A recommendation ranking method and system based on a graph convolution network and a factorization machine

Abstract

The application discloses a recommendation ranking method and system based on a graph convolution network and a factorization machine, and first performs necessary data cleaning and preprocessing on acquired user attribute data, item attribute data and user and item interaction behavior data; then, graph representation of the user and item interaction data is performed on a hyperparameter space through introduction of the graph convolution network to generate graph embedding vectors through graph representation learning, and meanwhile, feature extraction is performed in combination with context information, user and item multidimensional attributes and the like, so that the training and prediction of a machine learning ranking model are performed after feature splicing, more accurate recommendation results are generated, and the recommendation effect and user satisfaction are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of personalized recommendation, in particular to a recommendation ranking method and system based on a graph convolution network and a factorization machine. BACKGROUND

[0002] In the DT era, with the continuous development of the Internet and the diversification of product layout, everyone can enjoy the explosive growth of various network service information. Whether it is an Internet company or a traditional enterprise, they are trying to occupy as much of the user's fragmented time as possible by providing very fast-paced products and services, so as to win the long-term favor of users. From another perspective, this phenomenon also means that the attention rate of a large number of users is getting shorter and shorter every day. Almost everyone can immediately switch screens and services. There is no doubt that a product that can retain users so that the retention time is longer than the average time is the real winner. One way to achieve this goal is personalization, which is to arrange customized goods, services or information for individuals so that the user's various information needs can be met as much as possible. A personalized recommendation system can help users discover information and make quick choices without having to learn too much about the content being displayed. By filtering personalized information based on user preferences from a large amount of information, recommendations are made based on user interests to discover new things that are suitable for them.

[0003] Generally, the business process of a recommendation system is generally divided into two stages of recall and ranking. Recall is to find a part of the results that users may be interested in from a large set of candidate recommendations, and ranking is to accurately predict the user's preference for the recall results. The recommendation system aims to provide a personalized list of items ranked according to user preferences. Without a doubt, the ranking method is the core of many recommendation algorithms. However, there are also some challenges in the ranking learning task.

[0004] 1) Due to the huge amount of data of users and items, and uneven hot and cold, there is also a long tail phenomenon, so the interaction behavior between users and items is often very sparse. In addition, when doing feature engineering in machine learning tasks, discrete type data will be one-hot encoded, especially when using a large number of ID type features, the sparsity of data will be further amplified;

[0005] 2) Traditional machine learning models require a lot of feature engineering related work, especially a lot of features need to be continuously combined to optimize the model, which requires continuous trial and error, and the process is time-consuming and labor-intensive, affecting the iteration efficiency of the model.

[0006] 3) In the actual business field, users need to return the recommended results within a few hundred milliseconds or even tens of milliseconds when refreshing, so there is a high performance requirement for the high concurrency and high availability of the recommendation system, which poses new challenges to the overall performance of the model, and often requires distributed processing of massive data, lower memory occupation and faster training speed;

[0007] 4) The relationship between users and items and other data presents a complex and networked trend, and there are problems such as timeliness, diversified behavior types, uneven distribution of positive and negative feedback, and a large number of implicit feedback behaviors, which have restricted the expression ability of the vectorization method based on sequences;

[0008] 5) There are a large number of new user registrations and new item increases every day, and the lack of behavior data often greatly affects the recommendation effect;

[0009] 6) For deep learning models, only training a scene-specific ranking model based on data in each scene is not suitable for small or new scenes with very limited feedback data. SUMMARY

[0010] Therefore, the application provides a recommendation ranking method and system based on a graph convolution network and a factorization machine, which introduces a graph convolution network to perform graph representation on user and item interaction data, performs graph representation learning on a hyperparameter space to generate a graph embedding vector, and combines context information, user and item multi-dimensional attributes and the like to perform feature extraction, train and predict a recommendation ranking learning model, so as to generate more accurate recommendation results and improve the recommendation effect and user satisfaction.

[0011] A recommendation ranking method based on a graph convolution network and a factorization machine, specifically comprising the following steps:

[0012] S1, pre-processing user behavior data, user attribute data and item attribute data;

[0013] All the pre-processed data is sorted in descending order according to the interaction behavior occurrence time to obtain a set of items that any user has interacted with and a set of users that any item has interacted with;

[0014] S2, using a LightGCN model to construct a user graph structure for the set of items that any user has interacted with and the set of users that any item has interacted with, to obtain a user-side graph embedding representation and an item-side graph embedding representation;

[0015] S3, obtaining feature information of four dimensions of users, items, contexts and cross from the pre-processed user behavior data, and inputting the feature information of the four dimensions into an FM model to obtain user feature embedding representation, item feature embedding representation, context feature embedding representation and cross feature embedding representation.

[0016] S4, based on the user-side graph embedding representation, item-side graph embedding representation, user feature embedding representation, item feature embedding representation, context feature embedding representation and cross-feature embedding representation, train and test the LightGBM model to obtain a trained ranking learning model;

[0017] The trained ranking learning model outputs a candidate set of recommended items and predicts the score of each recommended item in the candidate set. The top K recommended items with the highest scores are used as the final recommendation results according to the size of the scores.

[0018] Preferably, the steps of pre-processing the user behavior data, user attribute data and item attribute data in step S1 are:

[0019] The user behavior data, user attribute data and item attribute data are sequentially subjected to field type check, outlier check, abnormal user filtering and anti-cheating cleaning processing.

[0020] Preferably, the user behavior data refers to the data of the user's interaction with the item in different business scenarios, including the user's unique identification ID, the item's unique identification ID, the interaction behavior type, and the time when the interaction behavior occurred;

[0021] User attribute data includes the user's unique ID, age, gender, portrait label, and occupation;

[0022] Item attribute data includes the item's unique identification ID, item title, item category, and item tags.

[0023] Preferably, the interactive behavior types include click, like, play, and purchase.

[0024] Preferably, the feature information of the four dimensions of user, item, context, and cross obtained from the pre-processed user behavior data in step S3 is:

[0025] The characteristic information of the user dimension includes the user's demographic information, portrait tags, and user statistical characteristics;

[0026] The feature information of the item dimension includes item attributes and item statistical features;

[0027] The feature information of the context dimension includes situation-related features and situational statistical features;

[0028] The feature information of the cross dimension includes the combined statistical information of a single feature.

[0029] Preferably, the specific steps of training and testing the LightGBM model based on the user-side graph embedding representation, the item-side graph embedding representation, the user feature embedding representation, the item feature embedding representation, the context feature embedding representation and the cross feature embedding representation in step S4 are as follows:

[0030] The user-side graph embedding representation and the item-side graph embedding representation are all spliced with the user feature embedding representation, the item feature embedding representation, the context feature embedding representation and the cross feature embedding representation to obtain complete feature representations;

[0031] All the complete feature representations are randomly divided into a training set and a test set according to a certain proportion, the LightGBM model is trained and tested, the parameters of the LightGBM model are continuously adjusted, and a trained ranking learning model is obtained.

[0032] Preferably, the parameters of the LightGBM model include the depth of a tree, the number of iterations and a regularization term coefficient.

[0033] A recommendation ranking system based on a graph convolution network and a factorization machine includes a data preprocessing module, a graph embedding vector generation module, a feature vector generation module and a ranking learning model training and prediction module,

[0034] The data preprocessing module is used for preprocessing user behavior data, user attribute data and item attribute data, and obtaining a set of items that any user has interacted with and a set of users that any item has interacted with from the preprocessed data.

[0035] The graph embedding vector generation module is used for constructing a user graph structure by using a LightGCN model on the set of items that any user has interacted with and the set of users that any item has interacted with, to obtain a user-side graph embedding representation and an item-side graph embedding representation.

[0036] The feature vector generation module is used for performing first-order and second-order cross learning on feature information in four dimensions of users, items, contexts and cross obtained from the preprocessed user behavior data by using an FM model, to obtain a user feature embedding representation, an item feature embedding representation, a context feature embedding representation and a cross feature embedding representation.

[0037] The ranking learning model training and prediction module is used for training and testing a LightGBM model based on the user-side graph embedding representation, the item-side graph embedding representation, the user feature embedding representation, the item feature embedding representation, the context feature embedding representation and the cross feature embedding representation, performing score prediction on a candidate set to be recommended, and selecting TopK as a final recommendation result according to the score size.

[0038] The present application has the following advantages:

[0039] 1、The application firstly carries out necessary data cleaning pretreatment on the obtained user attribute data, item attribute data, user and item interaction behavior data; then, the graph convolution network is introduced to perform graph representation on the user and item interaction data, graph representation learning is performed on the hyperparameter space to generate graph embedding vectors, and feature extraction is performed in combination with context information, user and item multi-dimensional attributes and the like, training and prediction of a machine learning ranking model are performed after feature splicing, so that more accurate recommendation results are generated, and the recommendation effect and user satisfaction are effectively improved.

[0040] 2、The LightGCN model, the FM model and the LightGBM model are creatively fused, so that the problems of data sparsity, cold start, feature cross, user and item interaction networking and the like can be well solved, and the recommendation accuracy and user satisfaction are greatly ensured.

[0041] 3、The LightGCN model constructs a graph network for the interaction behavior of the user and the item, learns the graph embedding representation of the user and the item respectively through linear propagation, and then weights the embedding of all layers as the final vector representation. First of all, the model is simple, linear, small and easy to implement and train. Secondly, not only the implicit feedback data is fully utilized to improve the recommendation effect, but also the problem of data sparsity can be well solved.

[0042] 4、The factorization machine is introduced to cross-learn the first-order and second-order dimensions of the user attribute, item attribute, context and feature combination and the like. Each feature learns an implicit weight vector, and the inner product of the two feature implicit vectors is used as the weight of the cross feature when the features are crossed. Not only the idea of matrix decomposition is fused, that is, the coefficients of the second-order cross feature are adjusted in the form of matrix decomposition, so that the coefficients are no longer independent and irrelevant, but also the problem of unable to train parameters caused by data sparsity is solved.

[0043] 5、The LightGBM model is used as a ranking learning model of the recommendation result. The LightGBM model not only supports efficient parallel training, but also has faster training speed, lower memory consumption, better accuracy, supports distribution, can quickly process massive data and the like, and lays a solid foundation for the real-time and stability of the recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0045] Figure 1 is the overall flowchart of the recommendation ranking method.

[0046] Figure 2 is the bipartite graph of user-item interaction and the schematic diagram of high-order connectivity.

[0047] Figure 3 is the LightGCN model architecture diagram.

[0048] Figure 4 is the method flowchart of the recommendation ranking method. DETAILED DESCRIPTION

[0049] In order to better understand the technical solutions of the present application, the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0050] It should be clear that the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] The present application will be described in further detail below through specific embodiments and in conjunction with the accompanying drawings.

[0052] The present application provides a recommendation ranking method and system based on graph convolution network and factorization machine. The present application introduces graph convolution network to represent user and item interaction data in graph form, performs graph representation learning on hyperparameter space to generate graph embedding vectors, and combines context information, user and item multi-dimensional attributes, etc. to extract features for training and prediction of the recommendation ranking learning model, so as to generate more accurate recommendation results and improve recommendation effect and user satisfaction.

[0053] First, the system for executing the recommendation ranking method based on graph convolution network and factorization machine given by the present application is introduced, which includes a data preprocessing module, a graph embedding vector generation module, a feature vector generation module, and a ranking learning model training and prediction module.

[0054] The data preprocessing module is used for preprocessing user behavior data, user attribute data and item attribute data, and obtaining a set of items that any user has interacted with and a set of users that any item has interacted with from the preprocessed data.

[0055] The graph embedding vector generation module is used for constructing a user graph structure by using a LightGCN model on the set of items that any user has interacted with and the set of users that any item has interacted with, to obtain a graph embedding representation on the user side and a graph embedding representation on the item side.

[0056] The feature vector generation module utilizes the FM model to perform first-order and second-order dimensional cross learning on the feature information of four dimensions of users, items, contexts and intersections obtained from the preprocessed user behavior data, to obtain user feature embedding representation, item feature embedding representation, context feature embedding representation and cross feature embedding representation.

[0057] The ranking learning model training and prediction module trains and tests the LightGBM model based on the user-side graph embedding representation, the item-side graph embedding representation, the user feature embedding representation, the item feature embedding representation, the context feature embedding representation and the cross feature embedding representation, performs score prediction on the candidate set to be recommended, and selects TopK as the final recommendation result according to the score size.

[0058] As shown in Figure 1 Figure 1 The overall flowchart of the recommendation ranking method based on the graph convolution network and the factorization machine is shown in the figure.The method first performs necessary data cleaning and preprocessing on the obtained user attribute data, item attribute data and user and item interaction behavior data (also referred to as user behavior data), to obtain temporary data required by each intermediate module; then, deep graph representation learning is performed to obtain graph embedding vectors, and the features required by the machine learning model (ranking learning model) are engineered, and after feature splicing, the machine learning ranking model is trained and the result is predicted.

[0059] Specifically, the recommendation ranking method based on the graph convolution network and the factorization machine comprises the following steps:

[0060] S1, preprocessing user behavior data, user attribute data and item attribute data.

[0061] The fields contained in the user behavior data, user attribute data and item attribute data under different business scenarios have great differences.

[0062] Generally, the user behavior data refers to the data of the interaction behavior between the user and the item in different business scenarios, including the unique ID (userid) of the user, the unique ID (itemid) of the item, the interaction behavior type (behavior_type), the time (behavior_time) of the interaction behavior, etc. The interaction behavior type (behavior_type) includes clicking, liking, playing, purchasing and other operation behaviors.

[0063] The user attribute data includes the unique ID (userid) of the user, the age (age), the gender (sex), the portrait label (tag), the occupation (occupation) and the like.

[0064] The item attribute data includes fields such as a unique ID (itemid) identifying the item, a title (title), an item category (category), and a tag (tag) applied to the item.

[0065] The above user behavior data, user attribute data, and item attribute data are subjected to field type checking, outlier checking, abnormal user filtering, anti-cheating, and other cleaning processes.

[0066] Then, all the preprocessed data is sorted in descending order according to the interaction behavior occurrence time (behavior_time), obtaining the item set that any user u has interacted with and the user set that any item i has interacted with The item set that any user u has interacted with is represented as [i1, i2, …, imn], and the user set that any item i has interacted with is represented as [u1, u2, …, umn]. ]. ks ].

[0067] S2, using the LightGCN model to construct a user graph structure based on the item set that any user has interacted with and the user set that any item has interacted with to obtain a graph embedding representation on the user side and a graph embedding representation on the item side

[0068] The LightGCN model has advantages such as easier interpretation, easier training and maintenance, easier analysis of the model process in technology, and continuous optimization.

[0069] The basic idea of GCN is to learn the representation of nodes by smoothing the features of the graph structure, so it iteratively performs graph convolution, i.e., aggregates the features of neighbors into a new representation of the target node. The LightGCN model constructs a graph structure representation based on the interaction behavior data of users and items to model the high-order connectivity between User-Item to improve the graph embedding effect, as shown in Figure 2 , which uses a simple weighted sum aggregator, i.e., only the normalized embedding sum of the next layer of neighboring nodes.

[0070] The architecture of the LightGCN model is shown in Figure 3 , in the layer combination (Layer Combination) stage, the embeddings of each layer are summed to obtain the final representation. represents the embedding of the userid corresponding to user u, represents the embedding of the itemid corresponding to item i, and ​are the new embedding representations of user u and item i after k layers of propagation, and the definition of the light graph convolution operation in LightGCN is as follows.

[0071]

[0072]

[0073] in is a symmetric normalization term that prevents the embedding size from increasing with graph convolution operations.

[0074] In LightGCN, the only trainable model parameters are the embedding representations of the 0th layer, i.e., the embedding representations of all users. and all items If both are given, the embedding of higher layers can be calculated by LGC defined in formula (3). After k layers of LGC, we further combine the embeddings obtained at each layer to form the final embedding representations of users and items as shown below.

[0075]

[0076]

[0077] where α k ≥ 0 indicates the importance of the k-th layer embedding in forming the final embedding representation.

[0078] Assume that the user-item interaction matrix is Where M and N represent the number of users and items respectively, and R ui The values ​​of are as follows.

[0079]

[0080] The adjacency matrix representation of the user-item graph structure is shown in (6).

[0081]

[0082] The embedding matrix of layer 0 is represented as E (0) ,but Where T represents the length of the embedding, the matrix representation of LGC is as follows.

[0083]

[0084] Where D is a diagonal matrix with (M+N)×(M+N) dimensions, D ii Represents the number of nonzero values ​​in the i-th row vector of the adjacency matrix.

[0085] Finally, the final embedding matrix representation used for model prediction is shown below.

[0086] E = a0E (0) + a1E (1) + a2E (2) +... + a K E (K) (8)

[0087] Further, the following formula is obtained.

[0088]

[0089] wherein is a symmetric normalized matrix.

[0090] The trainable parameters of LightGCN are only the embeddings of the 0th layer, i.e. Θ = {E (0)} using the Bayesian personalized ranking loss as the objective function, as follows.

[0091]

[0092] where λ controls the L2 regularization strength, and it is used with the Adam optimizer in a mini-batch fashion.

[0093] Using the above method, the set of items each user has interacted with and the set of users each item has interacted with are input into the LightGCN model for training, and during the training process, the learning rate, regularization system, layer combination system, dimension of embedding and other parameters of the model are constantly adjusted for optimization, and finally the user-side graph embedding representation of each user and the item-side graph embedding representation of each item

[0094] S3, the feature information of the four dimensions of users, items, contexts and cross from the preprocessed user behavior data is obtained, and the feature information of the four dimensions is input into the FM model to obtain the user feature embedding representation, item feature embedding representation, context feature embedding representation and cross feature embedding representation.

[0095] The FM model aims to solve the feature combination problem under sparse data, which combines the advantages of support vector machines and factorization machines, uses decomposition parameters to model all interactions between variables, and can handle any real-valued feature vector general predictor, even if the data is very sparse, it can still be calculated well.

[0096] Specifically, the factorization machine learns the first-order and second-order cross-learning of the feature information of the four dimensions of users, items, contexts, and cross, and learns a hidden weight vector for the feature information of each dimension. When performing cross combination between two features, the inner product of the hidden weight vectors of the two features is used as the weight of the cross feature, and the FM model equation is solved, and finally the user feature embedding representation item feature embedding representation context feature embedding representation and cross feature embedding representation

[0097] The factorization machine combines the idea of matrix decomposition, that is, the coefficients of the second-order cross feature are adjusted in the form of matrix decomposition, so that the coefficients are no longer independent and irrelevant, and the problem of unable to train parameters caused by data sparsity is solved.

[0098] The feature information of the above four dimensions of users, items, contexts, and cross is preprocessed including one-hot, multi-hot, attention mechanism, embedding representation, etc., and then input into the FM model to obtain the user feature embedding representation item feature embedding representation context feature embedding representation and cross feature embedding representation

[0099] The feature information of the user dimension (user feature) includes the demographic information, portrait label, and user statistical features of the user, such as age (age), gender (sex), bias, portrait label (tag), region, etc.

[0100] The feature information of the item dimension (item feature) includes the item attribute and item statistical features, such as title (title), label (tag), category (category), click rate (ctr), etc.

[0101] The feature information of the context dimension (context feature) includes the context-related features and context statistical features, such as the number of current page flips, exposure order, request time period, etc.

[0102] The feature information of the cross dimension (cross feature) includes the combined statistical information of single features, such as the click rate of the user on the category, the preference for the item label, and the similarity with the item title vector, etc.

[0103] S4, user-side graph embedding representation item-side graph embedding representation user feature embedding representation item feature embedding representation context feature embedding representation and cross feature embedding representation Training and testing of the LightGBM model are performed to obtain the trained ranking learning model.

[0104] The LightGBM model is a framework that implements the GBDT algorithm, supports efficient parallelism, and has the advantages of distributed processing of massive data, lower memory occupation, faster training speed, and better accuracy, etc. Experiments on multiple public datasets show that LightGBM speeds up the training process of traditional GBDT by more than 20 times, while achieving almost the same accuracy. Based on this, the embodiment selects LightGBM as the ranking learning model.

[0105] Gradient Boosting Decision Tree (GBDT) is a popular machine learning algorithm and has many effective implementations. Although many engineering optimizations are adopted in these implementations, the efficiency and scalability are still not satisfactory when the feature dimension is high and the data volume is large. One main reason is that for each feature, they need to scan all data instances to estimate the information gain of all possible split points, which is very time-consuming. LightGBM adopts two new techniques: gradient-based one-side sampling and exclusive feature bundling. Using GOSS, a large part of data instances with small gradients are excluded, and only the remaining part is used to estimate the information gain. Since data instances with large gradients play a more important role in the calculation of information gain, GOSS can obtain a relatively accurate information gain estimate with a smaller amount of data. Using EFB, mutually exclusive features (i.e. they rarely take non-zero values at the same time) are bundled to reduce the number of features.

[0106] Specifically, based on the embedding vectors obtained in steps S2 and S3 above, when training and testing the LightGBM model, first, the user-side graph embedding representation of all users and the item-side graph embedding representation of all items user feature embedding representation item feature embedding representation context feature embedding representation and cross feature embedding representation are spliced to obtain complete feature representation

[0107] Then, the complete feature representation of all users According to a certain proportion, the training set and the test set are randomly divided, the LightGBM model is trained and tested, the parameters of the LightGBM model are continuously adjusted, the parameters include the depth of the tree, the number of iterations, the regularization term coefficient and the like, and a trained ranking learning model is obtained.

[0108] In the recommendation result prediction stage, the trained ranking learning model is used to output a candidate set of recommended items and predict the scores of each recommended item in the candidate set, and the top K recommended items with the highest scores are displayed as the final recommendation results according to the score size.

[0109] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A recommendation ranking method based on graph convolutional network and factorization machine, characterized in that: The specific steps include: S1, preprocessing user behavior data, user attribute data and item attribute data; Sort all preprocessed data in descending order according to the time of interaction, and obtain the set of items that any user has interacted with and the set of users that any item has interacted with; S2: Use the LightGCN model to build a user graph structure for the set of items that any user has interacted with and the set of users that any item has interacted with, and obtain the graph embedding representation on the user side and the graph embedding representation on the item side; S3: Obtain feature information in four dimensions: user, item, context, and cross from the preprocessed user behavior data. Input these four dimensions into the FM model to obtain user feature embedding representation, item feature embedding representation, context feature embedding representation, and cross feature embedding representation. Specifically, the characteristic information of the user dimension includes the user's demographic information, portrait tags, and user statistical characteristics; The feature information of the item dimension includes item attributes and item statistical features; The feature information of the context dimension includes situation-related features and situational statistical features; The feature information of cross-dimensions includes the combined statistical information of single features; S4, based on the user-side graph embedding representation, item-side graph embedding representation, user feature embedding representation, item feature embedding representation, context feature embedding representation and cross-feature embedding representation, train and test the LightGBM model to obtain a trained ranking learning model; The trained ranking learning model outputs a candidate set of recommended items and predicts the score of each recommended item in the candidate set. The top K recommended items with the highest scores are used as the final recommendation results according to the size of the scores.

2. The recommendation ranking method based on graph convolutional network and factorization machine according to claim 1, characterized in that: The steps for pre-processing user behavior data, user attribute data and item attribute data in step S1 are as follows: The user behavior data, user attribute data and item attribute data are sequentially subjected to field type check, outlier check, abnormal user filtering and anti-cheating cleaning processing.

3. The recommendation ranking method based on graph convolutional network and factorization machine according to claim 1 or 2, characterized in that: User behavior data refers to the data on the user's interactions with items in different business scenarios, including the user's unique identification ID, the item's unique identification ID, the type of interaction behavior, and the time when the interaction behavior occurred; User attribute data includes the user's unique ID, age, gender, portrait label, and occupation; Item attribute data includes the item's unique identification ID, item title, item category, and item tags.

4. The recommendation ranking method based on graph convolutional network and factorization machine according to claim 3, characterized in that: The interactive behavior types include click, like, play, and purchase.

5. The recommendation ranking method based on graph convolutional network and factorization machine according to claim 1, characterized in that: The specific steps for training and testing the LightGBM model based on the user-side graph embedding representation, item-side graph embedding representation, user feature embedding representation, item feature embedding representation, context feature embedding representation, and cross-feature embedding representation in step S4 are as follows: All the user-side graph embedding representations and item-side graph embedding representations are concatenated with the user feature embedding representation, item feature embedding representation, context feature embedding representation, and cross-feature embedding representation to obtain a complete feature representation; All complete feature representations are randomly divided into training sets and test sets according to a certain ratio, the LightGBM model is trained and tested, and the parameters of the LightGBM model are continuously adjusted to obtain a trained ranking learning model.

6. The recommendation ranking method based on graph convolutional network and factorization machine according to claim 5, characterized in that: The parameters of the LightGBM model include the depth of the tree, the number of iterations, and the regularization coefficient.

7. A recommendation ranking system based on the method according to any one of claims 1 to 6, characterized in that: It includes data preprocessing module, graph embedding vector generation module, feature vector generation module and ranking learning model training and prediction module. The data preprocessing module is used to preprocess the user behavior data, user attribute data and item attribute data, and obtain the set of items that any user has interacted with and the set of users that any item has interacted with from the preprocessed data; The graph embedding vector generation module is used to use the LightGCN model to construct a user graph structure for the set of items that any user has interacted with and the set of users that any item has interacted with, and obtain the graph embedding representation on the user side and the graph embedding representation on the item side; The feature vector generation module uses the FM model to perform first-order and second-order cross-learning on the feature information of the four dimensions of user, item, context, and cross obtained from the pre-processed user behavior data to obtain user feature embedding representation, item feature embedding representation, context feature embedding representation, and cross feature embedding representation; The ranking learning model training and prediction module trains and tests the LightGBM model based on the graph embedding representation on the user side, the graph embedding representation on the item side, the user feature embedding representation, the item feature embedding representation, the context feature embedding representation and the cross-feature embedding representation, predicts the scores of the recommendation candidate set, and selects the TopK as the final recommendation result according to the score size.

Images