Amazon Applies AI To Help Comprehend Customer Search Queries
Amazon has always had an advantage in adopting AI to improve its business efficiencies because it was an early adopter of artificial intelligence and automation. It has not only been using AI to enhance its client experience, but it has also been putting a lot of effort into it inside.
Amazon’s AI skills range from predicting the number of customers who will buy a new product to operating a grocery store without a cashier. According to one report, Amazon’s recommendation engine is responsible for 35% of the company’s overall sales.
Amazon has continually deployed AI to comprehend their customers’ search queries, discovering what specific products their customers are searching for. When it comes to making relevant product recommendations to their customers, e-commerce companies need to know what their customers looked for and why they searched. Amazon plans to solve this puzzle by applying artificial intelligence to it.
A recent blog post on Amazon included mentions of artificial intelligence and machine learning used to identify search terms in context. This search system has been designed to aid customers on Amazon.com. This search system was developed to provide customers with better search results on Amazon.com.
In other words, according to the researchers from Amazon, retailers usually search for correlations between queries and products. However, for Amazon, the AI recommended matching products based on how they are typically used. The system anticipates these kinds of activities from inquiries that contain the terms “work out” when searching “Under Armour,” for instance. It will basically analyze the activity of the consumers to give them better options.
Predicting the query’s purpose, according to Amazon, is an important part of information retrieval, which increases the relevance of the results by comprehending latent user intents in addition to explicit query keywords. It is assumed that matching only high-quality products to search queries will make people’s purchasing experience better.
The initial phase in the procedure was to train the system, which required the creation of a data set by the team. Based on common product inquiries, the team created a list of 173 context-of-use categories grouped into 112 activities (such as reading, cleaning, and running) and 61 audiences (such as a child, daughter, male, and professional). They created aliases for the terms they used to designate the groups using standard reference literature. For example, for the category “father,” they included “dad,” “daddy,” “pops,” and “mother,” they included “mum,” “mommy,” “mom,” and so on, and then they used their in-house dataset to co-relate million of their items to specific query strings. They then combed through internet reviews of their items for category names and aliases, a process called simple binary classification.
Amazon’s in-house dataset links query strings with products based on an affinity score ranging from 1 to 15, with a low score indicating a poor correlation. However, to train their context-of-use predictor system, Amazon researchers built a new data set with three data pieces labeled on each entry: a query, a product ID added by context-of-use categories, and the affinity score derived from the in-house dataset. This data set was then partitioned into two smaller sets — one annotated by activity and the other by the audience — and two new datasets were created from each of those smaller datasets — one with a high-affinity score of 15 and one with a low-affinity score of 8. Following that, six distinct machine learning models were trained using the data set that resulted.
The method separated the models based on the affinity score of their training data once it was determined to train six different models. The ones with an affinity threshold of 15 were trained using binary cross-entropy, which penalizes inaccurate classifications with high confidence scores more harshly. However, for the data items with an affinity score of 8, Amazon researchers utilized both binary cross-entropy and B-weighted binary cross-entropy, which weights the penalty experienced by each data item based on its affinity score.
Six models were developed to estimate context-of-use based on query strings provided by clients. The best-performing algorithm predicted product annotations with 97 percent accuracy for activity categories and 92 percent accuracy for audience categories in tests. When human reviewers were asked to mark the classifications, they agreed on, and they claimed the system’s per-item predictions were correct 81 percent of the time.
- Fire TV Stick 4K streaming device with Alexa Voice Remote
- BUENOS NINOS Women’s V Neck Floral Maxi Dress
“The contexts-of-use discovered by Amazon’s technology could assist product discovery algorithms in producing more relevant results, increasing the consumer experience,” says Adrian Boteanu, an applied scientist in Amazon Search’s customer experience division. Furthermore, because no minimal human supervision is necessary to generate training data, the system might be easily expanded to additional categories.”
Final Thoughts
Customers shopping on Amazon will get increasingly relevant purchasing recommendations as Amazon’s algorithms improve. Such research, according to Amazon, might pave the way for customized digital shopping assistants. In this fast-paced world, when tech behemoths are still grappling with internal bureaucracy and technology silos, it’s remarkable to see Amazon continue to innovate to improve the consumer experience.