zq论文51

Summary

With the popularization of Internet technology, online shopping has gradually become an

important way for people to consume. In Amazon's online marketplace, customers can rate and

review after purchase, and these ratings and reviews help other customers make purchasing

decisions.

First, we screened the indicators in the three data sets provided to obtain seven important

indicators related to product review analysis. For the review indicators, we used a classifier

based on Naive Bayesian theory to perform sentiment analysis on the review indicators. In order

to achieve the purpose of quantifying indicators and considering the importance of review

three effective indicators that reflect product quality: enthusiasm for comment, credibility of

comment, and product popularity. Combined with time, a comprehensive evaluation level of the

product is obtained. Using the SRR model, we can score a product.

Then, in order to reflect whether the rating of star ratings can lead to more reviews, we

conducted an analysis of variance on the rating star index and the comment density index, and

concluded that there is a significant relationship between the rating star rating and the comment

density. Different star ratings have an impact on the number of reviews. Considering that

triggering is a long process, we analyze the correlation between the amount of individual reviews

and the total of reviews next month. It turns out that 1-star reviews can trigger more reviews,

while 5-star reviews can't.

Keywords: sentiment analysis, analytic hierarchy process, analysis of variance, TF-IDF

1. Introduction

1.1 问题背景

阳光公司计划在网络上推出和销售三种产品：微波炉、婴儿奶嘴和吹风机。网络市场

是一个竞争力极大的地方，且网络市场信息繁杂，如何从中提取出有用的信息且制定有

效的模式处理信息，是公司改进产品、制定发展计划的重中之重。

亚马逊是一个有名的网络市场，顾客可以在产品销售网店内对产品打星，最低为 1，

最高为 5，他们也可以在评论中发表他们对该产品的看法。

评论是以文字形式储存信息的，如何让计算机从文字中提取出信息并量化成可处理的

数据是这篇论文要解决的问题，除此之外，我们还要从所给的数据中提取出各项指标，

并建立相关模型帮助阳光公司在竞争中占据有利位置。

结合时间，分配权重，并得出所有评论中顾客的反馈，进而得出产品的质量。在第（2）节

中，我们陈述了 SRR 模型的基本假设。在第（4）节中，我们给出了模型中使用的每个指标

的详细解释和计算。第（6）节提供了对 SRR 模型的全面分析。

为了明确顾客对产品的评星等级能否引发其他客户的评价，我们对单个产品额外计算了

月评论总量，并进行单因素相关性分析。为了明确某些特定的单词与特定的星级指标是否有

关联，我们提取出 12 个频率较高的单词，并与时间进行相关性分析。

2.假设

1、一个产品的评论数越多，它的关注度越大

一个产品的评论数往往与它的销量正相关，评论越多，代表越多人关注过它，同时在网络市

场，销量多的产品往往被置于商品列表的前列，更容易获得人们关注。

4、最有价值的单词应该是在特定文档中出现频率最高的单词，同时在所有文档中出现频率

最低的单词。

3 数据初步分析

3.1 数据预处理

本文采用多元统计处理方法，由于数据来源于真实生活，存在一定误差，且数据间量纲不同，

需要对数据进行预处理。

（1）无关变量剔除

表格中一共给给出 15 个相关指标，观察数据不难发现。所有数据来源于美国；product_id

与评论涉及的产品唯一对应，不同 product_id 可能分属同一个 product_parent，产品分类时

仅使用 product_parent 指标就足够；product_title 是产品描述；product_category 为消费者类

型；这些指标对于基于星级和评论的分析用处不大，因而在实际数据使用过程中，我们将

考虑到本文的个别指标数据变化范围非常大，直接处理计算量大，且不好分析指标和因素间

的相关关系。为了方便起见，本文将所有数据进行相对化处理，即

以后无特殊说明均省略‘*’符号，用 x

表示相对化后的数据。相对化后的数据并不影响指

标间的相关关系，但处理起来更加方便和直观。

3.2 Quantitative processing of comments

3.2.1. Establishment of comment text indicators

When a product is purchased, reviews from other customers can often play an effective guiding

role. Considering that reviews are unstructured data, it is difficult to directly apply them to our

model. In order to effectively extract the information that a comment can contain, we split the

text information of the comment into length and emotion. The length can reflect the reliability of

a review. The longer the length, the more authentic the emotional expression of customers in

general. Emotions indicate the probability of this review being a positive review, and represent

the degree of customer dislike of the product.