CN113378062A - Graph collaborative filtering recommendation method based on decoupling and memory - Google Patents
Graph collaborative filtering recommendation method based on decoupling and memory Download PDFInfo
- Publication number
- CN113378062A CN113378062A CN202110761998.7A CN202110761998A CN113378062A CN 113378062 A CN113378062 A CN 113378062A CN 202110761998 A CN202110761998 A CN 202110761998A CN 113378062 A CN113378062 A CN 113378062A
- Authority
- CN
- China
- Prior art keywords
- user
- decoupling
- representation
- intention
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a recommendation method for collaborative filtering of a graph based on decoupling and memory, and belongs to the field of data mining. Firstly, the method represents the user and the article by using ID and historical interaction, obtains the representation vectors of the user and the article, embeds the representation vectors into the integration layer, and fuses the user and the article to embed the representation vectors. Secondly, each embedded user and item vector passes through a collaborative filtering framework based on a decoupling graph neural network to obtain a decoupling representation of the user about potential intention, and the intention blocking decoupling characteristic of the user item after the aggregation characteristic is updated. Then, in the integration layer, a gating cycle unit is introduced, and complete characteristic information is aggregated from high-order neighbourhoods of user and article nodes. Independent modules are then introduced to encourage independence between the different intents. And finally, constructing a loss function to optimize the model and predicting the corresponding situation of the user intention and the article. The method combines a decoupling and remembering graph collaborative filtering method, can effectively separate different intentions of the user for the object, and can learn more interpretable characteristics by combining a gating cycle unit.
Description
Technical Field
The invention belongs to the field of data mining, and particularly relates to a recommendation method for collaborative filtering of a neural network of a graph based on decoupling and memory.
Background
Recommendation systems (recommendation systems) have been widely used in real life, and their main purpose is to infer user preferences to help users find their preferred personalized items and recommend them to users. Collaborative Filtering (Collaborative Filtering) is a classic and widely applied algorithm in a recommendation system, and learns and recommends items preferred by a user based on user-item historical interaction behavior. For example, users with similar behavior may have similar preferences for items.
For a user-item interaction graph, although conventional collaborative filtering algorithms can effectively recommend relevant items for a user. However, the user's interaction with the item may be based on a variety of different intentions that may drive different behaviors. While the traditional collaborative filtering algorithm uses a uniform intention to process the relationship between the user and the article, the influence of different intentions of the user on the article is ignored, and the embedded representation of the intention of the mixture cannot well provide the interpretability of the personal intention. In addition, since noise or other adverse factors often exist in the user-item interaction graph, the mixed-together intention can reduce the interaction robustness and stability of the embedded representation and noise and the like, and can also reduce the interpretability. Therefore, when establishing the user-object interaction relationship, the user intention embedded in different objects should be considered, and the different intentions in the different objects should be separated. While taking into account the independence of each intent, by iterating the separation operations, the model is enabled to obtain a more interpretable representation.
Based on the method for the collaborative filtering of the decoupling graph, in a graph aggregation layer, the problem of high-order information loss and information loss between nodes easily occurs, so that a user and an article node cannot completely acquire complete information representation from high-order neighbors. Based on the above problems, a Gated Round Unit (GRU) is introduced, so that the node embedding representation can capture more complete high-order neighbor information, and the loss of information transmission between nodes is reduced.
Disclosure of Invention
The invention aims to separate a plurality of different intentions of an item interaction graph of a user, reduce the loss of information transmission between high-order nodes and obtain better embedded representation, thereby providing a graph collaborative filtering recommendation method based on decoupling and memory. The content comprises the following steps:
a recommendation method for graph collaborative filtering based on decoupling and memory is characterized by comprising the following steps:
s1: representing the user and the article by using ID and history interaction to obtain a representation vector of the user and the article, embedding the representation vector into an integration layer, and fusing the representation vectors into a user and article embedded representation vector;
s2: obtaining a decoupling representation of the user about potential intention and updating intention blocking decoupling characteristics of the user object after the aggregation characteristics by each embedded user and object vector through a collaborative filtering framework based on a decoupling graph neural network;
s3: in the integration layer, a gating cycle unit is introduced, and complete characteristic information is aggregated from high-order neighbors of user and article nodes;
s4: introducing an independent module to encourage independence between different intents;
s5: and constructing a loss function to optimize the model and predicting the corresponding situation of the user intention and the article.
