CN115880024A - Method for recommending lightweight graph convolutional neural network based on pre-training - Google Patents
Method for recommending lightweight graph convolutional neural network based on pre-training Download PDFInfo
- Publication number
- CN115880024A CN115880024A CN202211532515.7A CN202211532515A CN115880024A CN 115880024 A CN115880024 A CN 115880024A CN 202211532515 A CN202211532515 A CN 202211532515A CN 115880024 A CN115880024 A CN 115880024A
- Authority
- CN
- China
- Prior art keywords
- nodes
- graph
- user
- training
- article
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention discloses a method for recommending a graph convolution neural network based on pre-training, which comprises the following steps: constructing a heterogeneous graph and describing a path mode of the heterogeneous graph by using a meta-path; pre-training based on the meta-path to obtain low-dimensional embedded representation of user nodes and article nodes; and taking the low-dimensional embedded representations of the user nodes and the article nodes as initialization parameters of the graph convolution neural network, updating the low-dimensional embedded representations of the nodes on the bipartite graph of the user and the article, carrying out weighted summation on the embedded vectors of each layer of the graph convolution neural network to obtain the final low-dimensional embedded representation of the nodes, predicting scores through the inner product of the embedded vectors of the user and the article, and sequencing according to the scores to obtain the final recommendation result. By using the method, the high-quality recommendation result which is suitable for the preference of the user can be generated, and the adverse effect of overhigh model reasoning time delay on the user experience is avoided, so that the user experience and the traffic conversion rate can be improved.
Description
Technical Field
The invention relates to a method for recommending a lightweight graph convolutional neural network based on pre-training.
Background
The recommendation system is widely applied to business scenes such as social media, mobile apps and video websites, can utilize historical behavior data of users to mine interest preference of the users, provides commodity information and suggestions for the users, and helps the users to find high-quality content quickly, so that user stickiness and flow conversion rate are improved. The mainstream algorithms used by recommendation systems can be roughly divided into two categories: content-based recommendation algorithms and collaborative filtering-based recommendation algorithms.
The content-based recommendation algorithm analyzes the relationships between the items and recommends similar items based on the user's historical information. The disadvantages are that repeated items are easily recommended and are difficult to apply to multimedia scenes because the feature extraction of multimedia information such as music, pictures, video, etc. is difficult. The recommendation algorithm based on collaborative filtering assumes that similar users have similar preferences for candidate items, and measures the similarity of the users by adopting different measurement means such as Pearson correlation coefficients, so as to provide the target user with the items preferred by similar neighbor users. The disadvantage is that it is difficult to cope with the cold start problem of new users and that the recommendation result may show a significant head effect.
Disclosure of Invention
Content-based recommendation algorithms have the disadvantage of easily recommending duplicate items and are difficult to apply to multimedia scenes. The recommendation algorithm based on collaborative filtering has the disadvantages that it is difficult to cope with the cold start problem of the new user, and the recommendation result may have a significant head effect.
In order to solve the problems, the invention provides a method for recommending a lightweight graph convolutional neural network based on pre-training.
The invention uses a lightweight graph convolution neural network based on pre-training to complete the recommendation task in the online shopping scene. On one hand, the invention fuses data of different service scenes by using a heterogeneous graph structure, and extracts key words in the text by using a maximum entropy rule, thereby flexibly unifying various information into the heterogeneous graph and enabling the model to have expandability. On the other hand, the online shopping scene heterogeneous graph information is integrated into the reasoning process of the downstream graph convolutional neural network through the pre-training technology based on the meta-path, so that the evaluation index of the model is improved, and the high-efficiency characteristic of the original lightweight model is retained.
A method for recommendation based on a pre-trained atlas neural network, the method comprising:
constructing a heterogeneous graph and describing a path mode of the heterogeneous graph by using a meta-path;
pre-training based on the meta-path to obtain low-dimensional embedded representation of user nodes and article nodes; and
and taking the low-dimensional embedded representations of the user nodes and the article nodes as initialization parameters of the graph convolution neural network, updating the low-dimensional embedded representations of the nodes on a bipartite graph of the user and the article, carrying out weighted summation on embedded vectors of each layer of the graph convolution neural network to obtain final low-dimensional embedded representations of the nodes, predicting scores through inner products of the embedded vectors of the user and the article, and sequencing according to the scores to obtain a final recommendation result.
In the method, the step of constructing the heterogeneous map may include:
and extracting data items from the data set of each scene as nodes of the heterogeneous graph, wherein the data items comprise long text information, performing word segmentation and statistics on the long text information, and screening out the most valuable keywords as the nodes in the heterogeneous graph according to the maximum entropy rule.
In the method, the pre-training step may include:
and pre-training by adopting a Skip-Gram model, randomly walking on the heterogeneous graph according to the meta-path to obtain a large number of path sequences, obtaining low-dimensional embedded representations of the user nodes and the article nodes through the pre-training, and then transferring the pre-trained model to a downstream LightGCN network to be used as an initialization weight for further fine adjustment, thereby finally completing the training of the whole model.
In the method, the Skip-Gram algorithm maximizes an objective function as follows:
where V is a heterogeneous graph node, V is a set of nodes, A is a set of different types of nodes V, c t Is a neighbor node of node type t of v, N t (v) Is a set of neighboring nodes, and θ is a model parameter.
In the method, the step of updating a low-dimensional embedded representation of nodes on the bipartite graph of users and items may comprise:
the following update rules are used:
wherein e u And e i Low-dimensional embedded representation, N, representing users and items, respectively u And N i Respectively representing the neighbor nodes of the current user and the article node, and k is the layer number of the graph convolution neural network.
In the method, the PyTorch-based recommendation system open source framework RecBole may be employed.
By using the method, the high-quality recommendation result which is suitable for the user preference can be generated, and the adverse effect of the overhigh model reasoning time delay on the user experience is avoided, so that the user experience and the flow conversion rate can be improved.
Drawings
Specific embodiments of the present invention will now be described with reference to the accompanying drawings, which are illustrative and not restrictive.
FIG. 1 shows a diagram of meta-paths for describing path patterns;
FIG. 2 shows an overall framework diagram of a model of an embodiment of the invention;
FIG. 3 shows a frame diagram of a lightweight graph convolutional network recommendation algorithm; and
fig. 4 shows a flow chart of a method of an embodiment of the invention.
Detailed Description
The operator App is a mobile business hall App developed by the operator, and is dedicated to providing a warmer and more intimate experience for users in the aspects of service, life, entertainment, wealth and the like besides the main business of a telephone network. The online shopping scene is mainly embodied in an operator shopping mall and comprises application scenes such as life shopping, movie ticket subscription, an electronic book mall and the like. The high-quality recommendation result adaptive to the user preference can improve the user experience and the traffic conversion rate, and is beneficial to improving the commercial income and market competitiveness of the operator App. Aiming at a plurality of challenges faced by a recommendation system in an operator business scene, the invention designs a reasonable and efficient recommendation algorithm framework to adapt to actual requirements, and simultaneously ensures the lightweight characteristic of a model to facilitate deployment to a production environment.
In recent years, many deep learning models are used to implement collaborative filtering algorithms in recommendation systems, and graph neural networks are one of the efficient schemes. The invention aims to complete the construction of an algorithm framework based on a recommendation method of a graph structure and considering the requirements of an industrial practical application scene on algorithm performance, and by combining practical services on the basis of a light-weight graph convolution neural network (LightGCN).
Graph convolutional neural network is a popular graph network architecture, originally designed for the node classification problem, on which LightGCN makes a lot of simplifications, while making the model more suitable for collaborative filtering recommendation scenarios. Formally, consider updating a low-dimensional embedded representation of a node on a bipartite graph of users and items, using the following update rules:
wherein e u And e i Low-dimensional embedded representation, N, representing users and items, respectively u And N i And k is the number of layers of the graph convolution neural network. And after the propagation of a plurality of layers is completed, carrying out weighted summation on the embedding vector of each layer of the graph convolutional neural network to obtain the final low-dimensional embedding representation of the node. And predicting the scores through the inner product of the embedded vectors of the user and the articles, and sequencing according to the scores to obtain a final recommendation result.
This is merely an exemplary method of updating a low-dimensional embedded representation of a node on a bipartite graph of users and items, and the present invention may employ other methods known to those skilled in the art.
The node updating process of the model can be completely written into a matrix product form, the model can realize rapid reasoning under the hardware optimization acceleration aiming at the matrix multiplication, and the model is suitable for the actual application scene with high throughput and high concurrency requirements.
Furthermore, heterogeneous graph structures are also commonly used to construct graph-based recommendation methods. The heterogeneous graph is modified with respect to the conventional homogeneous graph structure by introducing an object type mapping function and a relationship type mapping function to map nodes and relationships to node classes and relationship classes, respectively. For example, for a knowledge graph structure related to a movie, a heterogeneous graph may have different node types such as "movie", "director", "actor", and the like, and different relationship types such as "lead actor", "director", and the like, while a homogeneous graph does not distinguish the node types and the relationship types at all.
Fig. 1 shows a schematic diagram of meta-paths for describing path patterns.
Meta-paths are often used to describe path patterns on heterogeneous graphs. For example, as shown in fig. 1, the one-length paths "(cameron, director, avanda)" and "(norland, director, interplanetary crossing)" follow the same path pattern, i.e., the meta-path "(director name, director, movie name)". The graph structure is viewed from the perspective of the meta-path, and mode information which cannot be obtained by the homogeneous graph can be fully utilized, so that the description capacity of the model for the graph structure is improved, and the performance of the recommendation system is further improved.
Designing and implementing a recommendation system adapted to an operator App online shopping scenario mainly faces two major challenges: firstly, the design of the lightweight graph convolutional neural network LightGCN only considers the user interaction records in a single-service scene, and cannot be directly popularized to an online shopping scene with various types of interaction records, so that the comprehensive utilization of the interaction data in each scene is difficult to obtain the comprehensive understanding of the user. Secondly, the online shopping scene of the operator App serves massive users, a lightweight and efficient deep learning model needs to be designed, and the requirements of high-throughput and high-concurrency scenes on the model reasoning performance are met.
In order to keep the excellent performance of the original lightweight graph convolution neural network on the reasoning performance, the invention does not modify the network architecture, but adopts the pre-training technology to initialize the network weight so as to fuse the information of the recommended multi-service scene with gain. Specifically, the method constructs a unified heterogeneous graph based on data of each service scene, designs a pre-training task from the perspective of a meta-path, and takes parameters obtained by training as initialization weight parameters of a downstream inference network.
Fig. 2 shows an overall framework diagram of a model of an embodiment of the present invention. The method comprises three parts of heterogeneous graph construction, pre-training based on element paths and graph convolution neural network recommendation based on training parameters according to data flow.
Firstly, constructing a heterogeneous graph based on online shopping scenes, extracting necessary data items from data sets of all the scenes to serve as nodes of the heterogeneous graph, and correspondingly designing edges of different types to connect the nodes. The nodes comprise user nodes and article nodes, and after pre-training, the low-dimensional embedded representation of the user nodes and the article nodes can be used as initialization parameters of corresponding nodes of the convolutional neural network of the downstream graph. Meanwhile, it is noted that the data items contain some long text information, the processing method can be used for performing word segmentation and statistics on the long texts, and the most valuable keywords are screened out according to the maximum entropy rule to serve as nodes in the heterogeneous graph.
Of course, the present invention may also employ other methods known to those skilled in the art to construct a heterogeneous map.
Secondly, a pre-training task is carried out, the invention adopts an extended Skip-Gram (Skip model) algorithm under the scene of a heterogeneous graph for pre-training, firstly, random walk is carried out on the heterogeneous graph according to a designed meta-path, and a large number of path sequences are obtained, wherein the path sequences comprise the associated information of nodes on the graph and meta-path semantic information. The basic idea of the Skip-Gram algorithm is to maximize the objective function:
where V is a heterogeneous graph node, V is a set of nodes, A is a set of different types of nodes V, c t Is a neighbor node of node type t of v, N t (v) Is a set of neighboring nodes, and θ is a model parameter. In a specific implementation, techniques such as negative sampling and the like are used to improve the training speed and the training effect. The model, namely the expression (3), is pre-trained to obtain a low-dimensional embedded representation of the user and the article node, and then the pre-trained model is transferred to a downstream LightGCN network to be used as an initialization weight to be further fine-tuned, namely, the LightGCN is initialized by using the network weight obtained by pre-training, and finally the training of the whole model is completed. The fine adjustment is to further optimize the recommendation effect by using the communicated shopping order data and the like. The fine tuning process follows the normal training method of the recommendation systemThe process is carried out. Because the network weight obtained by pre-training can not completely fit with the actual recommendation scene, adjustment is needed according to data, and the recommendation effect is improved.
Of course, the present invention may also be pre-trained using other algorithms known to those skilled in the art.
A frame diagram of a complete lightweight graph convolution network recommendation algorithm based on meta-path pre-training is shown in fig. 3.
In fig. 3, the loss function is actually the objective function in expression (3). The penalty used by the graph-convolution network is a default cross-entropy penalty function, with the goal of enabling the weights to fit the distribution of the current data set to complete the recommendation task.
The model parameters are vector representations of the user and the article, and then inner product operation is needed to measure similarity, so that recommendation is completed.
Fig. 4 shows a flow chart of a method of an embodiment of the invention. In this embodiment, a recommendation method based on a pre-trained atlas neural network includes the following steps:
s1: constructing a heterogeneous graph and describing a path mode of the heterogeneous graph by using a meta-path;
s2: pre-training based on the element path to obtain low-dimensional embedded representation of user nodes and article nodes; and
s3: the low-dimensional embedded representation of the user nodes and the article nodes is used as an initialization parameter of the graph convolution neural network, the low-dimensional embedded representation of the nodes is updated on the bipartite graph of the user and the article, the embedded vectors of each layer of the graph convolution neural network are weighted and summed to obtain the final low-dimensional embedded representation of the nodes, scores are predicted through the inner product of the embedded vectors of the user and the article, and the final recommendation result is obtained through sorting according to the scores.
The invention uses a lightweight graph convolution neural network based on pre-training to complete a recommendation task in an online shopping scene.
The invention uses the heterogeneous graph structure to fuse the data of different service scenes, and simultaneously uses the maximum entropy rule to extract the key words in the text, thereby flexibly unifying various information into the heterogeneous graph and leading the model to have expandability.
According to the method, heterogeneous graph information of an online shopping scene is integrated into the reasoning process of a downstream graph convolutional neural network through a pre-training technology based on a meta path, so that the evaluation index of the model is improved, and the high-efficiency characteristic of the original lightweight model is retained.
In order to embody the effect of the recommendation algorithm of the present invention, the present invention selects three algorithms as baseline methods for comparison: itemKNN, BPR and NeuMF.
ItemKNN is a collaborative filtering based on articles, and the core idea is to select the neighbor articles most similar to the articles and recommend the similar articles of the articles liked before to the user. Since the information of the items is relatively fixed, the similarity between different items can be calculated in advance, and then the similarities of different items are combined to calculate the similarity between the items and the candidate recommended items.
The BPR is a bayesian personalized ranking algorithm, and recommendations are made using implicit feedback from the user (e.g., clicks, favorites, etc.). Bayesian probability analysis is carried out on the problems, the maximum posterior probability is calculated, personalized sorting of all articles for the user is obtained, and then a recommendation result is generated.
NeuMF is a collaborative filtering algorithm based on a deep neural network. It can be divided into two parts: generic matrix factorization and multi-layer perceptrons. The input of the universal matrix decomposition is the serial numbers of users and articles, firstly, the user characteristic vector and the article characteristic vector are obtained through an embedding layer, and then the user characteristic and the article characteristic are multiplied by each other according to bits and are decomposed; the input of the multilayer perceptron is also the serial numbers of the user and the article, the characteristic vectors of the user and the article are obtained through the embedding layer, and then the characteristic vectors are processed through the full connection layer and the nonlinear activation function. And finally, connecting the results of the two parts by the model, and outputting a recommended value through a full connection layer.
The data sets used by the invention are data sets in four fields of operator App, namely shopping order data, shopping click data, reading data and movie ticket order data, and the formats of the data sets are csv files. The following is a basic information analysis on several data sets:
table 1: online shopping scenario dataset analysis
According to the data analysis results shown in the table, the shopping order data, the shopping click data and the reading data have higher sparsity, but the data scale is relatively larger, and the average interaction number of the user and the article is relatively higher; while the size of the movie ticket order data is relatively small, where the number of items is only 20, the average number of interactions by the user is very low. This characteristic of the movie ticket order data set results in underperformance of the model thereon, as described by the results analysis.
The training of the model of the invention is mainly divided into two stages of pre-training and fine tuning of the graph convolution neural network based on the element path.
In the pre-training stage, an example path of the meta-path needs to be acquired through a random walk algorithm, wherein 1000 paths are sampled at each node on the graph, the length of the random walk is set to be 100, the vector dimension adopted by the pre-training model is 64, the learning rate of the pre-training is 0.025, the batch processing size is 50, and 5 rounds of training are performed.
In the fine tuning stage, the vector dimension adopted is still 64, the number of layers of the graph convolution network is 2, an Adam optimizer is used for optimization, the learning rate is 0.001, the batch processing size in the fine tuning training stage is 2048, and the batch processing size in the fine tuning testing stage is 4096. The data samples in the fine tuning stage are divided into a training set, a verification set and a test set according to the proportion of 8.
The method is mainly realized by adopting an open source framework RecBole of the PyTorch-based recommendation system. RecBole is a uniform, comprehensive and efficient recommendation system library, and has a universal and extensible data structure, a comprehensive benchmark model and a data set, and large-scale standard evaluation. The model effect provided by the invention can be rapidly verified by using the tool.
The invention compares and analyzes a lightweight graph convolutional network recommendation algorithm based on heterogeneous information network element path pre-training with the baseline method under each service scene, and five evaluation indexes are selected: recall @10, MRR @10, NDCG @10, hit @10 and precision @10. The results were as follows:
table 2: shopping order data
Table 3: shopping click data
Table 4: reading data
Table 5: movie ticket order data
From the above results, it can be seen that all the indicators are improved by at least 17.8% in the shopping order data, compared with the baseline method with the best effect; on shopping click data, all indexes are improved by at least 20.8 percent compared with a baseline method with the best effect; on reading data, all indices were improved by at least 20.8% relative to the baseline method with the best results. The above results illustrate the effectiveness of the convolutional network recommendation algorithm based on the heterogeneous information network meta-path pre-training graph. The method not only exerts the reasoning performance of the lightweight graph convolutional neural network, but also fully utilizes the information gain of the heterogeneous graph, and is more suitable for the scene of shopping recommendation on the line of the operator mall.
The method is not good in effect on movie ticket order data. The analysis of the data set by the inventor shows that the size of the data set of the movie ticket order is small, only 3,292 interaction records are in total, the number of items (namely, the number of movies) is only 20, and the average number of interactions of the user is only 1.02, so that the model is difficult to capture the user preference from the data. Under the condition, the inventor finds that better results can be obtained by recommending according to the frequency of the occurrence of the articles, namely the heat degree of the articles, and the ItemKNN and NeuMF model structures determine that the ItemKNN and the NeuMF can sense the heat degree of the articles to a certain extent, so that the effect of the method is better than that of the method.
The invention provides a convolutional network recommendation algorithm based on heterogeneous information network meta-path pre-training graph aiming at a recommendation scene of online shopping of App of a Chinese operator. The expected target of the method is that the index performance is improved by more than 10% compared with the index performance of three baseline methods of ItemKNN, BPR and NeuMF, and the expected target is achieved according to the experimental results.
The method is based on the actual requirements of the online shopping scene of the Chinese operator App, and a lightweight graph convolutional neural network recommendation algorithm framework based on meta-path pre-training is constructed. The algorithm comprehensively utilizes data of actual service scenes such as shopping, electronic reading, movie ticket subscription and the like, and improves the performance of a recommendation result on each evaluation index; meanwhile, the design of the algorithm considers the requirements of high throughput and high concurrency of practical application scenes, and a lightweight model architecture is adopted, so that adverse effects on user experience caused by overhigh model reasoning time delay are avoided. As can be seen from the evaluation analysis of the actual data sets of the four online shopping scenes, the recommendation algorithm framework constructed by the invention achieves the expected target.
It will be understood by those of ordinary skill in the art that all or some of the steps of the methods, systems, functional modules/units in the devices disclosed above may be implemented as software, firmware, hardware, and suitable combinations thereof. In a hardware implementation, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be performed by several physical components in cooperation. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). The term computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, as is well known to those skilled in the art. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by a computer. In addition, communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media as known to those skilled in the art.
Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.
Claims (6)
1. A method for recommendation based on a pre-trained atlas neural network, the method comprising:
constructing a heterogeneous graph and describing a path mode of the heterogeneous graph by using a meta-path;
pre-training based on the meta-path to obtain low-dimensional embedded representation of user nodes and article nodes; and
and taking the low-dimensional embedded representations of the user nodes and the article nodes as initialization parameters of the graph convolution neural network, updating the low-dimensional embedded representations of the nodes on a bipartite graph of the user and the article, carrying out weighted summation on embedded vectors of each layer of the graph convolution neural network to obtain final low-dimensional embedded representations of the nodes, predicting scores through inner products of the embedded vectors of the user and the article, and sequencing according to the scores to obtain a final recommendation result.
2. The method of claim 1, wherein the step of constructing the heterogeneous map comprises:
and extracting data items from the data set of each scene as nodes of the heterogeneous graph, wherein the data items comprise long text information, performing word segmentation and statistics on the long text information, and screening out most valuable keywords as the nodes in the heterogeneous graph according to a maximum entropy rule.
3. The method of claim 1 or 2, wherein the pre-training step comprises:
and pre-training by adopting a Skip-Gram model, randomly walking on the heterogeneous graph according to the meta-path to obtain a large number of path sequences, obtaining low-dimensional embedded representations of the user nodes and the article nodes through the pre-training, and then transferring the pre-trained model to a downstream LightGCN network to be used as an initialization weight for further fine adjustment, thereby finally completing the training of the whole model.
4. A method according to claim 3, wherein the Skip-Gram algorithm maximizes an objective function as follows:
where V is a heterogeneous graph node, V is a set of nodes, A is a set of different types of nodes V, c t Is a neighbor node of node type t of v, N t (v) Is a set of neighboring nodes, and θ is a model parameter.
5. The method of claim 1 or 2, wherein the step of updating the low-dimensional embedded representation of nodes on the bipartite graph of users and items comprises:
the following update rules are used:
wherein e u And e i Low-dimensional embedded representation, N, representing users and items, respectively u And N i Respectively representing the neighbor nodes of the current user and the article node, and k is the layer number of the graph convolution neural network.
6. The method according to claim 1 or 2, wherein a PyTorch based recommender system open source framework RecBole is employed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211532515.7A CN115880024A (en) | 2022-11-29 | 2022-11-29 | Method for recommending lightweight graph convolutional neural network based on pre-training |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211532515.7A CN115880024A (en) | 2022-11-29 | 2022-11-29 | Method for recommending lightweight graph convolutional neural network based on pre-training |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115880024A true CN115880024A (en) | 2023-03-31 |
Family
ID=85765388
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211532515.7A Pending CN115880024A (en) | 2022-11-29 | 2022-11-29 | Method for recommending lightweight graph convolutional neural network based on pre-training |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115880024A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117575744A (en) * | 2024-01-15 | 2024-02-20 | 成都帆点创想科技有限公司 | Article recommendation method and system based on user association relation |
-
2022
- 2022-11-29 CN CN202211532515.7A patent/CN115880024A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117575744A (en) * | 2024-01-15 | 2024-02-20 | 成都帆点创想科技有限公司 | Article recommendation method and system based on user association relation |
| CN117575744B (en) * | 2024-01-15 | 2024-03-26 | 成都帆点创想科技有限公司 | Article recommendation method and system based on user association relation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110969516B (en) | Commodity recommendation method and device | |
| CN110717098B (en) | Meta-path-based context-aware user modeling method and sequence recommendation method | |
| CN110321422B (en) | Method for training model on line, pushing method, device and equipment | |
| Qu et al. | An end-to-end neighborhood-based interaction model for knowledge-enhanced recommendation | |
| US20170161644A1 (en) | Media recommendation using internet media stream modeling | |
| US9864951B1 (en) | Randomized latent feature learning | |
| CN112765373B (en) | Resource recommendation method and device, electronic equipment and storage medium | |
| CN109902235B (en) | User preference clustering collaborative filtering recommendation algorithm based on bat optimization | |
| CN105159910A (en) | Information recommendation method and device | |
| US20220253722A1 (en) | Recommendation system with adaptive thresholds for neighborhood selection | |
| WO2022036494A1 (en) | Graph structure aware incremental learning for recommender system | |
| Huang et al. | Position-enhanced and time-aware graph convolutional network for sequential recommendations | |
| Hazrati et al. | Addressing the New Item problem in video recommender systems by incorporation of visual features with restricted Boltzmann machines | |
| Daneshvar et al. | A social hybrid recommendation system using LSTM and CNN | |
| CN115880024A (en) | Method for recommending lightweight graph convolutional neural network based on pre-training | |
| Sridhar et al. | Content-Based Movie Recommendation System Using MBO with DBN. | |
| CN115618024A (en) | Multimedia recommendation method and device and electronic equipment | |
| Nguyen et al. | Geometric matrix completion with deep conditional random fields | |
| CN113342994B (en) | Recommendation system based on non-sampling cooperative knowledge graph network | |
| CN116910375B (en) | Cross-domain recommendation method and system based on user preference diversity | |
| Blöchl et al. | Co-clustering via information-theoretic Markov aggregation | |
| Jaffali et al. | Like-tasted user groups to predict ratings in recommender systems | |
| CN113095901A (en) | Recommendation method, training method of related model, electronic equipment and storage device | |
| Kvifte et al. | Hybrid recommendation of movies based on deep content features | |
| CN111931058B (en) | Sequence recommendation method and system based on self-adaptive network depth |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |