CN112818248A - Emotion-based article recommendation model construction and recommendation method and system - Google Patents

Emotion-based article recommendation model construction and recommendation method and system Download PDF

Info

Publication number
CN112818248A
CN112818248A CN202110219998.4A CN202110219998A CN112818248A CN 112818248 A CN112818248 A CN 112818248A CN 202110219998 A CN202110219998 A CN 202110219998A CN 112818248 A CN112818248 A CN 112818248A
Authority
CN
China
Prior art keywords
information
user
article
meta
vector
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.)
Granted
Application number
CN202110219998.4A
Other languages
Chinese (zh)
Other versions
CN112818248B (en
Inventor
贺小伟
唐可昕
王宾
侯榆青
吴昊
仝硕阳
候晨
李得栋
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.)
Northwestern University
Original Assignee
Northwestern University
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 Northwestern University filed Critical Northwestern University
Priority to CN202110219998.4A priority Critical patent/CN112818248B/en
Publication of CN112818248A publication Critical patent/CN112818248A/en
Application granted granted Critical
Publication of CN112818248B publication Critical patent/CN112818248B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • 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
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Economics (AREA)
  • Evolutionary Computation (AREA)
  • Strategic Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

An emotion-based item recommendation model building and recommending method and system predict user emotion classification of a user text through an emotion classification neural network, and mine associated information of users and items, so that the use of comment information is enriched, and the problem of data sparsity is relieved; by combining the semantic enhanced meta-path of the item context information, the scoring accuracy of the item scoring model is improved, the accuracy of the recommendation method is improved, and the recommendation process is clear and visible due to the use of the meta-path information, so that the prediction result is more interpretable; meanwhile, a multi-layer perceptron model is adopted to predict the rating value, and the rating prediction performance is improved.

Description

Emotion-based article recommendation model construction and recommendation method and system
Technical Field
The invention relates to a commodity recommendation method, in particular to an emotion-based article recommendation model construction and recommendation method and system.
Background
The goal of the recommendation system is to provide users with items that meet the user preference criteria, and since Netflix and Amazon successfully implemented the recommendation system, various countries have made various efforts to recommend items that users have associated with their preferences. The core of the recommendation system is a recommendation algorithm, and the traditional recommendation algorithm comprises project-based collaborative filtering, user-based collaborative filtering and a matrix decomposition method. Nowadays, various auxiliary data in online services are increasing, and many methods further propose to utilize contextual information to improve recommendation performance. However, due to the heterogeneity and complexity of the assistance data, efficient use of contextual information in a recommendation system remains a challenge.
In real life, a large number of components which are different in types and are mutually associated are formed, the components which are mutually interacted and are mutually associated can be abstracted into an information network, and most work is to model the information network into a homogeneous information network, namely, the network contains the same type of associated objects. However, isomorphic information network modeling often extracts only part of information in an actual interactive system and does not show the difference of the relationship among various objects. Most real world networks are heterogeneous, where nodes and relationships are of different types. For example, in a smart museum network, the nodes may be a collection knowledge base, an audience image base, a management subject base, an operation and maintenance subject base, and the like. On the one hand, treating all nodes as being of the same type (e.g., a homogeneous information network) may lose significant semantic information.
Recently, some researchers find that heterogeneous networks (HIN) formed by various types of nodes and links can flexibly process various heterogeneous data, for example, in an intelligent museum network, heterogeneous data is used for a big data model of a data center, a network system is built, a museum's data island' is opened, and repeated construction is reduced. The information comprehensiveness of the HIN, as well as the importance of rich semantics in the recommendation field-allow better recommendation results to be produced. In particular, meta-paths, a sequence of relationships connecting pairs of objects in the HIN, are widely used to extract structural features that are recommended to capture relevant semantic information.
Although various technologies have been proposed in the prior art to improve the accuracy of the recommendation model, there are still many problems, and a conventional heterogeneous network recommendation algorithm generally recommends through user scoring or item content, and the accuracy rate of occurrence is not high, so that users and items cannot be completely understood, and the overall performance of the recommendation model is affected.
Disclosure of Invention
The invention aims to provide a method and a system for building and recommending an article recommendation model based on emotion, which predict user emotion classification of a user text through an emotion classification neural network, mine associated information of a user and an article, enrich the use of comment information and relieve the problem of data sparsity; by combining the semantic enhanced meta-path of the item context information, the scoring accuracy of the item scoring model is improved, the accuracy of the recommendation method is improved, and the recommendation process is clear and visible due to the use of the meta-path information, so that the prediction result is more interpretable; meanwhile, a multi-layer perceptron model is adopted to predict the rating value, and the rating prediction performance is improved.
In order to realize the task, the invention adopts the following technical scheme:
an emotion-based item scoring model construction method is implemented according to the following steps:
s1, obtaining user information, user interest groups, article information and user texts, wherein the article information comprises article context information, and carrying out emotion classification processing on the user texts to obtain user emotion classification;
acquiring a user information set, wherein the acquired user information set comprises a plurality of user information;
acquiring an article information set, wherein the article information set comprises a plurality of article information;
acquiring an evaluation information set, wherein the evaluation information set comprises a plurality of evaluation information, and the evaluation information comprises user information, user interest groups, article information and user emotion classification;
obtaining the grade of a user on an article, obtaining the grade value and obtaining a tag set;
the user emotion is classified into negative emotions or positive emotions, and when the user emotion is classified into the negative emotions, the score value is a low score value; when the user emotion is classified into forward emotion, the score value is a high score value;
s2, taking the evaluation information set, the user information set and the article information set as input, taking the label set as output, and training a heterogeneous network to obtain an article scoring model;
the heterogeneous network comprises an input layer, a meta path planning layer, a prediction result layer and an output layer which are sequentially arranged, wherein the output end of the input layer is also connected with the input end of the prediction result layer;
the input layer comprises 4 parallel input modules which are respectively used for inputting user information, user interest groups, article information and user emotion classification;
the element path planning layer is a convolutional neural network and a semantic enhancement constructor which are sequentially connected in series, and the convolutional neural network comprises an embedded layer, a convolutional layer, a pooling layer and an output layer which are sequentially connected in series;
the prediction result layer is a multilayer perceptron.
Preferably, the training step of S2 specifically includes:
s21, embedding and representing the user information set, the article information set and the evaluation information set respectively to obtain a user information vector set, an article information vector set and an evaluation information vector set;
the user information vector set comprises a plurality of user information vectors;
the article information vector set comprises a plurality of article information vectors;
the evaluation information vector set comprises a plurality of evaluation information vectors, and the evaluation information vectors comprise user information vectors, user interest group vectors, article information vectors and user emotion classification vectors;
the item information vector comprises an item context information vector;
s22, sequentially inputting each evaluation information vector in the evaluation information vector set obtained in the S21 into an embedding layer, a convolutional layer and a pooling layer of the convolutional neural network of the S2 to carry out feature extraction and maximum pool operation, and obtaining vector representations of a plurality of meta-path types;
s23, obtaining the distribution weight of each meta path type according to the formula (3);
Figure BDA0002954398310000041
wherein alpha isjAn assigned weight representing the jth meta path type, f () representing a ReLU function, u representing user information, i representing item information, e user sentiment classification, ρ representing the meta path type, ρjRepresents the jth meta-path category, b is the bias term;
wherein, WuWeight matrix, W, for user informationiWeight matrix, W, for item informationeA weight matrix for classifying the user's emotion,
Figure BDA0002954398310000042
a weight matrix for the jth meta-path class;
wherein S isuAs a vector of user information, SiAs an item information vector, SeFor the user's emotion classification vector,
Figure BDA0002954398310000051
a vector representation representing the jth meta-path class;
s24, combining the vector representation of any meta-path type obtained in S22 with the context information vector of the article to obtain vector representations of a plurality of semantic enhanced meta-path types and obtain a vector representation set of a comprehensive meta-path type, wherein the vector representation set of the comprehensive meta-path type comprises the vector representation of one meta-path type and the vector representations of the plurality of semantic enhanced meta-path types;
respectively combining the vector representation of each meta-path type obtained in the step S22 with the context information vector of the article to obtain a plurality of vector representation sets of comprehensive meta-path types;
s25, summing the vector representation of one meta-path type in the vector representation set of each comprehensive meta-path type obtained in S24 and the vector representations of the multiple semantic enhanced meta-path types, and then performing weight distribution on the summed vector representations and the distributed weights of the meta-path types obtained in S23 to obtain a weight vector representation set of the comprehensive meta-path types, wherein the weight vector representation set of the comprehensive meta-path types comprises the weight vector representations of the multiple whole meta-path types;
and S26, sequentially splicing the user information vector set obtained in the S21, the weight vector representation set of the comprehensive meta-path types obtained in the S25 and the article information vector set obtained in the S21 to obtain learned comprehensive characteristics, and inputting the learned comprehensive characteristics into a multilayer perceptron in the S2 to obtain a score value.
Preferably, the value of the offset term b is 0.0010 ± 0.0002.
Preferably, the emotion classification processing in S1 is to preprocess the user text to obtain a plurality of word vectors with text emotion, input the plurality of word vectors with text emotion to the emotion classification neural network for emotion classification processing, and output the user emotion classification;
the value range of the low score value is {1,2}, and the value range of the high score value is {3,4,5 }.
An emotion-based item recommendation method is implemented according to the following method:
step A, obtaining user information, a user interest group and a user text of a user, and carrying out emotion classification processing on the user text to obtain user emotion classification, wherein the user emotion classification is divided into negative emotion or positive emotion;
obtaining item information of each item, wherein the item information comprises item context information;
step B, collecting the article information of each article obtained in the step A, the user information of the user, the user interest group and the user emotion classification to obtain evaluation information of each article;
step C, inputting the user information of the user and the article information of each article obtained in the step A and the evaluation information of each article obtained in the step B into an article grading model obtained by the emotion-based article grading model building method disclosed by the invention to obtain the grading value of each article;
and D, arranging the scoring values of the articles obtained in the step C from large to small to obtain an article recommendation sequence.
An emotion-based item scoring model construction system comprises a data acquisition device and a model construction device;
the data acquisition device is used for acquiring user information, a user interest group, article information and a user text, wherein the article information comprises article context information, and emotion classification processing is carried out on the user text to obtain user emotion classification;
acquiring a user information set, wherein the acquired user information set comprises a plurality of user information;
acquiring an article information set, wherein the article information set comprises a plurality of article information;
acquiring an evaluation information set, wherein the evaluation information set comprises a plurality of evaluation information, and the evaluation information comprises user information, user interest groups, article information and user emotion classification;
obtaining the grade of a user on an article, obtaining the grade value and obtaining a tag set;
the user emotion is classified into negative emotions or positive emotions, and when the user emotion is classified into the negative emotions, the score value is a low score value; when the user emotion is classified into forward emotion, the score value is a high score value;
the model construction device is used for taking the evaluation information set as input and the label set as output, and training a heterogeneous network to obtain an article scoring model;
the heterogeneous network comprises an input layer, a meta path planning layer, a prediction result layer and an output layer which are sequentially arranged, wherein the output end of the input layer is also connected with the input end of the prediction result layer;
the input layer comprises 4 parallel input modules which are respectively used for inputting user information, user interest groups, article information and user emotion classification;
the element path planning layer is a convolutional neural network and a semantic enhancement constructor which are sequentially connected in series, and the convolutional neural network comprises an embedded layer, a convolutional layer, a pooling layer and an output layer which are sequentially connected in series;
the prediction result layer is a multilayer perceptron.
Preferably, the training step of the heterogeneous network specifically includes:
step 1, respectively embedding and representing a user information set, an article information set and an evaluation information set to obtain a user information vector set, an article information vector set and an evaluation information vector set;
the user information vector set comprises a plurality of user information vectors;
the article information vector set comprises a plurality of article information vectors;
the evaluation information vector set comprises a plurality of evaluation information vectors, and the evaluation information vectors comprise user information vectors, user interest group vectors, article information vectors and user emotion classification vectors;
the item information vector comprises an item context information vector;
step 2, sequentially inputting each evaluation information vector in the evaluation information vector set obtained in the step 1 into an embedding layer, a convolution layer and a pooling layer of the convolutional neural network for feature extraction and maximum pool operation to obtain vector representations of a plurality of element path types;
step 3, obtaining the distribution weight of each meta-path type according to the formula (3);
Figure BDA0002954398310000081
wherein alpha isjAn assigned weight representing the jth meta path type, f () representing a ReLU function, u representing user information, i representing item information, e user sentiment classification, ρ representing the meta path type, ρjRepresents the jth meta-path category, b is the bias term;
wherein, WuFor the userWeight matrix of information, WiWeight matrix, W, for item informationeA weight matrix for classifying the user's emotion,
Figure BDA0002954398310000082
a weight matrix for the jth meta-path class;
wherein S isuAs a vector of user information, SiAs an item information vector, SeFor the user's emotion classification vector,
Figure BDA0002954398310000083
a vector representation representing the jth meta-path class;
step 4, combining the vector representation of any meta-path type obtained in the step 2 with the context information vector of the article to obtain vector representations of a plurality of semantic enhanced meta-path types and obtain a vector representation set of a comprehensive meta-path type, wherein the vector representation set of the comprehensive meta-path type comprises the vector representation of one meta-path type and the vector representations of the plurality of semantic enhanced meta-path types;
combining the vector representation of each meta-path type obtained in the step (2) with the context information vector of the article respectively to obtain a plurality of vector representation sets of comprehensive meta-path types;
step 5, summing the vector representation of one meta-path type in the vector representation set of each comprehensive meta-path type obtained in the step 4 and the vector representations of the plurality of semantic enhanced meta-path types, and then performing weight distribution on the summed vector representations and the distributed weights of the meta-path types obtained in the step 3 to obtain a weight vector representation set of the comprehensive meta-path types, wherein the weight vector representation set of the comprehensive meta-path types comprises the weight vector representations of the plurality of full meta-path types;
and 6, sequentially splicing the user information vector set obtained in the step 1, the weight vector representation set of the comprehensive meta-path types obtained in the step 5 and the article information vector set obtained in the step 1 to obtain a learned comprehensive characteristic, and inputting the learned comprehensive characteristic into a multilayer perceptron to obtain a score value.
Preferably, the value of the offset term b is 0.0010 ± 0.0002.
Preferably, the emotion classification processing in the data acquisition device is to preprocess a user text to obtain a plurality of word vectors with text emotions, input the word vectors with the text emotions into an emotion classification neural network for emotion classification processing, and output user emotion classification;
the value range of the low and high scores is {1,2}, and the value range of the high scores is {3,4,5 }.
An article recommendation system based on emotion comprises an information acquisition device, an information collection device, a scoring device and a sequencing device;
the information acquisition device is used for acquiring user information, a user interest group and a user text of a user, and performing emotion classification processing on the user text to obtain a user emotion classification which is divided into negative emotion or positive emotion;
obtaining item information of each item, wherein the item information comprises item context information;
the information collecting device is used for collecting the article information of each article, the user information of the user, the user interest group and the user emotion classification to obtain the evaluation information of each article;
the scoring device is used for inputting user information of a user, article information of each article and evaluation information of each article into an article scoring model obtained by the emotion-based article scoring model construction system disclosed by the invention to obtain a scoring value of each article;
the sorting device is used for sorting the scoring value of each article from large to small to obtain the article recommendation sequence.
Compared with the prior art, the invention has the following technical effects:
1. according to the method and the system for building and recommending the article recommendation model based on the emotion, the emotion classification of the user text is predicted based on the emotion classification neural network to obtain the user emotion classification, the associated information of the user and the article is mined, the use of comment information is enriched, the problem of data sparsity is relieved, the accuracy of the article scoring model is improved, and therefore the accuracy of the recommendation method is improved.
2. According to the emotion-based item recommendation model building and recommending method and system, the semantic enhanced meta path combined with the item context information is built, recommendation and the item context information are closely combined, the scoring accuracy of the scoring model is improved, the accuracy of the recommendation method is improved, the recommendation process is clear and visible due to the use of the meta path information, and the prediction result is more interpretable.
3. According to the emotion-based item recommendation model construction and recommendation method and system, the multi-layer perceptron model is adopted to predict the rating value, and the rating prediction performance is improved.
Drawings
Fig. 1 is a diagram illustrating an internal structure of an item scoring model according to the present invention.
Fig. 2 is a diagram of a heterogeneous network example of a meta path planning layer in embodiment 1.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and examples. So that those skilled in the art can better understand the present invention. It is to be expressly noted that in the following description, a detailed description of known functions and designs will be omitted when it may obscure the subject matter of the present invention.
The following definitions or conceptual connotations relating to the present invention are provided for illustration:
the semantic enhancement constructor combines various semantic contexts with meta-paths, such as: a set of meta-paths P, such that each path user starts, ends with an item, and has a maximum length of i, and if i is set to 3, the meta-path that the path P can express is: UMAM, UMDM, UMUM, see literature: patent publication No. CN111222049A, 6/2/2020.
Multilayer perceptrons (MLP), also called artificial neural networks, can be found in the literature: W.Z.Lu, H.Y.Fan, S.M.Lo.application of evaporative neural network methods in predicting polar levels in downtown area of Hong Kong [ J ]. Neurocompressing, 2003,51
The emotion classification neural network is an existing GRU neural network model, and can be referred to documents: patent publication No. CN109165387A, publication No. 1/8/2019.
The Convolutional Neural Network (CNN) of the present invention is an existing Network model, and can be referred to as a literature: hu B, Shi C, ZHao W, et al.Levering Meta-path based Context for Top-N Recommendation with A Neural Co-orientation Model [ J ].2018: 1531-.
In the embodiment of the present invention, the object may be a physical object that exists specifically, such as a museum collection object, a tourist attraction, or some virtual object, such as a movie, an electronic journal, etc. The user evaluation information of the article includes the user evaluation of the article. For example, after visiting the treasures of each museum, the user may give a rating to each theme house, comment information to the theme house, or after watching a movie, the user may give a movie rating and comment information on the movie website according to the rating of the movie.
Example one
The embodiment provides an emotion-based item scoring model construction method, which is specifically executed according to the following steps as shown in fig. 1:
s1, obtaining user information, user interest groups, article information and user texts, wherein the article information comprises article context information, and carrying out emotion classification processing on the user texts to obtain user emotion classification;
acquiring a user information set, wherein the acquired user information set comprises a plurality of user information;
acquiring an article information set, wherein the article information set comprises a plurality of article information;
acquiring an evaluation information set, wherein the evaluation information set comprises a plurality of evaluation information, and the evaluation information comprises user information, user interest groups, article information and user emotion classification;
obtaining the grade of a user on an article, obtaining the grade value and obtaining a tag set;
the user emotion is classified into negative emotions or positive emotions, and when the user emotion is classified into the negative emotions, the score value is a low score value; when the user emotion is classified as a forward emotion, the score value is a high score value.
In this embodiment, an operator may input, through a related interface provided by the system, collected evaluation information of a user on a plurality of items, a plurality of user information, and a plurality of item information, where each evaluation information includes user information, a user interest group, item information, and a user emotion classification; the user information is structured information capable of identifying a user, the user interest group is other users with the same preference as the user, the article information is structured information capable of identifying articles, and the user emotion classification is positive or negative emotion judgment obtained after emotion classification is carried out on a user text; the user text is text information describing user attributes or preferences thereof.
In this embodiment, one piece of evaluation information is evaluation information of one user on one article, and includes user information, user interest groups, article information and user emotion classification, and the piece of evaluation information corresponds to one score value, that is, the score value of the user on the article; and collecting a plurality of pieces of evaluation information to obtain an evaluation information set, and collecting the score values corresponding to the plurality of pieces of evaluation information to obtain a tag set.
Taking a movie recommendation as an example, the user information includes a user ID, the user interest group represents an interest discussion circle added by the user, the item information includes an ID of the movie, the item context information includes a director, an actor and a movie type of the movie, the user text represents a movie rating of the movie by the user, and the user emotion classification represents positive or negative judgment of the emotion of the movie rating by the user (for example, "2" represents positive emotion, and "1" represents negative emotion); the rating value refers to a rating score of the movie by the user.
When the user emotion is classified into '1', the value range of the corresponding score value is {1,2 }; when the user emotion is classified into 2, the value range of the corresponding score value is {3,4,5 }.
For example, the same piece of user data is [6:23:4:1:5] for the movie data of the broad bean; wherein 6 represents a user ID; 23 represents a user interest group; 4 item information (here, movie ID), where the item context information is that the director of the movie is chenkexin, the actors are scler, yellow bohai and wu-gang, and the movie type is a drama; 2, representing user emotion classification and representing forward emotion; and 5 represents a score value.
Specifically, the emotion classification processing is to preprocess a user text to obtain a plurality of word vectors with text emotions, input the word vectors with the text emotions into an emotion classification neural network to perform emotion classification processing, and output user emotion classification.
In the embodiment, the emotion classification of the user on the article comment is predicted based on the emotion classification neural network, the associated information of the user and the article is mined, the use of comment information is enriched, and the accuracy of the article scoring model scoring is improved, so that the accuracy of the recommendation method is improved.
In this embodiment, extracting user comment data from a webpage as a user text, preprocessing the user comment data to obtain a plurality of word vectors with text emotion, and performing emotion classification processing on the plurality of word vectors with text emotion to obtain a user emotion classification specifically includes the following steps:
(1) and (3) exception data processing: the method comprises the steps that a user comment data set extracted from a webpage comprises a plurality of user comment data, each user comment data possibly has a phenomenon that individual comments are vacant, so that missing value processing needs to be carried out on the comment data, meanwhile, the condition that symbols such as quotation marks in a single English format and quotation marks in a single Chinese format are wrong possibly occurs in one comment, and the problem that data entry does not exist due to symbol mismatching can occur, so that in order to avoid the problem, the abnormal data processing is carried out on the user comment data by adopting a method of replacing double quotation marks in the comment with space marks;
(2) inputting the comment data of the user after abnormal data processing into a word segmentation tool (such as a jieba word segmentation tool), and firstly adopting "" "," "and" ". ","! ","? ","; the punctuation marks are used as a sentence segmentation standard to perform sentence segmentation processing on each user comment data to obtain a plurality of sentences, and then each sentence is subjected to word segmentation processing to obtain a plurality of user comment words;
for example, the user text is: "Chenkexin is still serious, and is a new height of domestic sports subject. ", the preprocessed user comment words are: "New height of old and harsh domestic sports subject matter";
(3) inputting all user comment words obtained by word segmentation into a Chinese-word-vectors model, and outputting a plurality of word vectors with text emotion;
the Chinese-word-vectors model is an open-source 'Chinese-word-vectors model' developed by the cooperation of the Beijing university Chinese information processing research institute and the DBIIR laboratory of the Chinese university;
for example, the preprocessed user comment words are processed by using a Chinese-word-vectors model, and the obtained word vector with text emotion is as follows: "Chenkexin 1 and 2 domestic 1 sports 1 subject 1 new height 2";
(4) inputting all word vectors with text emotion into an emotion classification neural network for emotion classification processing, and outputting user emotion classification; the emotion classification neural network comprises a reset gate, an update gate and a hidden layer which are sequentially arranged;
for example, the word vector with text emotion is subjected to emotion classification processing to obtain user emotion classification: and 2, positive emotion.
S2, taking the evaluation information set, the user information set and the article information set as input, taking the label set as output, and training a heterogeneous network to obtain an article scoring model;
the heterogeneous network comprises an input layer, a meta path planning layer, a prediction result layer and an output layer which are sequentially arranged, wherein the output end of the input layer is also connected with the input end of the prediction result layer;
the input layer comprises 4 parallel input modules which are respectively used for inputting user information, user interest groups, article information and user emotion classification;
the element path planning layer is a convolutional neural network and a semantic enhancement constructor which are sequentially connected in series, and the convolutional neural network comprises an embedded layer, a convolutional layer, a pooling layer and an output layer which are sequentially connected in series;
the prediction result layer is a multilayer perceptron.
Specifically, the training step of S2 specifically includes:
s21, embedding and representing the user information set, the article information set and the evaluation information set respectively to obtain a user information vector set, an article information vector set and an evaluation information vector set;
the user information vector set comprises a plurality of user information vectors;
the article information vector set comprises a plurality of article information vectors;
the evaluation information vector set comprises a plurality of evaluation information vectors, and the evaluation information vectors comprise user information vectors, user interest group vectors, article information vectors and user emotion classification vectors;
the user information, the user interest group, the article information and the user emotion classification are all in a digital form, and because the digital form cannot be directly processed by a computer, the corresponding feature vectors of the input characters and the input numbers need to be found, namely, the characters or the numbers in a high-dimensional space are embedded into a continuous low-dimensional vector space, and then the user information vector, the user interest group vector, the article information vector and the user emotion classification vector are obtained; wherein the item information vector comprises an item context information vector.
S22, sequentially inputting each evaluation information vector in the evaluation information vector set obtained in S21 into an embedding layer, a convolutional layer and a pooling layer of the convolutional neural network in S2 to carry out feature extraction and maximum pool operation, and obtaining vector representations of a plurality of meta-path types;
in the embodiment, because the user information vector, the user interest group vector, the article information vector and the user emotion classification vector are non-text data, the features are extracted through an embedded layer of a convolutional neural network; at the embedding layer, features are represented as an embedding matrix to extract information.
In this embodiment, vector representations of meta-path instances are obtained by learning an embedding matrix of users and items (including user information vectors, user interest group vectors, article information vectors, and user emotion classification vectors) obtained by an embedding layer through a convolutional layer; the meta-path is a sequence of a plurality of groups of nodes, and the sequence of the nodes is embedded into a low-dimensional vector by adopting CNN (convolutional neural network) to obtain a vector representation of a meta-path instance;
obtaining vector representations of multiple element path types through the maximum pool operation of the pooling layer; the meta path comprises a plurality of meta path instances p, and the meta path instances p are classified to obtain N meta path types rho; the method comprises the following steps that k element path instances p exist in each element path type rho, and vector representation of each element path type is obtained through a formula (1);
Figure BDA0002954398310000171
wherein,
Figure BDA0002954398310000172
vector characterization, p, representing the jth meta-path classjRepresents the jth meta-path category, j ∈ [1,2, …, N](ii) a max-polling represents the maximum pool operation, hpFor the vector characterization of the meta-path instance,
Figure BDA0002954398310000173
a vector characterization for k meta-path instances.
S23, obtaining the distribution weight of each meta path type according to the formula (3);
Figure BDA0002954398310000174
wherein alpha isjThe assigned weight of the jth meta path type is represented, f () represents a ReLU function, u represents user information, i represents article information, e represents user emotion classification, rho represents the meta path type, and b is an offset term; wuWeight matrix, W, for user informationiWeight matrix, W, for item informationeA weight matrix for classifying the user's emotion,
Figure BDA0002954398310000181
is as followsA weight matrix of j meta-path classes; suAs a vector of user information, SiAs an item information vector, SeClassifying vectors for the user emotion;
because different meta-path types may have different semantics in the interaction, different users indicate different preferences through different meta-path types, and even if the same user and different articles pass through the same meta-path type, semantic information in the meta-path type may be different; therefore, to better characterize the semantic information of the user and the item, equation (3) is used to derive the assigned weight for each meta-path category.
Wherein, the weight matrix W of the user informationu13129, wherein the elements of the first column are all 1's and the remaining elements are all 0's;
weight matrix W of article informationiA matrix of 12345 by 12345, where the elements of the first column are all 1 and the remaining elements are all 0;
weight matrix W for user emotion classificationeA matrix of 12345 by 12345, where the elements of the first column are all 1 and the remaining elements are all 0;
weight matrix for jth meta-path class
Figure BDA0002954398310000182
A 128 x 128 matrix in which each element of each row and each column is 1.
Wherein, the value of the offset item b is 0.0010.
S24, combining the vector representation of any meta-path type obtained in S22 with the context information vector of the article to obtain vector representations of a plurality of semantic enhanced meta-path types and obtain a vector representation set of a comprehensive meta-path type, wherein the vector representation set of the comprehensive meta-path type comprises the vector representation of one meta-path type and the vector representations of the plurality of semantic enhanced meta-path types;
and combining the vector representation of each meta-path type obtained in the step S22 with the context information vector of the article respectively to obtain a plurality of vector representation sets of comprehensive meta-path types.
In the embodiment, the vector representation of each meta-path type obtained in step S22 is input to a semantic enhancement constructor to obtain vector representations of a plurality of semantic enhancement type meta-path types, respectively, in combination with the context information vector of the article; in order to fully consider the interaction between each user and the article, obtaining a plurality of vector characteristic sets of comprehensive meta-path categories;
wherein the item context information vector includes a movie type vector (G), an actor vector (a), and a director vector (D).
In the present embodiment, as shown in fig. 2, an example of a heterogeneous network (HIN) of a movie recommendation system is composed of various types of nodes, i.e., { user vector (U), movie vector (M), movie type vector (G), actor vector (a), and director vector (D) }.
As shown in table 1, different semantic relationships are defined for the meta-path categories and the corresponding meta-path examples thereof in the present embodiment, for example, two meta-path categories, i.e., UMAM and UMDM, which respectively represent movies stared by the same actor or movies guided by the same director; u3m3a2m2 is a meta-path instance for UMAM, u3m3d1m1 is a meta-path instance for UMDM; UGU represents the users in the same interest group with the target user, UMDM represents the movie shot by the same director as the target user watches the movie, and UGUMDM represents the movie shot by the same director as the user in the same interest group with the target user watches.
TABLE 1 Meta-Path types and corresponding Meta-Path instances
Connection type 1 Connection type 2 Connection classType 3 Connection type 4
Example of Path u1m2a2m3 u1m1d1m3 u3g1u2m1 u3g1u2m1d1m3
u2m2a2m3 u3m3d1m1
Meta path category UMAM UMDM UGUM UGUMDM
And S25, summing the vector representation of one meta-path type in the vector representation set of each comprehensive meta-path type obtained in the step S24 and the vector representations of the multiple semantic enhanced meta-path types, and then performing weight distribution on the summed vector representations and the distributed weights of the meta-path types obtained in the step S23 to obtain a weight vector representation set of the comprehensive meta-path types, wherein the weight vector representation set of the comprehensive meta-path types comprises the weight vector representations of the multiple whole meta-path types.
Weighting by a formula (4) to obtain a weight vector representation of the comprehensive meta-path type;
Figure BDA0002954398310000201
wherein, CjA weight vector representation representing the jth comprehensive meta-path class,
Figure BDA0002954398310000202
for vector characterization of the full meta-path class, ρ0Indicates the overall meta-path type, QjA set of vector tokens representing a jth overall meta-path category (including vector tokens for the jth meta-path category and a plurality of semantically enhanced meta-path categories corresponding thereto);
the weight vector representation set of the comprehensive meta path category is shown in formula 5;
{C1,C2,……,Cj} formula (5)
And S26, sequentially splicing the user information vector set obtained in the S21, the weight vector representation set of the comprehensive meta-path types obtained in the S25 and the article information vector set obtained in the S21 to obtain learned comprehensive characteristics, and inputting the learned comprehensive characteristics into a multilayer perceptron in the S2 to obtain a score value.
For example: and 5, predicting that the value of the user scoring the article is 5 points.
In this embodiment, the user information vector set, the weight vector representation set of the comprehensive meta-path category, and the article information vector set are all in the form of a matrix, and the user information vector set, the weight vector representation set of the comprehensive meta-path category, and the article information vector set are sequentially spliced to obtain the learned comprehensive features.
In this embodiment, the learned comprehensive features are used as the input of the multi-layer perceptron, the training set is used to train the model, the parameters of the multi-layer perceptron in the model are continuously adjusted, and the model prediction performance is optimized.
For example, parameters in a multi-layer perceptron: the learning rate is 0.001, the iteration is 30, and the batch size is 256.
Comparative example
The emotion-based item scoring model construction method in example 1 is compared with the existing four classical methods by using the normalized breaking cumulative gain as an evaluation index, and the comparison result of the normalized breaking cumulative gain of the level K is shown in table 2, and the smaller the normalized breaking cumulative gain, the better the prediction performance of the method.
As can be seen from Table 2, the emotion-based article recommendation method provided by the invention is optimal on both the two data sets of the bean movie and the bean book, and the emotion-based article recommendation method based on the emotion classification neural network and the convolutional neural network provided by the invention is proved to have better prediction performance on the rating prediction problem in a recommendation system.
MF in Table 2 is Yehuda Koren, Robert Bell, and Chris volinsky.2009.matrix catalysis reagents for recommender systems.computer 42 (2009); SVD is Tianqi Chen, Weinan Zhang, Qiaxia Lu, Kailong Chen, ZHao Zheng, and Yong Yu.2012.SVDFearure a toolkit for feature-based collectivity filtering. journal of Machine Learning Research 13(2012),3619 + 3622; MFPR is Jian Liu, Chuan Shi, Binbin Hu, Shenghua Liu, and S Yu Philip.2017.personalized ranking conversion vision integrating multiple feedback. in Pacific-Asia Conference on Knowledge Discovery and Data mining.131-143; MCRec is Binbin Hu, Chuan Shi, Wayne Xin Zhuao, Philip S.Yu, Leveraging Meta-path based Context for Top-N Recommendation with A Neural Co-attachment model KDD 2018, August 19-23,2018, London, United Kingdom.
TABLE 2 comparison of the Performance of the scoring model construction method provided by the present invention with other classical methods
Figure BDA0002954398310000221
Example two
An emotion-based item recommendation method is implemented according to the following method:
step A, obtaining user information, a user interest group and a user text of a user, and carrying out emotion classification processing on the user text to obtain user emotion classification, wherein the user emotion classification is divided into negative emotion or positive emotion;
obtaining item information of each item, wherein the item information comprises item context information;
step B, collecting the article information of each article obtained in the step A, the user information of the user, the user interest group and the user emotion classification to obtain evaluation information of each article;
step C, inputting the user information of the user and the article information of each article obtained in the step A and the evaluation information of each article obtained in the step B into an article scoring model obtained by the emotion-based article scoring model construction method in the first embodiment to obtain a scoring value of each article;
and D, arranging the scoring values of the articles obtained in the step C from large to small to obtain an article recommendation sequence.
For example, the user scores for each item [ 1: 2.4212,2: 1.2003,3: 4.6987,4: 2.5432], sorting by numerical value from large to small, obtaining [ 3: 4.6987,4: 2.5432,1: 2.4212,2: 1.2003], and the commodity recommendation sequence of the corresponding user is [3, 4, 1,2 ].
EXAMPLE III
An emotion-based item scoring model construction system comprises a data acquisition device and a model construction device;
the data acquisition device is used for acquiring user information, a user interest group, article information and a user text, wherein the article information comprises article context information, and emotion classification processing is carried out on the user text to obtain user emotion classification;
acquiring a user information set, wherein the acquired user information set comprises a plurality of user information;
acquiring an article information set, wherein the article information set comprises a plurality of article information;
acquiring an evaluation information set, wherein the evaluation information set comprises a plurality of evaluation information, and the evaluation information comprises user information, user interest groups, article information and user emotion classification;
obtaining the grade of a user on an article, obtaining the grade value and obtaining a tag set;
the user emotion is classified into negative emotions or positive emotions, and when the user emotion is classified into the negative emotions, the score value is a low score value; when the user emotion is classified into forward emotion, the score value is a high score value;
the model construction device is used for taking the evaluation information set as input and the label set as output, and training a heterogeneous network to obtain an article scoring model;
the heterogeneous network comprises an input layer, a meta path planning layer, a prediction result layer and an output layer which are sequentially arranged, wherein the output end of the input layer is also connected with the input end of the prediction result layer;
the input layer comprises 4 parallel input modules which are respectively used for inputting user information, user interest groups, article information and user emotion classification;
the element path planning layer is a convolutional neural network and a semantic enhancement constructor which are sequentially connected in series, and the convolutional neural network comprises an embedded layer, a convolutional layer, a pooling layer and an output layer which are sequentially connected in series;
the prediction result layer is a multilayer perceptron.
Specifically, the training step of the heterogeneous network specifically includes:
step 1, respectively embedding and representing a user information set, an article information set and an evaluation information set to obtain a user information vector set, an article information vector set and an evaluation information vector set;
the user information vector set comprises a plurality of user information vectors; the article information vector set comprises a plurality of article information vectors; the evaluation information vector set comprises a plurality of evaluation information vectors, and the evaluation information vectors comprise user information vectors, user interest group vectors, article information vectors and user emotion classification vectors;
the item information vector comprises an item context information vector;
step 2, inputting each evaluation information vector in the evaluation information vector set obtained in the step 1 into an embedding layer, a convolution layer and a pooling layer of the convolutional neural network in sequence for feature extraction and maximum pool operation to obtain vector representations of a plurality of element path types;
step 3, obtaining the distribution weight of each meta-path type according to the formula (3);
Figure BDA0002954398310000251
wherein alpha isjAn assigned weight representing the jth meta path type, f () representing a ReLU function, u representing user information, i representing item information, e user sentiment classification, ρ representing the meta path type, ρjRepresents the jth meta-path category, b is the bias term;
wherein, WuWeight matrix, W, for user informationiWeight matrix, W, for item informationeA weight matrix for classifying the user's emotion,
Figure BDA0002954398310000252
a weight matrix for the jth meta-path class;
wherein S isuAs a vector of user information, SiAs an item information vector, SeFor the user's emotion classification vector,
Figure BDA0002954398310000253
a vector representation representing the jth meta-path class;
wherein, the value of the offset item b is 0.0010;
step 4, combining the vector representation of any meta-path type obtained in the step 2 with the context information vector of the article to obtain vector representations of a plurality of semantic enhanced meta-path types and obtain a vector representation set of a comprehensive meta-path type, wherein the vector representation set of the comprehensive meta-path type comprises the vector representation of one meta-path type and the vector representations of the plurality of semantic enhanced meta-path types;
combining the vector representation of each meta-path type obtained in the step (2) with the context information vector of the article respectively to obtain a plurality of vector representation sets of comprehensive meta-path types;
step 5, summing the vector representation of one meta-path type in the vector representation set of each comprehensive meta-path type obtained in the step 4 and the vector representations of the plurality of semantic enhanced meta-path types, and then performing weight distribution on the summed vector representations and the distributed weights of the meta-path types obtained in the step 3 to obtain a weight vector representation set of the comprehensive meta-path types, wherein the weight vector representation set of the comprehensive meta-path types comprises the weight vector representations of the plurality of full meta-path types;
and 6, sequentially splicing the user information vector set obtained in the step 1, the weight vector representation set of the comprehensive meta-path types obtained in the step 5 and the article information vector set obtained in the step 1 to obtain a learned comprehensive characteristic, and inputting the learned comprehensive characteristic into a multilayer perceptron to obtain a score value.
Specifically, the emotion classification processing in the data acquisition device is to preprocess a user text to obtain a plurality of word vectors with text emotions, input the word vectors with the text emotions into an emotion classification neural network for emotion classification processing, and output user emotion classification.
Specifically, the value range of the low score value is {1,2}, and the value range of the high score value is {3,4,5 }.
In this embodiment, the emotion-based item scoring model building system may be deployed in a local server or a remote cloud server.
Example four
An article recommendation system based on emotion comprises an information acquisition device, an information collection device, a scoring device and a sequencing device;
the information acquisition device is used for acquiring user information, a user interest group and a user text of a user, and performing emotion classification processing on the user text to obtain user emotion classification which is divided into negative emotion or positive emotion; obtaining item information of each item, wherein the item information comprises item context information;
the information collecting device is used for collecting the article information of each article, the user information of the user, the user interest group and the user emotion classification to obtain the evaluation information of each article;
the scoring device is used for inputting the user information of the user, the article information of each article and the evaluation information of each article into the article scoring model obtained by the emotion-based article scoring model construction system of the third embodiment to obtain the scoring value of each article;
the sequencing device is used for sequencing the scoring value of each article from large to small to obtain the article recommendation sequence.
From the foregoing description of the embodiments, it is clear for those skilled in the art that the technical solutions of the present invention may be embodied in the form of a software product, which is stored in a readable storage medium, such as a floppy disk, a hard disk, or an optical disk of a computer, and includes several instructions for enabling a computer device (which may be a personal computer, a server, or a network device) to execute the methods of the embodiments of the present invention.

Claims (10)

1. An emotion-based item scoring model construction method is characterized by comprising the following steps:
s1, obtaining user information, a user interest group, article information and a user text, wherein the article information comprises article context information, and carrying out emotion classification processing on the user text to obtain user emotion classification;
acquiring a user information set, wherein the acquired user information set comprises a plurality of user information;
acquiring an article information set, wherein the article information set comprises a plurality of article information;
acquiring an evaluation information set, wherein the evaluation information set comprises a plurality of evaluation information, and the evaluation information comprises user information, user interest groups, article information and user emotion classification;
obtaining the grade of a user on an article, obtaining the grade value and obtaining a tag set;
the user emotion is classified into negative emotions or positive emotions, and when the user emotion is classified into the negative emotions, the score value is a low score value; when the user emotion is classified into forward emotion, the score value is a high score value;
s2, taking the evaluation information set, the user information set and the article information set as input, taking the label set as output, and training a heterogeneous network to obtain an article scoring model;
the heterogeneous network comprises an input layer, a meta path planning layer, a prediction result layer and an output layer which are sequentially arranged, wherein the output end of the input layer is also connected with the input end of the prediction result layer;
the input layer comprises 4 parallel input modules which are respectively used for inputting user information, user interest groups, article information and user emotion classification;
the element path planning layer comprises a convolutional neural network and a semantic enhancement constructor which are sequentially connected in series, wherein the convolutional neural network comprises an embedded layer, a convolutional layer, a pooling layer and an output layer which are sequentially connected in series;
the prediction result layer is a multilayer perceptron.
2. The emotion-based item scoring model construction method of claim 1, wherein the training step of S2 specifically includes:
s21, embedding and representing the user information set, the article information set and the evaluation information set respectively to obtain a user information vector set, an article information vector set and an evaluation information vector set;
the user information vector set comprises a plurality of user information vectors;
the item information vector set comprises a plurality of item information vectors;
the evaluation information vector set comprises a plurality of evaluation information vectors, and the evaluation information vectors comprise user information vectors, user interest group vectors, article information vectors and user emotion classification vectors;
the item information vector comprises an item context information vector;
s22, sequentially inputting each evaluation information vector in the evaluation information vector set obtained in S21 into the embedding layer, the convolutional layer and the pooling layer of the convolutional neural network in S2 to carry out feature extraction and maximum pool operation, and obtaining vector representations of a plurality of meta-path types;
s23, obtaining the distribution weight of each meta path type according to the formula (3);
Figure FDA0002954398300000021
wherein alpha isjAn assigned weight representing the jth meta path type, f () representing a ReLU function, u representing user information, i representing item information, e user sentiment classification, ρ representing the meta path type, ρjRepresents the jth meta-path category, b is the bias term;
wherein, WuWeight matrix, W, for user informationiWeight matrix, W, for item informationeA weight matrix for classifying the user's emotion,
Figure FDA0002954398300000031
a weight matrix for the jth meta-path class;
wherein S isuAs a vector of user information, SiAs an item information vector, SeFor the user's emotion classification vector,
Figure FDA0002954398300000032
a vector representation representing the jth meta-path class;
s24, combining the vector representation of any meta-path type obtained in S22 with the context information vector of the article to obtain vector representations of a plurality of semantic enhanced meta-path types and obtain a vector representation set of a comprehensive meta-path type, wherein the vector representation set of the comprehensive meta-path type comprises the vector representation of one meta-path type and the vector representations of the plurality of semantic enhanced meta-path types;
respectively combining the vector representation of each meta-path type obtained in the step S22 with the context information vector of the article to obtain a plurality of vector representation sets of comprehensive meta-path types;
s25, summing the vector representation of one meta-path type in the vector representation set of each comprehensive meta-path type obtained in S24 and the vector representations of the multiple semantic enhanced meta-path types, and then performing weight distribution on the summed vector representations and the distributed weights of the meta-path types obtained in S23 to obtain a weight vector representation set of the comprehensive meta-path types, wherein the weight vector representation set of the comprehensive meta-path types comprises the weight vector representations of the multiple whole meta-path types;
and S26, sequentially splicing the user information vector set obtained in the S21, the weight vector representation set of the comprehensive meta-path types obtained in the S25 and the article information vector set obtained in the S21 to obtain learned comprehensive characteristics, and inputting the learned comprehensive characteristics into a multilayer perceptron in the S2 to obtain a score value.
3. The emotion-based item scoring model construction method of claim 2, wherein the bias term b takes a value of 0.0010 ± 0.0002.
4. The method for constructing an emotion-based item scoring model as defined in any one of claims 1 to 3, wherein the emotion classification processing in S1 is to preprocess a user text to obtain a plurality of word vectors with text emotion, input the word vectors with text emotion to an emotion classification neural network for emotion classification processing, and output the user emotion classification;
the value range of the low score value is {1,2}, and the value range of the high score value is {3,4,5 }.
5. An emotion-based item recommendation method is characterized by being executed according to the following method:
step A, obtaining user information, a user interest group and a user text of a user, and carrying out emotion classification processing on the user text to obtain a user emotion classification, wherein the user emotion classification is divided into negative emotion or positive emotion;
obtaining item information for each item, the item information including item context information;
step B, collecting the article information of each article obtained in the step A, the user information of the user, the user interest group and the user emotion classification to obtain evaluation information of each article;
step C, inputting the user information of the user and the article information of each article obtained in the step A and the evaluation information of each article obtained in the step B into an article grading model obtained by the emotion-based article grading model building method according to any one of claims 1 to 4, and obtaining a grading value of each article;
and D, arranging the scoring values of the articles obtained in the step C from large to small to obtain an article recommendation sequence.
6. An emotion-based item scoring model construction system is characterized by comprising a data acquisition device and a model construction device;
the data acquisition device is used for acquiring user information, a user interest group, article information and a user text, wherein the article information comprises article context information, and emotion classification processing is carried out on the user text to obtain user emotion classification;
acquiring a user information set, wherein the acquired user information set comprises a plurality of user information;
acquiring an article information set, wherein the article information set comprises a plurality of article information;
acquiring an evaluation information set, wherein the evaluation information set comprises a plurality of evaluation information, and the evaluation information comprises user information, user interest groups, article information and user emotion classification;
obtaining the grade of a user on an article, obtaining the grade value and obtaining a tag set;
the user emotion is classified into negative emotions or positive emotions, and when the user emotion is classified into the negative emotions, the score value is a low score value; when the user emotion is classified into forward emotion, the score value is a high score value;
the model construction device is used for taking the evaluation information set as input and the label set as output, and training a heterogeneous network to obtain an article scoring model;
the heterogeneous network comprises an input layer, a meta path planning layer, a prediction result layer and an output layer which are sequentially arranged, wherein the output end of the input layer is also connected with the input end of the prediction result layer;
the input layer comprises 4 parallel input modules which are respectively used for inputting user information, user interest groups, article information and user emotion classification;
the element path planning layer comprises a convolutional neural network and a semantic enhancement constructor which are sequentially connected in series, wherein the convolutional neural network comprises an embedded layer, a convolutional layer, a pooling layer and an output layer which are sequentially connected in series;
the prediction result layer is a multilayer perceptron.
7. The emotion-based item scoring model construction system of claim 6, wherein the training step of the heterogeneous network specifically comprises:
step 1, respectively embedding and representing a user information set, an article information set and an evaluation information set to obtain a user information vector set, an article information vector set and an evaluation information vector set;
the user information vector set comprises a plurality of user information vectors;
the item information vector set comprises a plurality of item information vectors;
the evaluation information vector set comprises a plurality of evaluation information vectors, and the evaluation information vectors comprise user information vectors, user interest group vectors, article information vectors and user emotion classification vectors;
the item information vector comprises an item context information vector;
step 2, sequentially inputting each evaluation information vector in the evaluation information vector set obtained in the step 1 into an embedding layer, a convolution layer and a pooling layer of the convolutional neural network for feature extraction and maximum pool operation to obtain vector representations of a plurality of element path types;
step 3, obtaining the distribution weight of each meta-path type according to the formula (3);
Figure FDA0002954398300000061
wherein alpha isjAn assigned weight representing the jth meta path type, f () representing a ReLU function, u representing user information, i representing item information, e user sentiment classification, ρ representing the meta path type, ρjRepresents the jth meta-path category, b is the bias term;
wherein,Wuweight matrix, W, for user informationiWeight matrix, W, for item informationeA weight matrix for classifying the user's emotion,
Figure FDA0002954398300000062
a weight matrix for the jth meta-path class;
wherein S isuAs a vector of user information, SiAs an item information vector, SeFor the user's emotion classification vector,
Figure FDA0002954398300000071
a vector representation representing the jth meta-path class;
step 4, combining the vector representation of any meta-path type obtained in the step 2 with the context information vector of the article to obtain vector representations of a plurality of semantic enhanced meta-path types and obtain a vector representation set of a comprehensive meta-path type, wherein the vector representation set of the comprehensive meta-path type comprises the vector representation of one meta-path type and the vector representations of the plurality of semantic enhanced meta-path types;
combining the vector representation of each meta-path type obtained in the step (2) with the context information vector of the article respectively to obtain a plurality of vector representation sets of comprehensive meta-path types;
step 5, summing the vector representation of one meta-path type in the vector representation set of each comprehensive meta-path type obtained in the step 4 and the vector representations of the plurality of semantic enhanced meta-path types, and then performing weight distribution on the summed vector representations and the distributed weights of the meta-path types obtained in the step 3 to obtain a weight vector representation set of the comprehensive meta-path types, wherein the weight vector representation set of the comprehensive meta-path types comprises the weight vector representations of the plurality of full meta-path types;
and 6, sequentially splicing the user information vector set obtained in the step 1, the weight vector representation set of the comprehensive meta-path types obtained in the step 5 and the article information vector set obtained in the step 1 to obtain a learned comprehensive characteristic, and inputting the learned comprehensive characteristic into the multilayer perceptron to obtain a score value.
8. The emotion-based item scoring model construction method of claim 7, wherein the bias term b takes a value of 0.0010 ± 0.0002.
9. The system for constructing the emotion-based item scoring model as defined in any one of claims 6 to 8, wherein emotion classification processing in the data acquisition device is to preprocess a user text to obtain a plurality of word vectors with text emotion, input the plurality of word vectors with text emotion to an emotion classification neural network for emotion classification processing, and output user emotion classification;
the value range of the low and high scores is {1,2}, and the value range of the high scores is {3,4,5 }.
10. An emotion-based item recommendation system is characterized by comprising an information acquisition device, an information collection device, a scoring device and a sequencing device;
the information acquisition device is used for acquiring user information, a user interest group and a user text of a user, and performing emotion classification processing on the user text to obtain a user emotion classification, wherein the user emotion classification is negative emotion or positive emotion;
obtaining item information for each item, the item information including item context information;
the information collecting device is used for collecting the article information of each article, the user information of the user, the user interest group and the user emotion classification to obtain the evaluation information of each article;
the scoring device is used for inputting user information of a user, article information of each article and evaluation information of each article into an article scoring model obtained by the emotion-based article scoring model building system according to any one of claims 6 to 9 to obtain a scoring value of each article;
the sorting device is used for sorting the scoring values of all the articles from large to small to obtain the article recommendation sequence.
CN202110219998.4A 2021-02-26 2021-02-26 Emotion-based article recommendation model construction and recommendation method and system Active CN112818248B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110219998.4A CN112818248B (en) 2021-02-26 2021-02-26 Emotion-based article recommendation model construction and recommendation method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110219998.4A CN112818248B (en) 2021-02-26 2021-02-26 Emotion-based article recommendation model construction and recommendation method and system

Publications (2)

Publication Number Publication Date
CN112818248A true CN112818248A (en) 2021-05-18
CN112818248B CN112818248B (en) 2023-04-18

Family

ID=75864126

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110219998.4A Active CN112818248B (en) 2021-02-26 2021-02-26 Emotion-based article recommendation model construction and recommendation method and system

Country Status (1)

Country Link
CN (1) CN112818248B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107403007A (en) * 2017-07-25 2017-11-28 北京工商大学 A kind of method of network Twitter message reliability discriminant model
CN108710680A (en) * 2018-05-18 2018-10-26 哈尔滨理工大学 It is a kind of to carry out the recommendation method of the film based on sentiment analysis using deep learning
CN110490686A (en) * 2019-07-08 2019-11-22 西北大学 A kind of building of commodity Rating Model, recommended method and system based on Time Perception
CN110598130A (en) * 2019-09-30 2019-12-20 重庆邮电大学 Movie recommendation method integrating heterogeneous information network and deep learning
US20210056982A1 (en) * 2019-08-23 2021-02-25 International Business Machines Corporation Sentimental and emotional analysis of electronic documents

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107403007A (en) * 2017-07-25 2017-11-28 北京工商大学 A kind of method of network Twitter message reliability discriminant model
CN108710680A (en) * 2018-05-18 2018-10-26 哈尔滨理工大学 It is a kind of to carry out the recommendation method of the film based on sentiment analysis using deep learning
CN110490686A (en) * 2019-07-08 2019-11-22 西北大学 A kind of building of commodity Rating Model, recommended method and system based on Time Perception
US20210056982A1 (en) * 2019-08-23 2021-02-25 International Business Machines Corporation Sentimental and emotional analysis of electronic documents
CN110598130A (en) * 2019-09-30 2019-12-20 重庆邮电大学 Movie recommendation method integrating heterogeneous information network and deep learning

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
SHAOHUA FAN: "Metapath-guided Heterogeneous Graph Neural Network for Intent Recommendation", 《KDD "19: PROCEEDINGS OF THE 25TH ACM SIGKDD INTERNATIONAL CONFERENCE ON KNOWLEDGE DISCOVERY & DATA MINING》 *
宋瀚章: "基于异构数据的推荐系统", 《中国优秀硕士学位论文全文数据库》 *
张志强: "异质网络中基于语义元路径的推荐系统研究与实现", 《中国优秀硕士学位论文全文数据库》 *

Also Published As

Publication number Publication date
CN112818248B (en) 2023-04-18

Similar Documents

Publication Publication Date Title
Fan et al. Product-aware helpfulness prediction of online reviews
CN102929873B (en) Method and device for extracting searching value terms based on context search
CN108073568A (en) keyword extracting method and device
CN111507350B (en) Text recognition method and device
CN110929166A (en) Content recommendation method, electronic device and storage medium
CN111861550B (en) Family portrait construction method and system based on OTT equipment
Pasupa et al. Hybrid deep learning models for thai sentiment analysis
CN111061939B (en) Scientific research academic news keyword matching recommendation method based on deep learning
CN112307164A (en) Information recommendation method and device, computer equipment and storage medium
Cantini et al. Learning sentence-to-hashtags semantic mapping for hashtag recommendation on microblogs
Zhao et al. Sentiment analysis on the online reviews based on hidden Markov model
CN110110220A (en) Merge the recommended models of social networks and user's evaluation
CN111754208A (en) Automatic screening method for recruitment resumes
Yao et al. A personalized recommendation system based on user portrait
Jiang et al. Structurizing misinformation stories via rationalizing fact-checks
Wei et al. Online education recommendation model based on user behavior data analysis
Alam et al. Social media content categorization using supervised based machine learning methods and natural language processing in bangla language
Chen et al. Exploiting aesthetic features in visual contents for movie recommendation
Walha et al. A Lexicon approach to multidimensional analysis of tweets opinion
CN112818248B (en) Emotion-based article recommendation model construction and recommendation method and system
Kaur et al. News classification using neural networks
CN112231547A (en) Content recommendation method and system based on knowledge graph
CN115964495A (en) News event vein combing method and system
Schmitt et al. Outlier detection on semantic space for sentiment analysis with convolutional neural networks
CN112765465B (en) User map-based recommendation method

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