ₁ and bundle b ₂ share item i ₁ , bundle b ₂ and bundle b ₃ share item i ₂ , although bundle b ₁ and bundle b ₃ do not share any item, they are similar to some extent, As shown in Figure 7. Specifically, the graph convolution propagation is iteratively performed L times to obtain L user/bundle embedding vectors. The embeddings of all layers are combined, and preset operations, such as concatenation or summation, are used to combine the information received from neighbors of different depths for prediction. The prediction process based on the two propagation perspectives of item and bundle is shown in Fig. 8.

Because the special design of hierarchical propagation can not only distinguish the interaction or affiliation between any two nodes in the unified heterogeneous graph, but also distinguish whether the item node i belongs to the bundle node b or the bundle node b belongs to the item node i, so it is necessary to Both levels of information are taken into account when making predictions. Specifically, the final prediction is made through the user and bundle embeddings, such as a simple inner product method, and combining the two-level perspectives of item and bundle, such as a simple summation method, for example:

Based on any of the above-mentioned embodiments, the use of the hard-negative sample learning strategy based on the bundling scenario to train the joint prediction expression to obtain the bundled recommendation model specifically includes:

Define observed user binding interaction data and unobserved user binding interaction data based on the joint prediction expression, and construct paired training data with negative sampling based on the observed user binding interaction data and the unobserved user binding interaction data;

The paired training data is introduced with a preset objective function as a model training target and a preset probability, and the training is performed based on the hard-negative sample learning strategy to obtain the bundled recommendation model.

Specifically, since bundles contain more items and have higher prices, users generally take great care when making decisions or spending money in bundled scenarios to avoid unnecessary risks. For example, even though the user likes most of the items in the bundle, the bundle may be rejected due to one disliked item. Furthermore, for two highly similar bundles, the key to influencing the user's final choice is their non-overlapping parts.

In order to optimize the model and consider the user's decision when interacting with bundling, a learning strategy based on hard negative samples in bundling scenarios is designed. First, pair-wise learning is adopted, which is widely used in implicit recommender systems. Then, after the model converges, hard negative samples are introduced with a certain probability for finer training. Therefore, the objective function is defined as follows:

表示具有负采样的成对训练数据。

和

分别表示观察到的和未观察到的用户捆绑交互。在难负采样器中，对于每个

是u未与之交互但与其内部大多数物品交互或与b重叠的捆绑。λ为L2正则化项的权重，Θ为可训练的参数集合。难负样本的构造方式如图9所示。in

Represents paired training data with negative sampling.

and

represent the observed and unobserved user bundling interactions, respectively. In the hard negative sampler, for each

is the bundle that u does not interact with but interacts with most of the items inside it or overlaps with b. λ is the weight of the L2 regularization term, and Θ is the set of trainable parameters. The construction of hard negative samples is shown in Figure 9.

Based on any of the above embodiments, the training of the bundled recommendation model further includes setting several model hyperparameters.

Specifically, when training the bundled recommendation model, it is also necessary to set the model hyperparameters, including the number of negative samples sample_number, batch size mini_batch_size, embedding size embedding_size, learning rate learning_rate, L2 regularization term L2_normalization, message loss rate message_dropout and node loss The rate node_dropout, the selection probability hard_ratio of hard negative samples. In the process of training the network, the weights and bias values of each layer of the network can be updated through the stochastic gradient descent method in the process of backpropagation.

In order to describe the embodiments of the present invention more clearly, the specific implementation of the embodiments of the present invention will be described below with reference to FIG. 10 .

Embodiment 1: As shown in the middle branch in FIG. 10 , the user wants to recommend a new bundle to the user by using the interaction between the user and the item and the bundle history tracked by the platform. The platform here can be any e-commerce and content platform, corresponding to any item that can form a bundle, such as goods, food, places, music, books, movies, news, etc.

由用户节点

捆绑节点b∈B和物品节点i∈I组成，边E是由对应x_ub＝1的用户-捆绑交互边缘(u,b)，对应y_ui＝1的用户-物品交互边缘(u,i)和对应z_bi＝1的捆绑-物品从属边缘(b,i)组成。对于构造图上的用户，物品和捆绑节点，采用独热编码对输入进行编码，然后将其压缩为密集的实值向量：

其中

表示用户u，物品i和捆绑b的独热特征向量。P，Q和R分别表示可学习的用户嵌入，物品嵌入和捆绑嵌入的矩阵。这里可以仅通过两种交互和从属数据来对输入进行独热编码，而当平台中用户、物品和捆绑的其他属性可以利用时(例如年龄、性别等用户画像，价格、名称、图片等物品/捆绑属性)，则可以利用这些额外特征增强编码表示。First, the historical interaction between the user and the item, the historical interaction between the user and the bundle, and the composition information of the bundle are formalized into matrices, and the user-bundling history interaction matrix X _M×N , the user-item historical interaction matrix Y _M×O and the bundle are obtained. - Item affiliation matrix Z _N×O , a unified heterogeneous graph can be described by three matrices. where node

by user node

The bundle node b∈B and the item node i∈I are composed, and the edge E is composed of the user-bundle interaction edge (u,b) corresponding to x _ub =1, and the user-item interaction edge (u,i) corresponding to y _ui =1 and the bundle-item dependent edge (b, i) corresponding to z _bi = 1. For users, items, and bundle nodes on the construction graph, the input is encoded with one-hot encoding and then compressed into a dense real-valued vector:

in

One-hot feature vector representing user u, item i and bundle b. P, Q, and R denote the matrices of learnable user embeddings, item embeddings, and bundle embeddings, respectively. Here the input can be one-hot encoded with only two kinds of interaction and dependent data, while other attributes of users, items, and bundles in the platform can be leveraged (such as user portraits such as age, gender, items such as price, name, picture, etc./ bundling properties), the encoded representation can be enhanced with these additional features.

及其物品邻居节点i的表示

聚合到用户节点u以获得第l+1层的用户节点u的表示

再将第l层的物品节点i的表示

及其用户邻居节点u的表示

聚合到物品节点i以获得第l+1层的物品节点i的表示

然后，第l+1层捆绑节点b的表示

则通过其物品邻居节点i的表示

聚合而来。聚合函数不限于诸如简单均值函数，带采样的均值函数，最大池化等函数等。基于物品层级图卷积传播的特征提取的过程如图5所示。对于捆绑层级的传播，首先将第l层的用户节点u的表示

及其捆绑邻居节点b的表示

聚合到用户节点u以获得第l+1层的用户节点u的表示

再将第l层的捆绑节点b的表示

用户邻居节点u的表示

以及捆绑-物品-捆绑元路径上的捆绑邻居节点b′的表示

聚合到捆绑节点b以获得第l+1层的捆绑节点b的表示

其中基于重叠的捆绑元路径邻居的聚合以捆绑之间的重叠程度作为权重。基于捆绑层级图卷积传播的特征提取的过程如图6所示。The input features of users, items and bundles are represented as layer 0 features of a graph neural network, and the graph structure information is captured by graph convolution propagation on the graph structure, and the entity features are updated from the perspective of representation learning. For the propagation of the item level, the representation of the user node u of the lth layer is firstly

and the representation of its item neighbor node i

Aggregate to user node u to obtain the representation of user node u at layer l+1

Then the representation of the item node i of the lth layer

and the representation of its user neighbor node u

Aggregate to item node i to get the representation of item node i at level l+1

Then, layer l+1 bundles the representation of node b

Then through the representation of its item neighbor node i

aggregated. Aggregation functions are not limited to functions such as simple mean function, mean function with sampling, max pooling, etc. The process of feature extraction based on item-level graph convolution propagation is shown in Figure 5. For the propagation of the bundled level, the representation of the user node u of the lth layer is first

and its bundled neighbor node b's representation

Aggregate to user node u to obtain the representation of user node u at layer l+1

Then the representation of the bundled node b of the lth layer

Representation of the user's neighbor node u

and a representation of the bundle neighbor node b' on the bundle-item-bundle meta-path

Aggregate to bundle node b to obtain the representation of bundle node b at layer l+1

Among them, the aggregation of overlapping bundle meta-path neighbors is weighted by the degree of overlap between bundles. The process of feature extraction based on bundled hierarchical graph convolution propagation is shown in Figure 6.

然后，通过用户和捆绑嵌入做出最终预测，包括但不限于内积的方式，并结合物品和捆绑两个层级的视角，包括但不限于求和的方式，获得用户u交互捆绑b的可能性

基于物品和捆绑两种传播视角的预测过程如图8所示。对于模型训练，首先采用广泛应用于隐式推荐系统的成对学习的方式，使所估计的观察到的用户捆绑交互的可能性分值大于未观察到的用户捆绑交互的可能性分值。然后在模型收敛之后，以一定的概率(如80％)引入基于捆绑场景下的难负样本来进行更精细的训练，对每一个正样本对中的用户，选择未与之交互但与其内部大多数物品交互的捆绑作为难负样本，或对每一个正样本对中的捆绑，选择与该捆绑具有重叠物品的其他捆绑作为难负样本。使所估计的观察到的用户捆绑交互的可能性分值大于未观察到但对于用户来说难以抉择的用户捆绑样本对的可能性分值。通过随机梯度下降法优化目标函数获得模型中所有可学习参数，至此得到一个端到端的捆绑推荐系统。After iteratively performing L times of graph convolution propagation, L user/bundle embedding vectors are obtained, and the embeddings of all layers are combined, including but not limited to concatenation or summation, to obtain users and bundles in both propagations. final representation in perspective

Then, make final predictions through user and bundle embeddings, including but not limited to inner product, and combine the two-level perspectives of item and bundle, including but not limited to summation, to obtain the possibility of user u interacting with bundle b

The prediction process based on the two propagation perspectives of item and bundle is shown in Fig. 8. For model training, the pair-wise learning method, which is widely used in implicit recommender systems, is firstly adopted, so that the estimated likelihood score of observed user bundling interaction is greater than that of unobserved user bundling interaction. Then, after the model converges, a certain probability (such as 80%) is used to introduce difficult and negative samples based on the bundled scene for more precise training. The bundles with most item interactions are taken as hard negative samples, or for each bundle in the positive sample pair, other bundles with overlapping items with the bundle are selected as hard negative samples. Make the estimated likelihood score for the observed user bundling interaction greater than the likelihood score for the unobserved but intractable user bundling sample pair. All learnable parameters in the model are obtained by optimizing the objective function through stochastic gradient descent, and an end-to-end bundled recommendation system is obtained.

Embodiment 2: As shown in the left branch in FIG. 10 , the user wants to use the user-bundling history tracked by the platform to recommend a new bundle to the user. The platform here can be any e-commerce and content platform, corresponding to any item that can form a bundle, such as goods, food, places, music, books, movies, news, etc.

由用户节点

捆绑节点b∈B和物品节点i∈I组成，边E是由对应x_ub＝1的用户-捆绑交互边缘(u,b)和对应z_bi＝1的捆绑-物品从属边缘(b,i)组成。对于构造图上的用户和捆绑节点，采用独热编码对输入进行编码，然后将其压缩为密集的实值向量：

其中

表示用户u和捆绑b的独热特征向量。P和R分别表示可学习的用户嵌入和捆绑嵌入的矩阵。这里可以仅通过交互和从属数据来对输入进行独热编码，而当平台中用户和捆绑的其他属性可以利用时(例如年龄、性别等用户画像，价格、名称、图片等捆绑属性)，则可以利用这些额外特征增强编码表示。Firstly, the historical interaction between the user and the bundle, and the composition information of the bundle are formalized into matrices, and the user-bundling history interaction matrix X _M×N and the bundle-item affiliation matrix Z _N×O are obtained. The two matrices can describe a unified heterogeneous graph. where node

by user node

Binding node b∈B and item node i∈I are composed, edge E is composed of user-bundling interaction edge (u,b) corresponding to x _ub =1 and binding-item dependent edge (b,i) corresponding to z _bi =1 composition. For the user and bundle nodes on the construction graph, the input is encoded with one-hot encoding and then compressed into a dense real-valued vector:

in

One-hot eigenvectors representing user u and bundle b. P and R denote the matrix of learnable user embeddings and bundled embeddings, respectively. Here, the input can be one-hot encoded only by interaction and subordinate data, and when other attributes of users and bundles in the platform are available (such as user portraits such as age, gender, bundled attributes such as price, name, picture, etc.), it can be Enhance the encoded representation with these additional features.

及其捆绑邻居节点b的表示

聚合到用户节点u以获得第l+1层的用户节点u的表示

再将第1层的捆绑节点b的表示

用户邻居节点u的表示

以及捆绑-物品-捆绑元路径上的捆绑邻居节点b′的表示

聚合到捆绑节点b以获得第l+1层的捆绑节点b的表示

其中基于重叠的捆绑元路径邻居的聚合以捆绑之间的重叠程度作为权重。基于捆绑层级图卷积传播的特征提取的过程如图6所示。The user and bundled input features are represented as layer 0 features of a graph neural network, and the graph structure information is captured by graph convolution propagation on the graph structure, and the entity features are updated from the perspective of representation learning. For the propagation of the bundle level, the representation of the user node u at level 1 is first

and its bundled neighbor node b's representation

Aggregate to user node u to obtain the representation of user node u at layer l+1

Then the representation of the bundled node b of the first layer

Representation of the user's neighbor node u

and a representation of the bundle neighbor node b' on the bundle-item-bundle meta-path

Aggregate to bundle node b to obtain the representation of bundle node b at layer l+1

Among them, the aggregation of overlapping bundle meta-path neighbors is weighted by the degree of overlap between bundles. The process of feature extraction based on bundled hierarchical graph convolution propagation is shown in Figure 6.

然后，通过用户和捆绑嵌入做出最终预测，包括但不限于内积的方式，获得用户u交互捆绑b的可能性

对于模型训练，首先采用广泛应用于隐式推荐系统的成对学习的方式，使所估计的观察到的用户捆绑交互的可能性分值大于未观察到的用户捆绑交互的可能性分值。然后在模型收敛之后，以一定的概率(如80％)引入基于捆绑场景下的难负样本来进行更精细的训练，对每一个正样本对中的捆绑，选择与该捆绑具有重叠物品的其他捆绑作为难负样本。使所估计的观察到的用户捆绑交互的可能性分值大于未观察到但对于用户来说难以抉择的用户捆绑样本对的可能性分值。通过随机梯度下降法优化目标函数获得模型中所有可学习参数，至此得到一个端到端的捆绑推荐系统。After L times of graph convolution propagation iteratively, L user/bundle embedding vectors are obtained. And combine the embeddings of all layers, including but not limited to concatenation or summation, to obtain the final representation of users and bundles in a propagation perspective

Then, through the user and bundle embeddings to make final predictions, including but not limited to inner products, get the likelihood of user u interacting with bundle b

For model training, the pair-wise learning method, which is widely used in implicit recommender systems, is firstly adopted, so that the estimated likelihood score of observed user bundling interaction is greater than that of unobserved user bundling interaction. Then, after the model converges, a certain probability (such as 80%) is used to introduce difficult and negative samples based on the bundling scene for more precise training. For each bundle in the positive sample pair, select other bundles with overlapping items with the bundle. Bundle as a hard negative sample. Make the estimated likelihood score for the observed user bundling interaction greater than the likelihood score for the unobserved but intractable user bundling sample pair. All learnable parameters in the model are obtained by optimizing the objective function through stochastic gradient descent, and an end-to-end bundled recommendation system is obtained.

Embodiment 3: As shown in the right branch in FIG. 10 , the user wants to recommend a new bundle to the user by using the user-item history tracked by the platform. The platform here can be any e-commerce and content platform, corresponding to any item that can form a bundle, such as goods, food, places, music, books, movies, news, etc.

由用户节点

捆绑节点b∈B和物品节点i∈I组成，边E是由对应y_ui＝1的用户-物品交互边缘(u,i)和对应z_bi＝1的捆绑-物品从属边缘(b,i)组成。对于构造图上的用户和物品节点，采用独热编码对输入进行编码，然后将其压缩为密集的实值向量：

其中

表示用户u，物品i的独热特征向量。P，Q分别表示可学习的用户嵌入和物品嵌入的矩阵。这里可以仅通过交互和从属数据来对输入进行独热编码，而当平台中用户和物品的其他属性可以利用时(例如年龄、性别等用户画像，价格、名称、图片等物品属性)，则可以利用这些额外特征增强编码表示。Firstly, the historical interaction between users and items and the composition information of bundling are formalized into matrices, and the user-item historical interaction matrix Y _M×O and the bundle-item affiliation matrix Z _N×O are obtained. The two matrices can describe a unified Heterogeneous graph. where node

by user node

Binding node b∈B and item node i∈I are composed, edge E is composed of user-item interaction edge (u, i) corresponding to y _ui = 1 and binding-item dependent edge (b, i) corresponding to z _bi = 1 composition. For the user and item nodes on the construction graph, the input is encoded with one-hot encoding and then compressed into a dense real-valued vector:

in

represents the one-hot feature vector of user u, item i. P, Q represent the learnable user and item embedding matrices, respectively. Here, the input can be one-hot encoded only by interaction and subordination data, and when other attributes of users and items in the platform are available (such as user portraits such as age, gender, and item attributes such as price, name, and picture), you can Enhance the encoded representation with these additional features.

及其物品邻居节点i的表示

聚合到用户节点u以获得第l+1层的用户节点u的表示

再将第l层的物品节点i的表示

及其用户邻居节点u的表示

聚合到物品节点i以获得第l+1层的物品节点i的表示

然后，第l+1层捆绑节点b的表示

则通过其物品邻居节点i的表示

聚合而来。聚合函数不限于诸如简单均值函数，带采样的均值函数，最大池化等函数等。基于物品层级图卷积传播的特征提取的过程如图5所示。The input features of users and items are represented as the 0th layer features of the graph neural network, and the graph structure information is captured by graph convolution propagation on the graph structure, and the entity features are updated from the perspective of representation learning. For the propagation of the item level, the representation of the user node u of the lth layer is firstly

and the representation of its item neighbor node i

Aggregate to user node u to obtain the representation of user node u at layer l+1

Then the representation of the item node i of the lth layer

and the representation of its user neighbor node u

Aggregate to item node i to get the representation of item node i at level l+1

Then, layer l+1 bundles the representation of node b

Then through the representation of its item neighbor node i

aggregated. Aggregation functions are not limited to functions such as simple mean function, mean function with sampling, max pooling, etc. The process of feature extraction based on item-level graph convolution propagation is shown in Figure 5.

然后，通过用户和捆绑嵌入做出最终预测，包括但不限于内积的方式，获得用户u交互捆绑b的可能性

对于模型训练，首先采用广泛应用于隐式推荐系统的成对学习的方式，使所估计的观察到的用户捆绑交互的可能性分值大于未观察到的用户捆绑交互的可能性分值。然后在模型收敛之后，以一定的概率(如80％)引入基于捆绑场景下的难负样本来进行更精细的训练，对每一个正样本对中的用户，选择未与之交互但与其内部大多数物品交互的捆绑作为难负样本。使所估计的观察到的用户捆绑交互的可能性分值大于未观察到但对于用户来说难以抉择的用户捆绑样本对的可能性分值。通过随机梯度下降法优化目标函数获得模型中所有可学习参数，至此得到一个端到端的捆绑推荐系统。After L times of graph convolution propagation iteratively, L user/bundle embedding vectors are obtained. And combine the embeddings of all layers, including but not limited to concatenation or summation, to obtain the final representation of users and bundles in a propagation perspective

Then, through the user and bundle embeddings to make final predictions, including but not limited to inner products, get the likelihood of user u interacting with bundle b

For model training, the pair-wise learning method, which is widely used in implicit recommender systems, is firstly adopted, so that the estimated likelihood score of observed user bundling interaction is greater than that of unobserved user bundling interaction. Then, after the model converges, a certain probability (such as 80%) is used to introduce difficult and negative samples based on the bundled scene for more precise training. The bundle of most item interactions is used as a hard negative sample. Make the estimated likelihood score for the observed user bundling interaction greater than the likelihood score for the unobserved but intractable user bundling sample pair. All learnable parameters in the model are obtained by optimizing the objective function through stochastic gradient descent, and an end-to-end bundled recommendation system is obtained.

FIG. 11 is a structural diagram of a bundled recommendation system based on a graph convolutional neural network provided by an embodiment of the present invention. As shown in FIG. 11 , it includes: an acquisition module 1101 and a processing module 1102; wherein:

The acquisition module 1101 is used to acquire user-bundling historical interaction data, user-item historical interaction data and bundle-item affiliation data; the processing module 1102 is used to acquire the user-bundling historical interaction data, the user-item historical interaction data and the bundling and item affiliation data are input into the pre-trained bundling recommendation model, and the user-bundling interaction possibility recommendation result output by the bundling recommendation model is obtained; wherein the bundling recommendation model is based on the user-bundling interaction data set , Binding and item interaction data sets and user-item interaction data sets to build a unified heterogeneous graph, extracting item-level graph convolution propagation features and bundled-level graph convolution propagation features After joint prediction and feature connection, and learning based on difficult negative samples obtained from policy training.

The system provided by the embodiment of the present invention is used to execute the above corresponding method, and its specific implementation is the same as that of the method, and the involved algorithm flow is the same as that of the corresponding method, which is not repeated here.

The embodiment of the present invention reconstructs the interaction and affiliation between users, bundles and items into graphs, and utilizes the powerful ability of graph neural networks to learn the representations of three associated entities from complex topological structures and higher-order connectivity , can achieve better recommendation performance.

FIG. 12 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 12 , the electronic device may include: a processor (processor) 1210, a communication interface (Communications Interface) 1220, a memory (memory) 1230 and a communication bus 1240, The processor 1210 , the communication interface 1220 , and the memory 1230 communicate with each other through the communication bus 1240 . The processor 1210 may invoke the logic instructions in the memory 1230 to perform the following methods: obtain the historical interaction data between the user and the binding, the historical interaction data between the user and the item, and the affiliation data between the binding and the item; The user-item historical interaction data and the bundling-item affiliation data are input into the pre-trained bundling recommendation model, and the user-bundling interaction possibility recommendation result output by the bundling recommendation model is obtained; wherein the bundling recommendation model Based on the user-bundle interaction data set, the bundle-item interaction data set, and the user-item interaction data set, a unified heterogeneous graph is constructed, and the convolution propagation features of the item-level graph and the convolutional propagation features of the bundle-level graph are extracted for joint prediction and feature connection. , and based on the hard negative sample learning strategy training.

In addition, the above-mentioned logic instructions in the memory 1230 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments 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), magnetic disk or optical disk and other media that can store program codes .

On the other hand, an embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored, and the computer program is implemented by a processor to execute the transmission method provided by the above embodiments, for example, including : Obtain the historical interaction data between the user and the bundling, the historical interaction data between the user and the item, and the affiliation data between the bundling and the item; Input into the pre-trained bundling recommendation model, and obtain the user-bundling interaction possibility recommendation result output by the bundling recommendation model; wherein the bundling recommendation model is based on the user-bundling interaction data set, the bundling-item interaction data set, and the user-bundling interaction data set. Construct a unified heterogeneous graph with the item interaction data set, extract the item-level graph convolutional propagation feature and the bundled-level graph convolutional propagation feature, perform joint prediction and feature connection, and train based on the hard-negative sample learning strategy.

The device embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

From the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and certainly can also be implemented by hardware. Based on this understanding, the above-mentioned technical solutions can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic A disc, an optical disc, etc., includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A binding recommendation method based on a graph convolution neural network is characterized by comprising the following steps:

acquiring historical interaction data of a user and a binding item, historical interaction data of the user and the item and binding and item dependency relationship data;

inputting the historical interaction data of the user and the binding, the historical interaction data of the user and the item and the dependency relationship data of the binding and the item into a pre-trained binding recommendation model to obtain a recommendation result of the interaction possibility between the user and the binding, which is output by the binding recommendation model; the binding recommendation model is obtained by constructing a unified heterogeneous graph based on a user-binding interaction data set, a binding-item interaction data set and a user-item interaction data set, extracting an item level graph convolution propagation feature and a binding level graph convolution propagation feature, then performing joint prediction and feature connection, and training based on a hard-to-negative sample learning strategy.

2. The graph convolution neural network-based bundled recommendation method according to claim 1, wherein the bundled recommendation model is obtained by:

acquiring the user and binding interaction data set, the binding and item interaction data set and the user and item interaction data set, and constructing the unified abnormal picture based on the user and binding interaction data set, the binding and item interaction data set and the user and item interaction data set;

extracting the article level map convolution propagation feature and the bundle level map convolution propagation feature based on the uniform abnormal map;

embedding and connecting all layers of the convolution propagation characteristics of the item level map and the convolution propagation characteristics of the binding level map to obtain joint prediction expression of an item propagation visual angle and a binding propagation visual angle;

and training the joint prediction expression by adopting the difficult-to-bear sample learning strategy based on a binding scene to obtain the binding recommendation model.

3. The graph convolutional neural network-based binding recommendation method of claim 2, wherein the obtaining the user-to-binding interaction data set, the binding-to-item interaction data set, and the user-to-item interaction data set, and the constructing the uniform heteromorphic graph based on the user-to-binding interaction data set, the binding-to-item interaction data set, and the user-to-item interaction data set specifically comprise:

acquiring a plurality of user information, a plurality of binding information and a plurality of item information, and respectively defining the interaction data of the plurality of user information and the plurality of binding information as the user and binding interaction data set, and the subordinate relationship of the plurality of binding information and the plurality of item information as the binding and item interaction data set and the interaction data of the plurality of user information and the plurality of item information as the user and item interaction data set;

representing the user and bundle interaction data set, the bundle and item interaction data set and the user and item interaction data set by using an undirected graph; wherein the undirected graph comprises nodes and edges, the nodes comprise user nodes, binding nodes and item nodes, and the edges comprise user-binding interaction edges, user-item interaction edges and binding-item dependent edges;

the inputs to the user node, the binding node, and the item node are encoded using one-hot encoding and compressed into dense real-valued vectors.

4. The graph convolution neural network-based binding recommendation method according to claim 3, wherein the extracting the item level graph convolution propagation feature and the binding level graph convolution propagation feature based on the uniform anomaly graph specifically includes:

constructing embedding propagation between a user and an article based on the dense real-value vector to obtain an article level embedding updating rule, and obtaining the convolution propagation characteristic of the article level graph according to the article level embedding updating rule;

and constructing embedding propagation between the binding and the user based on the dense real-value vector to obtain a binding level embedding updating rule, and obtaining the convolution propagation characteristic of the binding level graph according to the binding level embedding updating rule.

5. The graph convolution neural network-based binding recommendation method according to claim 2, wherein the embedding connection of all layers is performed on the item-level graph convolution propagation feature and the binding-level graph convolution propagation feature to obtain a joint prediction expression of an item propagation perspective and a binding propagation perspective, and specifically includes:

carrying out graph convolution propagation on the item level graph convolution propagation characteristics and the binding level graph convolution propagation characteristics for a plurality of times to obtain a plurality of user embedded vectors and a plurality of binding embedded vectors;

and embedding and combining all layers of the user embedded vectors and the binding embedded vectors according to a preset operation mode to obtain the joint prediction expression.

6. The graph convolution neural network-based binding recommendation method according to claim 2, wherein the training of the joint prediction expression by using a difficult-to-negative sample learning strategy based on a binding scenario to obtain the binding recommendation model specifically comprises:

defining observed user bundled interaction data and unobserved user bundled interaction data based on the joint predictive expression, constructing paired training data with negative samples based on the observed user bundled interaction data and the unobserved user bundled interaction data;

and taking a preset target function as a model training target, introducing the paired training data according to a preset probability, and training based on the difficult-to-bear sample learning strategy to obtain the binding recommendation model.

7. The graph-convolution neural network-based bundled recommendation method of any one of claims 1-6, wherein the training of the bundled recommendation model further includes setting a number of model hyper-parameters.

8. A graph convolution neural network-based binding recommendation system is characterized by comprising:

the acquisition module is used for acquiring historical interaction data of the user and the binding, historical interaction data of the user and the article and the binding and article dependency relationship data;

the processing module is used for inputting the historical interaction data of the user and the binding, the historical interaction data of the user and the article and the dependency relationship data of the binding and the article into a binding recommendation model which is trained in advance to obtain a recommendation result of the interaction possibility between the user and the binding which is output by the binding recommendation model; the binding recommendation model is obtained by constructing a unified heterogeneous graph based on a user-binding interaction data set, a binding-item interaction data set and a user-item interaction data set, extracting an item level graph convolution propagation feature and a binding level graph convolution propagation feature, then performing joint prediction and feature connection, and training based on a hard-to-negative sample learning strategy.

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program performs the steps of the graph convolution based neural network binding recommendation method of any one of claims 1 to 7.

10. A non-transitory computer readable storage medium, having stored thereon a computer program, wherein the computer program, when being executed by a processor, implements the steps of the graph convolution neural network-based bundling recommendation method according to any one of claims 1 to 7.

Images