What is the value of an online review?

What is the value of an online review?

Business

In recent years, as the network technology has been developing rapidly, online shopping is  increasingly becoming part of our lives. Now the e-commerce industry is gaining momentum.  For online retailers, they are not limited by inventory or space, while physical stores must store  products in a limited space. For customers, they can rate or comment on the goods they bought.  

 However, online shopping also has its limitations. With regard to the e-com merce, when  they need to launch new products, they have to consider the purchase information of other  similar competitive products. Therefore, it makes great sense to understand market needs and  consumer behaviors through customer rating and comments.  

Literature Review  

Many economists have confirmed through statistical experiments that customers will refer  to evaluation and star rating when they buy goods online. E-commerce has begun to analyze  customer comments through some models, such as Latent Dirichlet Allocation Model (LDA),  so as to develop a market for their products. We expected to find some convenient and intuitive  models for a specific company’s selling plan based on the previous wisdom and some necessary  math knowledge like LDA and Statistics.  

The Task at Hand  

● Analyze the data sets of three products by mathematical method to help the company  succeed in their three new online marketplace product offerings.  

● Address some specific questions and requests from the Sunshine Company Marketing  Director through our analysis.  

● Present our analysis by letter.  

The Basic Models  

Related Concept  

Latent Dirichlet Allocation Model (LDA)  

LDA is a kind of unsupervised machine learning technology, which can be used to identify  the hidden subject information in large-scale document collection or corpus. It adopts the method  of bag of words, which regards each document as a word frequency vector, thus transforming  the text information into digital information which is easy to model.

However, the word bag  method does not consider the order between words, which simplifies the complexity of the  problem and provides an opportunity for the improvement of the model. Each document  represents a probability distribution of some topics, and each topic represents a probability  distribution of many words. Because of the weak correlation among the components of the  random vector of Dirichlet distribution (the reason why there is a little “correlation” is that the  sum of the components must be 1), the potential topics we hypothesized are almost irrelevant, which is not consistent with many practical problems, resulting in another legacy problem of  LDA.  

For each document in the corpus, LDA defines the following general process  ● For each document, extract a topic from the topic distribution;  

LDA Model’s Generation Process  

We can use the following graph model to express:  

Figure 1 LDA Model Generation

This representation of the graph model is also called “plate notation”. The shadow circle in the  figure represents the observed variable, the non-shadow circle represents the potential variable,  the arrow represents the conditional dependency between the two variables, the box represents  the repeated sampling, and the number of repetitions is in the lower right corner of the box.  

The LDA model has now become a standard in topic modeling. As mentioned before, the LDA model  has been expanding rapidly since its birth, especially in the field of social network and social  media research.  

The General Characteristic of Some Data  

The Labels Needed and Their Meanings  

For the analysis of data, we think that four parts should be selected for the data label: “product  title”, “review date”, “review body” and “star rating”. Its purposes are as follows:  “star rating”: screening out high scores  

“product title”: helps to better determine which products have high scores “review date”:  record the comment time, which is more conducive to analyzing the change of the number of  buyers after the comment  

“review body”: a key word that attracts people’s attention by analyzing comment statements and  processing them with natural language  

The Characteristic of Data on Display  

According to the information in these labels, we can summarize the top ten products with  the most comments, as shown in the following figure:（Take hair dryer as an example）： 

Figure 2 Ranking of number of reviews

Next, we calculate the proportion of top 10 and top 5 product reviews to the total number of  reviews, as shown in the following figure:  

Figure 3 TOP10 and TOP5 proportion of product reviews It can be seen that the proportion of TOP10 in the total data is very high. Next, we  analyze the change of product quantity over time through a “review date”.  

Then we can get a comment on the trend of time change: 

Figure 4 Number of monthly reviews over time 

 Figure 5 TOP5 product monthly review counts change over time

In this way, we can understand the products that are more popular with customers, and then  filter the effective information for the comments of these products.  

Finally, we can know the top five popular hair dryer products： 

Figure 6 TOP5 Average product rating of hair dryer 

Through the above methods, we can also analyze pacifiers and microwave ovens, and get the  following results:  

Figure 7 TOP5 Average product rating of microwave 

Figure 8 TOP5 Average product rating of pacifier

At the same time of getting top 5, we can also get the worst five kinds of goods through the  above methods.  

The Relationship Between Stars, Reviews and Help Ratings 

We have screened out the good and bad products of the same kind.  

Next, we need to use LDA model and QBRs (Quarterly Business Reviews) to select some key words, and use LDA to screen the key words of low evaluation goods and high evaluation goods respectively, so as to get which aspects consumers pay more  attention to (negative words in low evaluation key words, positive words in high evaluation goods)  and which aspects consumers don’t pay attentionto (positive words in low evaluation key words,  negative words in high evaluation goods).  

Obtain Data Information by LDA Model  

We take the hair dryer as an example. First, we need to delete punctuation, stop words, etc. After  the form of comments is unified as much as possible, the most common words in the data need to  be processed. As shown in the figure, the most common words are “the”, “and”, “to”, etc. these  words are not helpful to us, so we should also delete these words.  

Figure 9 The high frequency word list after deleting the stop words and punctuation

We can see that there have been changes, and the words related to hair dryer have appeared.  However, there are still some neutral words, such as “the” and so on. In order to further remove  the noise in the text, we can use the word shape reduction processing in the spice library. It can  restore words to their original form and reduce the repetition of words. It can be seen that this  process not only restores the morphology, but also leaves only nouns and adjectives. After that, we  change the word into normal form. In Figure 7, you can see that the common words are all related  

to hair dryers, then we can continue to build LDA models, through which we can see what aspects  of the product consumers are most concerned about.  

Then, we get the following results through LDA model:  

Figure 11 High frequency words list from LDA model

Figure 10 Product review common words list 

In fact, the optimization here is not the best, because the words obtained are all discrete units,  and they must be combined into a sentence to be more conducive to business analysis.  

Influence of Star Rating and Number of Monthly Comments on  Customers  

Firstly, we want to have a general command of the relevance between the rating scores and the  number of reviews in a month, so we use the sum of the comments in the previous month and the sum  of the current month reviews to draw the picture.  

Figure 12 Relationship between number of monthly reviews and previous month’s rate

There has been some relevance to some degree, but it is not quite obvious. In order to extract  information about product evaluation from the user’s comments, LDA analyzes the evaluation  content of each grade in advance, then uses NLP (Natural Language Processing) to extract the  emotionally relevant words in the comments, and further quantifies the comments. We will use  this quantified value as a measure of user satisfaction.  

Key Words in Different Star Ratings  

The comments are classified by evaluation star rating, the evaluation content of 1-2 stars and  4-5 stars are used for keyword filtering using LDA to observe which aspects of the commodity  consumers are more concerned about. We believe that negative reviews (1-2 stars) in the positive  words and positive comments (4-5 stars) in negative words can reflect what aspect about the goods  the customers pay little attention to, while the negative words in negative evaluation and positive  comments in the positive words can reflect the consumer’s concern about the commodity aspects.  Microwave:  

Table 1: keywords in microwave  

We can see that for microwave ovens, consumers are more concerned about the ease of use of  microwave ovens, the occupation of space and the appearance of these three aspects, and less attention  to operability, power.  

Hair dryer:  

Table 2: keywords in hair dryer  

We can see that for hair dryers, consumers pay more attention to the size, weight, price, voltage  and so on, less concerned about dryness, use time and so on.  

Pacifier  

Table 3:key words in pacifier  

We can see that for pacifiers, consumers pay more attention to the quality of nipples, easy to  clean, comfortable, material, price and so on, less attention to cuteness, color, use and so on.  

Simple Bayesian Method for Emotional Analysis  Through 2.3, we can know that there is a correlation between rate star and sales volume, but  it is very weak, so it is necessary to redesign a quantitative method to enhance this correlation. At  the same time, as the correlation becomes stronger, sales volume can also be predicted.  Therefore, we try to use another quantitative method. This redesigned quantitative method is  based on Naive Bayes, mapping the evaluation content into a number. 

The first goal we want is to determine whether the comment is positive or negative by  predicting the positive and negative probabilities of feelings in comments by using a word in  comments. When the simple Bayesian method is emotionally classified words: the input word $x$,  the posterior probability distribution is calculated by the learned model, and the class with the  highest post-mortem probability is used as the class output of $x$. 

After the training of our Bayesian classifier, we can get a positively rated word cloud and a  negative evaluation of the word cloud based on the accuracy of the classification.  

Figure 13 positively rated word cloud 

Figure 14 negatively rated word cloud

Based on these words with significant classification effects, we can quantify the emotional tone  of the comment. When we quantify a comment, for each word in the comment, we calculate the  probability whether the comment is positive or negative when existing in the comment. Then we  give a “user satisfaction” score to comments based on the following formula.  

We propose that this quantitative formula is based primarily on the following considerations  and assumptions:  

While not every word in the comments is quite helpful in the emotional judgment of the  commentary, the posterior probability calculated from these words is generally close to 0.5, smaller  than the posterior probability calculated based on the word with a significant effect (generally close  to 0.9 or 0.1). Generally, these insignificant words as a whole do not have a significant impact on  the value quantification.  

Apply a “reward” for an emotionally positive word, and a “punishment” for an emotionally  negative word. This is why we add this probability to the overall score when judged as a positive  posterior probability which is greater than 0.5. Similarly, we should subtract the corresponding  judgment as negative posterior probability when the probability is less than 0.5.  

To make this score gentler overall, we use the logarithmic function to reduce the effect of  extreme values on the whole, and we can also reduce the variance of the comment in general,  because it is difficult to quantify the text emotion accurately. In order to achieve this, we sum the  posterior probability of the word in comment and then add a constant C, so that all sums are greater  than 0. And then subtract the positive of the constant to make this score fluctuate up and down at  $x=0$. If the score is greater than 0, representing the comment tends to be positive, if the score is  less than 0, representing the comment tends to be negative.  

Then we use the comment score from the Naïve Bayes method and the comments in that month  to draw the picture: 

Figure 15 Relationship between number of monthly reviews and previous month’s rate 

The conclusion is that it has obvious relevance, so it is more likely that the customers can  be directed by content in comments, while the rating scores have less influence on customers.  

Conclusions  

Through LDA model experiment and simple Bayes method, we find that customers care more  about other customers’ comments on the product than the star level of the product, and customers  are more easily controlled by the content of the comments. Customers mainly pay attention to  positive evaluation and negative evaluation, while the focus on neutral evaluation is relatively  small. In these two types of evaluation, positive words in negative evaluation and negative words  in positive evaluation can reflect the aspects consumers don’t care about, while negative words in  negative evaluation and positive words in positive evaluation can reflect the aspects consumers  pay attention to.  

Strength  

● This model selects the representative products, which can effectively reduce the screening  numbers.  

● This model can effectively screen out customers’ concerns so that the company can  improve its products.  

● Some invalid comments or star ratings are excluded in this model.  

Weakness  

● In this model, some customers may make false comments for external reasons, which are  difficult to identify.  ● In this model, we get some key words, but it does not mean these keywords are the overall  solution, and we need to draw and guess some further sentences that affect the customer’s  decisions.

What is the value of an online review?