CN116489464B - Medical information recommendation method based on heterogeneous double-layer network in 5G application field - Google Patents
Medical information recommendation method based on heterogeneous double-layer network in 5G application field Download PDFInfo
- Publication number
- CN116489464B CN116489464B CN202310388247.4A CN202310388247A CN116489464B CN 116489464 B CN116489464 B CN 116489464B CN 202310388247 A CN202310388247 A CN 202310388247A CN 116489464 B CN116489464 B CN 116489464B
- Authority
- CN
- China
- Prior art keywords
- user
- vector
- layer
- item
- information
- 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
- 238000000034 method Methods 0.000 title claims abstract description 61
- 239000013598 vector Substances 0.000 claims abstract description 141
- 206010063659 Aversion Diseases 0.000 claims abstract description 26
- 230000007246 mechanism Effects 0.000 claims abstract description 8
- 230000006399 behavior Effects 0.000 claims description 85
- 230000006870 function Effects 0.000 claims description 35
- 230000015654 memory Effects 0.000 claims description 27
- 230000004913 activation Effects 0.000 claims description 26
- 238000006243 chemical reaction Methods 0.000 claims description 22
- 239000011159 matrix material Substances 0.000 claims description 19
- 238000012512 characterization method Methods 0.000 claims description 18
- 230000003542 behavioural effect Effects 0.000 claims description 11
- 230000009471 action Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 3
- 239000010410 layer Substances 0.000 description 76
- 239000002355 dual-layer Substances 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000000306 recurrent effect Effects 0.000 description 4
- 230000007787 long-term memory Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000006403 short-term memory Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
Classifications
-
- 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
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
- H04N21/4508—Management of client data or end-user data
- H04N21/4532—Management of client data or end-user data involving end-user characteristics, e.g. viewer profile, preferences
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
- H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
- H04N21/4662—Learning process for intelligent management, e.g. learning user preferences for recommending movies characterized by learning algorithms
- H04N21/4666—Learning process for intelligent management, e.g. learning user preferences for recommending movies characterized by learning algorithms using neural networks, e.g. processing the feedback provided by the user
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
- H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
- H04N21/4667—Processing of monitored end-user data, e.g. trend analysis based on the log file of viewer selections
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/45—Management operations performed by the client for facilitating the reception of or the interaction with the content or administrating data related to the end-user or to the client device itself, e.g. learning user preferences for recommending movies, resolving scheduling conflicts
- H04N21/466—Learning process for intelligent management, e.g. learning user preferences for recommending movies
- H04N21/4668—Learning process for intelligent management, e.g. learning user preferences for recommending movies for recommending content, e.g. movies
Abstract
The invention discloses a medical information recommendation method based on a heterogeneous double-layer network in the field of 5G application, which predicts the probability of clicking a target object by a user based on a heterogeneous behavior sequence of the user. The invention is divided into five parts: modeling a heterogeneous behavior sequence of a user by adopting a heterogeneous double-layer network to obtain hidden vectors of each article in the sequence; the second part is to obtain interest information and aversion information of the user by adopting a target user-based attention mechanism; the third part is to combine the interest information and aversion information of the user to obtain the user vector representation; the fourth part is to predict the click rate of the user on the object; the fifth part is to design a loss function based on model characteristics.
Description
Technical Field
The invention belongs to the technical field of Internet services, and particularly relates to a medical information recommendation method based on a heterogeneous double-layer network in the 5G application field.
Background
The 5G network is gradually replacing the 4G communication network as an emerging network, and the 5G network refers to a fifth generation mobile communication network, and compared with a fourth generation mobile network, the 5G network has the characteristics of high speed, low time delay and the like in the practical application process. The low latency nature of 5G networks has made it popular in the medical arts, such as users increasingly prefer to view medical information at the mobile device. However, the volume of medical information in the medical information platform is too large, and it is difficult for users to find useful information from the medical information platform. And many times, the user does not have obvious purposes when browsing information on the medical information platform, but only wants to send out time. Therefore, the recommendation system is particularly important. The recommendation system captures the interests of the user according to the historical behavior information of the user and predicts the next medical information of interest of the user.
