CN108829763B - Deep neural network-based attribute prediction method for film evaluation website users - Google Patents

Deep neural network-based attribute prediction method for film evaluation website users Download PDF

Info

Publication number
CN108829763B
CN108829763B CN201810524543.1A CN201810524543A CN108829763B CN 108829763 B CN108829763 B CN 108829763B CN 201810524543 A CN201810524543 A CN 201810524543A CN 108829763 B CN108829763 B CN 108829763B
Authority
CN
China
Prior art keywords
user
task
neural network
deep neural
matrix
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
Application number
CN201810524543.1A
Other languages
Chinese (zh)
Other versions
CN108829763A (en
Inventor
屈鸿
刘永胜
房展
舒扬
杨舰
邓悟
季江舟
张晓敏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN201810524543.1A priority Critical patent/CN108829763B/en
Publication of CN108829763A publication Critical patent/CN108829763A/en
Application granted granted Critical
Publication of CN108829763B publication Critical patent/CN108829763B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses an attribute prediction method for film evaluation website users based on a deep neural network, and belongs to the fields of data mining, machine learning and the like. The method carries out matrix decomposition on sparse user scoring data in the film evaluation website, decomposes the user scoring matrix into a user-feature matrix and a film-feature matrix, partially fills the user vector by utilizing the two matrices, and takes the filled user scoring data as the feature vector of the user. Inputting the feature vector data of a user into a fully-connected deep neural network, and training a network model by combining a multi-task learning technology to obtain a label score of each task; and performing decision analysis according to the scores of the task labels to obtain a final classification result. According to the invention, through a deep neural network technology, automatic attribute prediction is carried out on film evaluation website users, the problem of data sparsity is solved, and the classification accuracy and efficiency are higher.

Description

Deep neural network-based attribute prediction method for film evaluation website users
Technical Field
An attribute prediction method of film evaluation website users based on a deep neural network is used for predicting the preference degree of each attribute-people to films according to film evaluation and belongs to the technical fields of data mining, deep learning, matrix decomposition and the like.
Background
The deep learning is originated from an artificial neural network, and an artificial neural network model can be built by means of the thought of modern neurology and simulating the structure and mechanism of biological cerebral neurons. Deep learning is to establish a plurality of hidden layers between an input layer and an output layer of a network to form a complex neural network, and to optimize and adjust parameters between networks by learning sample data without a label or with a label, so as to solve the classification problem after regression.
The essence of deep learning is that a multi-layer neural network model is built to learn useful features from massive data, and sample data is accurately classified or predicted. Different from shallow learning, deep learning has a deeper network structure, and feature representation of training data in an original space is transformed into a new space with stronger expression capacity through layer-by-layer transformation of data features, so that more accurate classification is realized. Feature learning is an important link in deep learning.
The basic idea is as follows: for a system S with an n-layer network: (S)1,S2,...Sn) I represents the input of the network, O represents the output of the network, and the specific network structure is:
Figure GDA0001709342480000011
if the final objective is to equalize the inputs and outputs, then the parameters of each pair of networks are adjusted in order to bring O as close to I as possible, so that the series-level characteristics S of I are obtained1,S2,...Sn. The output of each layer in the network serves as the input to the next layer to achieve a hierarchical representation of the input data.
Matrix decomposition is an important technology in the field of personalized recommendation, and the mathematical theoretical basis of matrix decomposition algorithm is the transformation of rows and columns of a matrix. The idea on which matrix decomposition is based is: each user and each article have own unique characteristics, and a user-characteristic matrix and an article-characteristic matrix can be decomposed from a scoring matrix of the user for the items by using a matrix decomposition technology, so that the preference characteristics of the user and the characteristics of each article can be directly obtained from scoring data, and the dimensionality of the characteristic matrix is also obtained.
For example, in a movie website, each viewing user has their own preferences for movies, which can be intuitively understood as the types that the user likes: inspirational, comedy, action, science fiction, etc. The user-attribute matrix represents how much the user likes these potential factors, and the more the user likes this type of movie, the larger the corresponding value in the matrix. Similarly, each movie can also be described by these characteristic factors, so the characteristic-item matrix represents the weight of these factors in each movie, i.e., the genre of the movie. If the movie is for a certain type of material, the larger the corresponding values in the matrix to which those materials are mapped. Thus, the user-property matrix and the property-item matrix are combined to obtain a user-item matrix, and each corresponding value in the matrix represents a degree of a user's liking for a movie.
However, conventional matrix decomposition techniques do not accomplish the decomposition of the matrix because each user scores only a small fraction of the movies, which results in an extremely sparse scoring matrix. The commonly used method is to calculate the prediction error by using the existing score, then adjust the parameters by using a random gradient descent method or a least square method to minimize the error, and finally obtain the prediction score of the user on the movie.
In the traditional algorithms such as classification regression in machine learning, the class or the value of a single attribute is predicted through the characteristics of input samples. Namely, a single-task learning method for obtaining a learning function under a unified model, in which case the samples in the training set obtain only information of a single task. In the practical research and application process of the user attributes, more than one attribute is needed to be predicted, and a certain correlation exists between the attributes. Many researches have proved that some attribute prediction tasks with relatively large relevance are learned at the same time, certain related information is shared among all tasks in the learning process, and the learned model can achieve a better learning effect. Most of the existing multi-task learning methods are realized by using the joint probability distribution of a plurality of categories, and the prediction of samples with a plurality of label quantities and category quantities has certain limitations.
Disclosure of Invention
The invention aims to provide an attribute prediction method of film evaluation website users based on a deep neural network; the problem of the prior art that the prediction effect is poor due to the limitation on the number of the labels or the number of the categories of the tasks and the data sparsity is solved.
The technical scheme adopted by the invention is as follows:
a method for predicting attributes of film evaluation website users based on a deep neural network is characterized by comprising the following steps:
(1) decomposing the historical scoring data of the user into a user-feature matrix and a movie-feature matrix based on a matrix decomposition technology, and densely filling a user vector based on the two matrixes, wherein the user vector scores each movie for the user;
(2) carrying out fully-connected deep neural network model training based on the filled user vector, a plurality of attribute structural characteristics of the sample user and a multi-task learning technology to obtain a trained deep neural network model;
(3) and inputting the user vector of the sample user to be predicted into the trained deep neural network model for prediction to obtain the score of each task attribute of the predicted user, thereby obtaining the final user attribute category.
Further, the specific steps of the step (1) are as follows:
(11) based on a matrix decomposition technology, a matrix R of the historical scoring data of the user is decomposed into a user implicit preference matrix P and an article implicit feedback matrix Q, and the formula is as follows:
R＝PTQ
wherein P ∈ Rf×mAnd Q ∈ Rf×nRepresenting the mapping of users and articles in a potential factor space, wherein f is the dimensionality of a potential factor, T represents the transposition of a matrix, m is the number of the users, and n is the number of the movies;
(12) filling each user vector in a fixed quantity according to the user implicit preference matrix P and the article implicit feedback matrix Q, wherein the filling position is the position corresponding to the data with the highest value in the score data predicted by each user vector, and the specific formula of the filling position is as follows:
Figure GDA0001709342480000031
s.t.||D||2＝Z
where Z is the number of elements that need to be converted to nonzero values, indexjIndicating the position of the jth user that needs to be filled with non-zero value elements, D being the set of data that needs to be filled in each user vector, pj,iRepresenting the predicted scores of the j users for the i movies.
Further, the step (2) comprises the following specific steps:
(21) the label of the sample user is represented as a multi-classification label of a plurality of tasks which are combined into a structured vector form, and the label of each task can be represented as the following form:
Figure GDA0001709342480000032
wherein, N (t) is the category number of the t-th task, wherein, the task is the attribute of the user, if the task t belongs to the category corresponding to the l-th label, the corresponding task t belongs to the category corresponding to the l-th label
Figure GDA0001709342480000033
Is 1, otherwise is 0;
the final structured label of the sample is:
Y＝[Y1,Y2,...YT]
in the formula, T represents that T task exists, and Y represents a label of the T task;
(22) dividing a user set consisting of the filled user vectors into a training set and a testing set, and initializing a deep neural network model parameter w;
(23) randomly selecting samples from a training set, and calculating values of each unit of a node output layer and an output layer of a deep neural network model intermediate layer after initialization parameters are calculated, wherein the activation function of the intermediate layer neuron adopts ReLU (ReLU), namely a node output function, and the formula is as follows:
f(x)＝max(0,x)
wherein x is the input to the neuron;
the activation function of the output layer is a softmax function, namely a function of the output layer; for each task t, the formula for the activation value of the kth output neuron is:
Figure GDA0001709342480000041
in the formula, netk、netsIs the net output value of the neuron, and N (t) is the class number of the t-th task. (24) Performing cross entropy on the numerical values of the node output and each unit of the output layer of the middle layer obtained in the step (23) and the final structured label of the sample obtained in the step (21), calculating the total loss by using a cross entropy loss function, stopping training if the total loss value is smaller than a set threshold value or the iteration times are larger than a set maximum iteration time, otherwise, turning to the step (25);
(25) and (4) updating the deep neural network model parameters w in the step (22) by using the SGD, and then turning to the step (23).
Further, the specific steps of the step (3) are as follows:
(31) inputting the user vector of the sample user to be predicted into the trained deep neural network model, and calculating the output of the deep neural network model to obtain the output form [ Y [)1,Y2,...YT]Wherein the classification form of the task t is
Figure GDA0001709342480000042
Each element yi (t)'A probability score representing the sample user belonging to the category label;
(32) for each task t obtained in step (31), the classification Y of the tasktThe method is a multi-label classification, and the label category corresponding to the element with the highest score is selected as the classification result of the sample user under the task
Figure GDA0001709342480000043
The method comprises the following specific steps:
Figure GDA0001709342480000044
combining a plurality of tasks, wherein the final classification result of the sample is as follows:
Figure GDA0001709342480000045
in summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:
1. in the invention, because the user only interacts with part of the movies, the scoring data has high sparsity, and the scoring data of the user is partially filled by adopting a matrix decomposition technology, so that the feature vector of the user is more complete, and the risk of overfitting is avoided to a certain extent;
2. according to the invention, the deep neural network learning theory is utilized to learn data from a deeper level, the characteristic data of the user is expressed more essentially, and finally the decision analysis is carried out on the attribute classification of the user, so that the classification effect is improved.
3. In the invention, the multi-task learning is carried out on a plurality of attributes of the user at the same time, the correlation among different attributes is fully utilized, and the classification accuracy is improved.
Drawings
FIG. 1 is a general flow diagram of user attribute prediction in accordance with the present invention;
FIG. 2 is a flow chart of user vector filling in the present invention;
FIG. 3 is a diagram illustrating a parameter sharing mechanism in multi-task learning according to the present invention;
FIG. 4 is a schematic diagram of a structured representation of user attributes in multi-task learning according to the present invention;
FIG. 5 is a diagram of the overall network structure of the user attribute prediction based on deep neural network learning in the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Referring to fig. 1, a method for predicting attributes of users of a film evaluation website based on a deep neural network first extracts historical rating data of the users from the film evaluation website, wherein the historical rating data can be expressed as the likes or interests of the users in various movies. The method comprises the steps of preprocessing historical scoring data of a user by means of technologies such as matrix decomposition and the like to obtain a high-dimensional user characteristic vector, taking the characteristic vector of the user (the user vector, namely the score of the user on all movies) as the input of a deep neural network, automatically learning the characteristics of the user by combining a plurality of attribute structured representations of the user to obtain the label score of each task of the user, and finally performing classification decision on the attributes of the user. The method comprises the following specific steps:
a method for predicting attributes of film evaluation website users based on a deep neural network comprises the following steps:
(1) decomposing the historical scoring data of the user into a user-feature matrix and a movie-feature matrix based on a matrix decomposition technology, densely filling user vectors based on the two matrices, and then standardizing the matrices, wherein the user vectors are the scores of the user on all movies; referring to fig. 2, in the preprocessing process of the user feature data, the historical scoring data of the user on each movie can reflect the interest degree of the user on the movie, and can be used as the feature vector of the user, so as to predict the attribute information of the user. Because of the limited effort of the user, it is not possible to score all movies in the web site, and a movie that is not scored does not necessarily represent a user's disinterest in the movie. Therefore, it is necessary to predict the possible scoring situation of each user for all movies based on the historical scoring data of the user and the scores of other users.
The method comprises the following specific steps:
(11) based on a matrix decomposition technology, a matrix R of the historical scoring data of the user is decomposed into a user implicit preference matrix P and an article implicit feedback matrix Q, and the formula is as follows:
R＝PTQ
wherein P ∈ Rf×mAnd Q ∈ Rf×nRepresenting the mapping of users and articles in a potential factor space, wherein f is the dimensionality of a potential factor, T represents the transposition of a matrix, m is the number of the users, and n is the number of the movies;
(12) filling each user vector in a fixed quantity according to the user implicit preference matrix P and the article implicit feedback matrix Q, wherein the filling position is the position corresponding to the data with the highest value in the score data predicted by each user vector, and the specific formula of the filling position is as follows:
Figure GDA0001709342480000061
s.t.||D||2＝Z
where Z is the number of elements that need to be converted to nonzero values, indexjIndicating the position of the jth user that needs to be filled with non-zero value elements, D being the set of data that needs to be filled in each user vector, pj,iRepresenting the predicted scores of the j users for the i movies.
(2) Carrying out fully-connected deep neural network model training based on the filled user vector, a plurality of attribute structural characteristics of the sample user and a multi-task learning technology to obtain a trained deep neural network model;
fig. 3 is a parameter sharing mechanism in multi-task learning, in which all tasks share parameters of hidden layers in a network and output layers related to each task are reserved, as shown in fig. 4-5. The hard sharing mechanism reduces the risk of overfitting to some extent. Therefore, the more tasks are learned at the same time, the more tasks the same representation can be captured by the model, so that the overfitting risk on the original task is smaller. The network can adapt to different attributes of users through the training of a many-to-many task model, meanwhile, the generalization capability of extracting features is improved, and the advantages of the network are greatly exerted;
fig. 4 is a structural representation of attributes of a user in multitasking learning, and shows the results of structural representation of three attributes of the user, namely gender, age and occupation.
Fig. 5 is an overall network structure of user attribute prediction based on deep learning, which includes an input layer, two hidden layers and an output layer. The neurons between every two adjacent layers in the network are all connected, and the neurons of all the layers are not connected. The data after standardization processing is directly used as the characteristic vector of a user, high-dimensional user characteristics are input into a hidden layer in a full connection mode, the input characteristic vector of the user is converted into higher-level abstract characteristics through a multi-layer neural network structure, and finally the higher-level abstract characteristics are input into an output layer to serve as a classification result. The input layer neuron number, namely the feature vector dimension of the user, is the total number of the movies in the website, the two hidden layer neuron numbers are 400 and 40 respectively, and the output layer neuron number is related to the total category number of each attribute of the user.
The method comprises the following specific steps:
(21) the label of the sample user is represented as a multi-classification label of a plurality of tasks which are combined into a structured vector form, and the label of each task can be represented as the following form:
Figure GDA0001709342480000062
wherein, N (t) is the category number of the t-th task, wherein, the task is the attribute of the user, if the task t belongs to the category corresponding to the l-th label, the corresponding task t belongs to the category corresponding to the l-th label
Figure GDA0001709342480000063
Is 1, otherwise is 0;
the final structured label of the sample is:
Y＝[Y1,Y2,...YT]
in the formula, T represents that T task exists, and Y represents a label of the T task;
(22) dividing a user set consisting of the filled user vectors into a training set and a testing set, and initializing a deep neural network model parameter w;
(23) randomly selecting samples from a training set, and calculating values of each unit of a node output layer and an output layer of a deep neural network model intermediate layer after initialization parameters are calculated, wherein the activation function of the intermediate layer neuron adopts ReLU (ReLU), namely a node output function, and the formula is as follows:
f(x)＝max(0,x)
wherein x is the input to the neuron;
the activation function of the output layer is a softmax function, namely a function of the output layer; for each task t, the formula for the activation value of the kth output neuron is:
Figure GDA0001709342480000071
in the formula, netk、netsIs the net output value of the neuron, and N (t) is the class number of the t-th task. (24) Performing cross entropy on the numerical values of the node output and each unit of the output layer of the middle layer obtained in the step (23) and the final structured label of the sample obtained in the step (21), calculating the total loss by using a cross entropy loss function, stopping training if the total loss value is smaller than a set threshold value or the iteration times are larger than a set maximum iteration time, otherwise, turning to the step (25);
(25) and (4) updating the deep neural network model parameters w in the step (22) by using the SGD, and then turning to the step (23).
(3) Inputting the user vector of the sample user to be predicted into the trained deep neural network model for prediction to obtain the scores of each task attribute of the predicted user, so as to obtain the final user attribute category, and performing decision analysis according to the task scores (namely label scores) to obtain the final user attribute category. The method comprises the following specific steps:
(31) inputting the user vector of the sample user to be predicted into the trained deep neural network model, and calculating the output of the deep neural network model to obtain the output form [ Y [)1,Y2,...YT]Wherein the classification form of the task t is
Figure GDA0001709342480000072
Each element yi (t)'A probability score representing the sample user belonging to the category label;
(32) for each task t obtained in step (31), the taskClass Y of affairstThe method is a multi-label classification, and the label category corresponding to the element with the highest score is selected as the classification result of the sample user under the task
Figure GDA0001709342480000073
The method comprises the following specific steps:
Figure GDA0001709342480000074
combining a plurality of tasks, wherein the final classification result of the sample is as follows:
Figure GDA0001709342480000081
the grading data of the user to the movie (namely the feature vector of the user) can be abstracted into the interest and hobby vector of the user, and the implicit hobby degrees of the user to the like of praise and collection of the movie can be converted into numerical values to express the features of the user, so that the method can adapt to different operation objects and various data types. Moreover, a plurality of attributes of the user may have some relationship, for example, the gender of the user is related to the occupation of the user, and the age of the user is also related to the occupation of the user, so that in order to better utilize the correlation between the attributes of the user, the invention introduces a multi-task learning method into the user attribute prediction process, and provides a multi-task learning model for learning a plurality of attributes together. In summary, aiming at the problem of sparsity of user data of the film evaluation website, the matrix decomposition technology is used for filling data to obtain relatively dense user feature vectors, the deep neural network model is used for learning and classifying the features of the users, the basic attributes of the users are predicted, and the accuracy of prediction and classification is improved.
The specific implementation is as follows:
(1) the invention is applied to a Movielens1M data set, each attribute and corresponding value of the Movielens data set are shown in Table 1, the number of users of the data set is 6040, and the number of movies is 3952. Firstly, expressing scores of all movies by users in a data set into a user vector form;
table 1 user attributes and corresponding values in Movielens dataset
Figure GDA0001709342480000082
(2) Based on a matrix decomposition technology, a scoring matrix of a user is decomposed into a user-characteristic matrix P and a movie-characteristic matrix Q with a hiding factor dimension of 35, and the formula is as follows:
R＝PTQ
wherein P ∈ R35×6040，Q∈R35×3952；
(3) Predicting a scoring value of each user vector missing part according to the user implicit preference matrix P and the article implicit feedback matrix Q, and selecting the first 15 data with higher prediction scores to fill corresponding positions;
(4) the attribute tags of the sample user are represented in a structured form of a combination of a plurality of attribute tags, and the specific form is shown in fig. 4.
(5) Randomly selecting 90% of sample users as a training set and 10% of sample users as a testing set;
(6) inputting training samples into the fully-connected deep neural network models of the two hidden layers to carry out weight training adjustment to obtain a trained deep neural network model;
(7) and (4) testing the samples in the test set by using the network weight values obtained by training in the step (6), wherein the prediction accuracy rates of the three attributes of the gender, the age and the occupation of the user are respectively 80.28%, 50.98% and 19.6%.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims (2)

1. A method for predicting attributes of film evaluation website users based on a deep neural network is characterized by comprising the following steps:
(1) decomposing the historical scoring data of the user into a user-feature matrix and a movie-feature matrix based on a matrix decomposition technology, and densely filling a user vector based on the two matrixes, wherein the user vector scores each movie for the user;
(2) carrying out fully-connected deep neural network model training based on the filled user vector, a plurality of attribute structural characteristics of the sample user and a multi-task learning technology to obtain a trained deep neural network model;
(3) inputting the user vector of a sample user to be predicted into the trained deep neural network model for prediction to obtain the score of each task attribute of the predicted user, so as to obtain the final user attribute category;
the specific steps of the step (1) are as follows:
(11) based on a matrix decomposition technology, a matrix R of the historical scoring data of the user is decomposed into a user implicit preference matrix P and an article implicit feedback matrix Q, and the formula is as follows:
R＝PTQ
wherein P ∈ Rf×mAnd Q ∈ Rf×nRepresenting the mapping of users and articles in a potential factor space, wherein f is the dimensionality of a potential factor, T represents the transposition of a matrix, m is the number of the users, and n is the number of the movies;
(12) filling each user vector in a fixed quantity according to the user implicit preference matrix P and the article implicit feedback matrix Q, wherein the filling position is the position corresponding to the data with the highest value in the score data predicted by each user vector, and the specific formula of the filling position is as follows:
Figure FDA0003109863690000011
s.t.||D||2＝Z
where Z is the number of elements that need to be converted to nonzero values, indexjIndicating the position of the jth user that needs to be filled with non-zero value elements, D being the set of data that needs to be filled in each user vector, pj,iRepresents the predictedScoring of i movies by j users
The step (2) comprises the following specific steps:
(21) the label of the sample user is represented as a multi-classification label of a plurality of tasks which are combined into a structured vector form, and the label of each task can be represented as the following form:
Figure FDA0003109863690000012
wherein, N (t) is the category number of the t-th task, wherein, the task is the attribute of the user, if the task t belongs to the category corresponding to the l-th label, the corresponding task t belongs to the category corresponding to the l-th label
Figure FDA0003109863690000013
Is 1, otherwise is 0;
the final structured label of the sample is:
Y＝[Y1,Y2,...YT]
in the formula, T represents that T task exists, and Y represents a label of the T task;
(22) dividing a user set consisting of the filled user vectors into a training set and a testing set, and initializing a deep neural network model parameter w;
(23) randomly selecting samples from a training set, and calculating values of each unit of a node output layer and an output layer of a deep neural network model intermediate layer after initialization parameters are calculated, wherein the activation function of the intermediate layer neuron adopts ReLU (ReLU), namely a node output function, and the formula is as follows:
f(x)＝max(0,x)
wherein x is the input to the neuron;
the activation function of the output layer is a softmax function, namely a function of the output layer; for each task t, the formula for the activation value of the kth output neuron is:
Figure FDA0003109863690000021
in the formula, netk、netsIs the net output value of the neuron, and N (t) is the category number of the t task; (24) performing cross entropy on the numerical values of the node output and each unit of the output layer of the middle layer obtained in the step (23) and the final structured label of the sample obtained in the step (21), calculating the total loss by using a cross entropy loss function, stopping training if the total loss value is smaller than a set threshold value or the iteration times are larger than a set maximum iteration time, otherwise, turning to the step (25);
(25) and (4) updating the deep neural network model parameters w in the step (22) by using the SGD, and then turning to the step (23).
2. The method for predicting the attributes of the users of the film review website based on the deep neural network as claimed in claim 1, wherein the specific steps of the step (3) are as follows:
(31) inputting the user vector of the sample user to be predicted into the trained deep neural network model, and calculating the output of the deep neural network model to obtain the output form [ Y [)1,Y2,...YT]Wherein the classification form of the task t is
Figure FDA0003109863690000022
Each element
Figure FDA0003109863690000023
A probability score representing the sample user belonging to the category label;
(32) for each task t obtained in step (31), the classification Y of the tasktThe method is a multi-label classification, and the label category corresponding to the element with the highest score is selected as the classification result of the sample user under the task
Figure FDA0003109863690000024
The method comprises the following specific steps:
Figure FDA0003109863690000025
combining a plurality of tasks, wherein the final classification result of the sample is as follows:
Figure FDA0003109863690000026
CN201810524543.1A 2018-05-28 2018-05-28 Deep neural network-based attribute prediction method for film evaluation website users Active CN108829763B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810524543.1A CN108829763B (en) 2018-05-28 2018-05-28 Deep neural network-based attribute prediction method for film evaluation website users

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810524543.1A CN108829763B (en) 2018-05-28 2018-05-28 Deep neural network-based attribute prediction method for film evaluation website users

Publications (2)

Publication Number Publication Date
CN108829763A CN108829763A (en) 2018-11-16
CN108829763B true CN108829763B (en) 2021-09-07

Family

ID=64146384

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810524543.1A Active CN108829763B (en) 2018-05-28 2018-05-28 Deep neural network-based attribute prediction method for film evaluation website users

Country Status (1)

Country Link
CN (1) CN108829763B (en)

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108520155B (en) * 2018-04-11 2020-04-28 大连理工大学 Vehicle behavior simulation method based on neural network
CN109800424B (en) * 2018-12-21 2023-05-02 广东工业大学 Recommendation method based on improved matrix decomposition and cross-channel convolutional neural network
CN111523005B (en) * 2019-02-02 2023-08-11 北京嘀嘀无限科技发展有限公司 Network contract user analysis method and device and electronic equipment
CN109933602B (en) * 2019-02-28 2021-05-04 武汉大学 Method and device for converting natural language and structured query language
CN111832782B (en) * 2019-04-17 2024-06-18 北京京东振世信息技术有限公司 Method and device for determining physical distribution properties of articles
CN110096526A (en) * 2019-04-30 2019-08-06 秒针信息技术有限公司 A kind of prediction technique and prediction meanss of user property label
CN110222258A (en) * 2019-05-20 2019-09-10 山东科技大学 Eigenmatrix initial method based on attribute mapping and autocoding neural network
US11604980B2 (en) * 2019-05-22 2023-03-14 At&T Intellectual Property I, L.P. Targeted crowd sourcing for metadata management across data sets
CN110263236B (en) * 2019-06-06 2022-11-08 太原理工大学 Social network user multi-label classification method based on dynamic multi-view learning model
CN110472817B (en) * 2019-07-03 2023-03-24 西北大学 XGboost integrated credit evaluation system and method combined with deep neural network
CN110633368A (en) * 2019-09-12 2019-12-31 淮阴工学院 Deep learning classification method for early colorectal cancer unstructured data
CN111079015B (en) * 2019-12-17 2021-08-31 腾讯科技（深圳）有限公司 Recommendation method and device, computer equipment and storage medium
CN111104552B (en) * 2019-12-24 2023-06-06 浙江大学 Method for predicting scoring category of movie based on movie structured information and brief introduction
CN111160929B (en) * 2019-12-26 2024-02-09 深圳前海微众银行股份有限公司 Method and device for determining client type
CN111639714B (en) * 2020-06-01 2021-07-23 贝壳找房(北京)科技有限公司 Method, device and equipment for determining attributes of users
CN111785328B (en) * 2020-06-12 2021-11-23 中国人民解放军军事科学院军事医学研究院 Coronavirus sequence identification method based on gated cyclic unit neural network
CN115098931B (en) * 2022-07-20 2022-12-16 江苏艾佳家居用品有限公司 Small sample analysis method for mining personalized requirements of indoor design of user
CN117057852B (en) * 2023-10-09 2024-01-26 头流(杭州)网络科技有限公司 Internet marketing system and method based on artificial intelligence technology

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9721202B2 (en) * 2014-02-21 2017-08-01 Adobe Systems Incorporated Non-negative matrix factorization regularized by recurrent neural networks for audio processing
CN106529721B (en) * 2016-11-08 2018-12-25 安徽大学 A kind of ad click rate forecasting system and its prediction technique that depth characteristic is extracted
CN107679250B (en) * 2017-11-01 2020-12-01 浙江工业大学 Multi-task layered image retrieval method based on deep self-coding convolutional neural network
CN108038629A (en) * 2017-12-30 2018-05-15 北京工业大学 A kind of optimization method based on collaborative filtering

Also Published As

Publication number Publication date
CN108829763A (en) 2018-11-16

Similar Documents

Publication Publication Date Title
CN108829763B (en) Deep neural network-based attribute prediction method for film evaluation website users
CN109299396B (en) Convolutional neural network collaborative filtering recommendation method and system fusing attention model
CN111523047B (en) Multi-relation collaborative filtering algorithm based on graph neural network
CN110717098B (en) Meta-path-based context-aware user modeling method and sequence recommendation method
CN111881342A (en) Recommendation method based on graph twin network
Jain et al. A comparative study of machine learning and deep learning techniques for sentiment analysis
Rivas et al. Students performance analysis based on machine learning techniques
CN112967088A (en) Marketing activity prediction model structure and prediction method based on knowledge distillation
Zhou et al. Deep learning modeling for top-n recommendation with interests exploring
CN111949885B (en) Personalized recommendation method for scenic spots
Garg et al. Comparative analysis of various data mining techniques on educational datasets
CN115438732A (en) Cross-domain recommendation method for cold start user based on classification preference migration
CN115186097A (en) Knowledge graph and reinforcement learning based interactive recommendation method
Mohanapriya et al. Comparative study between decision tree and knn of data mining classification technique
Alfarhood et al. DeepHCF: a deep learning based hybrid collaborative filtering approach for recommendation systems
Pham et al. Unsupervised training of Bayesian networks for data clustering
Chou et al. Takagi-Sugeno-Kang type collaborative fuzzy rule based system
CN115577283A (en) Entity classification method and device, electronic equipment and storage medium
Chen et al. Poverty/investment slow distribution effect analysis based on Hopfield neural network
CN111414555A (en) Personalized recommendation method based on collaborative filtering
Duarte et al. Machine Learning and Marketing: A Literature Review.
Sharma et al. Review on text mining algorithms
Zhou et al. Online recommendation based on incremental-input self-organizing map
Zhang et al. Multi-view dynamic heterogeneous information network embedding
Rawat et al. Approaches towards AI-based recommender system

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