CN112507248A - Tourist attraction recommendation method based on user comment data and trust relationship - Google Patents

Tourist attraction recommendation method based on user comment data and trust relationship Download PDF

Info

Publication number
CN112507248A
CN112507248A CN202010984598.8A CN202010984598A CN112507248A CN 112507248 A CN112507248 A CN 112507248A CN 202010984598 A CN202010984598 A CN 202010984598A CN 112507248 A CN112507248 A CN 112507248A
Authority
CN
China
Prior art keywords
user
trust
words
attribute
score
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
CN202010984598.8A
Other languages
Chinese (zh)
Inventor
高岭
牛秀娟
黄帝淞
高全力
曹瑞
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 CN202010984598.8A priority Critical patent/CN112507248A/en
Publication of CN112507248A publication Critical patent/CN112507248A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/14—Travel agencies

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Artificial Intelligence (AREA)
  • Tourism & Hospitality (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种基于用户评论数据和信任关系的旅游景点推荐方法，包括获取用户对景点的评论和评分信息、用户和景点属性信息以及用户好友关注列表；对评论信息进行文本预处理；提取评论中的属性‑情感词对并对情感词进行量化，构建主题情感向量；结合用户评分和主题情感向量建立用户偏好分布；计算不同用户之间偏好分布的相似度；结合用户关注列表信息建立目标用户的信任关系集合并计算信任度；结合用户相似度和信任度进行推荐；本发明充分利用用户评论信息进行主题提取，并对用户的评论进行情感分析与量化，同时结合评分数据，时间信息来更好建模用户偏好模型，另外通过构建用户信任关系集合来优化最近邻居选择，从而达到提高推荐准确率的目的。

Figure 202010984598

A method for recommending tourist attractions based on user review data and trust relationship, including obtaining user's review and rating information on scenic spots, user and scenic spot attribute information, and user's friend follow list; text preprocessing on review information; and extracting attributes in reviews ‑Sentiment words are paired and quantified to construct topic sentiment vectors; combine user ratings and topic sentiment vectors to establish user preference distributions; calculate the similarity of preference distributions between different users; combine user follow list information to establish target users’ trust relationships Set and calculate the trust degree; combine user similarity and trust degree for recommendation; the present invention makes full use of user comment information for topic extraction, performs sentiment analysis and quantification on user comments, and combines scoring data and time information to better model User preference model, in addition to optimize the nearest neighbor selection by building a user trust relationship set, so as to achieve the purpose of improving the recommendation accuracy.

Figure 202010984598

Description

一种基于用户评论数据和信任关系的旅游景点推荐方法A tourist attraction recommendation method based on user review data and trust relationship

技术领域technical field

本发明属于数据处理技术领域，具体涉及一种基于用户评论数据和信任关系的旅游景点推荐方法；The invention belongs to the technical field of data processing, and in particular relates to a method for recommending tourist attractions based on user comment data and trust relationship;

背景技术Background technique

互联网与大数据技术的迅速发展丰富了人们获取信息的方式和途径，与之相伴随的，信息的迅速膨胀使得人们在享受信息高速流通带来的便利的同时不得不消耗更多的精力去寻找自己想要获取的信息，甚至使得人们在面对海量信息的过程中无法清晰辨认出自己的目标和明确自己的需求，这就是“信息过载”现象；“信息过载”不但使人们获取自己感兴趣内容的效率变得更低外，而且使得信息空间中大量具有价值的信息无法被及时利用而造成浪费；面对以上这些现象，目前主要采用个性化推荐系统来缓解；The rapid development of the Internet and big data technology has enriched the ways and means for people to obtain information. Accompanying this, the rapid expansion of information makes people have to spend more energy to find the convenience while enjoying the convenience brought by the high-speed circulation of information. The information they want to obtain even makes people unable to clearly identify their own goals and clarify their needs in the process of facing massive amounts of information. This is the phenomenon of "information overload"; "information overload" not only makes people interested in obtaining information about themselves The efficiency of the content has become lower, and a large amount of valuable information in the information space cannot be used in time, resulting in waste; in the face of the above phenomena, the personalized recommendation system is mainly used to alleviate it;

个性化推荐技术目前常用的算法有基于协同过滤的推荐、基于内容的推荐和基于关联规则的推荐；其中，协同过滤算法由于其与项目的具体内容无关的特性，大大简化了技术人员所需要的项目相关的专业知识，已经成为推荐系统的一个主流技术并且得到了广泛的应用；但由于该算法过于依赖历史评分数据的特性使得本技术存在一些局限性；由于每个用户所能接触到的信息有限以及大部分冷门项目很少被浏览而产生历史数据稀疏的问题，从而导致推荐精度严重下降；另外新加入系统的用户或项目由于没有历史信息而产生冷启动的问题；以及随着用户量和商品量的增加，推荐系统的性能也会下降，影响了系统的扩展性；在当今互联网大数据技术不断发展的背景下，数据稀疏的情况是无法避免的，导致产生的目标用户的邻近用户集合选择出现误差，进而会造成推荐质量不高，随着数据规模的扩大，这个问题会更加突出；Personalized recommendation technology currently commonly used algorithms include collaborative filtering-based recommendation, content-based recommendation, and association rule-based recommendation; among them, collaborative filtering algorithm greatly simplifies the needs of technical personnel due to its characteristics that it has nothing to do with the specific content of the project. Project-related expertise has become a mainstream technology in recommender systems and has been widely used; however, due to the fact that the algorithm relies too much on historical scoring data, this technology has some limitations; due to the information that each user can access Limited and most unpopular items are rarely viewed, resulting in sparse historical data, resulting in a serious drop in recommendation accuracy; in addition, new users or items that join the system have a cold start problem because they have no historical information; With the increase in the quantity of commodities, the performance of the recommendation system will also decline, which affects the scalability of the system; under the background of the continuous development of Internet big data technology, the situation of data sparseness is unavoidable, resulting in the generation of a set of adjacent users of the target user. Errors in selection will result in poor recommendation quality. With the expansion of data scale, this problem will become more prominent;

发明内容SUMMARY OF THE INVENTION

为了克服上述现有技术的不足，本发明的目的是提供一种基于用户评论数据和信任关系的旅游景点推荐方法，为了提高推荐系统的准确率，弥补传统协同过滤中严重依赖评分数据导致应用过程中存在诸如数据稀疏性、冷启动、用户兴趣偏移问题；本发明充分利用用户评论信息进行主题提取，并对用户的评论进行情感分析与量化，同时结合评分数据，时间信息来更好建模用户偏好模型，另外通过构建用户信任关系集合来优化最近邻居选择，从而达到提高推荐准确率的目的；In order to overcome the above-mentioned deficiencies of the prior art, the purpose of the present invention is to provide a method for recommending tourist attractions based on user review data and trust relationship, in order to improve the accuracy of the recommendation system, to make up for the application process caused by the heavy reliance on scoring data in traditional collaborative filtering There are problems such as data sparsity, cold start, and user interest offset in the system; the present invention makes full use of user comment information for topic extraction, performs sentiment analysis and quantification on user comments, and combines scoring data and time information to better model User preference model, in addition to optimize the selection of nearest neighbors by building a user trust relationship set, so as to achieve the purpose of improving the recommendation accuracy;

一种基于用户评论数据和信任关系的旅游景点推荐方法，其特征在于，包括以下步骤：A method for recommending tourist attractions based on user review data and trust relationship, characterized in that it includes the following steps:

步骤1，获取用户对景点的评论信息和评分信息、用户属性信息和景点属性信息以及用户好友关注列表；Step 1, obtain the user's comment information and rating information on the scenic spot, user attribute information and scenic spot attribute information, and the user's friend attention list;

步骤2，对评论信息进行文本预处理；Step 2, perform text preprocessing on the comment information;

首先利用Jieba分词对旅游景点的在线评论进行分词和词性标注以及停用词删除.分词和词性标注是将以句子形式表示的评论都分解成以若干词语的形式表示，并且在每个词后标注相应的词性；停用词删除是指将出现频率高，但又没有实际意义的词删除，参照中文停用词表中的停用词(保留标点符号)进行删除；Firstly, Jieba word segmentation is used to perform word segmentation and part-of-speech tagging and stop word removal for online reviews of tourist attractions. Word segmentation and part-of-speech tagging are to decompose comments expressed in sentence form into several words, and label each word after Corresponding part of speech; stop word deletion refers to the deletion of words with high frequency but no actual meaning, referring to the stop words in the Chinese stop word list (retaining punctuation marks) to delete;

步骤3，提取评论中的属性-情感词对并对情感词进行量化，构建主题情感向量；包括：Step 3: Extract attribute-sentiment word pairs in the comments and quantify the sentiment words to construct topic sentiment vectors; including:

1)提取评论中的属性词-情感词对：给出属性词和情感词的定义，属性词表示用户评论中用于描述景点某一特性的词，例如“景色”、“价格”；情感词表示用户对属性词的积极或者消极的情绪、态度和情感，例如“满意”、“失望”等；1) Extract attribute word-sentiment word pairs in comments: give definitions of attribute words and sentiment words, attribute words represent words used to describe a certain feature of scenic spots in user reviews, such as "scenery", "price"; sentiment words Indicates the user's positive or negative emotions, attitudes and emotions towards attribute words, such as "satisfaction", "disappointment", etc.;

首先，用户评论中的属性词和情感词往往是成对出现的，所以情感词通常出现在属性词的附近；其次，注意挖掘修饰情感词的副词，它是情感量化的关键，副词是区分正面情感和负面情感的重要依据，也是判断情感程度的重要信息源；在提取过程中加入了情感词的种子词库，采用句法分析器对句子进行依存句法解析，然后根据所需要的依存关系进行过滤；在整个提取过程中主要涉及两种依存关系：一是副词修饰形容词的状中关系，二是形容词修饰名词的定中关系；First, attribute words and sentiment words in user reviews often appear in pairs, so sentiment words usually appear near the attribute words; second, pay attention to the adverbs that modify sentiment words, which are the key to sentiment quantification, and adverbs are used to distinguish positive It is an important basis for emotion and negative emotion, and also an important information source for judging the degree of emotion; in the extraction process, a seed vocabulary of emotion words is added, and a syntactic analyzer is used to analyze the dependency syntax of the sentence, and then filter according to the required dependencies. ; There are mainly two kinds of dependencies involved in the whole extraction process: one is the adverb-modified adjective's central relationship, and the other is the adjective-modified noun's central relationship;

2)量化评论中属性词对应情感词的情感倾向：首先引入一个用户评论中常用副词的词库，并为每一类副词设定一个程度的值，用于描述用户的情感倾向的程度；设定情感词的词性为两个方面，正面倾向和负面倾向，正面倾向的情感词赋值为1，负面倾向的情感词赋值为-1，否定词赋值为一1，对应属性情感倾向量化值为：V＝词性值×副词的词性程序百分比×否定词词性值；最终求得用户ui对景点rj的属性wn的情感倾向值Sijn；2) Quantify the sentimental tendency of attribute words in comments corresponding to sentimental words: firstly introduce a thesaurus of common adverbs in user comments, and set a degree value for each type of adverbs to describe the degree of user's emotional tendency; There are two parts of speech to determine the sentiment words, positive tendency and negative tendency, the positive tendency sentiment words are assigned as 1, the negative tendency sentiment words are assigned as -1, the negative words are assigned as one 1, and the corresponding attribute sentiment tendency quantification value is: V = part-of-speech value × part-of-speech program percentage of adverbs × negative part-of-speech value; finally, the emotional tendency value S ijn of user ui to attribute wn of scenic spot r j is obtained;

3)构建用户的主题情感向量：给出主题的概念，不同用户在进行评论时描述旅游景点的同一属性可能使用不同的词，如景致、景观和风景等都表示景色，因此需要对描述旅游景点同一属性的词进行合并；利用LDA主题模型可以将属性词聚集成潜在主题，并量化用户在对应主题下情感倾向分值进而量化用户主题情感词向量；首先将提取出的属性词作为LDA主题模型中的特征词汇，利用LDA算法将相关词汇归属到相应的主题，从而属性词被聚集成K个潜在主题，其中每个主题表现为属性词的概率分布，从而将属性词和主题相关联，然后根据主题-属性词分布得到用户ui对景点rj的主题fk的情感倾向值Vijk，具体计算公式如下：3) Constructing the user's topic emotion vector: Given the concept of topic, different users may use different words to describe the same attribute of tourist attractions when commenting, such as scenery, landscape and scenery, etc. The words of the same attribute are combined; the attribute words can be aggregated into potential topics by using the LDA topic model, and the user's emotional tendency score under the corresponding topic can be quantified to quantify the user's topic sentiment word vector; first, the extracted attribute words are used as the LDA topic model The feature words in , use the LDA algorithm to attribute the related words to the corresponding topics, so that the attribute words are aggregated into K potential topics, each of which is expressed as a probability distribution of the attribute words, so as to associate the attribute words with the topic, and then According to the topic-attribute word distribution, the sentiment tendency value V ijk of the user ui to the topic f k of the scenic spot r j is obtained, and the specific calculation formula is as follows:

Figure BDA0002688788140000041
Figure BDA0002688788140000041

其中φkn表示属性词wn属于主题fk的概率；where φ kn represents the probability that the attribute word w n belongs to the topic f k ;

步骤4，结合用户评分和主题情感向量建立用户偏好分布，包括：Step 4, combine user ratings and topic sentiment vectors to establish user preference distribution, including:

为了获得更准确的用户偏好，将用户在评分集合上的平均评分值作为衡量用户对不同景点偏好的标准，同时引入sigmoid函数来计算用户主观评分可信度，用户主观评分可信度的定义为：In order to obtain a more accurate user preference, the average rating value of the user on the rating set is used as the criterion to measure the user's preference for different scenic spots, and the sigmoid function is introduced to calculate the user's subjective rating reliability. The definition of the user's subjective rating reliability is :

Figure BDA0002688788140000042
Figure BDA0002688788140000042

其中

Figure BDA0002688788140000043
表示用户ui对景点j的评分，
Figure BDA0002688788140000044
表示用户评价的所有景点的平均分值；in
Figure BDA0002688788140000043
represents the rating of user u i to the scenic spot j,
Figure BDA0002688788140000044
Indicates the average score of all attractions rated by users;

用户主观评分可信度的取值范围是从0到1；当用户对景点的评分与用户在景点评分集合上的平均评分值之间的差值越大，说明用户对该景点越喜好/厌恶，景点对应的属性分布越符合/不符合用户的偏好；The value range of the user's subjective rating reliability is from 0 to 1; when the difference between the user's rating on the scenic spot and the user's average rating value on the scenic spot rating set is larger, it indicates that the user likes/dislikes the scenic spot more. , the attribute distribution corresponding to the scenic spot is more in line with/in line with the user's preference;

另一方面，利用用户客观评分可信度来衡量用户评分的可靠性；如果用户对景点的评分接近景点的平均得分，则表示用户相对客观，反之亦然；用户客观评分可信度定义为：On the other hand, the reliability of the user's objective rating is used to measure the reliability of the user's rating; if the user's rating of the scenic spot is close to the average score of the scenic spot, it means that the user is relatively objective, and vice versa; the user's objective rating reliability is defined as:

Figure BDA0002688788140000051
Figure BDA0002688788140000051

用户客观评分可信度的取值范围为0到1，用户对景点的评分和景点的平均评分之间的差异越小，说明用户越客观；换言之，用户的客观评分可信度相对较高；The user's objective rating reliability ranges from 0 to 1. The smaller the difference between the user's rating of the scenic spot and the average rating of the scenic spot, the more objective the user is; in other words, the user's objective rating reliability is relatively high;

考虑到用户兴趣偏移的问题，加入时间修正权重来计算用户偏好分布时，使得越靠近当前时间的评分与评论的权重增加，因为其更能反映用户当下的偏好；首先将用户评分和评论的时间跨度通过标准化转换的方法映射到[-1，1]的变化范围，然后将映射后的时间用logistic函数处理后得到用户在不同时间段评分和评论的权重，不同时间段的权值计算公式如下：Considering the problem of user interest offset, when adding time correction weights to calculate user preference distribution, the weights of ratings and comments closer to the current time are increased, because they can better reflect the user's current preferences; The time span is mapped to the variation range of [-1, 1] through the method of standardized transformation, and then the mapped time is processed by the logistic function to obtain the weight of the user's ratings and comments in different time periods, and the calculation formula of the weights of different time periods. as follows:

Figure BDA0002688788140000052
Figure BDA0002688788140000052

其中

Figure BDA0002688788140000053
表示用户ui对于景点j评论的时间标准化后的结果。
Figure BDA0002688788140000054
表示的时间权重的范围在0到1之间且权值随着时间
Figure BDA0002688788140000061
的增加而增加；in
Figure BDA0002688788140000053
Represents the time-standardized result of user ui 's comments on scenic spot j.
Figure BDA0002688788140000054
The time weight represented ranges from 0 to 1 and the weight varies with time
Figure BDA0002688788140000061
increase with the increase;

将用户主题情感向量、用户评分、时间加权因子、用户主观评分可信度、用户客观评分可信度结合起来，得到最终的用户偏好分布，计算公式如下：Combining the user topic sentiment vector, user rating, time weighting factor, user subjective rating credibility, and user objective rating credibility to obtain the final user preference distribution, the calculation formula is as follows:

Figure BDA0002688788140000062
Figure BDA0002688788140000062

其中

Figure BDA0002688788140000063
表示用户ui对景点j的评分，
Figure BDA0002688788140000064
表示用户ui对景点j的主题情感向量，表示用户ui评论或者评分的所有景点的集合，
Figure BDA0002688788140000065
表示被用户ui评论过的景点数量。加入用户主观评分可信度和用户客观评分可信度这两个部分，可以有效地避免用户的错误或随机评论，从而获得更为真实的用户偏好分布；in
Figure BDA0002688788140000063
represents the rating of user u i to the scenic spot j,
Figure BDA0002688788140000064
represents the theme sentiment vector of user ui to scenic spot j, represents the set of all scenic spots that user ui commented or rated,
Figure BDA0002688788140000065
Indicates the number of attractions reviewed by user ui . Adding the two parts of user subjective rating reliability and user objective rating reliability can effectively avoid user errors or random comments, so as to obtain a more realistic user preference distribution;

步骤5，计算不同用户之间偏好分布的相似度，包括：Step 5: Calculate the similarity of preference distributions between different users, including:

在获得用户对景点不同特征的偏好分布后，采用JS散度来计算用户之间的相似性；根据JS散度可以得到用户ui和uj之间的用户相似度，计算公式如下：After obtaining the user's preference distribution for different features of scenic spots, JS divergence is used to calculate the similarity between users; according to the JS divergence, the user similarity between users u i and u j can be obtained, and the calculation formula is as follows:

Figure BDA0002688788140000066
Figure BDA0002688788140000066

其中，

Figure BDA0002688788140000067
in,
Figure BDA0002688788140000067

其中，

Figure BDA0002688788140000068
in,
Figure BDA0002688788140000068

步骤6，结合用户关注列表信息建立目标用户的信任关系集合并计算信任度，包括：Step 6, establishing a trust relationship set of the target user in combination with the user attention list information and calculating the trust degree, including:

获取用户关注列表信息，得到全局信任网络包括用户结点的属性与结点之间的信任关系，将全局信任网络中的用户属性以及信任关系信息存储为树结构，获取以目标用户为根结点的子树，即目标用户局部信任网络；以目标结点为A结点为例，根据“六度分割理论”，将距离节点A的最短路径小于6的所有节点视作A的信任节点(包括直接信任和间接信任)，从而得到用户A的信任节点集合；Obtain user watchlist information, obtain the global trust network including the attributes of user nodes and the trust relationship between nodes, store the user attributes and trust relationship information in the global trust network as a tree structure, and obtain the target user as the root node. The subtree of the target user, that is, the local trust network of the target user; taking the target node as node A as an example, according to the "six degrees of division theory", all nodes with the shortest path from node A less than 6 are regarded as trust nodes of A (including Direct trust and indirect trust), so as to obtain the set of trust nodes of user A;

计算信任度的方法主要分为两步：The method of calculating the trust degree is mainly divided into two steps:

步骤1：将目标用户设置为起始节点，假设目标用户完全信任自己，设置初始信任度为1；Step 1: Set the target user as the starting node, assume that the target user completely trusts himself, and set the initial trust degree to 1;

步骤2：为目标用户(起始节点)的直接信任用户和间接信任用户分配信任度，计算公式如下：Step 2: Assign the trust degree to the direct trust user and indirect trust user of the target user (starting node), the calculation formula is as follows:

trustcur＝trustpre*wi,j trust cur = trust pre *w i,j

Figure BDA0002688788140000071
Figure BDA0002688788140000071

当前结点的信任度的值是前一结点的信任度与权重的乘积，其中l表示用户在网络中的层数；The value of the trust degree of the current node is the product of the trust degree and the weight of the previous node, where l represents the number of layers of the user in the network;

步骤7，结合用户相似度和信任度进行推荐，包括：Step 7, combine user similarity and trust to recommend, including:

1)生成邻居集合，采用在当前用户的信任列表中，选择信任度最高的N个用户作为当前用户的邻居，若该用户的信任列表中用户数量不足N，则采用JS散度来计算当前用户ui与其他用户uj的相似度，选择相似度最高的K个用户补充进来，作为该用户的最近邻居；1) Generate a neighbor set, use the current user's trust list, and select the N users with the highest trust degree as the current user's neighbors. If the number of users in the user's trust list is less than N, the JS divergence is used to calculate the current user. The similarity between u i and other users u j , select the K users with the highest similarity as the nearest neighbors of the user;

2)结合相似度和用户信任度来计算推荐的权重，计算公式如下：2) Combine the similarity and user trust to calculate the recommended weight. The calculation formula is as follows:

Figure BDA0002688788140000081
Figure BDA0002688788140000081

其中

Figure BDA0002688788140000082
表示用户ui与用户uj之间用户偏好的相似度，
Figure BDA0002688788140000083
表示用户ui对用户uj之间的信任度，
Figure BDA0002688788140000084
表示推荐权重；in
Figure BDA0002688788140000082
represents the similarity of user preferences between user u i and user u j ,
Figure BDA0002688788140000083
represents the degree of trust between user u i and user u j ,
Figure BDA0002688788140000084
Indicates the recommendation weight;

3)采用协同过滤中的Resnick公式来计算当前用户对目标景点的预测评分值；3) Using the Resnick formula in collaborative filtering to calculate the current user's predicted rating value for the target scenic spot;

Figure BDA0002688788140000085
Figure BDA0002688788140000085

其中

Figure BDA0002688788140000086
和
Figure BDA0002688788140000087
分别表示用户ui和uk对已评分景点的评分均值，Nei(i)表示用户ui的邻居集合。ak,j表示用户ui对景点rj的评分；in
Figure BDA0002688788140000086
and
Figure BDA0002688788140000087
are the mean scores of users ui and uk to the rated scenic spots, respectively, and Nei(i) represents the set of neighbors of user ui . a k,j represents user ui 's rating of scenic spot r j ;

进一步地，所述的用户属性信息包括用户ID和用户名；用户评论评分信息包括用户对景点的所有评论与评分信息以及评论和评分的时间信息；景点属性信息包括景点名称和景点ID；用户好友关注信息包括用户关注的所有好友的id和名称；获取数据的方式为利用网络爬虫。Further, the user attribute information includes user ID and user name; user comment rating information includes all comments and rating information of users on scenic spots and time information of comments and ratings; scenic spot attribute information includes scenic spot name and scenic spot ID; user friends The following information includes the ids and names of all friends followed by the user; the way to obtain the data is to use a web crawler.

本发明的有益效果是：The beneficial effects of the present invention are:

1)本发明结合用户评分和用户评论，对评论中的情感信息进行提取与量化，同时引入用户主观评分可信度和客观评分可信度来避免错误评论和恶意随机评论的影响，运用LDA概率模型挖掘出用户潜在的主题情感倾向，构建准确的用户偏好分布。本发明还考虑了用户偏好随时间偏移的问题，将用户评分和评论的时间跨度用logistic函数处理后得到用户在不同时间段评分和评论的权重，时间距当下越近，权重越大。1) The present invention combines user ratings and user reviews to extract and quantify emotional information in reviews, and introduces user subjective rating reliability and objective rating reliability to avoid the influence of erroneous reviews and malicious random reviews, and uses LDA probability. The model digs out the user's potential topic emotional tendency and constructs an accurate user preference distribution. The invention also considers the problem of user preference shifting with time, and processes the time span of user ratings and comments with a logistic function to obtain the weights of user ratings and comments in different time periods. The closer the time is to the present, the greater the weight.

2)通过本发明可有效解决传统协同过滤算法应用时存在的数据稀疏性的问题。即在数据稀疏情况下，一方面，若相似度无法计算，系统就无法对这些用户进行推荐；另一方面，当用户间共同评价的景点数量很少时，计算出的相似度也存在一定的误差。本发明通过获取用户的关注列表，构建目标用户的信任结点集合并计算信任度，通过融合相似度和信任度用来计算推荐权重，提高了推荐的准确性。2) The present invention can effectively solve the problem of data sparsity existing in the application of traditional collaborative filtering algorithms. That is, in the case of sparse data, on the one hand, if the similarity cannot be calculated, the system cannot recommend these users; on the other hand, when the number of scenic spots jointly evaluated by users is small, the calculated similarity also has a certain degree of similarity. error. The invention obtains the user's attention list, constructs the trust node set of the target user, calculates the trust degree, and uses the similarity and the trust degree to calculate the recommendation weight, thereby improving the accuracy of the recommendation.

附图说明Description of drawings

图1为本发明的流程示意图；Fig. 1 is the schematic flow chart of the present invention;

图2为全局信任网络示意图；Figure 2 is a schematic diagram of a global trust network;

图3为以目标节点为根结点的局部信任网络示意图；3 is a schematic diagram of a local trust network with a target node as a root node;

图4为目标用户与其信任用户之间以同心圆的形式展示的网络结构示意图。FIG. 4 is a schematic diagram of a network structure shown in the form of concentric circles between a target user and its trusted users.

具体实施方式Detailed ways

以下结合附图对本发明进一步叙述。The present invention is further described below in conjunction with the accompanying drawings.

如图1至图3所示，如图1至图3所示，包括以下步骤：As shown in Figure 1 to Figure 3, as shown in Figure 1 to Figure 3, including the following steps:

步骤1，获取用户对景点的评论信息和评分信息、用户和景点属性信息以及用户好友关注列表。In step 1, the user's comment information and rating information on the scenic spot, the attribute information of the user and the scenic spot, and the user's friend attention list are obtained.

所获取的用户属性信息包括用户ID和用户名；用户评论评分信息包括用户对景点的所有评论与评分信息以及评论和评分的时间信息；景点属性信息包括景点名称和景点ID；用户关注好友信息包括用户关注的所有好友的id和名称；获取数据的方式为利用网络爬虫。The acquired user attribute information includes user ID and user name; user comment rating information includes all user comments and ratings information on scenic spots and time information of comments and ratings; scenic spot attribute information includes scenic spot name and scenic spot ID; user follow friend information includes The ids and names of all friends the user follows; the way to obtain data is to use web crawlers.

步骤2，对评论信息进行文本预处理。Step 2, perform text preprocessing on the comment information.

文本预处理首先利用Jieba分词对旅游景点的在线评论进行分词和词性标注以及停用词删除。分词和词性标注是将以句子形式表示的评论都分解成以若干词语的形式表示，并且在每个词后标注相应的词性。停用词删除是指将出现频率高，但又没有实际意义的词删除，参照中文停用词表中的停用词(保留标点符号)进行删除。The text preprocessing first uses Jieba word segmentation to perform word segmentation and part-of-speech tagging and stop word removal for online reviews of tourist attractions. Word segmentation and part-of-speech tagging is to decompose comments expressed in sentence form into several words, and mark the corresponding part-of-speech after each word. Stop word deletion refers to the deletion of words with high frequency but no actual meaning, referring to the stop words in the Chinese stop word list (retaining punctuation marks) for deletion.

步骤3，提取评论中的属性-情感词对并对情感词进行量化，构建主题情感向量。Step 3: Extract attribute-sentiment word pairs in the comments and quantify the sentiment words to construct topic sentiment vectors.

1)提取评论中的属性词-情感词对：给出属性词和情感词的定义，属性词表示用户评论中用于描述景点某一特性的词，例如“景色”、“价格”等。情感词表示用户对属性词的积极或者消极的情绪、态度和情感，例如“满意”、“失望”等。1) Extract attribute word-sentiment word pairs in comments: give definitions of attribute words and sentiment words, attribute words represent words used to describe a certain feature of scenic spots in user reviews, such as "scenery", "price", etc. Sentiment words represent the user's positive or negative emotions, attitudes and emotions towards attribute words, such as "satisfaction", "disappointment" and so on.

首先，用户评论中的属性词和情感词往往是成对出现的，所以情感词通常出现在属性词的附近；其次，注意挖掘修饰情感词的副词，它是情感量化的关键，副词是区分正面情感和负面情感的重要依据，也是判断情感程度的重要信息源。在提取过程中加入了情感词的种子词库，采用句法分析器对句子进行依存句法解析，然后根据所需要的依存关系进行过滤。在整个提取过程中主要涉及两种依存关系：一是副词修饰形容词的状中关系，二是形容词修饰名词的定中关系。First, attribute words and sentiment words in user reviews often appear in pairs, so sentiment words usually appear near the attribute words; second, pay attention to the adverbs that modify sentiment words, which are the key to sentiment quantification, and adverbs are used to distinguish positive It is an important basis for emotion and negative emotion, and it is also an important information source for judging the degree of emotion. In the extraction process, a seed lexicon of sentiment words is added, and a syntactic analyzer is used to parse the sentence dependencies, and then filter according to the required dependencies. The whole extraction process mainly involves two kinds of dependencies: one is the adverb-modified adjective's central relationship, and the other is the adjective-modified noun's central relationship.

2)量化评论中属性词对应情感词的情感倾向：首先引入一个用户评论中常用副词的词库，并为每一类副词设定一个程度的值，用于描述用户的情感倾向的程度。设定情感词的词性为两个方面，正面倾向和负面倾向，正面倾向的情感词赋值为1，负面倾向的情感词赋值为-1，否定词赋值为-1，对应属性情感倾向量化值为：V＝词性值×副词的词性程序百分比×否定词词性值。最终求得用户ui对景点rj的属性wn的情感倾向值Sijn。2) Quantify the sentiment tendency of the attribute words in the comments corresponding to the sentiment words: firstly, a thesaurus of common adverbs in user comments is introduced, and a degree value is set for each type of adverbs to describe the degree of the user's emotional tendency. The part of speech of sentiment words is set as two aspects, positive tendency and negative tendency, the positive tendency sentiment word is assigned as 1, the negative tendency sentiment word is assigned as -1, the negative word is assigned as -1, and the corresponding attribute sentiment tendency is quantified as 1. : V = part-of-speech value × part-of-speech program percentage of adverbs × negative part-of-speech value. Finally, the emotional tendency value S ijn of the user ui to the attribute wn of the scenic spot r j is obtained.

3)构建用户的主题情感向量：给出主题的概念，不同用户在进行评论时描述旅游景点的同一属性可能使用不同的词，如景致、景观和风景等都表示景色，因此需要对描述旅游景点同一属性的词进行合并。利用LDA主题模型可以将属性词聚集成潜在主题，并量化用户在对应主题下情感倾向分值进而量化用户主题情感词向量。首先将提取出的属性词作为LDA主题模型中的特征词汇，利用LDA算法将相关词汇归属到相应的主题，从而属性词被聚集成K个潜在主题，其中每个主题表现为属性词的概率分布，从而将属性词和主题相关联，然后根据主题-属性词分布得到用户ui对景点rj的fk主题的情感倾向值Vijk，具体计算公式如下：3) Constructing the user's topic emotion vector: Given the concept of topic, different users may use different words to describe the same attribute of tourist attractions when commenting, such as scenery, landscape and scenery, etc. Words of the same attribute are merged. Using the LDA topic model, attribute words can be aggregated into potential topics, and the user's sentiment tendency score under the corresponding topic can be quantified to quantify the user's topic sentiment word vector. First, the extracted attribute words are used as the feature words in the LDA topic model, and the related words are attributed to the corresponding topics by using the LDA algorithm, so that the attribute words are aggregated into K potential topics, and each topic is represented by the probability distribution of the attribute words. , so as to associate the attribute word with the topic, and then according to the topic-attribute word distribution, the sentiment tendency value V ijk of the user ui to the f k topic of the scenic spot r j is obtained. The specific calculation formula is as follows:

Figure BDA0002688788140000111
Figure BDA0002688788140000111

其中φkn表示属性词wn属于主题fk的概率。where φ kn represents the probability that the attribute word w n belongs to the topic f k .

步骤4，结合用户评分和主题情感向量建立用户偏好分布。Step 4, combining user ratings and topic sentiment vectors to establish user preference distribution.

为了获得更准确的用户偏好，将用户在评分集合上的平均评分值作为衡量用户对不同景点偏好的标准，同时引入sigmoid函数来计算用户主观评分可信度，用户主观评分可信度的定义为：In order to obtain a more accurate user preference, the average rating value of the user on the rating set is used as the criterion to measure the user's preference for different scenic spots, and the sigmoid function is introduced to calculate the user's subjective rating reliability. The definition of the user's subjective rating reliability is :

Figure BDA0002688788140000121
Figure BDA0002688788140000121

其中

Figure BDA0002688788140000122
表示用户ui对景点j的评分，
Figure BDA0002688788140000123
表示用户评价的所有景点的平均分值。in
Figure BDA0002688788140000122
represents the rating of user u i to the scenic spot j,
Figure BDA0002688788140000123
Represents the average score of all attractions rated by users.

用户主观评分可信度的取值范围是从0到1。当用户对景点的评分与用户在景点评分集合上的平均评分值之间的差值越大，说明用户对该景点越喜好/厌恶，景点对应的属性分布越符合/不符合用户的偏好。The user's subjective rating reliability ranges from 0 to 1. When the difference between the user's rating on the scenic spot and the user's average rating value on the scenic spot rating set is larger, it indicates that the user likes/dislikes the scenic spot more, and the attribute distribution corresponding to the scenic spot is more in line with/dislike the user's preference.

另一方面，利用用户客观评分可信度来衡量用户评分的可靠性。如果用户对景点的评分接近景点的平均得分，则表示用户相对客观，反之亦然。用户客观评分可信度定义为：On the other hand, the reliability of user ratings is measured by the reliability of user objective ratings. If the user's rating of the attraction is close to the average score of the attraction, it means that the user is relatively objective, and vice versa. User objective rating reliability is defined as:

Figure BDA0002688788140000124
Figure BDA0002688788140000124

用户客观评分可信度的取值范围为0到1，用户对景点的评分和景点的平均评分之间的差异越小，说明用户越客观。换言之，用户的客观评分可信度相对较高。The value range of the user's objective rating reliability is 0 to 1. The smaller the difference between the user's rating of the scenic spot and the average rating of the scenic spot, the more objective the user is. In other words, the user's objective rating reliability is relatively high.

考虑到用户兴趣偏移的问题，加入时间修正权重来计算用户偏好分布时，使得越靠近当前时间的评分与评论的权重增加，因为其更能反映用户当下的偏好。首先将用户评分和评论的时间跨度通过标准化转换的方法映射到[-1，1]的变化范围，然后将映射后的时间用logistic函数处理后得到用户在不同时间段评分和评论的权重，不同时间段的权值计算公式如下：Considering the problem of user interest offset, when adding time correction weights to calculate user preference distribution, the weights of ratings and comments closer to the current time are increased, because they can better reflect the user's current preferences. First, the time span of user ratings and comments is mapped to the variation range of [-1, 1] by the method of standardized transformation, and then the mapped time is processed with the logistic function to obtain the weights of user ratings and comments in different time periods. The formula for calculating the weight of the time period is as follows:

Figure BDA0002688788140000131
Figure BDA0002688788140000131

其中

Figure BDA0002688788140000132
表示用户ui对于景点j评论的时间标准化后的结果。
Figure BDA0002688788140000133
表示的时间权重的范围在0到1之间且权值随着时间
Figure BDA0002688788140000134
的增加而增加。in
Figure BDA0002688788140000132
Represents the time-standardized result of user ui 's comments on scenic spot j.
Figure BDA0002688788140000133
The time weight represented ranges from 0 to 1 and the weight varies with time
Figure BDA0002688788140000134
increases with the increase.

将用户主题情感向量、用户评分、时间加权因子、用户主观评分可信度、用户客观评分可信度结合起来，得到最终的用户偏好分布，计算公式如下：Combining the user topic sentiment vector, user rating, time weighting factor, user subjective rating credibility, and user objective rating credibility to obtain the final user preference distribution, the calculation formula is as follows:

Figure BDA0002688788140000135
Figure BDA0002688788140000135

其中

Figure BDA0002688788140000136
表示用户ui对景点j的评分，
Figure BDA0002688788140000137
表示用户ui对景点j的主题情感向量，
Figure BDA0002688788140000138
表示用户ui评论或者评分的所有景点的集合，
Figure BDA0002688788140000139
表示被用户ui评论过的景点数量。加入用户主观评分可信度和用户客观评分可信度这两个部分，可以有效地避免用户的错误或随机评论，从而获得更为真实的用户偏好分布。in
Figure BDA0002688788140000136
represents the rating of user u i to the scenic spot j,
Figure BDA0002688788140000137
represents the subject sentiment vector of user u i to the scenic spot j,
Figure BDA0002688788140000138
represents the set of all attractions reviewed or rated by user ui ,
Figure BDA0002688788140000139
Indicates the number of attractions reviewed by user ui . Adding the two parts of user subjective rating reliability and user objective rating reliability can effectively avoid user errors or random comments, so as to obtain a more realistic user preference distribution.

步骤5，计算不同用户之间偏好分布的相似度。Step 5: Calculate the similarity of preference distributions among different users.

在获得用户对景点不同特征的偏好分布后，采用JS散度来计算用户之间的相似性。根据JS散度可以得到用户ui和uj之间的用户相似度，计算公式如下：After obtaining the user's preference distribution for different features of scenic spots, JS divergence is used to calculate the similarity between users. According to the JS divergence, the user similarity between users u i and u j can be obtained. The calculation formula is as follows:

Figure BDA0002688788140000141
Figure BDA0002688788140000141

其中，

Figure BDA0002688788140000142
in,
Figure BDA0002688788140000142

其中，

Figure BDA0002688788140000143
in,
Figure BDA0002688788140000143

步骤6，结合用户关注列表信息建立目标用户的信任关系集合并计算信任度。Step 6, establishing a trust relationship set of the target user in combination with the user attention list information and calculating the trust degree.

如果用户A关注了用户B，则用户A和用户B有直接信任关系，如果用户B关注了用户C，则用户B和用户C也有直接信任关系，同时用户A和用户C之间通过用户B产生间接信任关系。所以信任网络中各节点的信任性质上可分为两种：直接信任和间接信任。根据提取的各个用户的关注列表信息，构建初始的信任网络。在信任网络中，为了形式化表示用户之间的信任关系，采用节点代表用户并使用有向边代表一个用户对另一个用户的直接信任关系，则用户的信任网络可记为G(V，E)，V表示用户节点的集合，E表示信任关系的集合。若A信任B，则可用A→B表示。两用户互相信任的情况，可以被看做是单向信任的特殊情况，可用双向有向边表示，即

Figure BDA0002688788140000144
全局信任网络的示意图如图2所示。If user A follows user B, then user A and user B have a direct trust relationship; if user B follows user C, then user B and user C also have a direct trust relationship, and the relationship between user A and user C is generated through user B Indirect trust relationship. Therefore, the trust properties of each node in the trust network can be divided into two types: direct trust and indirect trust. According to the extracted attention list information of each user, an initial trust network is constructed. In the trust network, in order to formally represent the trust relationship between users, nodes are used to represent users and directed edges are used to represent the direct trust relationship between one user and another user, then the user's trust network can be recorded as G(V, E ), V represents the set of user nodes, and E represents the set of trust relationships. If A trusts B, it can be represented by A→B. The situation where two users trust each other can be regarded as a special case of one-way trust, which can be represented by a two-way directed edge, that is,
Figure BDA0002688788140000144
A schematic diagram of the global trust network is shown in Figure 2.

将全局信任网络中的用户属性以及信任关系信息存储为树结构，获取以目标用户为根结点的子树，即目标用户局部信任网络，如图3所示。以目标结点为A结点为例，根据“六度分割理论”，将距离节点A的最短路径小于6的所有节点视作A的信任节点(包括直接信任和间接信任)，从而得到用户A的信任节点集合。为了更清晰地表示，将用户A与其信任用户之间的网络结构以同心圆的形式表示，如图4所示，其中第1层为用户A，第2层为用户A的直接信赖的用户，第3层及以外的节点为用户A通过其他用户而间接信任的用户。The user attributes and trust relationship information in the global trust network are stored as a tree structure, and the subtree with the target user as the root node is obtained, that is, the local trust network of the target user, as shown in Figure 3. Taking the target node as node A as an example, according to the "six degrees of division theory", all nodes whose shortest path from node A is less than 6 are regarded as trust nodes of A (including direct trust and indirect trust), so as to obtain user A. set of trusted nodes. For a clearer representation, the network structure between user A and its trusted users is represented in the form of concentric circles, as shown in Figure 4, where the first layer is user A, and the second layer is user A's directly trusted users. Nodes at layer 3 and beyond are users that User A indirectly trusts through other users.

计算信任度的方法主要分为两步：The method of calculating the trust degree is mainly divided into two steps:

步骤1：将目标用户设置为起始节点，假设目标用户完全信任自己，设置初始信任度为1。Step 1: Set the target user as the starting node, assuming that the target user completely trusts himself, and set the initial trust degree to 1.

步骤2：为目标用户(起始节点)的直接信任用户和间接信任用户分配信任度，计算公式如下：Step 2: Assign the trust degree to the direct trust user and indirect trust user of the target user (starting node), the calculation formula is as follows:

trustcur＝trustpre*wi,j trust cur = trust pre *w i,j

Figure BDA0002688788140000151
Figure BDA0002688788140000151

当前结点的信任度的值是前一结点的信任度与权重的乘积，其中l表示用户在网络中的层数。The value of the trust degree of the current node is the product of the trust degree of the previous node and the weight, where l represents the number of layers of the user in the network.

步骤7，结合用户相似度和信任度进行推荐。Step 7, combine user similarity and trust to recommend.

1)生成邻居集合，采用在当前用户的信任列表中，选择信任度最高的N个用户作为当前用户的邻居，若该用户的信任列表中用户数量不足N，则采用JS散度来计算当前用户ui与其他用户uj的相似度，选择相似度最高的K个用户补充进来，作为该用户的最近邻居。1) Generate a neighbor set, use the current user's trust list, and select the N users with the highest trust degree as the current user's neighbors. If the number of users in the user's trust list is less than N, the JS divergence is used to calculate the current user. The similarity between u i and other users u j is determined by selecting the K users with the highest similarity as the nearest neighbors of the user.

2)结合相似度和用户信任度来计算推荐的权重，计算公式如下：2) Combine the similarity and user trust to calculate the recommended weight. The calculation formula is as follows:

Figure BDA0002688788140000161
Figure BDA0002688788140000161

其中

Figure BDA0002688788140000162
表示用户ui与用户uj之间用户偏好的相似度，
Figure BDA0002688788140000163
表示用户ui对用户uj之间的信任度，
Figure BDA0002688788140000164
表示推荐权重。in
Figure BDA0002688788140000162
represents the similarity of user preferences between user u i and user u j ,
Figure BDA0002688788140000163
represents the degree of trust between user u i and user u j ,
Figure BDA0002688788140000164
Indicates the recommendation weight.

3)采用协同过滤中的Resnick公式来计算当前用户对目标景点的预测评分值。3) Using the Resnick formula in collaborative filtering to calculate the current user's predicted rating value for the target scenic spot.

Figure BDA0002688788140000165
Figure BDA0002688788140000165

其中

Figure BDA0002688788140000166
和
Figure BDA0002688788140000167
分别表示用户ui和uk对已评分景点的评分均值，Nei(i)表示用户ui的邻居集合。ak,j表示用户ui对景点rj的评分。in
Figure BDA0002688788140000166
and
Figure BDA0002688788140000167
are the mean scores of users ui and uk to the rated scenic spots, respectively, and Nei(i) represents the set of neighbors of user ui . a k,j represents the user ui 's rating of the scenic spot r j .

Claims (2)

1. A tourist attraction recommendation method based on user comment data and trust relationship is characterized by comprising the following steps:
step 1, obtaining comment information and scoring information of a user on a scenic spot, user attribute information and scenic spot attribute information and a user friend attention list;
step 2, text preprocessing is carried out on the comment information;
the method comprises the following steps of firstly, utilizing a Jieba word segmentation to carry out word segmentation, part-of-speech tagging and stop word deletion on online comments of tourist attractions, wherein the word segmentation and part-of-speech tagging are to decompose the comments expressed in the form of sentences into expressions in the form of a plurality of words, and tag corresponding part-of-speech after each word, wherein the stop word deletion refers to deleting words which have high occurrence frequency but no actual meaning and delete stop words (reserved punctuations) in a Chinese stop word list;
step 3, extracting attribute-emotion word pairs in the comments, quantizing the emotion words, and constructing a theme emotion vector; the method comprises the following steps:
1) extracting attribute word-emotion word pairs in the comment: the method comprises the following steps of giving definitions of attribute words and emotion words, wherein the attribute words represent words used for describing certain characteristics of scenic spots in user comments, such as 'scenery' and 'price', and the emotion words represent positive or negative emotions, attitudes and emotions of users on the attribute words;
firstly, attribute words and emotion words in user comments often appear in pairs, so that the emotion words usually appear near the attribute words; secondly, paying attention to the fact that adverbs for modifying emotion words are mined and are key for emotion quantification, the adverbs are important bases for distinguishing positive emotions from negative emotions and are also important information sources for judging emotion degrees, a seed word bank of the emotion words is added in the extraction process, a syntax analyzer is adopted for carrying out dependency syntax analysis on sentences, then filtering is carried out according to required dependency relations, and two dependency relations are mainly involved in the whole extraction process: firstly, the adverb modifies the middle-form relationship of the adjective, and secondly, the adjective modifies the middle-form relationship of the noun;
2) quantifying the emotional tendency of the attribute words in the comment corresponding to the emotional words: firstly, introducing a word bank of common adverbs in user comments, setting a degree value for each type of adverb for describing the emotional tendency degree of a user, setting the part of speech of the emotional words into two aspects, namely positive tendency and negative tendency, assigning the value of the emotional words with the positive tendency to be 1, assigning the value of the emotional words with the negative tendency to be-1, assigning the value of the negative words to be-1, and setting the corresponding attribute emotional tendency quantization value to be: v is the part of speech value x part of speech program percentage of adverb x part of speech value of negative word, finally, user u is obtainediTo the scenery spot rjProperty w ofnEmotional tendency value S ofijn；
3) Constructing a theme emotion vector of the user: the concept of the theme is given, different words may be used for describing the same attribute of the scenic spots when different users comment, for example, the scene, the landscape and the landscape all represent the scene, so the words for describing the same attribute of the scenic spots need to be combined, the attribute words can be gathered into the potential theme by using the LDA theme model, and the emotional tendency scores of the users under the corresponding theme are quantized so as to quantize the emotional word vector of the user theme; firstly, the extracted attribute words are used as feature words in an LDA topic model, the related words are attributed to the corresponding topics by utilizing an LDA algorithm, so that the attribute words are gathered into K potential topics, each topic is represented as the probability distribution of the attribute words, the attribute words are associated with the topics, and then the user u is obtained according to the distribution of the topic-attribute wordsiTo the scenery spot rjSubject f ofkValue of emotional tendency VijkThe specific calculation formula is as follows:
Figure FDA0002688788130000021
wherein phiknRepresenting an attribute word wnBelonging to the subject fkThe probability of (d);
step 4, establishing user preference distribution by combining user scores and theme emotion vectors, wherein the method comprises the following steps:
in order to obtain more accurate user preference, the average score value of the user on a score set is used as a standard for measuring the preference of the user to different scenic spots, and a sigmoid function is introduced to calculate the user subjective score reliability, wherein the user subjective score reliability is defined as follows:
Figure FDA0002688788130000031
wherein,
Figure FDA0002688788130000032
representing user uiThe score for the sight j is determined,
Figure FDA0002688788130000033
average scores representing all sights rated by the user;
the value range of the user subjective score credibility is from 0 to 1; when the difference value between the score of the user for the scenery spot and the average score value of the user on the scenery spot score set is larger, the scenery is more like or dislike by the user, and the attribute distribution corresponding to the scenery is more accordant with or not accordant with the preference of the user;
on the other hand, the reliability of the user score is measured by using the reliability of the user objective score; if the user's score for the attraction is close to the average score for the attraction, it indicates that the user is relatively objective, and vice versa; the credibility of the objective scoring of the user is defined as:
Figure FDA0002688788130000034
the value range of the objective scoring credibility of the user is 0 to 1, and the difference between the score of the user on the scenic spot and the average score of the scenic spot is smaller, so that the objective scoring of the user is more objective; in other words, the reliability of the objective scoring of the user is relatively high;
when the time correction weight is added to calculate the preference distribution of the user in consideration of the problem of the bias of the interest of the user, the weights of the scores and the comments which are closer to the current time are increased, because the weights can reflect the current preference of the user better; firstly, mapping the time span of user scores and comments to the variation range of [ -1, 1] by a standardized conversion method, then processing the mapped time by a logistic function to obtain the weights of the scores and comments of the user in different time periods, wherein the weight calculation formulas of the different time periods are as follows:
Figure FDA0002688788130000041
wherein,
Figure FDA0002688788130000042
representing user uiThe result after the time of the review of sight j is normalized;
Figure FDA0002688788130000043
Figure FDA0002688788130000044
the temporal weight represented ranges between 0 and 1 and the weight is over time
Figure FDA0002688788130000045
Is increased with an increase in;
combining the user theme emotion vector, the user score, the time weighting factor, the user subjective score credibility and the user objective score credibility to obtain the final user preference distribution, wherein the calculation formula is as follows:
Figure FDA0002688788130000046
wherein,
Figure FDA0002688788130000047
representing user uiThe score for the sight j is determined,
Figure FDA0002688788130000048
representing user uiFor the subject emotion vector for sight j,
Figure FDA0002688788130000049
representing user uiA collection of all sights that are commented or scored,
Figure FDA00026887881300000410
represents the user uiThe number of the scenic spots which are commented is added with two parts of user subjective rating reliability and user objective rating reliability, so that errors or random comments of the user can be effectively avoided, and more real user preference distribution is obtained;
step 5, calculating the similarity of preference distribution among different users, comprising:
after preference distribution of the user on different characteristics of the scenic spots is obtained, calculating the similarity between the users by adopting JS divergence; obtaining the user u according to the JS divergenceiAnd ujThe user similarity between the two is calculated according to the following formula:
Figure FDA00026887881300000411
wherein,
Figure FDA0002688788130000051
wherein,
Figure FDA0002688788130000052
step 6, establishing a trust relationship set of the target user and calculating the trust degree by combining the user attention list information, wherein the trust relationship set comprises the following steps:
acquiring user attention list information, acquiring a trust relationship between the global trust network and the nodes including the attributes of the user nodes, storing the user attributes and the trust relationship information in the global trust network as a tree structure, and acquiring a subtree taking a target user as a root node, namely a target user local trust network; taking a target node as a node A as an example, according to a six-degree segmentation theory, all nodes with the shortest path from the node A being less than 6 are taken as trust nodes (including direct trust and indirect trust) of the node A, and therefore a trust node set of the user A is obtained;
the method for calculating the trust degree mainly comprises two steps:
step 1: setting a target user as an initial node, and setting an initial trust level to be 1 on the assumption that the target user trusts the target user completely;
step 2: distributing the trust degrees for the direct trust user and the indirect trust user of the target user (the starting node), wherein the calculation formula is as follows:
trustcur＝trustpre*wi,j
Figure FDA0002688788130000053
the value of the trust degree of the current node is the product of the trust degree of the previous node and the weight, wherein l represents the number of layers of the user in the network;
and 7, recommending by combining the user similarity and the trust degree, wherein the recommending comprises the following steps:
1) generating a neighbor set, selecting N users with highest trust degree from a trust list of the current user as the neighbors of the current user, and if the number of the users in the trust list of the current user is less than N, calculating the u current user by adopting JS divergenceiWith other users ujSelecting K users with the highest similarity to supplement the K users as the nearest neighbors of the users;
2) and calculating the recommended weight by combining the similarity and the user trust, wherein the calculation formula is as follows:
Figure FDA0002688788130000061
wherein
Figure FDA0002688788130000062
Representing user uiWith user ujThe similarity of the user preferences between them,
Figure FDA0002688788130000063
representing user uiFor user ujThe degree of trust between the two devices,
Figure FDA0002688788130000064
representing a recommendation weight;
3) calculating the prediction score value of the current user to the target scenic spot by adopting a Resnick formula in the collaborative filtering;
Figure FDA0002688788130000065
wherein
Figure FDA0002688788130000066
And
Figure FDA0002688788130000067
respectively represent users uiAnd ukMean score for scored sights, Nei (i) represents user uiA neighbor set of (2); a isk,jRepresenting user uiTo the scenery spot rjThe score of (1).
2. The tourist attraction recommendation method based on user comment data and trust relationship as claimed in claim 1, wherein said user attribute information comprises a user ID and a user name; the user comment scoring information comprises all comments and scoring information of the scenic spot and time information of the comments and the scoring of the user; the scenic spot attribute information comprises a scenic spot name and a scenic spot ID; the friend attention information of the user comprises the ids and the names of all friends attended by the user; the data is acquired by using a web crawler.
CN202010984598.8A 2020-09-18 2020-09-18 Tourist attraction recommendation method based on user comment data and trust relationship Pending CN112507248A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010984598.8A CN112507248A (en) 2020-09-18 2020-09-18 Tourist attraction recommendation method based on user comment data and trust relationship

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010984598.8A CN112507248A (en) 2020-09-18 2020-09-18 Tourist attraction recommendation method based on user comment data and trust relationship

Publications (1)

Publication Number Publication Date
CN112507248A true CN112507248A (en) 2021-03-16

Family

ID=74953472

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010984598.8A Pending CN112507248A (en) 2020-09-18 2020-09-18 Tourist attraction recommendation method based on user comment data and trust relationship

Country Status (1)

Country Link
CN (1) CN112507248A (en)

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112883282A (en) * 2021-03-30 2021-06-01 辽宁工程技术大学 Group recommendation method based on sparrow search optimization clustering
CN113254800A (en) * 2021-06-03 2021-08-13 武汉卓尔数字传媒科技有限公司 Information recommendation method and device, electronic equipment and storage medium
CN113298367A (en) * 2021-05-12 2021-08-24 北京信息科技大学 Theme park perception value evaluation method
CN113407842A (en) * 2021-06-28 2021-09-17 携程旅游信息技术（上海）有限公司 Model training method, method and system for obtaining theme recommendation reason and electronic equipment
CN113792118A (en) * 2021-09-08 2021-12-14 浙江力石科技股份有限公司 Satisfaction improving system and method based on scenic spot evaluation
CN113821727A (en) * 2021-09-24 2021-12-21 武汉卓尔数字传媒科技有限公司 Item recommendation method, computer device, and computer-readable storage medium
CN114048378A (en) * 2021-11-10 2022-02-15 四川泛华航空仪表电器有限公司 BTM model-based personalized recommendation method
CN114429384A (en) * 2021-12-30 2022-05-03 杭州盟码科技有限公司 Intelligent product recommendation method and system based on e-commerce platform
CN114931753A (en) * 2022-03-30 2022-08-23 网易（杭州）网络有限公司 Friend recommendation method and device in game
CN114997723A (en) * 2022-06-30 2022-09-02 辽宁大学 Construction method of multi-source heterogeneous blockchain quality assessment model for enterprise business activities
CN115018584A (en) * 2022-06-13 2022-09-06 浙江理工大学 Recommendation method integrating comment text subject word emotional tendency and user trust relationship
CN115187361A (en) * 2022-06-16 2022-10-14 平安银行股份有限公司 Credit card recommendation method, device, electronic device and storage medium
CN116628317A (en) * 2023-04-19 2023-08-22 上海顺多网络科技有限公司 Method for analyzing user group preference by using small amount of information
CN116662556A (en) * 2023-08-02 2023-08-29 天河超级计算淮海分中心 Text data processing method integrating user attributes
CN117333203A (en) * 2023-12-01 2024-01-02 广东付惠吧数据服务有限公司 An affiliate marketing platform combined with business marketing solutions
CN119537704A (en) * 2025-01-22 2025-02-28 山东科技大学 A time-aware user portrait modeling method based on sentiment analysis
CN120354011A (en) * 2025-06-26 2025-07-22 山东征途信息科技股份有限公司 Behavior analysis and medical service pushing method and system based on big data

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102929931A (en) * 2012-09-24 2013-02-13 上海师范大学 Information credibility assessment method based on body in semantic net
CN105049354A (en) * 2015-08-14 2015-11-11 电子科技大学 Trustful routing method based on OSPF
CN107330461A (en) * 2017-06-27 2017-11-07 安徽师范大学 Collaborative filtering recommending method based on emotion with trust
CN111061962A (en) * 2019-11-25 2020-04-24 上海海事大学 A Recommendation Method Based on User Score Analysis

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102929931A (en) * 2012-09-24 2013-02-13 上海师范大学 Information credibility assessment method based on body in semantic net
CN105049354A (en) * 2015-08-14 2015-11-11 电子科技大学 Trustful routing method based on OSPF
CN107330461A (en) * 2017-06-27 2017-11-07 安徽师范大学 Collaborative filtering recommending method based on emotion with trust
CN111061962A (en) * 2019-11-25 2020-04-24 上海海事大学 A Recommendation Method Based on User Score Analysis

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
王春旭: "基于用户评论的公园推荐算法研究", 《中国优秀硕士学位论文全文数据库》 *
赵文涛: "基于Logistic时间函数和用户特征的协同过滤算法", 《计算机应用与软件》 *

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112883282A (en) * 2021-03-30 2021-06-01 辽宁工程技术大学 Group recommendation method based on sparrow search optimization clustering
CN112883282B (en) * 2021-03-30 2023-12-22 辽宁工程技术大学 A group recommendation method based on sparrow search optimization clustering
CN113298367A (en) * 2021-05-12 2021-08-24 北京信息科技大学 Theme park perception value evaluation method
CN113298367B (en) * 2021-05-12 2023-12-12 北京信息科技大学 Theme park perception value evaluation method
CN113254800A (en) * 2021-06-03 2021-08-13 武汉卓尔数字传媒科技有限公司 Information recommendation method and device, electronic equipment and storage medium
CN113407842A (en) * 2021-06-28 2021-09-17 携程旅游信息技术（上海）有限公司 Model training method, method and system for obtaining theme recommendation reason and electronic equipment
CN113407842B (en) * 2021-06-28 2024-03-22 携程旅游信息技术（上海）有限公司 Model training method, theme recommendation reason acquisition method and system and electronic equipment
CN113792118A (en) * 2021-09-08 2021-12-14 浙江力石科技股份有限公司 Satisfaction improving system and method based on scenic spot evaluation
CN113821727A (en) * 2021-09-24 2021-12-21 武汉卓尔数字传媒科技有限公司 Item recommendation method, computer device, and computer-readable storage medium
CN114048378A (en) * 2021-11-10 2022-02-15 四川泛华航空仪表电器有限公司 BTM model-based personalized recommendation method
CN114048378B (en) * 2021-11-10 2024-05-14 四川泛华航空仪表电器有限公司 Personalized recommendation method based on BTM model
CN114429384B (en) * 2021-12-30 2022-12-09 杭州盟码科技有限公司 Intelligent product recommendation method and system based on e-commerce platform
CN114429384A (en) * 2021-12-30 2022-05-03 杭州盟码科技有限公司 Intelligent product recommendation method and system based on e-commerce platform
CN114931753A (en) * 2022-03-30 2022-08-23 网易（杭州）网络有限公司 Friend recommendation method and device in game
CN115018584A (en) * 2022-06-13 2022-09-06 浙江理工大学 Recommendation method integrating comment text subject word emotional tendency and user trust relationship
CN115187361A (en) * 2022-06-16 2022-10-14 平安银行股份有限公司 Credit card recommendation method, device, electronic device and storage medium
CN114997723A (en) * 2022-06-30 2022-09-02 辽宁大学 Construction method of multi-source heterogeneous blockchain quality assessment model for enterprise business activities
CN116628317A (en) * 2023-04-19 2023-08-22 上海顺多网络科技有限公司 Method for analyzing user group preference by using small amount of information
CN116662556A (en) * 2023-08-02 2023-08-29 天河超级计算淮海分中心 Text data processing method integrating user attributes
CN116662556B (en) * 2023-08-02 2023-10-20 天河超级计算淮海分中心 Text data processing method integrating user attributes
CN117333203A (en) * 2023-12-01 2024-01-02 广东付惠吧数据服务有限公司 An affiliate marketing platform combined with business marketing solutions
CN117333203B (en) * 2023-12-01 2024-04-16 广东付惠吧数据服务有限公司 Member marketing platform combined with business marketing solution
CN119537704A (en) * 2025-01-22 2025-02-28 山东科技大学 A time-aware user portrait modeling method based on sentiment analysis
CN120354011A (en) * 2025-06-26 2025-07-22 山东征途信息科技股份有限公司 Behavior analysis and medical service pushing method and system based on big data

Similar Documents

Publication Publication Date Title
CN112507248A (en) Tourist attraction recommendation method based on user comment data and trust relationship
Dhelim et al. Personality-aware product recommendation system based on user interests mining and metapath discovery
Sang et al. Context-dependent propagating-based video recommendation in multimodal heterogeneous information networks
CN108256093B (en) A Collaborative Filtering Recommendation Algorithm Based on User's Multi-interest and Interest Change
CN110377840B (en) Music list recommendation method and system based on long-term and short-term preference of user
Tsur et al. What's in a hashtag? Content based prediction of the spread of ideas in microblogging communities
CN114201682B (en) Graph neural network recommendation method and system fusing social relations and semantic relations
Lin et al. Heterogeneous knowledge-based attentive neural networks for short-term music recommendations
CN111061962B (en) A recommendation method based on user rating analysis
Zhang et al. Joint Personalized Markov Chains with social network embedding for cold-start recommendation
CN112507246B (en) Social recommendation method fusing global and local social interest influence
CN112069290B (en) A Recommendation Method for Academic Papers Based on Graph Local Structure and Text Semantic Similarity
CN110275964A (en) Recommendation Model Based on Knowledge Graph and Recurrent Neural Network
CN109933664A (en) A kind of fine granularity mood analysis improved method based on emotion word insertion
CN112861541A (en) Commodity comment sentiment analysis method based on multi-feature fusion
CN109598586B (en) Recommendation method based on attention model
Sharma et al. A multi-criteria review-based hotel recommendation system
Zhou et al. Domain-constrained advertising keyword generation
CN110069713A (en) A kind of personalized recommendation method based on user's context perception
CN107818183A (en) A kind of Party building video pushing method based on three stage combination recommended technologies
CN112464108A (en) Resource recommendation method for crowdsourcing knowledge sharing community
CN114153965B (en) A method, system, and terminal for recommending public opinion events that combines content and graphs.
CN111061958A (en) Information recommendation method and system based on user viewpoint and emotional tendency
Mariani et al. NLP4NLP+ 5: The deep (r) evolution in speech and language processing
CN113392319A (en) Academic paper recommendation method based on network representation and auxiliary information embedding

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
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20210316