Based on the above, the specific steps of S1 are as follows:
s11: using the ID and history interaction of the user and the article as original input, and performing one-hot coding to obtain a representation vector of the user and the article;
s12: linearly embedding the high-dimensional sparse representation vector to obtain a low-dimensional dense representation vector;
s13: and integrating the user and the object into a representation vector of the interaction of the user and the object by utilizing the integration layer.
Based on the above, the specific steps of S2 are as follows:
s21: for the user and article interaction graph, each user and article is divided into K blocks by using a decoupling graph neural network, the divided K blocks output an intention of a decoupling representation consisting of K independent components corresponding to a user u, wherein the K-th intention decoupling representation of the user u is as follows:
where K is 1, …, K indicates the K-th intention, ekuIs a decoupled representation of the kth potential intention of user u, ekiIs a decoupled representation of the kth potential intention of item i, l represents the number of aggregation layers, g represents the aggregation function, NuRepresenting a first order neighbor set of user u.
S22: decoupling each intention of the user item graph, and adding and combining decoupling characteristics obtained after the kth intention of the user u is output to obtain a potential intention decoupling representation e of aggregating all neighbor characteristics of the user u under the kth intentionkuFor the same reason, can obtain ekiExpressed as:
where L represents the total number of layers in the diagram.
S23: updating the aggregated feature of the intent block decoupling feature of the user item, the intent block decoupling feature of the user itemThe update of (a) is represented as:
wherein T is 1, …, T represents iteration round number, T represents final iteration round number,an embedded representation representing that user u aggregates first order neighbors under the kth intent in the tth iteration,representing historical input of item I, initialization
S24: degree of interaction between user u and item i at k-th intention, i.e. interaction similarity(u, i) item History input with kth intentUnder interactive conditions, it can be aggregated into an embedded representation of the user u aggregating first-order neighbors under the kth intention in the tth iterationExpressed as:
wherein the content of the first and second substances,respectively representing the values of the user u and the item i,after K iterations is
Based on the above, the specific steps of S3 are as follows:
s31: kth intent decoupling representation of user uPolymerized by a polymerization function g. The formula is known from S21 and is expressed asThe detailed representation is as follows:
where LeakyReLu (-) is the activation function, fGRU(. cndot.) is a gated cycle unit.
Based on the above, the specific steps of S4 are as follows:
s41: an algorithm framework based on decoupling and remembering graph collaborative filtering aims at that different intents can be mutually independent and do not influence each other, so that an independent module is established, and loss among intents is reducedintendExpressed as:
wherein the content of the first and second substances,the representation is embedded as the intent of all users and items, COV (-) as the inter-matrix distance covariance, and VAR (-) as the distance variance of each matrix.
Based on the above, the specific steps of S5 are as follows:
s51: establishing a predicted user based on the final representations of the user and the itemA function of the likelihood of interaction between the items, the prediction functionExpressed as:
s52: constructing a BPR loss function to optimize a model, predicting the corresponding situation of the user intention and the article, wherein the BPR loss function is expressed as:
wherein O { (u, i, j) | (u, i) ∈ O+,(u,j)∈O-Denotes the interacted training data O+Training data O corresponding to undetected data-The collection of (1) | · | non-conducting phosphor2Represents the L2 regular term, and σ represents the sigmoid activation function.
Further, k is [1, 2, 4, 8, 16 ].
Further, L ═ 0, 1, 2, 3.
The invention provides a recommendation method for collaborative filtering of a graph based on decoupling and memory, and starts with three aspects of decoupling and separating user-article interaction intentions, aggregating high-order neighbor information among nodes and constructing intention independence. Compared with the prior art, the invention has the advantages that:
1) by adopting a graph collaborative filtering recommendation method based on decoupling, compared with the original traditional graph collaborative filtering method, the method can reduce adverse factors such as noise in user-article historical interaction and learn a more robust and interpretable representation;
2) in the graph aggregation layer, the gated circulation unit with a memory function is adopted, and compared with the traditional aggregation unit, the method can learn more complete high-order neighbor information among nodes, reduce loss during high-order information transmission in the aggregation process and improve generalization capability of the model.
Drawings
FIG. 1 is a detailed flow diagram of a graph collaborative filtering recommendation method based on decoupling and memory;
FIG. 2 is a diagram of the polymerization of a gated cyclic unit according to the present invention.
Detailed description of the preferred embodiments
In order to make the technology, content and advantages of the invention more apparent, the invention will be further described with reference to the accompanying drawings and specific examples.
The invention relates to a technology based on a graph collaborative filtering structure, which is used for the fields of data mining and the like. In the invention, given historical interaction information of a user and an article, the decomposition of the user intention is completed through a network structure based on the decoupling and memory graph collaborative filtering, so that the task of recommending the article is completed. The following describes in detail a specific implementation of the present invention.
Fig. 1 is a schematic method flow diagram provided by an embodiment of the present invention, and provides a recommendation method for graph collaborative filtering based on decoupling and memory, which includes the following basic steps:
s1: representing the user and the article by using ID and history interaction to obtain a representation vector of the user and the article, embedding the representation vector into an integration layer, and fusing the representation vectors into a user and article embedded representation vector;
s2: obtaining a decoupling representation of the user about potential intention and updating intention blocking decoupling characteristics of the user object after the aggregation characteristics by each embedded user and object vector through a collaborative filtering framework based on a decoupling graph neural network;
s3: in the integration layer, a gating cycle unit is introduced, and complete characteristic information is aggregated from high-order neighbors of user and article nodes;
s4: introducing an independent module to encourage independence between different intents;
s5: and constructing a loss function to optimize the model and predicting the corresponding situation of the user intention and the article.
In the steps, the user-item interaction graph is subjected to intention decomposition mainly through a graph collaborative filtering network based on decoupling and memory, and a more interpretable embedded representation is obtained. Meanwhile, a gating cycle unit is introduced into an aggregation layer, and information such as different and high-order neighbor relations is aggregated into different block intention characteristics. And then, establishing a loss function through an independent module so as to improve the anti-interference and generalization capability of the whole model. Therefore, the steps S1 to S5 can have various implementations, and the following describes the implementation process of the above steps in the present embodiment:
in this embodiment, the specific steps for implementing S1 are as follows:
s11: the user and item ID and history interactions are used as raw inputs. Firstly, one-hot coding is carried out on user u, wherein the one-hot coding of the ID is a binary vector x with the length of Mu∈RMOnly the u-th element is 1 and the other elements are 0. Similarly, one-hot code of ID of item i is xi∈RN。
The historical interactions of the user and the item are then encoded. Is provided withRepresenting the collection of items that user u has interacted with,representing a set of users with whom item i has interacted. Firstly, historical interactive coding is carried out on a user u, wherein the historical interactive coding is a binary vector x 'with the length of N'u∈RNOnly the position corresponding to the ID of the item interacted with the history of the user u is 1, and the other positions are 0. Similarly, the historical recoding of item i is a binary vector x 'of length M'i∈RM。
S12: and linearly embedding the high-dimensional sparse representation vector to obtain a low-dimensional dense representation vector, wherein the embedding process is as the following formula.
Wherein, PuAnd PiThe expression vectors m of the user u and the item i are obtained from the IDuAnd miRepresenting vectors of the user u and the item i obtained from historical interaction respectively;andrespectively representAndthe number of (2); eu、Ei、E′u、E′iRepresenting a transformation matrix.
S13: and at the embedding integration layer, embedding and integrating the user and the object into a representation vector of the interaction of the user and the object. In this embodiment, a hadamard product is used to embed and integrate the expression vectors, and the specific formula is as follows.
eu=Pu⊙mu
ei=Pi⊙mi
Wherein e isuAnd eiRespectively representing final representation vectors of the user u and the article i after the Hadamard product operation; as indicates a hadamard product.
Therefore, after the steps of S11-S13, the final representation vector of the user u and the item i is obtained, and the vector can be used for further processing in a subsequent decoupling and memory map collaborative filtering network.
In this embodiment, the specific steps for implementing S2 are as follows:
s21: for the user and article interaction graph, each user and article is divided into K blocks by using a decoupling graph neural network, the divided K blocks output an intention of a decoupling representation consisting of K independent components corresponding to a user u, wherein the K-th intention decoupling representation of the user u is as follows:
where K is 1, …, K indicates the K-th intention, ekuIs a decoupled representation of the kth potential intention of user u, ekiIs a decoupled representation of the kth potential intention of item i, l represents the number of aggregation layers, g represents the aggregation function, NuRepresenting a first order neighbor set of user u.
S22: decoupling each intention of the user item graph, and adding and combining decoupling characteristics obtained after the kth intention of the user u is output to obtain a potential intention decoupling representation e of aggregating all neighbor characteristics of the user u under the kth intentionkuFor the same reason, can obtain ekiExpressed as:
where L represents the total number of layers in the diagram.
S23: updating the aggregated feature of the intent block decoupling feature of the user item, the intent block decoupling feature of the user itemThe update of (a) is represented as:
wherein T is 1, …, T represents iteration round number, T represents final iteration round number,an embedded representation representing that user u aggregates first order neighbors under the kth intent in the tth iteration,representing historical input of item I, initialization
S24: degree of interaction between user u and item i at k-th intention, i.e. interaction similarity Item history input with kth intentUnder interactive conditions, it can be aggregated into an embedded representation of the user u aggregating first-order neighbors under the kth intention in the tth iterationExpressed as:
wherein the content of the first and second substances,respectively representing the values of the user u and the item i,after K iterations is
The S21-S24 form the basic framework of the neural network of the decoupling graph. The number of intents k in this framework is 1 or 2 or 4 or 6 or 8 or 16, which is adjusted as the case may be, and in the present embodiment, the optimum number of intents k is 4.
The number of layers of the neural network of the decoupling graph adopts a multi-layer network form to obtain the output of each layer of intention, and after the iteration of the neural network of the multi-layer decoupling graph is finished, the decoupling intents of each layer are added, so that all intention characteristics are obtained. In this embodiment, the most number of layers L is set to 3.
In this embodiment, the specific steps for implementing S3 are as follows:
s31: kth intent decoupling representation of user uPolymerized by a polymerization function g. The formula is known from S21 and is expressed asThe detailed representation is as follows:
where LeakyReLu (-) is the activation function, fGRU(. cndot.) is a gated cycle unit.
S32: the gate control circulation unit GRU controls the stored history information quantity by using the reset gate, the update gate determines the quantity of the forgetting information, the latest state information is added, and finally the output of the GRU is obtained, and the formula f thereofGRU(. cndot.) is shown below.
Wherein, Wr,Wz,WeIs a weight matrix;an output state representing the kth intention of the user u at the previous time;output representing a time instant on all first-order item neighbors of user u; [,]representing the stitching of two vectors; as indicates a hadamard product.
From the above, the steps S31 to S32 constitute the basic structure of the polymeric layer. In the aggregation process, the problem of high-order information loss and node-node information loss easily occurs, so that the user and the article node cannot completely acquire complete information representation from high-order neighbors. After a gating cycle unit is introduced, the nodes are embedded to show that more complete high-order neighbor information can be captured, and the loss of information transmission between the nodes is reduced.
In this embodiment, the specific steps of implementing S4 are as follows:
s41: an algorithm framework based on decoupling and remembering graph collaborative filtering aims at that different intents can be mutually independent and do not influence each other, so that an independent module is established, and loss among intents is reducedintendExpressed as:
wherein the content of the first and second substances,the representation is embedded as the intent of all users and items, COV (-) as the inter-matrix distance covariance, and VAR (-) as the distance variance of each matrix.
From the above step of S41, the entire intent network builds independent modules such that each intent is independent of the other, reducing redundancy.
In this embodiment, the specific steps of implementing S5 are as follows:
s51: establishing a function for predicting interaction possibility between user and article according to the obtained final representation of user and article, wherein the prediction functionExpressed as:
s52: constructing a BPR loss function to optimize a model, predicting the corresponding situation of the user intention and the article, wherein the BPR loss function is expressed as:
wherein O { (u, i, j) | (u, i) ∈ O+,(u,j)∈O-Denotes the interacted training data O+Training data O corresponding to undetected data-The collection of (1) | · | non-conducting phosphor2Represents the L2 regular term, and σ represents the sigmoid activation function.
In summary, in the present embodiment, the steps S1 to S5 provide a decoupling and remembered graph collaborative filtering recommendation algorithm, which can reduce adverse factors such as noise in the historical interaction between the user and the object, and learn a more robust and interpretable embedded representation. In addition, the model can learn more complete high-order neighbor information among the nodes, reduce loss in the process of aggregation and during high-order information transmission, and improve generalization capability of the model.
Claims (8)
1. A recommendation method for graph collaborative filtering based on decoupling and memory is characterized by comprising the following steps:
s1: representing the user and the article by using ID and history interaction to obtain a representation vector of the user and the article, embedding the representation vector into an integration layer, and fusing the representation vectors into a user and article embedded representation vector;
s2: obtaining a decoupling representation of the user about potential intention and updating intention blocking decoupling characteristics of the user object after the aggregation characteristics by each embedded user and object vector through a collaborative filtering framework based on a decoupling graph neural network;
s3: in the integration layer, a gating cycle unit is introduced, and complete characteristic information is aggregated from high-order neighbors of user and article nodes;
s4: introducing an independent module to encourage independence between different intents;
s5: and constructing a loss function to optimize the model and predicting the corresponding situation of the user intention and the article.
2. The method for recommending graph collaborative filtering based on decoupling and memory according to claim 1, wherein the step S1 is as follows:
s11: using the ID and history interaction of the user and the article as original input, and performing one-hot coding to obtain a representation vector of the user and the article;
s12: linearly embedding the high-dimensional sparse representation vector to obtain a low-dimensional dense representation vector;
s13: and integrating the user and the object into a representation vector of the interaction of the user and the object by utilizing the integration layer.
3. The method for recommending graph collaborative filtering based on decoupling and memory according to claim 1, wherein the step S2 is as follows:
s21: for the user and article interaction graph, each user and article is divided into K blocks by using a decoupling graph neural network, the divided K blocks output an intention of a decoupling representation consisting of K independent components corresponding to a user u, wherein the K-th intention decoupling representation of the user u is as follows:
where K is 1, …, K indicates the K-th intention, ekuIs a decoupled representation of the kth potential intention of user u, ekiIs a decoupled representation of the kth potential intention of item i, l represents the number of aggregation layers, g represents the aggregation function, NuRepresenting a first order neighbor set of user u.
S22: decoupling each intention of the user item graph, and adding and combining decoupling characteristics obtained after the kth intention of the user u is output to obtain a potential intention decoupling representation e of aggregating all neighbor characteristics of the user u under the kth intentionkuFor the same reason, can obtain ekiExpressed as:
where L represents the total number of layers in the diagram.
S23: updating the aggregated feature of the intent block decoupling feature of the user item, the intent block decoupling feature of the user itemThe update of (a) is represented as:
wherein T is 1, …, T represents iteration round number, T represents final iteration round number,an embedded representation representing that user u aggregates first order neighbors under the kth intent in the tth iteration,representing historical input of item I, initialization
S24: degree of interaction between user u and item i at k-th intention, i.e. interaction similarityItem history input with kth intentUnder interactive conditions, it can be aggregated into an embedded representation of the user u aggregating first-order neighbors under the kth intention in the tth iterationExpressed as:
4. The recommendation algorithm based on the graph collaborative filtering for decoupling and memorizing as claimed in claim 1, wherein the step S3 is as follows:
s31: kth intent decoupling representation of user uPolymerized by a polymerization function g. The formula is known from S21 and is expressed asThe detailed representation is as follows:
where LeakyReLu (-) is the activation function, fGRU(. cndot.) is a gated cycle unit.
5. The recommendation algorithm based on the graph collaborative filtering for decoupling and memorizing as claimed in claim 1, wherein the step S4 is as follows:
s41: an algorithm framework based on decoupling and remembering graph collaborative filtering aims at that different intents can be mutually independent and do not influence each other, so that an independent module is established, and loss among intents is reducedintendExpressed as:
6. The recommendation algorithm based on the graph collaborative filtering for decoupling and memorizing as claimed in claim 1, wherein the step S5 is as follows:
s51: establishing a function for predicting interaction possibility between user and article according to the obtained final representation of user and article, wherein the prediction functionExpressed as:
s52: constructing a BPR loss function to optimize a model, predicting the corresponding situation of the user intention and the article, wherein the BPR loss function is expressed as:
wherein O { (u, i, j) | (u, i) ∈ O+,(u,j)∈O-Denotes the interacted training data O+Training data O corresponding to undetected data-The collection of (1) | · | non-conducting phosphor2Represents the L2 regular term and σ represents the siqmoid activation function.
7. The recommendation algorithm based on graph collaborative filtering for decoupling and memory according to claim 2, wherein k ═ 1, 2, 4, 8, 16.
8. The recommendation algorithm for graph collaborative filtering based on decoupling and memory according to claim 2, wherein L ═ 0, 1, 2, 3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110761998.7A CN113378062A (en) | 2021-07-06 | 2021-07-06 | Graph collaborative filtering recommendation method based on decoupling and memory |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110761998.7A CN113378062A (en) | 2021-07-06 | 2021-07-06 | Graph collaborative filtering recommendation method based on decoupling and memory |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113378062A true CN113378062A (en) | 2021-09-10 |
Family
ID=77581053
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110761998.7A Pending CN113378062A (en) | 2021-07-06 | 2021-07-06 | Graph collaborative filtering recommendation method based on decoupling and memory |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113378062A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113722603A (en) * | 2021-11-02 | 2021-11-30 | 阿里巴巴达摩院(杭州)科技有限公司 | Object pushing method, product pushing method, computer terminal and storage medium |
-
2021
- 2021-07-06 CN CN202110761998.7A patent/CN113378062A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113722603A (en) * | 2021-11-02 | 2021-11-30 | 阿里巴巴达摩院(杭州)科技有限公司 | Object pushing method, product pushing method, computer terminal and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112905900B (en) | Collaborative filtering recommendation method based on graph convolution attention mechanism | |
Sun et al. | What and how: generalized lifelong spectral clustering via dual memory | |
CN112989064B (en) | Recommendation method for aggregating knowledge graph neural network and self-adaptive attention | |
CN111310063B (en) | Neural network-based article recommendation method for memory perception gated factorization machine | |
CN112232925A (en) | Method for carrying out personalized recommendation on commodities by fusing knowledge maps | |
CN110866145B (en) | Co-preference-assisted deep single-class collaborative filtering recommendation method | |
CN111881342A (en) | Recommendation method based on graph twin network | |
CN110955826B (en) | Recommendation system based on improved cyclic neural network unit | |
CN111611488B (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
Nguyen et al. | Extendable neural matrix completion | |
WO2022166115A1 (en) | Recommendation system with adaptive thresholds for neighborhood selection | |
CN115964568A (en) | Personalized recommendation method based on edge cache | |
CN112632296A (en) | Knowledge graph-based paper recommendation method and system with interpretability and terminal | |
CN113378062A (en) | Graph collaborative filtering recommendation method based on decoupling and memory | |
Sheng et al. | Personalized recommendation of location-based services using spatio-temporal-aware long and short term neural network | |
CN116821519A (en) | Intelligent recommendation method for system filtering and noise reduction based on graph structure | |
CN116434347A (en) | Skeleton sequence identification method and system based on mask pattern self-encoder | |
CN114996584B (en) | Diversity perception interaction recommendation method based on deep reinforcement learning | |
CN116452293A (en) | Deep learning recommendation method and system integrating audience characteristics of articles | |
CN113343121B (en) | Lightweight graph convolution collaborative filtering recommendation method based on multi-granularity popularity characteristics | |
Gao et al. | Deep learning with consumer preferences for recommender system | |
Flores-Munguía et al. | Leveraging a neuroevolutionary approach for classifying violent behavior in video | |
CN115203557A (en) | Method, device, equipment, storage medium and product for generating content recommendation model | |
CN113449193A (en) | Information recommendation method and device based on multi-classification images | |
CN116932862A (en) | Cold start object recommendation method, cold start object recommendation device, computer equipment and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication |