2
$\begingroup$

I was not sure about posting this question with mentioning the name of the company, which I quite respect and admire. However, I've figured that a wider exposure might help the team to fix this and similar problems faster as well as increase the quality of the machine learning (ML) engine of their website.

The problem exposes itself by too many occurrences of a quite trivial misclassification error on Amazon's book categories classification (which I'm a frequent visitor of). In the following example, the underlying reason of such behavior is quite clear, but in other cases the reasons might be different. I am curious about what could be other potential reasons for misclassification and what are the strategies/approaches to avoiding such problems. Without much further ado, here's how the problem appears in real life.

I was reviewing some books, related to transitioning from graduate programs (Ph.D., in particular) to work environment in academia. Among several other books, I ran across the following one:

enter image description here

So far, so good. However, let's scroll down a bit further to see the the books ratings in relevant categories. We should expect Amazon to figure out categories, relevant to the book's discipline, topic and contents. How surprised was I (and that's an understatement!) to see the following result of Amazon.com's sophisticated ML engine and algorithms:

enter image description here

Clearly, the only fuzzy fact that connects this book with the subject "Audiology and Speech Pathology" (!) is IMHO the author's last name (Boice), which, is close to the word "voice". If my guess is correct, Amazon's ML engine, for some reason, decided to take into account the book's lexicographical attribute instead of the book's most important and most relevant attributes, such as title, topic and contents. I've seen multiple occurrences of similar absolutely incorrect ML-based decision making on Amazon.com and some other websites. So, hopefully my question makes sense as well as interesting and important enough to spark a discussion: What could be other potential reasons for misclassification and what are the strategies/approaches to avoiding such problems? (Any related thoughts will also be appreciated.)

$\endgroup$
2
  • $\begingroup$ What's the category of the "Advice for new faculty members"? $\endgroup$ – LauriK Feb 11 '15 at 8:15
  • $\begingroup$ @LauriK: Definitely not the one that has been selected. $\endgroup$ – Aleksandr Blekh Feb 11 '15 at 8:22
2
$\begingroup$

By looking at the ratings of the book, it looks like they are doing some sort of hierarchical clustering. First of all it's a book, then a text book, then in medicine & health science.

This is a pretty difficult problem because:

  1. You don't know how many and which topics you have, and the topics are constantly changing.
  2. Since this is a clustering problem (or a semi-supervised problem in the better case), a closed feedback loop for these errors cannot be applied. Clustering is always more difficult than a supervised learning problem.

Apparently they're doing a pretty good job in general. This book in particular is a tough one...

$\endgroup$
2
  • 1
    $\begingroup$ Thank you for the answer (+1). I don't think that the hierarchical structure of the classification path necessarily implies the use of hierarchical clustering (HC) approach. They might be using some fancy ontology-based or other technique, different from HC. I'm not sure that they are doing a good job - as I said, I've seen multiple similar errors in the books section, which is one of Amazon's major e-commerce channels. $\endgroup$ – Aleksandr Blekh Feb 11 '15 at 13:34
  • 1
    $\begingroup$ @AleksandrBlekh Yep, it's just a thought. $\endgroup$ – Omri374 Feb 11 '15 at 14:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.