CN112800326B - Improved Ripp-MKR recommendation method combining multitask learning and knowledge graph - Google Patents
Improved Ripp-MKR recommendation method combining multitask learning and knowledge graph Download PDFInfo
- Publication number
- CN112800326B CN112800326B CN202110060998.4A CN202110060998A CN112800326B CN 112800326 B CN112800326 B CN 112800326B CN 202110060998 A CN202110060998 A CN 202110060998A CN 112800326 B CN112800326 B CN 112800326B
- Authority
- CN
- China
- Prior art keywords
- user
- item
- vector
- knowledge graph
- project
- 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.)
- Active
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- 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
-
- 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
Abstract
The invention discloses an improved Ripp-MKR recommendation method combining multitask learning and a knowledge graph, which comprises the following steps: step one, inputting user click history, a knowledge graph and an initialization vector of a project; performing recursion aiming at the user click history and the knowledge graph to acquire the historical interest attribute and the weight of the user; weighting the feature vectors of all tails to obtain primary response; step four, weighting and summing all responses to obtain a user vector matrix; fifthly, cross training of the project vector matrix and the knowledge map head vector is carried out; step six, iteratively updating the user vector, the project vector and the head and tail vectors of the knowledge graph; and seventhly, learning a loss function. Has the advantages that: and automatically discovering the level potential interest of the user, and connecting the recommendation module and the KGE module through a specially designed cross and compression unit. And adopting a multi-task learning framework to alternately learn.
Description
Technical Field
The invention relates to a Ripp-MKR recommendation method, in particular to an improved Ripp-MKR recommendation method combining multitask learning and a knowledge graph.
Background
Currently, recommendation systems are known as growth engines for the internet. A better recommendation system model can facilitate users to efficiently acquire high-interest information under the condition of information overload, improve the user conversion rate of products and achieve the aim of continuously increasing the enterprise operation target. In order to improve the accuracy of the recommendation system, a large amount of side information is added into a model of the recommendation system, hidden contents of the information are extracted, and the association between a user and an item is enhanced. Side information may be understood as auxiliary information such as item attributes, item reviews, user social networks, etc. For example, products of the same brand or category should be more similar when side information is used to understand product embedding. Recommendation methods can be divided into two categories according to the data type: the first category is recommendation methods based on user behavior data, also known as collaborative filtering. Collaborative filtering can be divided into two categories, memory-based and model-based. Representative of these memory-based collaborative filtering methods are UserCF (user-based) and ItemCF (item-based), which function to directly compute the similarity of user-user or item-item to behavioral data. The representative method of model-based collaborative filtering mainly includes some implicit variable models, such as SVD, matrix decomposition (MF), etc. These models use the behavioral data to compute implicit vectors for users and items, and then compute user-user or item-item matches to make recommendations. In the second category of methods, the most common model is the CTR model. The CTR model is essentially a binary classifier, typically using LR, XgBoost, lightGBM and other classifiers. For the two types of models, different side information is used to improve the accuracy of the recommendation. For the first method, in addition to the behavior data of the user, portrait data of the user and the subject, such as sex, age, region, label, category, title, text, etc., may be used. The second category of methods uses behavioral data and side information to construct features and classification criteria of training samples.
In addition to the above attribute data, other data structure information may be used as side information, such as social networks, attributes, multimedia (e.g., text, images, Knowledge Graphs (KGs), etc. since the user set, project set, and user scoring matrix of the recommendation system are highly integrated and related to the knowledge graphs, many trainees are interested in studying the knowledge graphs as a recommendation model for side information.
To obtain the potential information in the knowledge graph and maximize the information content of the knowledge graph, we use the Ripp-MKR model, which combines the preference propagation idea of RippleNet and the cross-training idea of the MKR model. And combining the KG side information with the user interaction historical information to represent the feature vector of the user, and training the maximum KG mining of the hidden information by taking the KG side information as an embedded item. Ripp-MKR differs from the existing literature mainly in that Ripp-MKR adopts two modes of joint learning and alternate learning in the learning process of the characteristics of the knowledge graph applied to the recommendation system: (1) the KGE method is integrated into the recommendation through preference propagation, and the feature vector of the user is represented by information and preference information. (2) The feature learning and recommendation algorithm of the knowledge graph is regarded as two independent but related tasks, and a multi-task learning framework is adopted for alternate learning.
Disclosure of Invention
The invention mainly aims to provide an improved Ripp-MKR recommendation method combining multitask learning and a knowledge graph for acquiring potential information in the knowledge graph and maximizing the information content of the knowledge graph.
The invention provides an improved Ripp-MKR recommendation method combining multitask learning and a knowledge graph, which comprises the following steps:
step one, inputting user click history, a knowledge graph and an initialization vector of a project:
the recommendation system comprises a user set U and an item set I, wherein the user set is represented as: U-U1, U2 … … un, and the item set is denoted as I-I1, I2 … … im, and the user item interaction set may be denoted as: y belongs to Rm multiplied by n, wherein n users and m items exist, when yuv belongs to Y as 1, the user u participates in the item v, otherwise, the value is 0, the user u does not participate in the item v, the knowledge graph G is represented by a triple (h, R, t), wherein h belongs to E, R belongs to R, t belongs to the entity set, the relation R belongs to the relation set, in the recommendation system combining the knowledge graph, the interaction set Y of the user items and the knowledge graph G are given, the purpose is to obtain whether the user u is interested in the non-interactive item v, and the formula representation is that:whereinRepresenting the probability of the user u interacting with the item v, and theta represents the model parameter of the function F;
step two, recursion is carried out aiming at the user click history and the knowledge graph, and the historical interest attribute and the weight of the user are obtained:
the method comprises the following steps that an existing knowledge graph G and a user project history interaction matrix Y are represented by a knowledge graph set as described above, the interaction matrix is combined with a knowledge graph triple, the interaction history Vu and items V of a user U are used as input, a seed set is obtained from the history click record of the user U, and the items are used as the existing preference information of the user, so that the following logic specifically exists:
that is, for any historical interaction of the user u, the historical interaction belongs to a set of historical click items of the user and also belongs to a set of items, the fact of any historical interaction of the user u is an item, in the knowledge graph, any item v of the item can be set as a head of a triple, and a tail of the triple is an item attribute, so that any historical interaction of the user u can be converted into a set of tails, that is, a set of item attributes, and the following relations exist:
the set of attribute feature vectors of the kth user u along the knowledge graph and the historical interaction set is obtained similarly as follows:
for each (h, r, t) in the set, multiplying the item V by (h × r) yields the item V and each (h, r, t) in 1-hopi,ri) Then P of the correlation score by SoftMaxiNormalization is carried out, and the correlation calculation formula is as follows:
step three, weighting the feature vectors of all tails to obtain a first-level response:
weighting all tail vectors to represent the interest result of the user u after 1 jump, which is specifically as follows:
step four, weighting and summing all responses to obtain a user vector matrix:
finally, summing all the hop counts to obtain a vector representation of the user, which is specifically as follows:
step five, performing cross training of the project vector matrix and the knowledge map head vector:
initializing all the project vectors, performing one-hot coded representation on all the projects at the beginning of training, initializing the user vectors temporarily through the projects, and performing iterative modification on the user vectors by combining an interaction matrix and a knowledge graph; and performing cross training through the project vector and the initialized knowledge map triple head vector, and performing updating iteration. Each update of the item vector updates the user vector, which is expressed as follows:
VL=Ee~S(v)[CL(v,e)[v]] (7)
s (v) is a set of project v association entities, and after potential features of a user u and a project v are obtained, the final prediction probability of the user u participating in the project v can be obtained through a prediction function;
in the aspect of knowledge graph, in a KGE module of Ripp-MKR, a head and a relation are used as input, a multi-layer perceptron MLP, a cross unit and a compression unit are respectively used for extracting features of the head and the relation, a head of a KG ternary is embedded into an item ID corresponding to a recommendation system, a relation is embedded into the item attribute, a tail of the KG ternary is embedded into a specific item attribute value, and the process of obtaining a k-layer MLP for predicting a tail t is as follows:
hL=Ev~s(h)[CL(v,h)[e]]
rL=ML(r)
s (h) represents a set of h in the knowledge graph, v represents item data corresponding to h in the knowledge graph, CLIs a cross compression unit, M(x)σ (Wx + b) is a fully-connected neural network layer, W represents weights, b bias, and a prediction vector t of a nonlinear activation function σ (·) and a knowledge graph tail vector;
the structural unit of the cross compression unit is a link module between the item v and the entity e, and the v vector and the e vector are expressed as a matrix as follows:
projecting the cross feature matrix into a potential representation space, and outputting feature vectors of the next-layer items and entities, wherein the details are as follows:
step six, iteratively updating the user vector, the project vector and the head and tail vectors of the knowledge graph:
the prediction formulas of the recommendation module and the knowledge graph module are obtained through the formulas, and the prediction formulas are as follows:
the click probability formula for the recommendation module is shown in equation 11, and the predicted click formula can be finally expressed as:
in the knowledge-graph unit, the prediction formula for the tail t vector is as follows:
hL=Ev~s(h)[CL(v,h)[e]]
rL=ML(r)
step seven, learning a loss function:
the loss function is divided into three parts, namely a loss function of the recommendation module, a loss function of the KGE module and a regularization term for preventing overfitting, and is specifically represented as follows:
the penalty function for the recommendation module is expressed as:
the penalty function for the knowledge-graph module is expressed as:
the regularization term for preventing overfitting is expressed as:
the potential preference of the user for item, namely the prediction matrix, is obtained through the above training.