In the medical information platform, the behavior of the user can be classified into click and non-click behaviors. Click behavior refers to the fact that after the platform presents the title and thumbnail information of the medical information to the user, the user clicks the information. The click-through behavior refers to the fact that after the platform presents medical information to the user, the user does not click, but skips the information. In the existing recommendation method, when the clicking and non-clicking behaviors of a user are modeled, the clicking behaviors and the non-clicking behaviors of the user are independently modeled, namely, information which is interested by the user is extracted from a clicking behavior sequence of the user, and information which is not interested is extracted from a non-clicking behavior sequence of the user. The method ignores the integrity of the user behavior, ignores the influence of the non-clicked behavior on the clicked behavior and ignores the influence of the clicked behavior on the non-clicked behavior. For example, a user not clicking on a piece of medical information may not be because the topic is not of interest, but rather because the medical information has been viewed on the relevant topic.
Therefore, the method provides a medical information recommendation method based on the heterogeneous double-layer network applied to the field of 5G application. The method creatively considers the clicking action and the non-clicking action of the user as a whole. The traditional sequence modeling method has a recurrent neural network RNN, but the sequence processed by the recurrent neural network is homogenous, i.e. the items in the sequence are of the same nature. The behavior sequence in the method comprises both clicked items and non-clicked items, and is a heterogeneous behavior sequence. Therefore, the conventional recurrent neural network is not applicable to the present scenario. Heterogeneous dual-layer networks are presented herein to model the heterogeneous behavior sequence, to model information of interest and information of no interest to a user, and to predict the probability of the user clicking on a target item.
Disclosure of Invention
The problem of the method is defined as predicting the probability of a user clicking on a target item based on the heterogeneous behavior sequence of the user. The mathematical symbols involved are: the user set is U, the article setThe sum is V. User u i The heterogeneous behavior sequence of (2) isWherein the subscript l denotes user u i Behavior sequence->The t-th item in the sequence is denoted v t ，v t E V. In the existing recommendation method, when the clicking and non-clicking behaviors of a user are modeled, the clicking behaviors and the non-clicking behaviors of the user are independently modeled, namely, information which is interested by the user is extracted from a clicking behavior sequence of the user, and information which is not interested is extracted from a non-clicking behavior sequence of the user. The method ignores the integrity of the user behavior, ignores the influence of the non-clicked behavior on the clicked behavior and ignores the influence of the clicked behavior on the non-clicked behavior. For example, a user not clicking on a piece of medical information may not be because the topic is not of interest, but rather because the medical information has been viewed on the relevant topic. Therefore, the method provides a medical information recommendation method based on the heterogeneous double-layer network applied to the field of 5G application. The method creatively considers the clicking action and the non-clicking action of the user as a whole. The behavior sequence in the method comprises both clicked items and non-clicked items, and is a heterogeneous behavior sequence. Therefore, the conventional recurrent neural network is not applicable to the present scenario. For this purpose, the invention adopts the following technical scheme:
a medical information recommendation method based on heterogeneous double-layer network in the 5G application field comprises the following steps:
modeling the heterogeneous behavior sequence of the user by adopting a heterogeneous double-layer network to obtain the hidden vector of each article in the sequence. The user set is U and the item set is V. User u i The heterogeneous behavior sequence of (2) isWherein the subscript l denotes user u i Behavior sequence->The t-th item in the sequence is denoted v t ，v t E V. For any article v t Its vector representation is x t . The article herein represents medical information. The heterogeneous behavior sequence of the heterogeneous double-layer network to the user is proposed in the methodModeling is carried out to obtain hidden vectors of each article in the sequence. The heterogeneous double-layer network consists of two layers, wherein the first layer is a full exposure layer and represents all medical information displayed to a user by a platform, and the medical information comprises clicking behaviors and non-clicking behaviors of the user. The second layer is a conversion layer which represents the clicking action of the user, namely, after the platform displays the medical information to the user, the user clicks to see the information. By gate variable g t Representing user u i Is->Medium v t Whether it is clicked by the user. If g t =1, indicating that the user clicks on item v t The method comprises the steps of carrying out a first treatment on the surface of the If g t =0, indicating that the user has not clicked on item v t . The heterogeneous double-layer network updates the hidden vector of the article in the two layers, and the specific steps are as follows:
the first layer is the full exposure layer, and the memory cell and hidden vector update formula of the nodes in the sequence is as follows:
wherein , and />Respectively an input door, a forget door and an output door. /> and />Memory cell and hidden vector, which are current node, < +.> and />Is the memory cell and hidden vector of the previous node. Here-> and />The superscript "1" of (c) indicates a first fully exposed layer. /> and />Is a matrix vector which is a vector of the matrix, and />Is a vector parameter. Sigma is a sigmoid activation function and tanh is a tanh activation function. X is x t Is the item v in the sequence t Is also the input to the current node. Input door formula +.> Middle->The input information representing the current node contains two: item vector x represented by the current node t And the hidden vector of the last node. If the last node is click action, the input of the current node is the hidden vector of the second layer conversion layer of the last node>If the last node is not clicked, the input of the current node is the hidden vector of the first full exposure layer of the last node>
The second layer is a conversion layer, and the memory units and hidden vector updating formulas of the nodes in the sequence are as follows:
wherein , and />Respectively an input door, a forget door and an output door. /> and />Memory cell and hidden vector, which are current node, < +.> and />Is the memory cell and hidden vector of the previous node. The superscript "2" for a parameter in the formula indicates that it is a parameter of the second conversion layer. />Is the input to the current node, from the hidden vector of the first full exposure layer. and />Is a matrix vector, +.> and />Is a vector parameter. Sigma is a sigmoid activation function and tanh is a tanh activation function. Memory cell->And hidden vector->Is expressed by the updated formula of (1), when g t When =1, i.e. the current item is clicked, +.>And hidden vector->Updating information; otherwise, directly copy the memory cell of a node on the layer +.>And hidden vector->
And obtaining interest information and aversion information of the user by adopting an attention mechanism based on the target object. The user heterogeneous behavior sequence is obtained in the last stepThe latent vector of any object in the first full exposure layer and the second conversion layer. For clicked articles, using hidden vectors of a second conversion layer as vector characterization; the article which is not clicked adopts the hidden vector of the first full exposure layer as the vector characterization. The object is v new Its vector representation is x new . The attention mechanism based on the target article is adopted to extract the interest information of the user from the clicked article, and the aversion information of the user is extracted from the non-clicked article. The method comprises the following specific steps:
from heterogeneous behavioral sequences using a target item-based attention mechanismInterest information p of the user is extracted from the clicked item like ：
wherein ,plike For the interest information of the user,representing heterogeneous behavioral sequencesItem set clicked in (2), ->Is the isomerism behavior sequence->Item v of (v) t Hidden vector, x, of second layer conversion layer of heterogeneous double-layer network new Is the target object v new Is described. W (W) 1 and W2 Is a matrix parameter> and />Is a vector parameter that is a function of the vector,superscript->Is the transposed symbol of the vector. Sigma is a sigmoid activation function. Alpha t Is an attention value representing the correlation of the clicked item and the target item.
From heterogeneous behavioural sequences using the same methodThe item not clicked in (1) extracts the aversion information p of the user dislike The method specifically comprises the following steps:
wherein ,pgislike For the aversion information of the user,representing heterogeneous behavioral sequencesItem set not clicked in (a), a->Is the isomerism behavior sequence->Item v of (v) t Hidden vector, x of full exposure layer at first layer of heterogeneous double-layer network new Is the object article x new Is described. W (W) 3 and W4 Is a matrix parameter> and />Is a vector parameter,/->Superscript->Is the transposed symbol of the vector. Sigma is a sigmoid activation function. Beta t Is an attention value representing the correlation of the non-clicked item and the target item.
And combining interest information and aversion information of the user to obtain the user vector characterization. The method adopts a simple vector weighted summation mode to process interest information and aversion information of the user to obtain user vector characterization, and specifically comprises the following steps:
p＝λ c ·p like +λ u ·p dislike
wherein ,λc and λu Representing interest information p like And aversion information p dis/ike Is a hyper-parameter. p is the user vector characterization.
The click rate of the user on the item is predicted. Connecting the user vector representation with the target object vector, and then transmitting the user vector representation into a double-layer perceptron network to predict the click rate of the user on the target object. The method comprises the following steps:
wherein ,xnew Is the target object v new P is the user vector token.Representing the two vectors p and x new And connecting. W (W) 5 Is a matrix parameter, b y and />Is a vector parameter,/->The superscript T of (a) is the transposed symbol of the vector. c y Is a scalar parameter. Sigma is a sigmoid activation function and tanh is a tanh activation function.
The loss function is designed based on model characteristics. Predicted value of click rate of target object by userCalculating predictive value +.>And the true value y, and then updating the model parameters using the error. The cross entropy loss function is adopted to guide the updating process of the model parameters:
where y ε {0,1} is a true value representing whether the user clicked on the target item. Model parameters are updated using Adam optimizer.
The beneficial technical effects of the invention are as follows:
(1) The method creatively considers the clicking behavior and the non-clicking behavior of the user as a whole, and considers the influence of the non-clicking behavior on the clicking behavior and the influence of the clicking behavior on the non-clicking behavior.
(2) The method provides a heterogeneous double-layer network for modeling a heterogeneous behavior sequence of a user. The heterogeneous bilayer network consists of two layers, the first layer being a fully exposed layer and the second layer being a conversion layer.
(3) According to the method, interest information and aversion information of the user are extracted from the heterogeneous behavior sequence of the user at the same time, so that more comprehensive and comprehensive user vector characterization is obtained.
Drawings
FIG. 1 is a flow chart of a medical information recommendation method based on a heterogeneous dual-layer network in the field of 5G application of the present invention;
FIG. 2 is a heterogeneous behavior sequence of a user in a medical information recommendation scenario;
FIG. 3 is a schematic diagram of a heterogeneous dual-layer network according to the present invention.
Detailed Description
In order to further understand the present invention, a medical information recommendation method based on a heterogeneous dual-layer network in the field of 5G application provided by the present invention is specifically described below with reference to the specific embodiments, but the present invention is not limited thereto, and the technical personnel in the field make insubstantial improvements and adjustments under the core guiding concept of the present invention, still fall within the protection scope of the present invention.
The problem of the method is defined as predicting the probability of a user clicking on a target item based on the heterogeneous behavior sequence of the user. The mathematical symbols involved are: the user set is U and the item set is V. User u i The behavior sequence of (2) isWherein the subscript l denotes user u i Behavior sequence->The t-th item in the sequence is denoted v t ，v t E V. For any article v t Its vector representation is x t . The article herein represents medical information. In order to use the more general description, the following articles are used to represent medical information. The method can be applied to the field of medical information recommendation and the recommendation fields of electronic commerce, short videos and the like. A medical information recommendation method based on heterogeneous double-layer network in the 5G application field mainly comprises five parts. Modeling a heterogeneous behavior sequence of a user by adopting a heterogeneous double-layer network to obtain hidden vectors of each article in the sequence; the second part is to obtain interest information and aversion information of the user by adopting a target user-based attention mechanism; the third part is combinedObtaining user vector characterization by interest information and aversion information of a user; the fourth part is to predict the click rate of the user on the object; the fifth part is to design a loss function based on model characteristics. The heterogeneous dual-layer network presented in the first part is the main innovation point of the present invention.
As shown in fig. 1, according to one embodiment of the invention, the method comprises the steps of:
and S100, modeling a heterogeneous behavior sequence of the user by adopting a heterogeneous double-layer network to obtain hidden vectors of each article in the sequence. The user set is U and the item set is V. User u i The heterogeneous behavior sequence of (2) isWherein the subscript f denotes user u i Behavior sequence->The t-th item in the sequence is denoted v t ，v t E V. FIG. 2 shows that the heterogeneous behavior sequence of a certain user is { v 1 ，v 2 ，v 3 ，v 4 ，v 5 ，v 6 ，v 7 ，v 8 ，v 9 Bold circles indicate that the user clicked on item v 3 ，v 4 and v7 . It can be seen that the behavior sequence contains both clicking behavior and non-clicking behavior of the user, so that the behavior sequence is a heterogeneous behavior sequence. For any article v t Its vector representation is x t . For a homogeneous behavior sequence, modeling is typically performed using a long and short term memory network LSTM. If the sequence { v 1 ，v 2 ，…，v l The sequence of homogeneous behaviors is modeled by using a long-short-term memory network LSTM as follows:
i t ＝σ(W i x t +U i h t-1 +b i )
f t ＝σ(W f x t +U f h t-1 +b f )
o t ＝σ(W o x t +U o h t-1 +b o )
c t ＝i t tanh(W c x t +U c h t-1 +b c )+f t c t-1
h t ＝o t tanh(c t )
wherein ,it 、f t and ot Respectively an input door, a forget door and an output door. W (W) i 、U i 、W f 、U f 、W o 、U o 、W c and Uc Is a matrix parameter, b i 、b f 、b o and bc Is a vector parameter. Sigma is a sigmoid activation function and tanh is a tanh activation function. X is x t Is the item v in the sequence t Is also the input of the current node, c t A memory unit h which is the current node t Is the hidden vector of the current node. c t-1 Is the memory cell of the previous node, h t-1 Is the hidden vector of the previous node.
However, long and short term memory networks LSTM are only suitable for homogeneous behavior sequences, which in this scenario are heterogeneous behavior sequences. The heterogeneous behavior sequence of the heterogeneous double-layer network to the user is proposed in the methodModeling is carried out to obtain hidden vectors of each article in the sequence. The heterogeneous double-layer network consists of two layers, wherein the first layer is a full exposure layer and represents all medical information displayed to a user by a platform, and the medical information comprises clicking behaviors and non-clicking behaviors of the user. The second layer is a conversion layer which represents the clicking action of the user, namely, after the platform displays the medical information to the user, the user clicks to see the information. By gate variable g t Representing user u i Is->Medium v t Whether it is clicked by the user. If g t =1, indicating that the user clicks on item v t The method comprises the steps of carrying out a first treatment on the surface of the If g t =0, indicating that the user has not clicked on item v t . As shown in fig. 3, the heterogeneous dual-layer network updates the hidden vector of the object in the two layers, and specifically comprises the following steps:
the first layer is the full exposure layer, and the memory cell and hidden vector update formula of the nodes in the sequence is as follows:
wherein , and />Respectively an input door, a forget door and an output door. /> and />Memory cell and hidden vector, which are current node, < +.> and />Is the memory cell and hidden vector of the previous node. Here-> and />The superscript "1" of (c) indicates a first fully exposed layer. /> and />Is a matrix vector which is a vector of the matrix, and />Is a vector parameter. Sigma is a sigmoid activation function and tanh is a tanh activation function. X is x t Is the item v in the sequence t Is also the input to the current node. Input door formula +.> Middle->The input information representing the current node contains two: item vector x represented by the current node t And the hidden vector of the last node. If the last node is click action, the input of the current node is the hidden vector of the second layer conversion layer of the last node>If the previous node is not clicked, the input of the current node is the hidden vector of the first full exposure layer of the previous node
The second layer is a conversion layer, and the memory units and hidden vector updating formulas of the nodes in the sequence are as follows:
wherein , and />Respectively an input door, a forget door and an output door. /> and />Memory cell and hidden vector, which are current node, < +.> and />Is the memory cell and hidden vector of the previous node. The superscript "2" for a parameter in the formula indicates that it is a parameter of the second conversion layer. />Is the input to the current node, from the hidden vector of the first full exposure layer. and />Is a matrix vector, +.> and />Is a vector parameter. Sigma is a sigmoid activation function and tanh is a tanh activation function. Memory cell->And hidden vector->Is expressed by the updated formula of (1), when g t When =1, i.e. the current item is clicked, +.>And hidden vector->Updating information; otherwise, directly copying the memory unit of the node on the layer/>And hidden vector->
S200, obtaining interest information and aversion information of the user by adopting an attention mechanism based on the target object. The user heterogeneous behavior sequence is obtained in the last stepThe latent vector of any object in the first full exposure layer and the second conversion layer. For clicked articles, using hidden vectors of a second conversion layer as vector characterization; the article which is not clicked adopts the hidden vector of the first full exposure layer as the vector characterization. The object is v new Its vector representation is x new . The attention mechanism based on the target article is adopted to extract the interest information of the user from the clicked article, and the aversion information of the user is extracted from the non-clicked article. The method comprises the following specific steps:
from heterogeneous behavioral sequences using a target item-based attention mechanismInterest information p of the user is extracted from the clicked item like ：
wherein ,plike For the interest information of the user,representing heterogeneous behavioral sequencesItem set clicked in (2), ->Is the isomerism behavior sequence->Item v of (v) t Hidden vector, x, of second layer conversion layer of heterogeneous double-layer network new Is the target object v new Is described. W (W) 1 and W2 Is a matrix parameter> and />Is a vector parameter that is a function of the vector,superscript->Is the transposed symbol of the vector. Sigma is a sigmoid activation function. Alpha t Is an attention value representing the correlation of the clicked item and the target item.
From heterogeneous behavioural sequences using the same methodThe item not clicked in (1) extracts the aversion information p of the user dislike The method specifically comprises the following steps:
wherein ,pdislike For the aversion information of the user,representing heterogeneous behavioral sequencesItem set not clicked in (a), a->Is the isomerism behavior sequence->Item v of (v) t Hidden vector, x of full exposure layer at first layer of heterogeneous double-layer network new Is the target object v new Is described. W (W) 3 and W4 Is a matrix parameter> and />Is a vector parameter,/->Superscript->Is the transposed symbol of the vector. Sigma is a sigmoid activation function. Beta t Is an attention value representing the correlation of the non-clicked item and the target item.
S300, combining interest information and aversion information of the user to obtain user vector characterization. The method adopts a simple vector weighted summation mode to process interest information and aversion information of the user to obtain user vector characterization, and specifically comprises the following steps:
p＝λ c ·p like +λ u ·p dislike
wherein ,λc and λu Representing interest information p like And aversion information p dislike Is a hyper-parameter. In the present method, lambda c ＝λ u =0.5. p is the user vector characterization.
S400, predicting the click rate of the user on the object. Connecting the user vector representation with the target object vector, and then transmitting the user vector representation into a double-layer perceptron network to predict the click rate of the user on the target object. The method comprises the following steps:
wherein ,xnew Is the target object v new P is the user vector token.Representing the two vectors p and x new And connecting. W (W) 5 Is a matrix parameter, b y and />Is a vector parameter,/->Superscript->Is the transposed symbol of the vector. c y Is a scalar parameter. Sigma is a sigmoid activation function and tanh is a tanh activation function.
S500, designing a loss function according to the model characteristics. Predicted value of click rate of target object by userCalculating predictive value +.>And the true value y, and then updating the model parameters using the error. The cross entropy loss function is adopted to guide the updating process of the model parameters:
where y ε {0,1} is a true value representing whether the user clicked on the target item. Model parameters are updated using Adam optimizer.
The foregoing description of the embodiments is provided to facilitate the understanding and application of the invention to those skilled in the art. It will be apparent to those having ordinary skill in the art that various modifications to the above-described embodiments may be readily made and the generic principles described herein may be applied to other embodiments without the use of inventive faculty. Therefore, the present invention is not limited to the above-described embodiments, and those skilled in the art, based on the present disclosure, should make improvements and modifications within the scope of the present invention.
Claims (1)
1.5G application field medical information recommendation method based on heterogeneous double-layer network, which is characterized in that:
modeling a heterogeneous behavior sequence of a user by adopting a heterogeneous double-layer network to obtain hidden vectors of each article in the sequence; the user set is U, and the article set is V; user u i The heterogeneous behavior sequence of (2) isWherein the subscript l denotes user u i Behavior sequence->The t-th item in the sequence is denoted v t ，v t E, V; for any article v t Its vector representation is x t The method comprises the steps of carrying out a first treatment on the surface of the The article herein represents medical information; the heterogeneous behavior sequence of the heterogeneous double-layer network to the user is proposed in the methodModeling is carried out, so that hidden vectors of each article in the sequence are obtained; heterogeneous double-layer network is composed of two layersThe first layer is a full exposure layer, and represents all medical information displayed to a user by a platform, wherein the medical information comprises clicking behaviors and non-clicking behaviors of the user; the second layer is a conversion layer which represents clicking behaviors of the user, namely after the platform displays the medical information to the user, the user clicks to see the information; by gate variable g t Representing user u i Is->Medium v t Whether or not clicked by the user; if g t =1, indicating that the user clicks on item v t The method comprises the steps of carrying out a first treatment on the surface of the If g t =0, indicating that the user has not clicked on item v t The method comprises the steps of carrying out a first treatment on the surface of the The heterogeneous double-layer network updates the hidden vector of the article in the two layers, and the specific steps are as follows:
the first layer is the full exposure layer, and the memory cell and hidden vector update formula of the nodes in the sequence is as follows:
wherein , and />An input door, a forget door and an output door respectively; /> and />Memory cell and hidden vector, which are current node, < +.> and />Is the memory unit and hidden vector of the previous node; here-> and />The superscript "1" of (1) indicates a first fully exposed layer; /> and />Is a matrix vector which is a vector of the matrix, and />Is a vector parameter; sigma is a sigmoid activation function, and tanh is a tanh activation function; x is x t Is the item v in the sequence t Is also an input to the current node; input door formula +.> Middle->The input information representing the current node contains two: item vector x represented by the current node t And the hidden vector of the last node; if the last node is click action, the input of the current node is the hidden vector of the second layer conversion layer of the last node>If the last node is not clicked, the input of the current node is the hidden vector of the first full exposure layer of the last node>
The second layer is a conversion layer, and the memory units and hidden vector updating formulas of the nodes in the sequence are as follows:
wherein , and />An input door, a forget door and an output door respectively; /> and />Memory cell and hidden vector, which are current node, < +.> and />Is the memory unit and hidden vector of the previous node; the superscript "2" for a parameter in the formula indicates that it is a parameter of the second conversion layer; />Is the input of the current node, and comes from the hidden vector of the first full exposure layer; and />Is a matrix vector, +.> and />Is a vector parameter; sigma is a sigmoid activation function, and tanh is a tanh activation function; memory cell->And hidden vector->Is expressed by the updated formula of (1), when g t When =1, i.e. the current item is clicked, +.>And hidden vector->Updating information; otherwise, directly copy the memory cell of a node on the layer +.>And hidden vector->
Obtaining interest information and aversion information of a user by adopting an attention mechanism based on a target object; the user heterogeneous behavior sequence is obtained in the last stepThe hidden vectors of any article in the first full exposure layer and the second conversion layer; for clicked articles, using hidden vectors of a second conversion layer as vector characterization; non-clicked objectThe hidden vector of the first full exposure layer is used as vector representation; the object is v new Its vector representation is x new The method comprises the steps of carrying out a first treatment on the surface of the Extracting interest information of a user from the clicked article and aversion information of the user from the non-clicked article by adopting an attention mechanism based on the target article; the method comprises the following specific steps:
from heterogeneous behavioral sequences using a target item-based attention mechanismInterest information p of the user is extracted from the clicked item like ：
wherein ,plike For the interest information of the user,representing the isomerism behavior sequence->Item set clicked in (2), ->Is the isomerism behavior sequence->Item v of (v) t Hidden vector, x, of second layer conversion layer of heterogeneous double-layer network new Is the target object v new Vector characterization of (2); w (W) + and W2 Is a matrix parameter> and />Is a vector parameter,/->The superscript T of (a) is the transposed symbol of the vector; sigma is a sigmoid activation function; alpha t Is an attention value representing the correlation of the clicked item and the target item;
from heterogeneous behavioural sequences using the same methodThe item not clicked in (1) extracts the aversion information p of the user dislike The method specifically comprises the following steps:
wherein ,pdislike For the aversion information of the user,representing heterogeneous behavioral sequencesItem set not clicked in (a), a->Is the isomerism behavior sequence->Item v of (v) t In the isomerismHidden vector, x of full exposure layer of first layer of double-layer network new Is the target object v new Vector characterization of (2); w (W) 3 and Wk Is a matrix parameter> and />Is a vector parameter,/->The superscript T of (a) is the transposed symbol of the vector; sigma is a sigmoid activation function; beta 3 Is an attention value representing the correlation of the non-clicked item and the target item;
combining interest information and aversion information of the user to obtain user vector characterization; the method adopts a simple vector weighted summation mode to process interest information and aversion information of the user to obtain user vector characterization, and specifically comprises the following steps:
p＝λ c ·p like +γ u ·p dislike
wherein ,γc and γu Representing interest information p like And aversion information p dislike Is a superparameter; p is a user vector characterization;
predicting the click rate of a user on an article; connecting the user vector representation with the target object vector, then transmitting the user vector representation into a double-layer perceptron network, and predicting the click rate of the user on the target object; the method comprises the following steps:
wherein ,xnew Is the target object v new P is the user vector token;representing twoVectors p and x new Are connected; w (W) p Is a matrix parameter, b y and />Is a vector parameter,/->The superscript T of (a) is the transposed symbol of the vector; c y Is a scalar parameter; sigma is a sigmoid activation function, and tanh is a tanh activation function;
designing a loss function according to model characteristics; predicted value of click rate of target object by userCalculating predictive value +.>And the true value y, and updating the model parameters using the error; the cross entropy loss function is adopted to guide the updating process of the model parameters:
wherein y E {0,1} is a true value representing whether the user clicked on the target item; model parameters are updated using Adam optimizer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310388247.4A CN116489464B (en) | 2023-04-12 | 2023-04-12 | Medical information recommendation method based on heterogeneous double-layer network in 5G application field |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310388247.4A CN116489464B (en) | 2023-04-12 | 2023-04-12 | Medical information recommendation method based on heterogeneous double-layer network in 5G application field |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116489464A CN116489464A (en) | 2023-07-25 |
| CN116489464B true CN116489464B (en) | 2023-10-17 |
Family
ID=87211261
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310388247.4A Active CN116489464B (en) | 2023-04-12 | 2023-04-12 | Medical information recommendation method based on heterogeneous double-layer network in 5G application field |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116489464B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112395504A (en) * | 2020-12-01 | 2021-02-23 | 中国计量大学 | Short video click rate prediction method based on sequence capsule network |
| CN112492396A (en) * | 2020-12-08 | 2021-03-12 | 中国计量大学 | Short video click rate prediction method based on fine-grained multi-aspect analysis |
| CN112765461A (en) * | 2021-01-12 | 2021-05-07 | 中国计量大学 | Session recommendation method based on multi-interest capsule network |
| CN115658936A (en) * | 2022-12-29 | 2023-01-31 | 中国传媒大学 | Personalized program recommendation method and system based on double-layer attention model |
| WO2023039681A1 (en) * | 2021-09-20 | 2023-03-23 | Applied Brain Research Inc. | Methods and systems for implicit attention with sub-quadratic complexity in artificial neural networks |
-
2023
- 2023-04-12 CN CN202310388247.4A patent/CN116489464B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112395504A (en) * | 2020-12-01 | 2021-02-23 | 中国计量大学 | Short video click rate prediction method based on sequence capsule network |
| CN112492396A (en) * | 2020-12-08 | 2021-03-12 | 中国计量大学 | Short video click rate prediction method based on fine-grained multi-aspect analysis |
| CN112765461A (en) * | 2021-01-12 | 2021-05-07 | 中国计量大学 | Session recommendation method based on multi-interest capsule network |
| WO2023039681A1 (en) * | 2021-09-20 | 2023-03-23 | Applied Brain Research Inc. | Methods and systems for implicit attention with sub-quadratic complexity in artificial neural networks |
| CN115658936A (en) * | 2022-12-29 | 2023-01-31 | 中国传媒大学 | Personalized program recommendation method and system based on double-layer attention model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116489464A (en) | 2023-07-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Parvin et al. | TCFACO: Trust-aware collaborative filtering method based on ant colony optimization | |
| CN107103000A (en) | It is a kind of based on correlation rule and the integrated recommended technology of Bayesian network | |
| CN112765461A (en) | Session recommendation method based on multi-interest capsule network | |
| CN112256916B (en) | Short video click rate prediction method based on graph capsule network | |
| Wang et al. | Multitask feature learning approach for knowledge graph enhanced recommendations with RippleNet | |
| CN112395504B (en) | Short video click rate prediction method based on sequence capsule network | |
| Zhao et al. | AGRE: A knowledge graph recommendation algorithm based on multiple paths embeddings RNN encoder | |
| Garcia-Vega et al. | Learning from data streams using kernel least-mean-square with multiple kernel-sizes and adaptive step-size | |
| Lin et al. | Knowledge-enhanced recommendation using item embedding and path attention | |
| CN115147192A (en) | Recommendation method and recommendation system based on double-view-angle deviation correction | |
| CN112819575B (en) | Session recommendation method considering repeated purchasing behavior | |
| CN112256918B (en) | Short video click rate prediction method based on multi-mode dynamic routing | |
| CN112199550B (en) | Short video click rate prediction method based on emotion capsule network | |
| CN113704438A (en) | Conversation recommendation method of abnormal picture based on layered attention mechanism | |
| Wang et al. | Session-based recommendation with time-aware neural attention network | |
| CN112307258B (en) | Short video click rate prediction method based on double-layer capsule network | |
| CN116489464B (en) | Medical information recommendation method based on heterogeneous double-layer network in 5G application field | |
| CN116257691A (en) | Recommendation method based on potential graph structure mining and user long-short-term interest fusion | |
| CN113704439B (en) | Conversation recommendation method based on multi-source information heteromorphic graph | |
| CN112905886B (en) | Session recommendation method based on multi-interest repeated network | |
| CN115660147A (en) | Information propagation prediction method and system based on influence modeling between propagation paths and in propagation paths | |
| CN112307257B (en) | Short video click rate prediction method based on multi-information node graph network | |
| CN113704440B (en) | Conversation recommendation method based on path representation in article graph network | |
| CN114036400B (en) | Hypergraph-based collaborative session recommendation method | |
| CN113704627B (en) | Session recommendation method based on time interval graph |
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 |