But what does this all mean for your company?
Although these advancements in artificial intelligence and machine learning are exciting, it’s difficult to ascertain how the AI revolution of our world, just in its infancy, will translate to businesses’ bottom line. These popular examples in the headlines, while great for publicity, provide little evidence of ROI. They are very specific use cases that required significant investment in time (years, potentially) and personnel.
For example, the AI that wins a game of chess will not be able to win a game of Settlers of Catan without the same investment to learn the strategy and rules of the game. When this drawn-out time investment is added to the costs of hiring the team of machine learning specialists required to monitor the AI product’s development, it’s clear that the immediate returns that enterprises need are just not existent.
This is not to say there is no value eventually; the “long game” may be the application of the lessons learned while developing the AI to real products, akin to the impact that Formula One had on your daily commuter. The technology and innovation developed by Formula One “trickles down” in some way to the mass market.
Learn more about how AI-powered search can deliver results for you business – and how it plays a role in the evolution of enterprise search into insight engines – by accessing your complimentary report:
Pragmatic Machine Learning – with Actual Business Impact
You may not see this in the headlines, but we believe it’s more exciting. Our team debuted the incorporation of an artificial intelligence program into our platform in 2015. Machine learning plays a crucial role in helping our customers’ businesses to scale.
What’s different? We started with a very specific problem that we wanted to solve for our customers. After deploying cloud and usage analytics in 2014, we saw how successful our customers were with our product, but tuning search results for relevance required constant attention. When our customers added more users, more content and more connectors, you can imagine how it would be easy to get out of hand.
Our research and development team spent a lot of time looking at usage analytics to identify how to improve our product. We ran into the same issues when answering the same questions for each customer: How can we predict what results will be relevant and useful to our customers end-users? How can customers better understand their users’ behavior and habits? How can we help them to understand their end-users’ context more comprehensively?
During an “innovation session,” one of our research and development working groups found the answer: machine learning (ML). Using ML, the group hypothesized, our team could automate the resource-intensive work of analyzing and applying insights from that analysis to tune the search engine.
The rest is history, as they say. Instead of creating publicity stunts, our team pulled together to create a pragmatic machine learning program that is uniquely positioned to have an impact today. Not at some point in the distant future. Not after certain conditions are met. We have enabled businesses to see real ROI from machine learning today.
What Sets Coveo Machine Learning Apart
Search is a journey. Everyone is different, there are multiple ways to formulate the same query depending on one’s own personal preferences, character, or vocabulary. The digital age has also made us “lazier,” as in expecting systems to understand the few clues we are providing them.
What sets Coveo Machine Learning apart is its mission to seamlessly reconcile a user’s personal, fuzzy query with his/her true intent and to provide the right information that will lead to a successful journey. In a nutshell, our approach to Machine Learning is highly efficient, flexible, and can automatically learn personalized models for all our customers.
Let’s start with a simple scenario to understand how machine learning works. The classic scenario favored by Machine Learning aficionados examines the dataset of Titanic passengers, to predict whether or not an individual would survive, given the set of features (name, age, ticket fare, class, number of relatives on board, sex, boarding port). While many in the general public believe that being in first class, a woman or a child significantly increased the chances of survival, applying machine learning reveals a more nuanced reality. When analyzing the dataset, it is clear that all those other features/variables held just as much, if not more, influence than class, age or gender.
Once a model has been created, data scientists keep tuning it, or testing its ability to predict outcome based on given features. They compare the prediction of the model (in the example above: the survival of a passenger) and the actual/expected outcome. It is easy to see how optimizing a model to perform in a certain environment becomes extremely complex and time consuming, even with this Titanic example. For many websites who are trying to meet the expectations of customers who have gotten used to the Amazon and Google search experiences, the payoff and benefits are worth the investment. A machine learning model that is completely personalized to the use case of your website is priceless for customer experience.
At Coveo, we recognize the uniqueness of each customer’s use case; we build hundreds of models daily to meet each customer’s needs. Our solution operates at scale by auto-tuning itself and rebuilding the models frequently to constantly reflect the ever changing user behaviors and information flow.
How does it work?
Coveo ML leverages the search sessions data to learn what the users are looking for and their journey to get to the right information. Features that are incorporated into the Coveo ML models include queries, clicks, pageviews, context (e.g. user role, country, device used, …) and other custom actions specific to each customer. The weight and the selection of features is specific to each use case.
In addition to features, Machine Learning algorithms use a specific kind of parameters called “HyperParameters” that need to be adjusted to improve performances. Again, to operate at scale, the Coveo team has developed a fully automated and efficient HyperParameters optimization process that is used by all our models for fine-tuning. The weight of the different HyperParameters is tweaked slightly to help more accurately predict the outcome.
While this may seem relatively simple when dealing with few thousands of rows of data, the shear scale at which our customers operate make innovation like Coveo ML necessary. Our customers don’t need to hire their own team of specialized data scientists. We do it for them, out-of-the-box.
After years of working in enterprise search and helping our customers understand their usage data, we understand that tuning models for each individual dataset is required to achieve the accuracy their customers/users expect when performing a search. Other vendors may choose to apply a standard algorithm across the board for the sake of simplicity; this, we believe, leads to search models that are not effective, returning poor relevance.
In this example from Logitech’s support site, you can see how basic keyword-matching results would frustrate the user. Prior to Coveo, searching for “mouse” brought up all of the results with mouse without any ranking for popularity. With Coveo, it’s the most popular models that appear first in the results list – because the machine learning has understood that the results for those mice are more frequently clicked.
With Coveo™ Machine Learning
In another example, without any human intervention, machine learning was able to identify that customers searching for the model “mx5500” meant the keyboard model MX™5500. The model had been indexed with the trademark symbol, so searching for “MX5500” did not return results for the keyboard.
With Coveo Machine Learning, however, the popular keyboard appears first in product results list.
Machine learning automatically upgrades the customer experience by learning from users. The system was able to realize that users were searching for a manufacturing part number on the dongle that comes with the mouse. This number was not in the index at all. Thanks to machine learning, however, the system learned the product that matched the manufacturing part number and surfaced the right results – all without human intervention.
These are just a few cases of how “pragmatic AI” is able to dramatically upgrade the experience automatically. Machine learning plays a crucial role in enabling businesses to scale their search experience.
Enabling scalability is partly fuelling the transformation from enterprise search to “insight engines” that deliver relevant insights to the right person at the right time. Learn more by accessing this complimentary report from Gartner: