“Machine learning is ubiquitous at Amazon today,” said Rajeev Rastogi, Vice President, Machine Learning at Amazon India, in an interview with Gadgets 360. “Within the retail business, we are using machine learning extensively to recommend products to customers, forecast future demand for products, and improve the quality of a product catalogue, both classifying products, and also eliminating duplicate products.”
One of the most basic examples of how Amazon is using machine learning (ML) is when you misspell a query on its search bar. The e-commerce site, Rastogi noted, looks at the phonetic distance between the typed misspelt query and the correct query instead of looking at their textual distance to provide accurate results — no matter whether you have spelt something incorrect.
For instance, if you type “geezer” on Amazon to look for the available geyser options, the marketplace will autocorrect the spellings and show you relevant results. Amazon is also using ML models to translate the content on its site to the Indian languages it now supports.
Of course, these kinds of uses of computers are now commonplace, and not something that most of us think about when we consider the phrases artificial intelligence (AI), or machine learning.
Rastogi revealed that his team is currently working on a seed initiative that is aimed to bring a conversational shopping experience. It is aimed at first-time online shoppers who are more familiar in communicating with offline shopkeepers over placing an order through an e-commerce site.
Conversational commerce, through chatbots, through smart assistants like Amazon’s own Alexa, is one of those ideas that keeps coming back every few years as the technology improves, and Rastogi talks about how it will start with text, in English, but grow to other languages, and to voice.
“A machine can read a document and then answer any question about the document, it is difficult. Today AI cannot generate a review for a movie, for example… Even summarising a set of documents is a challenging problem. It’s not solved by AI by any means,” underlined Rastogi.
AI has been used for analysing text and speech at various levels. But computer engineers and data scientists have not yet been able to find a relevant mix for using AI and machine learning to generate accurate assessments such as movie or product reviews. In a research article, published by researchers Gerit Wagner, Roman Lukyanenko, and Guy Paré of the Department of Information Technologies, HEC Montréal, on how AI can be used in the literature review review process, it is noted that even “technically perfect tools (like researchers)” sometimes struggle to evaluate information from sources which use ambiguous, confusing language, and presentation.
McKinsey Global Institute (MGI) partners Michael Chui, James Manyika, and Mehdi Miremadi also pointed out in an article that AI models have “difficulty carrying their experiences from one set of circumstances to another” and require companies to train models even when the use cases are very similar. This adds additional resource requirements.
Shreyas Sekar, an Assistant Professor of Operations Management at the Department of Management, University of Toronto Scarborough and Rotman School of Management, said that effectiveness of an AI-based bot communicating with humans and giving them appropriate results especially in markets including India is not certain. Sekar has done extensive research on how e-commerce platforms are using machine learning at both consumer end and warehouses to enhance their operations.
“When you ask these chatbots, simple questions, like no, is it going to rain tomorrow? Or can you play me the song from this movie? They do a great job. But as you start giving more and more complex questions, like hey, can you help me find a good shoe for my trek? I think it’s very hard for the chatbot or even Alexa to clearly break this question down into what is your intent? What do you want as a person, and how do you differ from other people? And what products match for you?” he said.
Dealing with biases and errors
One of the biggest challenges of using AI and ML nowadays is to limit biases and errors. Companies from Google and Facebook to Microsoft are dealing with these blunders on a regular basis. Amazon is also not foolproof at that front.
Sekar of University of Toronto Scarborough and Rotman School of Management noted that Amazon’s AI deployments include a lot of biases that the company is already aware of and is apparently working towards resolving them, but not clear how successfully it has accomplished desired results.
“For example, maybe historically, users have clicked on one particular brand of earphones, then what happens is that in the future, I keep amplifying that exact brand over and over again. So, this is usually called some sort of popularity bias where I try to spotlight products that are already popular, and I’m basically helping the rich get richer in the system,” he mentioned.
Rastogi staunchly disagreed, though, and said that Amazon’s goal is to assist the human workers, not replace them entirely.
Who does this help?
The use of AI and ML helps Amazon offer what you need by understanding your buying behaviour and purchase history. This, however, sometimes leads to impulse buying and simply convinces you to purchase something that you don’t actually require. Experts believe that it would grow further with a more conversational shopping experience.
“I think AI and ML can definitely increase conversion of window shoppers to regular shoppers,” said Sekar. “And this is definitely a good way to think of Amazon as a very persuasive salesperson.”
Consumers can themselves overcome this behaviour by understanding how algorithms can influence their choices.
“Even though we are the ones who go and click on a product to purchase at the end, we are kind of guided along the shopping funnel by the algorithm in different places whether it is the recommendation, or the reviews,” Sekar said.
Ankur Bisen, Senior Partner and Head of Consumer, Food, and Retail divisions at management consulting firm Technopak, said the nature of how Amazon uses its algorithms to entice consumers to buy more was exactly similar to what advertisements, marketing, and even discounts at a retail shop did.