The invention has the beneficial effects that:
according to the technical scheme provided by the invention, a RippleNet model and an MKR model are used as basic models, the knowledge map information is deeply mined, the expression of a user vector in a traditional recommendation system is banned, the user vector is completely processed by using the information of historical interactive projects, and the potential information of the user is better mined by the method. The present invention proposes Ripp-MKR, a framework that utilizes KGs to assist in the recommendation system. Ripp-MKR finds the potential interest of user's hierarchy automatically by propagating the user's preferences iteratively in units of knowledge-graph triples, and the recommendation module and KGE module are connected by specially designed cross and compression units. The KGE method is integrated into the recommendation through preference propagation, and the feature vector of the user is represented by information and preference information. The feature learning and recommendation algorithm of the knowledge graph is regarded as two independent but related tasks, and a multi-task learning framework is adopted for alternate learning.
Drawings
FIG. 1 is a schematic diagram of the overall operation of the method of the present invention.
FIG. 2 is a schematic diagram of the Ripp-MKR model architecture of the present invention.
FIG. 3 is a schematic diagram of a cross-compression unit in the Ripp-MKR model architecture according to the present invention.
Detailed Description
Please refer to fig. 1 to 3:
initializing a MovieLens-1M data set, wherein the MovieLens-1M data contains explicit feedback data, and the data is converted into implicit feedback data; each entry is labeled 1, indicating a user rating item (MovieLens-1M rating is a threshold of 4), indicating that the user is a positive rating for the movie when the user rating for the movie is greater than or equal to 4, and a negative rating when the user rating for the movie is less than the threshold. A knowledge graph was constructed for each dataset using Microsoft Satori. For MovieLens-1M, a triple subset with a relationship name containing "movie" and a confidence greater than 0.9 is first selected from the entire KG. Given the sub-KG, all valid movie ids are collected by matching the names of all valid movies with the tail of the triplet (tail). For simplicity, items without a matching entity or multiple matching entities are excluded. And then matching the id with the head and tail of all KG triples, selecting all well-matched triples from the sub KGs, and iteratively expanding the entity set to four hops at most.
Step two, setting experiment parameters: the parameter data set are d 16, H2, λ 1 10-7, λ 2 0.01, η 0.02. At Ripp-MKR, the number of advanced layers K is set to 1, and the ratio of training set, validation set, and test set is 6:2:2 for the data set partitioning.
And step three, acquiring interaction triple combinations and potentially interesting attribute combinations of the users by acquiring historical interaction data and knowledge map triples of the users. One-hot encoding of project features is performed.
And step four, continuously updating the vectors of the one-hot codes of the items and the one-hot codes of the head portrait amount in the knowledge graph in an iterative manner through the learning of a cross compression unit.
And step five, taking the loss function as an iteration condition, taking the difference between the click prediction probability in the training set and the predicted tail vector and the actual tail vector as a minimization target, and continuing training.
And step six, acquiring a prediction matrix, namely a preference matrix of the user to the items.
And seventhly, obtaining the values of AUC and ACC after the training is finished so as to judge the quality of the model.
Claims (1)
1. An improved multitask learning and knowledge graph combined Ripp-MKR recommendation method, characterized by: the method comprises the following steps:
step one, inputting user click history, a knowledge graph and an initialization vector of a project:
the recommendation system comprises a user set U and an item set I, wherein the user set is represented as: u ═ U1,u2,.....unIs expressed as V ═ V }, the set of items1,v2.....vmThe interaction matrix of the user and the item is represented as Y, if n users and m items exist, Y is an m × n matrix, scoring the item V by any user U is represented as yuv, yuv belongs to Y, U belongs to U and V belongs to V, when the value of yuv is 1, the user U participates in the item V, otherwise yuv is 0, the user U does not participate in the item V, the knowledge graph G is represented by a triple (h, R, t), h is a head entity, t is a tail entity, h and t both belong to an entity set E, and the relation R belongs to a relation set R, so that the relation is represented as h belongs to E, R belongs to R, and t belongs to E, in a recommendation system combining knowledge graphs, the interaction matrix Y of the user item and the knowledge graph G are given, the goal is to determine whether the user U is interested in the non-interacted item V, and the formula is represented as:
step two, recursion is carried out aiming at the user click history and the knowledge graph, and the historical interest attribute and the weight of the user are obtained:
combining an interaction matrix Y of a user and an item with a knowledge graph triple set, assuming that an item set of historical interaction of the user u is Vu, obtaining a seed set from a historical click record of the user u, using the historical click record of the user as the existing preference information of the user, namely, for any historical interaction item of the user u, the seed set belongs to the set Vu of the historical click item of the user, and also belongs to the item set V, any historical interaction of the user u is an item, the substance of the historical interaction is an item, in the knowledge graph, any item V can be set as a head h of the knowledge graph triple, and a tail t of the triple is an attribute of the item V, so that the historical interaction of any user u can be converted into a tail set, namely, the item attribute set has the following relation:
is the attribute feature vector set obtained after the k-th ripple propagation along the knowledge graph and the historical interaction set of the user u,
triple set obtained by item vector v and user u under 1-hopObtaining the related score of the tuple consisting of the item v in the 1 jump and each head vector and the relation vector; and using a SoftMax function to normalize the correlation score; the correlation calculation formula is as follows:
step three, weighting the feature vectors of all tails to obtain a first-level response:
weighting all tail vectors to represent the interest result of the user u after 1 jump, which is specifically as follows:
step four, weighting and summing all responses to obtain a user vector matrix:
finally, summing all the hop counts to obtain a vector representation of the user, which is specifically as follows:
adding the multi-hop user interest results to obtain a final vector representation result of the user;
step five, performing cross training of the project vector matrix and the knowledge map head vector:
initializing all the project vectors, performing one-hot coded representation on all the projects at the beginning of training, initializing the user vectors through the projects, and performing iterative modification on the user vectors by combining an interaction matrix and a knowledge graph; the user vector is updated every time the project vector is updated through cross training of the project vector and the initialized knowledge map triple head vector, and the project vector is expressed as follows:
VL=Ee~S(v)[CL(v,e)[v]] (7)
s (v) is a set of project v association entities, and after the feature vectors of the user u and the project v are obtained, the final prediction probability of the user u participating in the project v is obtained through a prediction function;
in the aspect of knowledge graph, in a knowledge graph feature extraction module of Ripp-MKR, a head entity h and a corresponding relation r are used as input, a multi-layer perceptron MLP, a cross unit and a compression unit are respectively used for extracting feature vectors of the head entity and the relation, embedding of a knowledge graph triple head corresponds to an item ID in a recommendation system, embedding of the relation corresponds to an item attribute, and embedding of a tail corresponds to a specific item attribute value, and the process of obtaining the multi-layer perceptron with K layers for predicting the tail t is as follows:
the structural unit for the cross-compression unit is a link module between the item v and the entity e
Cross feature matrix clProjecting the data into a potential representation space, and outputting the feature vectors of the next layer of items and entities, wherein the specific steps are as follows:
step six, iteratively updating the user vector, the project vector and the head and tail vectors of the knowledge graph:
the prediction formulas of the recommendation module and the knowledge graph module are obtained by the following formulas, specifically the following
The predicted click formula is ultimately expressed as follows:
step seven, learning a loss function:
the loss function is divided into three parts, namely a loss function of a recommendation module, a loss function of a knowledge graph module and a regularization item for preventing overfitting, and is specifically expressed as follows:
wherein L isRSFor recommending module loss functions, LKGLoss function for knowledge-graph modules, LREGIts regularization term;
The penalty function for the knowledge-graph module is expressed as:
the regularization term for preventing overfitting is expressed as:
