Double-end Node2vec graph migration feature extraction and multi-depth convolution model integration recommendation algorithm

Abstract

The invention relates to a double-end Node2vec graph migration feature extraction and multi-depth convolution model integrated recommendation algorithm, relates to the field of optimization of learning algorithms, and realizes dual extraction of features by performing graph migration in two modes of depth first and breadth first at the same time; inputting the double extracted features into a plurality of deep convolution recommendation grids, and forming a plurality of integrated model parameter sets through an integrated learning mode; and compressing the plurality of integrated models in a knowledge distillation mode. According to the method, local and global information in the graph data can be effectively captured, the feature expression capability is improved, and the recommendation precision is improved through integrated learning of multiple deep convolutional networks. And the knowledge distillation technology is used for compressing the integrated model, so that the model reasoning efficiency is improved on the premise of ensuring the recommendation effect.

Description

Technical Field

[0001] The present invention relates to the field of optimization of learning algorithms, and in particular to a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm. Background Art

[0002] With the development of Internet technology, recommendation systems have become the core modules of major application platforms. The accuracy and efficiency of recommendation algorithms directly affect the user experience and the business growth of the platform. At present, graph embedding technology is particularly important in recommendation systems, among which graph walk algorithms (such as DeepWalk and Node2vec) are widely used in feature extraction of graph structured data. However, these methods have their own advantages and disadvantages, and it is difficult to achieve comprehensive results in feature extraction.

[0003] Node2vec is a graph embedding algorithm that generates node sequences by simulating random walks and uses the Skip-gram model for embedding learning. Its core lies in controlling the ratio of depth-first search (DFS) and breadth-first search (BFS) during random walks, thereby achieving a balance between local and global structures. However, a single walk strategy is difficult to achieve the best results in different types of graph data. Summary of the invention

[0004] Purpose of the invention: To provide a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm to solve the above-mentioned problems existing in the prior art.

[0005] Technical solution: A dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm, which uses both depth-first and breadth-first methods to perform graph walks to achieve dual feature extraction; the dual-extracted features are input into multiple deep convolution recommendation grids, and multiple integrated model parameter sets are formed through integrated learning; multiple integrated models are compressed through knowledge distillation.

[0006] In a further embodiment, each deep convolutional recommendation grid is composed of a one-dimensional convolutional layer and a multi-layer perceptron layer; the one-dimensional convolutional layer learns local patterns of features, which is suitable for feature extraction of sequence data; the multi-layer perceptron layer performs nonlinear combination and global pattern learning on the convolved features to improve recommendation accuracy.

[0007] In a further embodiment, the depth-first path goes deep into the node, and the breadth-first path expands around the node, thereby acquiring more comprehensive node features.

[0008] In a further embodiment, the specific method of knowledge distillation is to train a lightweight student model to learn the output distribution of the integrated model, retain the model performance while quickly deleting invalid model parameters, thereby increasing the thrust speed of the recommendation algorithm.

[0009] In a further embodiment, multiple integrated model parameter sets integrate different feature expressions at the model level, and fuse multiple model feature expressions to reduce the deviation of a single model and enhance the generalization ability of the model.

[0010] Beneficial effects: The present invention relates to a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm, which relates to the field of optimization of learning algorithms. By simultaneously using depth-first and breadth-first methods to walk the graph, dual feature extraction is achieved; the dual extracted features are input into multiple deep convolution recommendation grids, and multiple integrated model parameter sets are formed through integrated learning; and multiple integrated models are compressed through knowledge distillation. The present invention can effectively capture local and global information in graph data, improve feature expression capabilities, and improve recommendation accuracy through integrated learning of multiple deep convolutional networks. The integrated model is compressed using knowledge distillation technology, which improves the model reasoning efficiency while ensuring the recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is the operation flow chart of the present invention. DETAILED DESCRIPTION

[0012] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known in the art are not described.

[0013] The applicant believes that a single walking strategy is difficult to achieve the best results in different types of graph data.

[0014] To this end, the applicant designed a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm, which can effectively capture local and global information in graph data, improve feature expression capabilities, and improve recommendation accuracy through integrated learning of multiple deep convolutional networks. The integrated model is compressed using knowledge distillation technology, which improves the model reasoning efficiency while ensuring the recommendation effect.

[0015] The dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm involved in the present invention realizes dual feature extraction by simultaneously using depth-first and breadth-first methods for graph walk; the dual-extracted features are input into multiple deep convolution recommendation grids, and multiple integrated model parameter sets are formed through integrated learning; and multiple integrated models are compressed through knowledge distillation.

[0016] Each deep convolutional recommendation grid consists of a one-dimensional convolutional layer and a multi-layer perceptron layer; the one-dimensional convolutional layer learns local patterns of features and is suitable for feature extraction of sequence data; the multi-layer perceptron layer performs nonlinear combination and global pattern learning on the convolutional features to improve recommendation accuracy. Specifically, the features extracted by the dual-headed Node2vec graph walk are input into multiple deep convolutional recommendation networks.

[0017] Each recommendation network consists of a one-dimensional convolutional layer and a multi-layer perceptron (MLP) layer:

[0018] 1. One-dimensional convolution layer: learns local patterns of features and is suitable for feature extraction of sequence data.

[0019] 2.MLP layer: Perform nonlinear combination and global pattern learning on the convolutional features to improve recommendation accuracy.

[0020] The depth-first path goes deep into the node, while the breadth-first path expands around the node, thereby obtaining more comprehensive node features. This method can not only capture local neighbor information, but also preserve global structural information.

[0021] Since the integrated model has a large number of parameters and a slow inference speed, the knowledge distillation method is used in this application to compress the integrated model. The specific method of knowledge distillation is to train a lightweight student model to learn the output distribution of the integrated model (teacher model), retain the model performance while quickly deleting invalid model parameters, and increase the thrust speed of the recommendation algorithm.

[0022] Multiple integrated model parameter sets integrate the expressions of different features at the model level and fuse the feature expressions of multiple models to reduce the deviation of a single model and enhance the generalization ability of the model.

[0023] The algorithm process and advantages are summarized as follows:

[0024] 1. The two-headed Node2vec algorithm combines depth-first search (DFS) and breadth-first search (BFS) to effectively capture local and global information in graph data and improve feature expression capabilities. 2. The integrated learning of multiple deep convolutional networks improves recommendation accuracy. 3. The integrated model is compressed using knowledge distillation technology, which improves model reasoning efficiency without significantly reducing the recommendation effect.

[0025] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm, characterized by: Both depth-first and breadth-first methods are used to walk the graph to achieve dual feature extraction; the dual-extracted features are input into multiple deep convolutional recommendation grids, and multiple integrated model parameter sets are formed through ensemble learning; multiple integrated models are compressed through knowledge distillation.

2. According to claim 1, a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm is characterized by: Each deep convolution recommendation grid consists of a one-dimensional convolution layer and a multi-layer perceptron layer; the one-dimensional convolution layer learns local patterns of features and is suitable for feature extraction of sequence data; the multi-layer perceptron layer performs nonlinear combination and global pattern learning on the convolved features to improve recommendation accuracy.

3. According to claim 2, a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm system is characterized by: The depth-first path goes deep into the node, while the breadth-first path expands around the node to obtain more comprehensive node features.

4. According to claim 1, a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm system is characterized by: The specific method of knowledge distillation is to train a lightweight student model to learn the output distribution of the integrated model, retain the model performance while quickly deleting invalid model parameters, and increase the thrust speed of the recommendation algorithm.

5. According to claim 1, a dual-headed Node2vec graph walk feature extraction and multi-depth convolution model integrated recommendation algorithm system is characterized by: Multiple integrated model parameter sets integrate the expressions of different features at the model level and fuse the feature expressions of multiple models to reduce the deviation of a single model and enhance the generalization ability of the model.

Images