Customer Insights from Amazon Alexa Reviews

Back Story

Amazon Alexa is a digital assistant that takes and executes a users’ voice commands. I received one of those as a Christmas gift. I really enjoyed it and quickly got used to Alexa style. Now this becomes the way how I typically start my morning: ‘Hi Alexa, can you play some music?’

Amazon Alex is a smart home device that similar to Google Home. Alexa has a great user rating, but Google Home still manages to beat it. Although Alexa is leading the marketing share currently, research firm projected that Google Home will be able to catch up by 2022 (see reference at the end).

With the curiosity of users’ experience about Amazon Alexa, I decided to look into users reviews. My main objective is for learning and exploration purposes. It would be great if any finding can be helpful to understand customers’ concerns and pain points. This may potentially be used to enhance customer satisfaction, eventually leading to better product experience overall.


Data Collection and Modeling

I collected about ~10K reviews for Amazon Alexa. Based on the rating of reviews, I broke down reviews into three sentiment groups: positive (3.0 + rating), neural (3.0 rating), and negative sentimental groups (<3.0 rating).

I would like to extract insights out of positive and negative ratings, and performed cluster analysis as first step. To begin with, I vectorized the reviews and ended up with ~2000 features or dimensions. One of common challenges of Natural Language Processing (NLP) problem is the huge size of dimensionality. Curse of dimensionality mentions that high dimension can result in high sparsity in data set and unnecessarily increase of storage space and processing time. Reducing dimensionality to a manageable number is critical, and SVD/NMF algorithms were used here to achieve the purpose.

Subsequently, I performed KMeans to cluster reviews. For both positive and negative reviews, NMF methodology with 100 remaining features produce best result. However, only limited information was provided by identified clusters. To better understand common topics in customers’ reviews,  I decided to further perform LDA topic modeling. As a result, I was able to summarize three common topics out of positive and negative reviews, respectively (see below).



Positive Highlights:

Topic 1: Simplicity

Topic 2: Smart

Topic 3: Best Gift Option


Improvement Areas:

Topic 1: Connection Issue

Topic 2: Speaker Issue

Topic 3: Customer Service



With the finding, a few strategic actions can be taken:

  1. Connection Issue. Alexa needs Internet connection to stay smart. Connection is essential for Alexa to function properly. Any hardware issue should be investigated and addressed. At the same time, video manual can be provided online to guide new users on connection setup.
  2. Speaker Issue. Many customers complain about volume and quality issue of speaker. Investigation is needed to fix and enhance speaker quality
  3. Customer Service. Customer service is the last line of defense to address users’ questions and concerns. Apparently there is zoom of improvement for customer service. More training can be provided to enhance customer experience.
  4. Best Gift Option.The insight can also be used to enhance the effectiveness of marketing strategy. For example, with Amazon Alexa listed as one of best gift options, Amazon can align next major release or promotion event with gift seasons.

Recommended Next Steps:

  1. Due to limitation of time, cluster analysis and topic modeling were performed separately in this project. They may be combined to potentially achieve better result
  2. Overall, the model can be generalized and applied to any product or service reviews. This will help companies know better about customers’ concerns and pain points. The end goal to enhance self-awareness, drive up product/service quality, and increase market competitiveness.



Leave A Reply

Your email address will not be published. Required fields are marked *

five × 1 =

Interested in learning more?