Being an early adopter of artificial intelligence and automation, Amazon always had an edge in using AI to improve its business efficiencies. Not only has it been using AI to enhance its customer experience but has been heavily focused internally. 

From using AI to predict the number of customers willing to buy a new product to running a cashier-less grocery store, Amazon’s AI capabilities are designed to provide customised recommendations to its customers. According to a report, Amazon’s recommendation engine is driving 35% of its total sales.

One of the main areas where Amazon is applying continuous AI is to better understand their customer search queries and what is the reason they are looking for a particular product. For an e-commerce company to make relevant recommendations to its customers, it is not only crucial for them to know what their customers searched for, but it is also critical to understand why a customer is searching for a product. Understanding the context can help the retailer to recommend complementary items to its customers, and Amazon is intent to work out this puzzle by applying AI to the problem.

In a recently posted blog, Amazon discussed using AI and machine learning to predict the context from their customers’ search queries. This system has been aimed to augment the quality of search results on Amazon.com platform, which indeed intended towards enhancing the overall Amazon’s shopping experience. 

Explaining further, in a paper accepted to the ACM SIGIR Conference on Human Information Interaction and Retrieval, Amazon researchers described how most retailers use product discovery algorithms to look for correlations between queries and products; however, Amazon used their AI to identify the best matches depending on the context of use. Therefore the system predicts activities like “running” from customer queries like “Adidas men’s pants” or if an Amazon customer enters the query “waterproof shoes”, is she looking to go for a weeklong hike?

According to Amazon, predicting the intent of the query is a significant component of information retrieval which in turn, improves the relevance of the results through an understanding of latent user intents in addition to explicit query keywords. The researchers believe this might improve people’s shopping experience by matching only high-quality products to search queries.

Training The System

The first step of the process was to train the system for which the team had to build a data set. In order to build the data set, the team assembled a list of 173 context-of-use categories divided into 112 activities — such as reading, cleaning, and running — and 61 audiences — like a child, daughter, man, and professional — based on common product queries. They used standard reference texts to create aliases for the terms they used to denote the categories. Such as for the category ‘father’ they included ‘dad’, ‘daddy’, ‘pops’ etc. or for ‘mother’ they included ‘mum’, ‘mommy’, ‘mom’ etc. and then they used their in-house dataset to co-relate million of their products to particular query strings. They also scoured online reviews of their products to label them with their category terms and their aliases — also known as simple binary classification.

The in-house dataset that Amazon used, correlates their query strings with products according to an affinity score — from 1 to 15, where a low score indicates a weak correlation. But, to train their context-of-use predictor system, Amazon researchers created another data set, where each entry was labelled with three data items — a query; a product ID, which has been added by context-of-use categories; and the affinity score derived from the in-house dataset. This data set was then divided into two smaller sets — one annotated according to activity and one according to the audience, and from each of those smaller datasets they constructed two more — one with high-affinity score of 15 and one which was low as 8. This resulting data set was then used to train six different machine learning models.

Multiple Metrics

Once decided to train six different models, the system segregated the models in terms of the affinity score of their training data. The ones had an affinity threshold of 15 were trained using binary cross-entropy, which imposes particularly stiff penalties on incorrect classifications that get high confidence scores. But the ones that had an affinity threshold of 8, Amazon researchers used both binary cross-entropy and B-weighted binary cross-entropy — where it also weights the penalty incurred by each data item according to its affinity score.

The resulted six models were trained to predict context-of-use based on customers’ query strings. In tests, the best-performing model managed to anticipate product annotations with 97% accuracy for activity categories and 92% for audience categories. And, when asked by human reviewers to indicate the classifications they agreed on, they said, an average of 81% of the time the system’s per-item predictions have been correct.

“This suggests, according to Adrian Boteanu, an applied scientist in the customer experience division of Amazon Search that, “The contexts-of-use identified by Amazon’s system could help product discovery algorithms to deliver more relevant results, improving the customer experience. Moreover, the minimal human supervision required to produce training data means that the method could be expanded to new categories with relatively little effort.”

Wrapping Up

As Amazon continues to improve its algorithms, customers shopping on Amazon will see increasingly relevant shopping recommendations. According to Amazon, such research could open a whole new prospect for personalised digital shopping assistants. In this dynamic world where the tech giants are still struggling with their internal bureaucracy and technology silos, it is exceptional to see how Amazon keeps emerging with encouraging innovations to enhance the customer experience.