the potential preference of the user for item, namely the prediction matrix, is obtained through the above training.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110060998.4A CN112800326B (en) | 2021-01-18 | 2021-01-18 | Improved Ripp-MKR recommendation method combining multitask learning and knowledge graph |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110060998.4A CN112800326B (en) | 2021-01-18 | 2021-01-18 | Improved Ripp-MKR recommendation method combining multitask learning and knowledge graph |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112800326A CN112800326A (en) | 2021-05-14 |
CN112800326B true CN112800326B (en) | 2022-03-15 |
Family
ID=75809992
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110060998.4A Active CN112800326B (en) | 2021-01-18 | 2021-01-18 | Improved Ripp-MKR recommendation method combining multitask learning and knowledge graph |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112800326B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112988135B (en) * | 2021-05-20 | 2021-07-27 | 中国人民解放军国防科技大学 | Task unit recommendation method and device for open source software and computer equipment |
CN113642804A (en) * | 2021-08-27 | 2021-11-12 | 西安交通大学 | Multi-component enhanced family graduate-going prediction and recommendation multitasking method and system |
CN113822776B (en) * | 2021-09-29 | 2023-11-03 | 中国平安财产保险股份有限公司 | Course recommendation method, device, equipment and storage medium |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110188208A (en) * | 2019-06-04 | 2019-08-30 | 河海大学 | A kind of the information resources inquiry recommended method and system of knowledge based map |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
IL243026A0 (en) * | 2015-12-10 | 2016-02-29 | Rokach Lior | Designing context aware recommendations systems based on latent contexts |
US20180052885A1 (en) * | 2016-08-16 | 2018-02-22 | Ebay Inc. | Generating next user prompts in an intelligent online personal assistant multi-turn dialog |
US11042922B2 (en) * | 2018-01-03 | 2021-06-22 | Nec Corporation | Method and system for multimodal recommendations |
CN111522886B (en) * | 2019-01-17 | 2023-05-09 | 中国移动通信有限公司研究院 | Information recommendation method, terminal and storage medium |
CN111061856B (en) * | 2019-06-06 | 2022-05-27 | 北京理工大学 | Knowledge perception-based news recommendation method |
CN111079018A (en) * | 2019-12-19 | 2020-04-28 | 深圳中兴网信科技有限公司 | Exercise personalized recommendation method, exercise personalized recommendation device, exercise personalized recommendation equipment and computer readable storage medium |
CN111538846A (en) * | 2020-04-16 | 2020-08-14 | 武汉大学 | Third-party library recommendation method based on mixed collaborative filtering |
CN111523029B (en) * | 2020-04-20 | 2022-03-25 | 浙江大学 | Personalized recommendation method based on knowledge graph representation learning |
-
2021
- 2021-01-18 CN CN202110060998.4A patent/CN112800326B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110188208A (en) * | 2019-06-04 | 2019-08-30 | 河海大学 | A kind of the information resources inquiry recommended method and system of knowledge based map |
Also Published As
Publication number | Publication date |
---|---|
CN112800326A (en) | 2021-05-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112800326B (en) | Improved Ripp-MKR recommendation method combining multitask learning and knowledge graph | |
CN108648049B (en) | Sequence recommendation method based on user behavior difference modeling | |
CN107423442B (en) | Application recommendation method and system based on user portrait behavior analysis, storage medium and computer equipment | |
Darban et al. | GHRS: Graph-based hybrid recommendation system with application to movie recommendation | |
CN112232925A (en) | Method for carrying out personalized recommendation on commodities by fusing knowledge maps | |
CN111339433B (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
CN111428147A (en) | Social recommendation method of heterogeneous graph volume network combining social and interest information | |
CN107562795A (en) | Recommendation method and device based on Heterogeneous Information network | |
CN111611488B (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
CN112950324B (en) | Knowledge graph assisted pairwise sorting personalized merchant recommendation method and system | |
CN108595533B (en) | Article recommendation method based on collaborative filtering, storage medium and server | |
CN112487199B (en) | User characteristic prediction method based on user purchasing behavior | |
CN112487200B (en) | Improved deep recommendation method containing multi-side information and multi-task learning | |
Zarzour et al. | RecDNNing: a recommender system using deep neural network with user and item embeddings | |
CN112085525A (en) | User network purchasing behavior prediction research method based on hybrid model | |
CN113850649A (en) | Customized recommendation method and recommendation system based on multi-platform user data | |
CN112231583A (en) | E-commerce recommendation method based on dynamic interest group identification and generation of countermeasure network | |
CN112765461A (en) | Session recommendation method based on multi-interest capsule network | |
CN114065048A (en) | Article recommendation method based on multi-different-pattern neural network | |
Yin et al. | An efficient recommendation algorithm based on heterogeneous information network | |
CN115618128A (en) | Collaborative filtering recommendation system and method based on graph attention neural network | |
CN113051468B (en) | Movie recommendation method and system based on knowledge graph and reinforcement learning | |
CN112883289B (en) | PMF recommendation method based on social trust and tag semantic similarity | |
Zhang et al. | Knowledge graph driven recommendation model of graph neural network | |
CN112559877A (en) | CTR (China railway) estimation method and system based on cross-platform heterogeneous data and behavior context |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |