CN115689687A - Dynamic graph neural network recommendation method based on cellular automaton - Google Patents
Dynamic graph neural network recommendation method based on cellular automaton Download PDFInfo
- Publication number
- CN115689687A CN115689687A CN202211376444.6A CN202211376444A CN115689687A CN 115689687 A CN115689687 A CN 115689687A CN 202211376444 A CN202211376444 A CN 202211376444A CN 115689687 A CN115689687 A CN 115689687A
- Authority
- CN
- China
- Prior art keywords
- user
- article
- information
- dynamic graph
- time
- 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
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a dynamic graph neural network recommendation method based on a cellular automaton. The user view dynamic graph and the article view dynamic graph are respectively learned, the processes are user interest deviation, user interest enhancement, user cellular state change deviation calculation and user interest aggregation, and the article view dynamic graph learning process is article neighbor information aggregation, article cellular state information judgment and article embedding information updating. And predicting the representation of the future time by using the user article embedded representation, calculating the interaction probability of the user and the articles, and sequencing according to the interaction probability for recommendation. The invention well holds the rule that the user and the article continuously evolve along with time in the recommendation scene, captures the change of the preference of the user and the update of the carried information of the article in time, finds the article which the user is interested in and carries out personalized recommendation.
Description
Technical Field
The invention relates to the technical field of computer information recommendation, in particular to a dynamic graph neural network recommendation method based on a cellular automaton.
Background
In recent years, with the continuous development of online recommendation services, dynamic recommendation becomes an important research direction; in the past work was primarily based on static graph perspectives, focusing on the sequence of interactions of users and items, and learning embedded representations of users and items. Meanwhile, because the positions of the user and the article in the actual recommendation scene are different, the user purchases the article in an active state; the article is often in a selected passive state. The existing dynamic graph neural network recommendation method has the following three problems to be solved:
(1) Discrete dynamic graph learning node representation is mostly used in time change;
(2) The users and the articles are in different positions in the recommendation and need to be treated respectively;
(3) The dynamic graph neural network indicates that the learning should be performed with fine granularity according to the interaction time.
Disclosure of Invention
The invention aims to make up for the defects of the prior art and provides a dynamic graph neural network recommendation method based on a cellular automaton.
The invention is realized by the following technical scheme:
a dynamic graph neural network recommendation method based on a cellular automaton specifically comprises the following steps:
step 1, constructing a user and article interaction dynamic graph according to interaction information of users and articles along with time;
step 2, dividing the interactive dynamic graph obtained in the step 1 into a user view dynamic graph and an article view dynamic graph;
step 3, learning the user view angle dynamic graph obtained in the step 2, wherein the process comprises user interest deviation, user interest enhancement, user cell state change deviation calculation and user interest aggregation;
and 5, determining an objective function according to the user and article embedded expression obtained in the steps 3 and 4, and recommending according to a prediction result.
In step 1, a user and article interaction dynamic graph is constructed according to interaction information of users and articles along with time, and the process is as follows:
constructing the interaction between the user and the object at the time t into a dynamic diagramAre respectively dynamic diagramsNode set and edge set.
In step 2, the interactive dynamic graph obtained in step 1 is divided into a user view dynamic graph and an article view dynamic graph, and the specific process is as follows:
taking the user and the article interacted at the current moment as the center, and the neighbor nodes as the nodes interacted at the current moment and before, and comparing the dynamic graphAnd dividing the dynamic images into a user view dynamic image and an article view dynamic image.
In step 3, learning the user view dynamic graph obtained in step 2, wherein the process includes user interest migration, user interest enhancement, user cell state change migration calculation and user interest aggregation, and the specific process is as follows:
step 3.1, taking the information of the user node interaction node at the current moment as input, calculating the user interest offset expression, and comparing the user interest offset expression with the previous moment expression to obtain a user interest offset state difference value, wherein the calculation formula is as follows:
is a matrix of parameters that is,is a state context. Phi is a u Is an activation function.f v Respectively at t - An indication of the time of day and characteristics of the article itself. Updating embedding of users and items by propagating current interaction information 1 Indicating that the user's state at the time of the interest bias is poor.
Step 3.2, aggregating the second-order neighbor information of the user node at the current moment to perform user interest-enhanced learning, wherein the basic principle of the second-order aggregation is to learn the cooperative relationship between the user and the article, and the calculation formula is as follows:
zeta is the polymerization process, where attention is paid to polymerization, gamma 2 Indicating that the user's state is poor when the interest is enhanced.
Step 3.3, calculating the historical preference and the new characteristics of the user node at the current moment, and performing attention aggregation on the historical preference and the new characteristics of the user node and the expression obtained in the step 3.1 and the step 3.2, wherein the calculation formula is as follows:
Δ t is the current time t and the previous interaction time t - Time interval of (a), theta u Is an activation function, w 0 Is a vector of parameters of the time interval at.
h (u,t) Is user node embedding after interactive updating of user u and item v at time t, F u Is an aggregation function, here a sigmod activation function is used.
In step 4, the dynamic view of the article obtained in step 2 is learned, and the process includes article neighbor information aggregation, article cellular state information judgment and article embedding information updating, and the specific process is as follows:
step 4.1, aggregating the user set information interacted with the article v at the current time t, wherein the calculation formula is as follows:
Step 4.2, combining the neighbor information of the article before the time t and the neighbor information of the article node v at the current time to judge the current state of the article node, wherein the judging process is shown as formulas 8 and 9:
W s converting article information changes into state parameters for state context parameters normalize As a normalization function, when the result of the determination rho t When the value is more than or equal to k, the state of the article is changed, interest offset information carried by the article is updated, and next article embedding information updating is carried out; when the result of determination ρ is t If < k, the article status information is not changed, and article embedding means that only the history information of the article is updated.
Step 4.3, aggregating the second-order neighbor information of the article node at the current moment, and enhancing the interest preference carried by the article, wherein the calculation formula is as follows:
ζ v for the polymerization method, attention was paid to polymerization.
Step 4.4, calculating the historical preference and the new characteristics of the article node at the current moment, and performing aggregation calculation according to the information obtained in the steps 4.1-4.3 to obtain the final representation of the article node, wherein the calculation formula is as follows:
θ v is an aggregation function.
In step 5, determining an objective function according to the user item embedding expression obtained in the step 3 and the step 4, and recommending according to a prediction result, wherein the specific process is as follows:
step 5.1, predicting future embedded representation of user and article, useRepresenting the predicted future user embedded representation, the calculation formula is as follows:
is a temporal context parameter for converting a time interval into a vector,is a vector of all elements 1 + Is the future time of the user's interaction with the next item;
and 5.2, calculating future state information of the article according to the predicted future embedded expression of the user, wherein the calculation formula is as follows:
step 5.3, determining an objective function according to the user future embedded expression and the article future state information, and performing loss calculation, wherein the calculation formula is as follows:
represents the user-item interaction time in chronological order u ,λ v User embedding and item status updating are constrained for smoothing coefficients.
The invention has the advantages that: according to the invention, the continuous time dynamic graph learning method is used for grasping time fine-grained information, updating node information of the user and the article in time, meanwhile, in order to prevent the fine-grained time learning from generating overfitting influence on the representation updating of the user and the article, the idea of a cellular automata is introduced, different conversion rules are formulated according to different states of the user and the article, and finally, the embedding of the user and the article at the future moment is predicted, so that personalized recommendation is carried out, and the accuracy of a recommendation result is improved.
Drawings
FIG. 1 is a diagram showing the overall structure of the model of the present invention.
FIG. 2 is a user representation learning diagram of the present invention.
FIG. 3 is an article representative learning diagram of the present invention.
Detailed Description
The invention is further illustrated with reference to the following figures and examples.
The invention discloses a dynamic graph neural network recommendation method based on a cellular automaton, which is characterized in that constructed dynamic graphs are divided, a continuous dynamic graph learning method is used for learning user interest offset information, user interest enhancement information and article neighbor information, and then state conversion calculation is respectively carried out on a user and an article according to an appointed state conversion rule to obtain final embedded representation of the user and the article, wherein the dynamic graphs are shown in figures 1, 2 and 3. The conversion rule of the user and the article is different according to the different positions of the user and the article, the state change of the user is continuous, and the state change of the article is discrete. And finally, predicting the embedded representation and state information of the user and the articles at the future time, calculating the probability closure of interaction between the user and the articles according to the prediction result, sequencing, and selecting the articles with higher probability to generate a recommendation list.
The invention specifically comprises the following steps:
step 1, constructing a user and article interaction dynamic graph according to interaction information of a user and an article along with time, wherein the specific process comprises the following steps:
constructing the interaction between the user and the object at the time t into a dynamic diagramAre respectively dynamic diagramsNode set and edge set in (1).
Step 2, dividing the interactive dynamic graph obtained in the step 1 into a user view dynamic graph and an article view dynamic graph, wherein the specific process is as follows:
taking the user and the article interacted at the current moment as the center, and the neighbor nodes as the nodes interacted at the current moment and before, and comparing the dynamic graphAnd dividing the images into a user view angle dynamic image and an article view angle dynamic image respectively.
Step 3, learning the user view angle dynamic graph obtained in the step 2, wherein the process comprises user interest deviation, user interest enhancement, user cell state change deviation calculation and user interest aggregation, and the specific process comprises the following steps:
step 3.1, taking the node interaction information of the user node at the current moment as input, calculating the user interest offset representation, and comparing the user interest offset representation with the previous moment representation to obtain a user interest offset state difference value, wherein the calculation formula is as follows:
is a matrix of parameters that is,is a state context. Phi is a unit of u Is an activation function.f v Respectively, the article is at t - An indication of the time of day and characteristics of the article itself. Updating user and item embeddings by propagating current interaction information gamma 1 Representing the state difference of the user at the time of interest bias;
step 3.2, aggregating the second-order neighbor information of the user node at the current moment to perform user interest-enhanced learning, wherein the basic principle of the second-order aggregation is to learn the cooperative relationship between the user and the article, and the calculation formula is as follows:
zeta is the polymerization process, where attention is paid to polymerization, gamma 2 Indicating that the user's status is poor when the interest is enhanced.
Step 3.3, calculating the historical preference and the new characteristics of the user node at the current moment, and performing attention aggregation on the historical preference and the new characteristics of the user node and the expression obtained in the step 3.1 and the step 3.2, wherein the calculation formula is as follows:
Δ t is the current time t and the previous interaction time t - Time interval of (a), theta u Is an activation function, w 0 Is a vector of parameters of the time interval at.
h (u,t) User u and item v are embedded into user node after interactive update at time t, F u Is an aggregation function, here a sigmod activation function is used.
step 4.1, aggregating the user set information interacted with the article v at the current time t, wherein the calculation formula is as follows:
Step 4.2, combining the neighbor information of the article before the time t and the neighbor information of the article node v at the current time to judge the current state of the article node, wherein the judging process is shown as formulas 8 and 9:
W s for status context parameters, change of item information into status parameters, F normalize As a normalization function, when the result of the determination rho t When the value is more than or equal to k, the state of the article is changed, interest offset information carried by the article is updated, and next article embedding information updating is carried out; when the result of judgment ρ is t If < k, the article status information is not changed, and article embedding means that only the history information of the article is updated.
And 4.3, aggregating the second-order neighbor information of the article node at the current moment, and enhancing the interest preference of the article carried by the article, wherein the calculation formula is as follows:
ζ v for the polymerization method, attention was paid to polymerization.
Step 4.4, calculating the historical preference and the new characteristics of the article node at the current moment, and performing aggregation calculation according to the information obtained in the steps 4.1-4.3 to obtain the final expression of the article node, wherein the calculation formula is as follows:
θ v is an aggregation function;
to control the parameters of the extent of influence of the update mechanism, F v Is an aggregation function.
Step 5, determining an objective function according to the embedded representation of the user object obtained in the step 3 and the step 4, and recommending according to a prediction result, wherein the specific process is as follows:
step 5.1,Predicting future embedded representations of users and items, usageRepresenting the predicted future user-embedded representation, the computational formula is as follows:
is a temporal context parameter for converting a time interval into a vector,is a vector with all elements 1, t + Is the future time of the user's interaction with the next item.
And 5.2, calculating future state information of the article according to the predicted future embedded expression of the user, wherein the calculation formula is as follows:
step 5.3, determining an objective function according to the user future embedded expression and the article future state information, and performing loss calculation, wherein the calculation formula is as follows:
representing user-item interaction time, λ, arranged in time sequence u ,λ v User embedding and item status updating are constrained for smoothing coefficients.
The embodiments of the present invention are described only for the preferred embodiments of the present invention, and not for the limitation of the concept and scope of the present invention, and various modifications and improvements made to the technical solution of the present invention by those skilled in the art without departing from the design concept of the present invention shall fall into the protection scope of the present invention, and the technical content of the present invention which is claimed is fully set forth in the claims.
Claims (6)
1. A dynamic graph neural network recommendation method based on cellular automata is characterized by comprising the following steps: the method specifically comprises the following steps:
step 1, constructing a user and article interaction dynamic graph according to interaction information of users and articles along with time;
step 2, dividing the interactive dynamic graph obtained in the step 1 into a user view dynamic graph and an article view dynamic graph;
step 3, learning the user view angle dynamic graph obtained in the step 2, wherein the process comprises user interest deviation, user interest enhancement, user cell state change deviation calculation and user interest aggregation;
step 4, learning the article visual angle dynamic graph obtained in the step 2, wherein the process comprises article neighbor information aggregation, article cellular state information judgment and article embedding information updating;
and 5, determining an objective function according to the user and article embedded expression obtained in the step 3 and the step 4, and recommending according to a prediction result.
2. The method for recommending the neural network of the dynamic graph based on the cellular automaton as claimed in claim 1, wherein: step 1, constructing a user and article interaction dynamic graph according to interaction information of users and articles over time, wherein the specific process is as follows:
3. The method for recommending the neural network of the dynamic graph based on the cellular automata as claimed in claim 2, wherein: step 2, the interactive dynamic graph obtained in the step 1 is divided into a user view dynamic graph and an article view dynamic graph, and the specific process is as follows:
taking the user and the article interacted at the current moment as the center, and the neighbor nodes as the nodes interacted at the current moment and before, and comparing the dynamic graphAnd dividing the images into a user view angle dynamic image and an article view angle dynamic image respectively.
4. The cellular automaton-based dynamic graph neural network recommendation method of claim 3, wherein: step 3, learning the user view dynamic graph obtained in step 2, wherein the learning process comprises user interest migration, user interest enhancement, user cell state change migration calculation and user interest aggregation, and the specific process comprises the following steps:
step 3.1, taking the information of the user node interaction node at the current moment as input, calculating the user interest offset expression, and comparing the user interest offset expression with the previous moment expression to obtain a user interest offset state difference value, wherein the calculation formula is as follows:
is a matrix of parameters that is,is the state context, phi u In order to activate the function(s),f v respectively at t - Representation of time of day and characteristics of the item itself; updating the embedding of users and objects by propagating current mutual information, gamma 1 Representing the user's state difference at the time of the interest bias;
step 3.2, aggregating the second-order neighbor information of the user node at the current moment to perform user interest-enhanced learning, wherein the second-order aggregation is a collaborative relationship between learning users and articles, and the calculation formula is as follows:
zeta is the polymerization process, using attention polymerization, gamma 2 Representing a poor state of the user when the interest is enhanced;
step 3.3, calculating the historical preference and the new characteristics of the user node at the current moment, and performing attention aggregation on the historical preference and the new characteristics of the user node and the expression obtained in the step 3.1 and the step 3.2, wherein the calculation formula is as follows:
Δ t is the current time t and the previous interaction time t - Time interval of (a), theta u Is an activation function, w 0 Is a vector of parameters of the time interval at,is a parameter matrix, f u Is the characteristic of the user himself or herself,
h (u,t) is user node embedding after interactive updating of user u and item v at time t, F u Is an aggregation function, the function is activated using sigmod.
5. The method of claim 4, wherein the method comprises the following steps: step 4, learning the article visual angle dynamic graph obtained in the step 2, wherein the process comprises article neighbor information aggregation, article cellular state information judgment and article embedding information updating, and the specific process comprises the following steps:
step 4.1, aggregating the user set information interacted with the article v at the current time t, wherein the calculation formula is as follows:
is a parameter matrix, phi v In order to activate the function(s),is a user set which purchases the items v at the time t, and n represents the number of users;
step 4.2, combining the neighbor information of the article before the time t and the neighbor information of the article node v at the current time to judge the current state of the article node, wherein the judging process is shown as formulas (8) and (9):
W s for the state context parameter, the article information change is converted into a state parameter, rho t A value representing a state map of the item v at time t, k being a threshold value for the change in state of the item v, F normalize As a normalization function, when the result ρ is determined t When the value is more than or equal to k, the state of the article is changed, interest offset information carried by the article is updated, and next article embedding information updating is carried out; when the result of judgment ρ is t When the value is less than k, the article state information is not changed, and article embedding means that only the history information of the article is updated;
and 4.3, aggregating the second-order neighbor information of the article node at the current moment, and enhancing the interest preference of the article carried by the article, wherein the calculation formula is as follows:
ζ v as a polymerization method, attention polymerization was used;
step 4.4, calculating the historical preference and the new characteristics of the article node at the current moment, and performing aggregation calculation according to the information obtained in the steps 4.1-4.3 to obtain the final representation of the article node, wherein the calculation formula is as follows:
θ v is an aggregation function.
6. The cellular automaton-based dynamic graph neural network recommendation method of claim 5, wherein: step 5, determining an objective function according to the user and article embedded representation obtained in step 3 and step 4, and recommending according to a prediction result, wherein the specific process is as follows:
step 5.1, predicting future embedded representation of user and article, useRepresenting the predicted future user embedded representation, the calculation formula is as follows:
is a temporal context parameter for converting a time interval into a vector,is a vector with all elements 1, t + Is the future time of the user's interaction with the next item, MLP u A fully connected network representing a representation of computing a user's future time;
and 5.2, calculating future state information of the article according to the predicted future embedded expression of the user, wherein the calculation formula is as follows:
step 5.3, determining an objective function according to the future embedded expression of the user and the future state information of the article, and performing loss calculation, wherein the calculation formula is as follows:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211376444.6A CN115689687A (en) | 2022-11-04 | 2022-11-04 | Dynamic graph neural network recommendation method based on cellular automaton |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211376444.6A CN115689687A (en) | 2022-11-04 | 2022-11-04 | Dynamic graph neural network recommendation method based on cellular automaton |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115689687A true CN115689687A (en) | 2023-02-03 |
Family
ID=85047700
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211376444.6A Pending CN115689687A (en) | 2022-11-04 | 2022-11-04 | Dynamic graph neural network recommendation method based on cellular automaton |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115689687A (en) |
-
2022
- 2022-11-04 CN CN202211376444.6A patent/CN115689687A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Luo et al. | NeuLFT: A novel approach to nonlinear canonical polyadic decomposition on high-dimensional incomplete tensors | |
Shen et al. | Wind speed prediction of unmanned sailboat based on CNN and LSTM hybrid neural network | |
CN109544306B (en) | Cross-domain recommendation method and device based on user behavior sequence characteristics | |
Chen et al. | Deep reinforcement learning in recommender systems: A survey and new perspectives | |
Luo et al. | Adjusting learning depth in nonnegative latent factorization of tensors for accurately modeling temporal patterns in dynamic QoS data | |
Tan et al. | Application of genetic algorithm for fuzzy rules optimization on semi expert judgment automation using Pittsburg approach | |
Wan et al. | Network traffic prediction based on LSTM and transfer learning | |
Pham et al. | Unsupervised training of Bayesian networks for data clustering | |
Luo et al. | Pseudo Gradient-Adjusted Particle Swarm Optimization for Accurate Adaptive Latent Factor Analysis | |
Wang et al. | Time-varying-aware network traffic prediction via deep learning in IIoT | |
Pang et al. | Efficient deep reinforcement learning-enabled recommendation | |
Peng et al. | Reservoir computing models based on spiking neural P systems for time series classification | |
Yuan et al. | Deep learning from a statistical perspective | |
Kong et al. | A novel ConvLSTM with multifeature fusion for financial intelligent trading | |
CN115689687A (en) | Dynamic graph neural network recommendation method based on cellular automaton | |
CN112559905B (en) | Conversation recommendation method based on dual-mode attention mechanism and social similarity | |
CN115982373A (en) | Knowledge graph recommendation method combining multi-level interactive contrast learning | |
CN112184341B (en) | Gourmet recommending method based on archive network | |
CN115470406A (en) | Graph neural network session recommendation method based on dual-channel information fusion | |
Pise et al. | Evolving learners’ behavior in data mining | |
Lin et al. | MSP-RNN: Multi-step piecewise recurrent neural network for predicting the tendency of services invocation | |
Hubin | Bayesian model configuration, selection and averaging in complex regression contexts | |
David et al. | HERMES: Hybrid Error-corrector Model with inclusion of External Signals for nonstationary fashion time series | |
Li et al. | Research on recommendation algorithm based on e-commerce user behavior sequence | |
Shen et al. | Long-term multivariate time series forecasting in data centers based on multi-factor separation evolutionary spatial–temporal graph neural networks |
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 |