Amazon customer service: Big data analytics

1. Introduction

As per Statistica, the online e-commerce industry saw a boost of 26% in purchase of essentials due to the covid imposition. With this boost, e-commerce giants such as Amazon, have had to strengthen their business model to cater their customers. In this project, we focus on the last step of a customer’s purchase journey. This implies the post sales customer service and review analysis.

We have broken down our business problem into two parts for a holistic perspective:

1.

Many a times, the numeric ratings of a product are not indicative of what the words say about the product. Since, the ratings serve as a gateway to help customers get a better idea of the product and its usage, the wrong ratings could end up hurting the sales of the product and ruin the customer experience. Thus, we provide a Sentiment Analysis on the text of the review and compare it with the original rating provided by the customers. There are two possible outcomes to the results we get,

a.

An increase in the predicted ratings than the original ratings may indicate an opportunity loss for Amazon, as the product is being undersold.

b.

A decrease in the predicted ratings than the original ratings could have a negative impact on the customers with a product that does not live up to its expectations, thus negatively impacting Amazon’s customer experience.

2.

In addition to this, people have varying personalities. This implies that even though two customers may have the same experience with a product, their numeric ratings may differ due different personality traits. Thus, normalizing the ratings is significant to our analysis in order to standardize the reviews to completely capture the sentiment around the product.

To get a more realistic and real-time result, we extract live data for Amazon customer service in the form of tweets by using the Google Cloud Platform (GCP). We have also aggregated datasets in a smaller size that capture the reviews and numeric ratings of customers on the Amazon website along with product details. As a part of our analysis, we developed a pipeline on the Google Cloud Platform brought an API connection to take the analysis to python for further text analysis on the reviews. While the live dataset was being populated in the Google BigQuery table, we worked on the Amazon’s review dataset to capture the true essence of the reviews into numeric ratings. This is based on our analysis wherein we observed a discrepancy between people talking about the product negatively but end up giving a higher rating, or vice versa. In an attempt to solve the discrepancy and identify customer pain points in various categories, we developed three prediction models which captures sentiment of the user.

Using those models, we are able to predict ratings which incorporate sentiment of the user with the user’s actual rating to provide an unbiased Rating. Every customer is unique and each one can rate the same experience differently on the basis of their personality traits. These different customers were then modeled into 3 separate categories:

•

Stringent Reviewers

•

Generous Reviewers

•

Standard Reviewers

These clusters can be used as a distinction factor while measuring different KPIs for Amazon. The end purpose would be to improve the targeted marketing for consumers.

2. Literature review

As per Wlodarczak et al. (2014) the use of opinion mining as a technique of machine learning is a method of utilizing the vast global opinions on Twitter to be able to achieve more targeted business decisions. They perform four phases as a part of their research which include data collection, data pre-processing, data mining and post-processing. They make use of an ensemble model which includes bagging, boosting and stacking. The challenge involved is application of these models requires high experience and research. A singular, simpler model which does not reduce prediction performance is more preferred. With respect to Amazon, the authors make use of web scraping to perform their POS tagging feature extraction analysis after data processing. The main drawback of this study is that it is performed on only one product and cannot be generalized to a broad variety of products. Jain, Vineet et al. (2020) propose a methodology to perform sentiment analysis on the product reviews of Amazon to capture the overall perspective of customers with respect to a product. They use the Cross Industry Standard Process for Data Mining as their methodology to conclude principal component analysis is the next step to develop a generalized model. The ongoing narrative suggested the research is in its preliminary stages. Furthermore, Ihsan Zul et al. (2018) made use of K-Means and Naïve Bayes algorithms as their fundamental models to perform sentiment analysis on social media. This study has mainly focused on comparing a variety of different models with accuracy as their metric to observe which algorithm performs the best after cleaning and processing the data from Facebook from Excel’s Power Query and Twitter API.

3. Research methodology

We performed a Sentiment Analysis using Sentiment Intensity Analyzer and computed the compounded scores for each review and ascertain each review as Positive and Negative Reviews. From Fig. 1, we observed a strong correlation (0.385) between the compounded scores and actual user ratings. Our objective, now, was to get the best curve/function according to the compound scores and incorporate the same to the actual user rating in order to capture the True Rating.

OPEN IN VIEWER

OPEN IN VIEWER

OPEN IN VIEWER

•

Rolling method

In this method compound scores which are scored from - 1 to 1 are being scaled between 1–5 Rating. Figure 2, Fig. 3 represent the functions graphically.

Rolling Backwards – Compound Scores

def Back (x):

return round (((x + 1)*5/2))

def standard_Back (x):

if x 1:

return 1

else:

return x

Rolling Front – Compound Scores

def Front (x):

return round (((x + 1)*5/2))+1

def standard_Front (x):

if x $>$ 5:

return 5

else:

return x

In this method, we remove the discrepancies and sort the compound scores to get a generalized graph of the population.

OPEN IN VIEWER

Figure 4, shows the sorted average rating and respective Sentiment Scores, we observe certain anomalies: Reviews with 5 rating have negative sentiment compound scores and Reviews with 1–2 Ratings have Positive Sentiment.

OPEN IN VIEWER

Moreover, we notice that the compound scores are more populated around the given ratings, but we observe anomalies which encircled in black (Fig. 5).

From Fig. 5, the main observation we can obtain is that the cluster of people who rate generously, have a neutral review description, give full rating to the product. In contradiction, we also observe the people who are the opposite to the abovementioned category. As people are unique and their reviews are subjective, they tend to rate the same product/service differently. Thus, the ratings need to be normalized using the sentiment of their reviews.

Choice of the best models from the proposed solutions is a difficult task. The graph generated by the compound score which shows the maximum ratings in correspondence to the actual user ratings deems appropriate. After performing a correlation analysis for the above methods, we can conclude that the Curve fitting Model has the highest correlation with the user ratings. Correlation of the Sentiment Score C (Scores of Curve Fitting Model) were better than the correlation of actual user ratings with the compound scores generated by the default Sentiment Analyzer library. Figure 6 summarizes the obtained scores for every model.

OPEN IN VIEWER

OPEN IN VIEWER

OPEN IN VIEWER

Let us plot the compound score against the ratings provided by the user. From Fig. 7, we can see 3 types of behaviors generally shown by the people. We got trend lines to understand the different kinds of cluster behaviors. Figure 8 gives us a clear understanding on how the scores can be used to classify the customer into the three different categories.

1)

Likely Behavior –

a.

If the description has a low compound score (Negative), the user is less likely to give higher rating (Identified from the Blue Cluster).

b.

If the description has a high compound score, then the user is more likely to give higher rating.

2)

Generous Reviewers – Customers with a low compound score and relatively higher user rating.

3)

Critical Reviewers – Customers with a high compound score and relatively low user rating.

6. Summary and conclusions

•

The sentiments of the text reviews are used to convert them into more accurate numeric ratings using the developed model. This can be used for the purpose of targeted and more direct consumer marketing.