CN112328900A - Deep learning recommendation method integrating scoring matrix and comment text - Google Patents

Deep learning recommendation method integrating scoring matrix and comment text Download PDF

Info

Publication number
CN112328900A
CN112328900A CN202011357879.7A CN202011357879A CN112328900A CN 112328900 A CN112328900 A CN 112328900A CN 202011357879 A CN202011357879 A CN 202011357879A CN 112328900 A CN112328900 A CN 112328900A
Authority
CN
China
Prior art keywords
comment
user
data
layer
features
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011357879.7A
Other languages
Chinese (zh)
Inventor
杜永萍
王艳
彭治
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.)
Beijing University of Technology
Original Assignee
Beijing University of Technology
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 Beijing University of Technology filed Critical Beijing University of Technology
Priority to CN202011357879.7A priority Critical patent/CN112328900A/en
Publication of CN112328900A publication Critical patent/CN112328900A/en
Pending legal-status Critical Current

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
    • G06F40/00—Handling natural language data
    • G06F40/10—Text processing
    • G06F40/103—Formatting, i.e. changing of presentation of documents
    • G06F40/117—Tagging; Marking up; Designating a block; Setting of attributes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/10—Text processing
    • G06F40/166—Editing, e.g. inserting or deleting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/10—Text processing
    • G06F40/189—Automatic justification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/284—Lexical analysis, e.g. tokenisation or collocates
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/047—Probabilistic or stochastic networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Machine Translation (AREA)

Abstract

The invention discloses a deep learning method for fusing score data and comment texts, which is used for relieving the influence of the sparsity of the score data by introducing auxiliary information. The comment text can be used for acquiring the preference information and the project characteristics of the user, and the score data comprises the potential association between the user and the project. Most of the existing fusion models adopt a matrix decomposition method for processing the rating data, and in order to better utilize effective information in the rating data, the method utilizes a convolutional neural network to process comment texts and introduces an attention mechanism to extract representative comments in the comment information, so that user preference and project characteristics are better represented. And processing the scoring data by using a deep neural network to extract the depth features in the scoring data, and fusing the features to predict the scoring of the user on the project.

Description

Deep learning recommendation method integrating scoring matrix and comment text
Technical Field
The invention relates to a recommendation model for extracting text features and fusing score data by utilizing deep learning, and belongs to the field of information recommendation.
Background
With the rapid development of networks, the data volume increases exponentially, and it becomes more and more difficult for people to acquire useful information from massive data. The recommendation system is proved to be effective in solving the problem of information overload, and can help people to quickly find needed information in mass data.
The traditional recommendation algorithm mainly comprises the following steps: collaborative filtering and content-based recommendation algorithms. Both of these methods recommend based on the user's historical behavior. The content-based recommendation algorithm determines a common feature among the items by using the rating information of the user on the items, and if the feature exists in the new item, the item is recommended to the user; collaborative filtering is to use historical data to find users who like the user and then recommend items to the user that are found to be liked by these similar users in the past. Or an item with the same characteristic as a certain item is found, and if the user likes the item, other items with the same characteristic are recommended to the user.
Many successful collaborative filtering techniques are based on Matrix Factorization (MF) algorithm, a scoring Matrix is used as input, the Matrix is decomposed into a form of multiplying a user Matrix and a project Matrix, so that hidden vector representations of users and projects are obtained, the hidden vectors are represented in space, similarity between users and projects close to each other is high, and projects close to each other are recommended to target users when recommendation is performed. The matrix decomposition algorithm can relieve the sparsity of data to a certain extent.
Text information is introduced to mitigate data sparsity. In a general recommendation system, a user can write comment information besides directly scoring, the general comment information contains the reason that the user gives the score, and compared with the score, the comment information contains more abundant content. On one hand, the reason for giving the score can be obtained, and on the other hand, preference information of the user or some characteristics of the commodity can be reflected.
In recent years, deep learning has been widely used in many fields, and particularly, breakthrough progress has been made in natural language processing. The advantages of natural language in the aspect of text content mining are utilized, and the natural language is applied to a recommendation system to process comment texts, so that a new research direction is provided for the recommendation system; the deep learning techniques currently used in the recommendation system are mainly CNN, RNN, attention mechanism, RBM, encoder, etc.
The CNN and RNN networks in deep learning can well retain word sequence information when processing text information. The ConvMF (constraint Matrix factorization) model combines the convolutional neural network with probability Matrix decomposition, and the convolutional neural network is used for processing comment text information, so that word order information is effectively reserved. But the model only utilizes the comment text information of the item and ignores the comment information of the user. The model is a DeepCoNN (deep collaborative Neural network) model, which uses two parallel CNNs to respectively process the comment information of a project and the comment information of a user, and uses a factorization machine to perform score prediction at the last layer. The NARRE (neural attribute Regression Model with Review Level extensions) Model takes the scoring matrix and the comment text as input, and simultaneously fuses the hidden vector in the scoring data, thereby greatly improving the prediction performance of the recommendation system.
In the model, only the potential scoring features are used when the comment text features are fused, the scoring data is not deeply modeled, and the scoring data is deeply learned. Therefore, the invention provides a deep recommendation model for fusing the score matrix and the comment text, utilizes deep learning to extract the deep features in the score, fuses with the user preference and the item features extracted from the comment text information, and improves the recommendation performance by better expressing the user item features. Experiments prove that the model can better extract the characteristics of the user and the project, and effectively improves the recommendation accuracy.
The problems to be solved and the beneficial effects are achieved:
the method and the device have the advantages that the influence of the sparsity of the scoring data is relieved by introducing auxiliary information, so that the recommendation quality is improved. The comment text is utilized to obtain the preference information and the project characteristics of the user, the score data contains the potential relation between the user and the project, in order to better utilize the effective information in the score data, the neural network is utilized to extract the depth characteristics in the score, and the user preference and the project characteristics are better represented, so that the recommendation quality is improved.
In order to obtain context information in a text, the invention utilizes a Convolutional Neural Network (CNN) to extract feature representations of users and items in the comment in an experiment.
The comment texts contain a large amount of useful information, some comment information can accurately describe relevant characteristics of the items, and some comments are not greatly related to the items. The contribution made by different review information is different. Therefore, a re-experiment introduced an attention mechanism to select review information that is more representative of the characteristics of the project.
The scoring data is always explicit feedback of the user on the item preference degree, and the neural network is used for acquiring the preference of the user and the characteristic representation of the item according to the scoring data.
And performing linear and high-order interaction on the user and item features extracted from the comment text and the user and item features acquired from the scoring data respectively to acquire final feature representation of the user and the items.
And performing element product on the obtained final feature representation of the user and the item, sending the final feature representation into a prediction layer to obtain a final prediction score, and fusing a score matrix and a deep learning recommendation model diagram of the comment text as shown in the figure I.
Disclosure of Invention
The deep learning recommendation method mainly designs and realizes a deep learning recommendation method for fusing the scoring matrix and the comment text, and achieves an effect superior to that of other models by comparing with a plurality of public baseline models and taking MSE as an evaluation index.
The main invention content of the invention is as follows:
firstly, selecting and processing a data set:
selecting Amazon 6 public data sets in different fields to evaluate the model, dividing the data sets into a training set, a verification set and a test set according to the ratio of 8:1: 1.
Secondly, the characteristics of the user and the project are respectively extracted from the comment text and the scoring data of the data set, and the main work is as follows:
(1) respectively modeling a user and a project, wherein the network structures used when modeling the user and the project are the same, only the input is different, and in the project modeling:
the method for extracting the features in the item comment text mainly comprises the following steps:
(1) the comment information of a given project is used as input, each comment is mapped into a d-dimensional vector through an embedding layer, and a word embedding matrix with a fixed length is obtained.
(2) And using the obtained word embedding matrix as the input of the convolutional neural network to obtain the characteristic vector.
(3) Enter the max pooling layer, capture important high value features, and compress the convolutional layer output into a fixed size vector
(4) In order to select comment information which can represent the characteristics of the item, an attention mechanism is introduced, attention scores are obtained by taking the obtained output as input, and the attention scores are normalized to obtain final comment weight.
(5) Sending the result to a full connection layer to obtain the feature representation of the project based on the comment information
Extracting features in the project scoring data mainly works as follows:
and taking the rating data of a given project as input, sending the data into a network layer, and outputting a project depth feature representation based on the rating at the last layer.
And (3) merging layers to obtain final feature representations of users and projects, wherein the specific method comprises the following steps:
and interacting the item feature representations obtained in the comment text and the scoring data, and acquiring the final features of the item through linear and high-order interactive modeling. The same way and the final feature representation of the user is obtained.
And the scoring prediction layer waits for the final scoring prediction through the information, and the specific process is as follows:
and carrying out element product operation on the obtained final item feature representation and the final user feature representation to obtain a vector, and sending the obtained vector to a prediction layer to obtain a final prediction score.
Drawings
FIG. 1 is a schematic diagram of an overall model of the present invention.
Detailed Description
In the existing model for fusing the scoring data and the comment text, a matrix decomposition algorithm is mostly adopted when the scoring data is processed, only potential features in the scoring data are used, and in order to better utilize the scoring information, the invention utilizes a deep neural network to process the scoring data and extract the depth features in the scoring data. And processing the comment text by using a convolutional neural network, and introducing an attention mechanism to extract representative comments in the comment information, so that the user preference and the item characteristics are better represented. And fusing the characteristics to obtain a final prediction score.
The invention comprises the following specific steps:
firstly, preprocessing comment texts in a data set, and mainly working as follows:
and loading data, carrying out standardized processing, and uniformly converting all letters in the comment information into lower case. Then, punctuation marks and special marks in the comment text are filtered, the text data are decomposed into words, and labels are generated. Stop words are removed and the sentence length of the text data is converted to the same length, which is defined by the longest sentence.
Establishing a sentence-based vocabulary mapping from words to indexes, mapping sentences and tags to vectors according to a vocabulary table, initializing word embedding using a 300-dimensional word2vec vector trained on 1000-hundred million words of Google news,
dividing the preprocessed data set into a training set, a verification set and a test set according to the proportion of 8:1:1, training a model by using the training set, performing parameter adjustment on the verification set, and finally verifying the performance of the model on the test set.
Given a set of comments { R for an item ii1,Ri2…RikK represents the maximum number of comments allowed to be input by the model, and after passing through the embedding layer, each piece of comment information is mapped into a d-dimensional vector to obtain a word embedding matrix M with a fixed lengthi。
And taking the word vector matrix obtained by the embedding layer as the input of the convolutional neural network layer to obtain the characteristic vector. Let the convolutional layer consist of m neurons, using a convolutional filter f over a sliding window of size sjTo extract contextual features. The method comprises the following specific steps:
kj＝Relu(Mi*fj+bj)
relu in the above formula represents the activation function, which is a convolution operation, bjThe deviation is indicated.
Entering the largest pooling layer in the convolution module, max-posing can capture important high-value features and compress the output of the convolution layer into a fixed-size vector, namely:
cj＝max(k1,k2...kd-s+1)
the feature vector of the final output of the convolutional layer can be expressed as: c. Ci1,ci2,...,cik。
For the same project, different users have different comment information, some user comments accurately describe the relevant characteristics of the project, and some user comments written by the users may not have great relevance with the project. The contribution made by different review information is different. In order to select comment information which can represent project characteristics better, an attention mechanism is introduced, an attention layer is used for corresponding comment texts, and attention scores are obtained as follows:
gik＝Relu(wi×cik+b)
relu in the above formula represents the activation function, wiAnd b is the weight and offset of the network automatic learning respectively.
Score g attention using softmax functionikNormalization processing is carried out to obtain the final comment weight aikAs shown in the following formula:
Figure BDA0002803118110000051
highlighting the influence of the nth comment in the current comment set on the characteristics of the item i by using the comment weight, and scoring the attention of the comment by aikAnd (3) acting on the feature vectors output by the corresponding convolution layers one by one to obtain the weighted sum of the feature vectors of the item i:
Figure BDA0002803118110000052
then, the method is sent to a full connection layer to obtain a final feature representation of the item i based on the text information:
Ti＝WiCi+bi
in the above formula, Wi，biRespectively the weight and the deviation amount of the fully connected layer. Similarly, the characteristic representation T of the user u based on the comment information can be obtained by modeling the useru。
The scoring data of the user for the items is explicit feedback, the preference degree of the user for the items can be directly reflected, different scoring shows that the preference degree of the user for the items is different, and therefore the preference of the user and the feature representation of the items are obtained by the neural network according to the scoring data.
Using the scoring data of item i as input pi＝{yi1,yi2,...,yinWhere the model of the network is defined as:
Figure BDA0002803118110000053
in the above formula
Figure BDA0002803118110000054
A weight matrix representing the x-th layer,
Figure BDA0002803118110000055
representing bias vectors of layer xF (-) denotes an activation function, here the Relu function is used, the output X of the last layeriRepresenting the depth feature representation of item i. Similarly, modeling the user can obtain the feature representation X of the user u based on the scoring datau。
And respectively fusing the features obtained in the comment text and the features obtained in the comment data to obtain final feature representation of the user and the project:
Figure BDA0002803118110000056
the feature vector I comprises project features obtained from the comment information and the scoring data, wherein ^ indicates splicing, and g indicates high-order interactive modeling. Here, we use the second order term in the factorization machine for obtaining the high-order features, and first convert the second order term in the factorization machine:
Figure BDA0002803118110000061
wherein
Figure BDA0002803118110000062
The product of the elements is represented by,<vi，vj>used to model the relationship between the ith and jth terms, computing the interaction between arbitrary two-dimensional features. All information of the second order interaction term is already contained before sum () is executed. Therefore, only the previous part is required to be spliced with the characteristics obtained from the comment information and the characteristics obtained from the comment data
Figure BDA0002803118110000063
The following network structure is used to obtain the high-order features:
Figure BDA0002803118110000064
g can further enhance the interaction of potential features, and the final feature representation U of the user can be obtained by the same method.
Performing element product operation on the obtained user final feature U and the project final feature I to obtain a vector H ═ multiplex (U, I), and transmitting the vector H ═ multiplex (U, I) to a prediction layer to obtain a prediction score
Figure BDA0002803118110000065
Figure BDA0002803118110000066
In the above formula, W represents a weight parameter of a prediction layer, bu、biRespectively representing the deviation of the user and the deviation of the item, bgRepresenting the amount of global deviation.
To score a prediction
Figure BDA0002803118110000067
And true score R in the datasetu,iThe difference between them is minimal, the expression for the loss function loss is given below:
Figure BDA0002803118110000068
in the above equation, O represents the number of samples in the training set,
Figure BDA0002803118110000069
Ru,ithe prediction score and the truth score are represented, respectively. To prevent overfitting, an L2 norm regularization term constraint is used in the loss function, where λ is used to control the degree of regularization, and Adam optimization function is used to minimize loss to obtain the relevant parameters in the model.
During the experiment, we use 300-dimensional word2vec vectors trained on Google news to initialize word embedding and fine-tune during the training.
The initial learning rate is set to 0.005 and the convolution kernel size is set to 3.
To prevent overfitting, we used the dropout strategy in the experiment, setting the value to 0.5.
The performance evaluation index of the model selects mean square error MSE, the smaller the value is, the better the performance of the model is represented, and the calculation is as follows:
Figure BDA00028031181100000610
in the above formula, K represents the number of samples,
Figure BDA0002803118110000071
represents a prediction score, Ru,iRepresenting a true score.
Through experimental verification, the experimental results of the method provided by the invention are shown in table 1:
TABLE 1 results of the evaluation of recommended Performance under the six data sets
Dataset MSE
Toys_and_Games 0.773
Musical_Instruments 0.869
office_product 0.709
Instant_Video 0.909
Digital_Music 0.789
Kindle_Store 0.626

Claims (4)

1. A deep learning recommendation method integrating a scoring matrix and comment texts is characterized by comprising the following steps of: comprises the following steps of (a) carrying out,
s1, the comment text in the data set is preprocessed:
loading data, carrying out standardized processing, and uniformly converting all letters in the comment information into lower case; then, punctuation marks and special marks in the comment text are filtered, text data are decomposed into words, and labels are generated; removing stop words, and then converting the sentence length of the text data into the same length, wherein the length is defined by the longest sentence;
s2 extracts features in the comment text:
given a set of reviews for item i { Ri1,Ri2…RikK represents the maximum number of comments allowed to be input by the model, and after passing through the embedding layer, each piece of comment information is mapped into a d-dimensional vector to obtain a word embedding matrix M with a fixed lengthi；
S3 extracts features in the score data:
the scoring data of the user to the items is explicit feedback, which directly reflects the preference degree of the user to the items, and different scores show that the preference degree of the user to the items is different, so that the neural network is used for acquiring the preference of the user and the feature representation of the items according to the scoring data;
and S4 feature fusion:
and respectively fusing the features obtained in the comment text and the features obtained in the comment data to obtain final feature representation of the user and the project:
Figure FDA0002803118100000011
the feature vector I comprises project features obtained from the comment information and the scoring data, wherein ^ indicates splicing, and g indicates high-order interactive modeling; the high-order characteristics are obtained by using a second order term in the factorization machine for reference, and the second order term in the factorization machine is firstly converted:
Figure FDA0002803118100000012
wherein
Figure FDA0002803118100000013
The product of the elements is represented by,<vi,vj>used to model the relationship between the ith and jth terms, computing the interaction between arbitrary two-dimensional features. Before sum () is executed, all information including the second-order interaction item is contained, and the features obtained from the comment information and the features obtained from the comment data are spliced
Figure FDA0002803118100000021
The following network structure is used to obtain the high-order features:
Figure FDA0002803118100000022
g, interaction of potential features can be further enhanced, and a final feature representation U of a user is obtained;
s5 score prediction:
performing element product operation on the obtained user final feature U and the project final feature I to obtain a vector H ═ multiplex (U, I), and transmitting the vector H ═ multiplex (U, I) to a prediction layer to obtain a prediction score
Figure FDA0002803118100000023
Figure FDA0002803118100000024
In the above formula, W represents a weight parameter of a prediction layer, bu、biRespectively representing the deviation of the user and the deviation of the item, bgRepresenting the amount of global deviation.
2. The deep learning recommendation method integrating the scoring matrix and the comment text as claimed in claim 1, wherein: in S1, a sentence-based vocabulary mapping from words to indexes is established, sentences and labels are mapped to vectors according to a vocabulary table, 300-dimensional word2vec vectors trained on 1000-hundred million words Google news are used for initializing word embedding, a preprocessed data set is divided into a training set, a verification set and a test set according to the proportion of 8:1:1, a training model of the training set is used for carrying out parameter adjustment on the verification set, and finally the performance of the model is verified on the test set.
3. The deep learning recommendation method integrating the scoring matrix and the comment text as claimed in claim 1, wherein: in S2, the word vector matrix obtained by the embedding layer is used as the input of the convolutional neural network layer to obtain a characteristic vector; let the convolutional layer consist of m neurons, using a convolutional filter f over a sliding window of size sjTo extract contextual features; the method comprises the following specific steps:
kj＝Relu(Mi*fj+bj)
wherein Relu is the activation function and convolution operation, bjIndicating a deviation;
entering the largest pooling layer in the convolution module, max-posing can capture important high-value features and compress the output of the convolution layer into a fixed-size vector, namely:
cj＝max(k1,k2...kd-s+1)
the feature vector of the final output of the convolutional layer can be expressed as: c. Ci1,ci2,...,cik；
For the same project, different users have different comment information, in order to select the comment information which can represent the project characteristics better, an attention mechanism is introduced, an attention layer is used for corresponding comment texts, and the attention score is obtained as follows:
gik＝Relu(wi×cik+b)
wherein Relu represents the activation function, wiB is the weight and offset of the network automatic learning respectively;
score g attention using softmax functionikNormalization processing is carried out to obtain the final comment weight aikAs shown in the following formula:
Figure FDA0002803118100000031
highlighting the influence of the nth comment in the current comment set on the characteristics of the item i by using the comment weight, and scoring the attention of the comment by aikActing on the feature vector output by the corresponding convolution layer to obtain a weighted sum of the feature vectors of the item i:
Figure FDA0002803118100000032
then, the method is sent to a full connection layer to obtain a final feature representation of the item i based on the text information:
Ti＝WiCi+bi
in the formula, Wi，biRespectively the weight and deviation amount of the full connection layer; similarly, the characteristic representation T of the user u based on the comment information can be obtained by modeling the useru。
4. The deep learning recommendation method integrating the scoring matrix and the comment text as claimed in claim 1, wherein: in S3, the score data of item i is used as input pi＝{yi1,yi2,...,yinWhere the model of the network is defined as:
Figure FDA0002803118100000033
Figure FDA0002803118100000034
....
Figure FDA0002803118100000035
in the above formula
Figure FDA0002803118100000041
A weight matrix representing the x-th layer,
Figure FDA0002803118100000042
representing the bias vector of layer X, f (-) representing the activation function, here the Relu function, the output X of the last layeriRepresenting a depth feature representation of item i; modeling the user can obtain the feature representation X of the user u based on the scoring datau。
CN202011357879.7A 2020-11-27 2020-11-27 Deep learning recommendation method integrating scoring matrix and comment text Pending CN112328900A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011357879.7A CN112328900A (en) 2020-11-27 2020-11-27 Deep learning recommendation method integrating scoring matrix and comment text

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011357879.7A CN112328900A (en) 2020-11-27 2020-11-27 Deep learning recommendation method integrating scoring matrix and comment text

Publications (1)

Publication Number Publication Date
CN112328900A true CN112328900A (en) 2021-02-05

Family

ID=74307847

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011357879.7A Pending CN112328900A (en) 2020-11-27 2020-11-27 Deep learning recommendation method integrating scoring matrix and comment text

Country Status (1)

Country Link
CN (1) CN112328900A (en)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112700342A (en) * 2020-12-29 2021-04-23 四川大学 Personalized grading prediction model and prediction method for social media information authenticity
CN112784173A (en) * 2021-02-26 2021-05-11 电子科技大学 Recommendation system scoring prediction method based on self-attention confrontation neural network
CN112861012A (en) * 2021-03-09 2021-05-28 河南工业大学 Recommendation method and device based on context and user long-term and short-term preference adaptive learning
CN112860880A (en) * 2021-03-23 2021-05-28 辽宁工程技术大学 Comment text and feature weighting fused four-dimensional tensor decomposition recommendation method
CN112966148A (en) * 2021-03-05 2021-06-15 安徽师范大学 Video recommendation method and system based on deep learning and feature fusion
CN113111257A (en) * 2021-03-30 2021-07-13 北京工业大学 Collaborative filtering-based recommendation method for fusing multi-source heterogeneous information
CN113127737A (en) * 2021-04-14 2021-07-16 江苏科技大学 Personalized search method and search system integrating attention mechanism
CN114154071A (en) * 2021-12-09 2022-03-08 电子科技大学 Emotion time sequence recommendation method based on attention mechanism
CN114723522A (en) * 2022-03-31 2022-07-08 合肥工业大学 Comment text-oriented graph neural network recommendation method
CN115905717A (en) * 2023-03-02 2023-04-04 中南大学 Scenic spot recommendation method and device based on mixed 3D convolution fusion rating and comment
CN115905617A (en) * 2023-03-02 2023-04-04 南京邮电大学 Video scoring prediction method based on deep neural network and double regularization
CN115980298A (en) * 2023-03-20 2023-04-18 山东思睿环境设备科技有限公司 Multi-parameter-based adaptive water quality detection and analysis method and device

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109753566A (en) * 2019-01-09 2019-05-14 大连民族大学 The model training method of cross-cutting sentiment analysis based on convolutional neural networks
CN110059262A (en) * 2019-04-19 2019-07-26 武汉大学 A kind of construction method and device, item recommendation method of the project recommendation model based on hybrid neural networks
CN111666496A (en) * 2020-06-09 2020-09-15 中森云链(成都)科技有限责任公司 Group recommendation method based on comment text

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109753566A (en) * 2019-01-09 2019-05-14 大连民族大学 The model training method of cross-cutting sentiment analysis based on convolutional neural networks
CN110059262A (en) * 2019-04-19 2019-07-26 武汉大学 A kind of construction method and device, item recommendation method of the project recommendation model based on hybrid neural networks
CN111666496A (en) * 2020-06-09 2020-09-15 中森云链(成都)科技有限责任公司 Group recommendation method based on comment text

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
冯兴杰等: "基于评分矩阵与评论文本的深度推荐模型", 《计算机学报》 *

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112700342A (en) * 2020-12-29 2021-04-23 四川大学 Personalized grading prediction model and prediction method for social media information authenticity
CN112784173A (en) * 2021-02-26 2021-05-11 电子科技大学 Recommendation system scoring prediction method based on self-attention confrontation neural network
CN112966148A (en) * 2021-03-05 2021-06-15 安徽师范大学 Video recommendation method and system based on deep learning and feature fusion
CN112861012A (en) * 2021-03-09 2021-05-28 河南工业大学 Recommendation method and device based on context and user long-term and short-term preference adaptive learning
CN112860880A (en) * 2021-03-23 2021-05-28 辽宁工程技术大学 Comment text and feature weighting fused four-dimensional tensor decomposition recommendation method
CN112860880B (en) * 2021-03-23 2024-03-08 辽宁工程技术大学 Four-dimensional tensor decomposition recommendation method integrating comment text and feature weighting
CN113111257A (en) * 2021-03-30 2021-07-13 北京工业大学 Collaborative filtering-based recommendation method for fusing multi-source heterogeneous information
WO2022218139A1 (en) * 2021-04-14 2022-10-20 江苏科技大学 Personalized search method and search system combined with attention mechanism
CN113127737A (en) * 2021-04-14 2021-07-16 江苏科技大学 Personalized search method and search system integrating attention mechanism
CN113127737B (en) * 2021-04-14 2021-09-14 江苏科技大学 Personalized search method and search system integrating attention mechanism
CN114154071A (en) * 2021-12-09 2022-03-08 电子科技大学 Emotion time sequence recommendation method based on attention mechanism
CN114723522A (en) * 2022-03-31 2022-07-08 合肥工业大学 Comment text-oriented graph neural network recommendation method
CN115905717A (en) * 2023-03-02 2023-04-04 中南大学 Scenic spot recommendation method and device based on mixed 3D convolution fusion rating and comment
CN115905617A (en) * 2023-03-02 2023-04-04 南京邮电大学 Video scoring prediction method based on deep neural network and double regularization
CN115980298A (en) * 2023-03-20 2023-04-18 山东思睿环境设备科技有限公司 Multi-parameter-based adaptive water quality detection and analysis method and device
CN115980298B (en) * 2023-03-20 2023-07-21 山东思睿环境设备科技有限公司 Multi-parameter adaptive water quality detection and analysis method and device

Similar Documents

Publication Publication Date Title
CN112328900A (en) Deep learning recommendation method integrating scoring matrix and comment text
CN109299396B (en) Convolutional neural network collaborative filtering recommendation method and system fusing attention model
CN109145112B (en) Commodity comment classification method based on global information attention mechanism
CN109977413B (en) Emotion analysis method based on improved CNN-LDA
CN111209738B (en) Multi-task named entity recognition method combining text classification
CN109325112B (en) A kind of across language sentiment analysis method and apparatus based on emoji
CN109241424B (en) A kind of recommended method
CN110188272B (en) Community question-answering website label recommendation method based on user background
CN112884551B (en) Commodity recommendation method based on neighbor users and comment information
CN110765260A (en) Information recommendation method based on convolutional neural network and joint attention mechanism
CN108363753A (en) Comment text sentiment classification model is trained and sensibility classification method, device and equipment
CN106844632B (en) Product comment emotion classification method and device based on improved support vector machine
CN111127146B (en) Information recommendation method and system based on convolutional neural network and noise reduction self-encoder
CN111274398A (en) Method and system for analyzing comment emotion of aspect-level user product
CN112800344B (en) Deep neural network-based movie recommendation method
CN112016002A (en) Mixed recommendation method integrating comment text level attention and time factors
CN113590970A (en) Personalized digital book recommendation system and method based on reader preference, computer and storage medium
Chen et al. Deep neural networks for multi-class sentiment classification
CN114693397A (en) Multi-view multi-modal commodity recommendation method based on attention neural network
CN114969304A (en) Case public opinion multi-document generation type abstract method based on element graph attention
CN111985243A (en) Emotion model training method, emotion analysis device and storage medium
CN111241410A (en) Industry news recommendation method and terminal
CN115481313A (en) News recommendation method based on text semantic mining
CN116976505A (en) Click rate prediction method of decoupling attention network based on information sharing
CN114579741B (en) GCN-RN aspect emotion analysis method and system for fusing syntax information

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210205