How AI Is Impacting Operations At LinkedIn

• Read original article here

LinkedIn has been at the cutting edge of AI for years and uses AI in many ways users may not be aware of. I recently had the opportunity totalk to Igor Perisic, Chief Data Officer (CDO) and VP of Engineering at LinkedIn to learn more about the evolution of AI at LinkedIn, how it’s being applied to daily activities, how worldwide data regulations impact the company, and some unique insight into the changing AI-related work landscape and job roles.

The Evolution of AI at LinkedIn

Very early on at LinkedIn, data was identified as one of the company’s core differentiating factors. Another differentiating factor was a core company value of “members first” (clarity, consistency, and control of how member data is used) and their vision to create economic opportunity for every member of the global workforce. 

As LinkedIn began finding more and more ways to weave AI into their products and services, they also recognized the importance of ensuring all employees were well-equipped to work with AI as needed in their jobs. To that end, they created an internal training program called the AI Academy.  It’s a program that teaches everyone from software engineers to sales teams about AI at the level most suited to them, in order for them to be prepared to work with these technologies.

One of the very first AI projects was the People You May Know (PYMK) recommendations. Essentially, this is an algorithm that recommends to members other people that they may know on the platform and helps them build their networks. It is a recommendation system that is still central to their products, although now it is much more sophisticated than it was in those early days. PYMK as a data product began around 2006. It was started by folks that would eventually be known as one of the first “data science” teams in the tech industry. Back in those early days, no one referred to PYMK as an “AI” project, as the term AI was not yet a back in favor buzz word”.  

The other significant project which we started around the same time was of course search ranking, which was a classic AI problem at that time due to the emergence of Google and competition in the search engine space. 

How AI is applied to daily activities

At LinkedIn, Igor says that “we compare AI to oxygen—it permeates everything we do. For example, for our members, it helps recommend job opportunities, organizes their feed, ensures that the notifications they receive are timely and informative, and suggests LinkedIn Learning content to help them learn new skills.” With respect to LinkedIn’s enterprise products, he says “AI helps salespeople reach members that have an interest in their products, marketers serve relevant sponsored content, and recruiters identify and reach out to new talent pools.” The benefits of AI at Linkedin also operate in the background, from helping protect members from fraudulent and harmful content to routing internet connections to ensure the best possible site speed for our members. 

Ensuring member safety on the platform is something that we take very seriously.  Being a social network with a very strong professional intent, it’s important to act quickly in identifying and preventing abuse. Because abuse and threats are constantly changing, AI is certainly at the core of these efforts. LinkedIn has found machine learning very helpful in detecting inappropriate profiles. 

Without AI, many of their products and services would simply not function. The “economic graph” they use to represent the global economy is simply too large and too nuanced to be understood without it.

AI is literally enhancing every experience. Starting from the notifications our members are getting about relevant items.  But, probably, one of the most prominent ways through which our members experience AI is in the feed, which sorts and ranks a heterogeneous inventory of activities (posts, news, videos, articles, etc.). To ensure relevance in the feed, it’s important that the algorithms consider the different nuances of content recommendations and member’s preferences.

One interesting example Igor shares is that at the start of 2018, they discovered an uneven distribution of engagement in the feed—gains in viral actions were accrued by the top 1% of power users, and the majority of creators were increasingly receiving zero feedback. The feed model was simply doing as it was told: sharing broad-interest, viral content that would generate lots of engagement. However, he says they realized that this optimization wasn’t necessarily the most beneficial for all members. To combat the negative ecosystem effect that the AI had created, they incorporated creator-side optimization in their feed relevance objective function to help their creators with smaller audiences. With this update, the ranking algorithms began taking into consideration the value that would result for both viewer and creator in surfacing a specific item. For the viewer,they wanted to surface relevant content based on their preferences, and for the creator, they wanted to encourage high-quality content and help them reach their audiences. Igor says “by tweaking our models to optimize for more than just viral sharing moments, our feed changed into a healthy mix of content from influencers as well as direct connections, which then improved engagement for both viewers and creators.”.

In recent years regions around the world have started to put in place laws around how companies are able to store and use user data. Laws such as the EU’s General Data Protection Regulation (GDPR) or the  California Consumer Privacy Act (CCPA) are  intended to enhance privacy rights and consumer protection. For some companies, becoming compliant meant having to totally chance how they approach data. Luckily for LinkedIn, data was always considered an asset to the company and approached with respect as one of the company’s core differentiating factors.

Even before GDPR, Igor says LinkedIn had an internal framework they call the 3C’s—clarity, consistency, and control. He says “We believed then and still do today that we owed it to our members to provide clarity about what we do with their data, to be consistent in only doing as we say, and to give our members control over their data:.  In that context, LinkedIn approached GDPR as an opportunity to reinforce their commitment to data privacy for all members globally. For example, LinkedIn extended GDPR Data Subject Rights to all members globally. They continue to be thoughtful in how they approach the use of members’ data throughout LinkedIn and in AI, and in how they review and update processes, to ensure privacy by design. Acting in the best interest of members continues to be LinkedIn’s north star, and they always felt that it’s their joint responsibility across the organization to protect members’ data.

As a very large professional social network, LinkedIn has the unique opportunity to see insights about changing job roles, popular positions, and regional popularity that other companies might not have as deep insights into. At the end of last year, LinkedIn released their third annual Emerging Jobs Report to identify the most rapidly growing jobs. AI specialist emerged as the #1 emerging job of that list, showing 74% annual growth over the past 4 years. It’s especially exciting to see this growth beyond the tech industry. In 2017, they found that the education sector had the second-highest numbers of core AI skills added by members, showing that AI’s growth is correlated with more research in the field.

More recently, amid the economic downturn caused by the pandemic, LinkedIn is still observing that the AI job market continues to grow. When normalized against overall job postings, AI jobs increased 8.3% in the ten weeks after the COVID-19 outbreak in the U.S. Even though AI job listings are growing slower than they did before the pandemic, and despite an overall slowdown in demand for talent, employers still appear to be open to hiring AI specialists. 

What’s interesting about the field of AI is that LinkedIn is seeing an entire ecosystem of technical roles that support different stages of the AI lifecycle. If you go back to the Emerging Jobs Report at the end of last year, AI specialist roles (people who build and train models, etc.) are up, but that so-called “AI-adjacent” jobs are also on the rise. This means that you’re seeing more demand for data scientists, data engineers, and cloud engineers. You’re also seeing this demand growing across multiple industries, not just the technology sector. It is across the entire spectrum. 

At the end of the day, AI is a tool, and its greatest potential lies in how it will augment human intelligence and how it will enable people to achieve more. LinkedIn’s current AI tools depend greatly on human input and can never fully be automated. 

Igor strongly believes that the future of AI is in applications and especially how we leverage that tool to make us all smarter and to enable us to do more. To do so, AI needs to be much more accessible to a wider set of individuals than just AI experts. AI needs to become more of a plug-and-play, almost a point-and-click interface. He’s seeing the major cloud players get into this space, developing tools that help lower the barrier of entry into AI. Once AI is application-driven, it opens up human creativity to develop really cool and interesting use cases.

In that context, AI technologies are really fascinating across the entire spectrum; from algorithmic and mathematical developments to hardware and AI systems. Just think about the ingenuity researchers have shown in attempting to make their deep neural nets simply converge. In the AI landscape, it seems that there are treasures behind every bush or under every rock.