At this stage in its development, one of machine learning’s best use cases is for automatically identifying patterns in a large data set, and using what it learns about those patterns to make predictions about future outcomes.
Companies use this capability for a number of business reasons. Instagram uses its predictive algorithms to recommend relevant content in the explore tab, with the intent of driving more engagement. Hopper, the mobile travel app that tells you when the best time to buy a plane ticket is, uses past data to make predictions about future prices. Music applications use information they have and continue to learn about songs and user preferences to create playlists they predict their listeners will enjoy.
But not all music recommendation apps are created using the same type of algorithm, or even the same fundamental approach to music. A service like iTunes relies solely on the behavior of others who appear to have a similar taste in music to you. They recommend songs they think you’ll like simply because other people “like you” also listen to them. The inherent challenge with this approach is that it becomes difficult to try or find something new; the most likely outcome is that you will keep hearing similar songs and albums over and over, causing listeners to cement their catalog of artists, without an opportunity for real discovery.
Spotify combines multiple machine learning approaches to recommend songs, including algorithms that connect songs to one another by looking at the raw audio files, as well as algorithms that connect user preferences to other users, to create large neural networks designed to recognize patterns. While this is highly advanced, it leaves a great deal of decision making power to the machines and removes the opportunity for nuance to take hold.
Pandora has an equally sophisticated but highly contextual approach to its prediction engine. Instead of only relying on the music habits of its listeners, it rather successfully attempts to understand, map, and classify the contents of every song in its collection based on predetermined, tangible characteristics defined by human inputs. This leads to highly personalized recommendations, all built around the intention of playing people what they love at the exact right moment, as well as introducing them to totally new music that expands their tastes and preferences without any effort or research.
The way in which Pandora treats its songs and its listeners is exactly the way we fundamentally treat our candidates and our partners at Suited. There exist many similarities in our approaches to making machine learning based predictions, and we believe it is because of these deliberate choices in our methodology that we are able to make the type of predictions we do — those that expand the way firms think about a candidate and afford them the opportunity to escape typical hiring patterns that often stifle diversity and the discovery for something potentially novel, but just as “good.”
Below is a comparative analysis of the Pandora and Suited approaches, looking specifically at the most relevant overlaps, in order to provide a greater sense of understanding as to how trustworthy predictions are made across both services.
The Consideration of A LOT OF Qualifying Traits
Pandora uses over 450 selection attributes that are combined into larger groups to form nearly 2,000 focus traits, some examples of which are rhythm syncopation, keytonality, vocal harmonies, and displayed instrumental proficiency. This project of classifying every song is what Pandora refers to as the Music Genome Project. As mentioned, some music services only consider the listening habits of other people who like the same songs or artists as you, or never explicitly define what is at the core of a song's “DNA.” It may not be obvious at a casual listen, but songs are incredibly complex things, and what you hear is the product of multiple factors and their interactions with one another.
When companies normally decide which early-talent candidates to bring in for interviews, they only look at three data points: their university, GPA, and major. However, only looking at these data points is not especially predictive of potential on the job performance. In order to reduce time wasted on interviewing ill-fitted candidates, or money spent hiring them, Suited analyzes over 50 traits to determine the probability of whether a person will be a good fit for a particular firm.
The ability to perform may seem simple but human behavior, and specifically performance on the job, as with songs, is incredibly complex. Suited measures all of these individual dimensions and accounts for their interactions as well.
The Combination of Human Input and Artificial Intelligence
Pandora employs “musician-analysts” (people who know about the technical and social aspects of music) to listen to songs and classify them using those 450 tags to indicate their internal characteristics and feel. This musical metadata is then inputted into their machine learning system to continuously teach it to automatically identify these traits.
In order to fully understand what traits and qualifications are necessary to fulfill a particular role within an industry, Suited employs our in-house Industrial Organizational Psychologists to interview and assess relevant professionals and perform specific job-analyses to quantify these particular qualities. For example, to determine which cognitive skills are critical for a job in investment banking, Suited interviewed multiple incumbents and supervisors to uncover the basic skills that enable someone to be successful. It was discovered that building complex financial models in Excel requires a lot of technical knowledge (e.g., Excel skills, understanding financial concepts, knowing financial formulas). However, it also requires a high degree of attention to detail and an ability to look at lots of numbers and computations and verify they look “right.” Working with multiple clients across various industries, we were able to identify three critical basic cognitive skills important for entry-level roles in banking, and that information is provided to our customers.
The Removal of Bias to Prevent Homogeneous Tendencies
If you’ve ever used Pandora, you know that the system does not allow you to choose the exact song you want to hear. Instead, you give the algorithm a direction to go in by selecting an artist or type of song you’re in the mood for. It then uses its own internal database of music characteristics, defined and attributed by human experts to determine what your preferences are and make good recommendations.
Other music services give their users the opportunity to curate their own music experiences. While that can be great for certain instances, it also places a burden on the user to make continued choices that are often repetitive and cause musical-ruts. With Pandora, listeners are not forced to spend tedious amounts of time building their own playlists or manually enter their preferences. Instead, the system is able to do the work as soon as you give it a place to start with a single song.
Suited also structures its model around “similarly-matching” (i.e., what candidates are most like the high performers at a particular firm), but does not take explicitly stated recruiter or firm preferences into account, because often these surface-level preferences are incredibly biased. We dive deeper into what we know makes someone successful at your firm and then match you to candidates who possess similar traits, tendencies, and skills to those who have proven themselves to be valuable to your business.
The Ability to Make it Customized and Personal
Pandora possesses years of data it’s been collecting from users about music preferences and listening habits. The algorithm also learns more about what you like and don’t like through your thumbs up or thumbs down ratings. Using all of this data, Pandora is able to make accurate predictions of not only what you want to hear, but when you want to hear it, and will adjust their recommendations accordingly.
As mentioned already, the models Suited creates are customized to each firm’s workforce and culture. Each candidate’s profile is put through a specific algorithm and may render as a high potential candidate for one firm and an average potential candidate for another. Suited’s algorithms can also learn over time. Feedback about successful performers or candidates who weren’t a good fit for one reason or another can help adjust the predictive algorithm to narrow in on what really matters for performance.
The Likelihood You Will Be Introduced to New Things
Pandora prides itself on being the place for music discovery. They serve their users recommendations that immediately appeal to their tastes while simultaneously broadening their musical scope.
Suited provides you the ability to discover candidates you may have otherwise overlooked because of their collegiate pedigree, grades, or other less predictive factors. Traditional scholastic indicators of success have been proven to be especially biased towards People of Color, and both POC and female candidates are often not actively recruited into finance roles based on what some people call the “pipeline problem,” believing that there just don’t exist qualified candidates to fill their desire for diversity. With Suited, you are intentionally served a collection of candidates who are guaranteed to be diverse, qualified, and different to what you may normally consider.
Using what worked before isn’t always bad. But that approach makes it too easy to perpetuate the “like-me bias” and further entrench biases. Finding diverse candidates who are a good match for your firm and who will bring some unique features or new perspectives can both capitalize on your previous hiring successes and introduce benefits at the same time.
The Chance For Divergence
Pandora wants you to like what they recommend to you, but they also want you to feel like you are expanding your musical horizons. To do this, Pandora includes necessary social information to make sure they don’t verge too drastically from what your tastes are. For example, the quality they label “head-nodic beats” includes a certain type of hip hop but not the many other genres that compel a listener to nod their head. Even so, sometimes a song comes along that you don’t particularly like, and you skip and give it a thumbs down to tell them so. Because Pandora sets these sorts of cultural boundaries for itself, you most likely won’t be introduced to something that makes you uncomfortable or is completely ill-fitting.
Suited also has the potential to serve you up someone who ultimately may not be a good fit for your firm. However, the screening process is made exponentially fairer by using a tool that systematically mitigates bias, and to do that requires a necessary margin of error. Even so, this type of error is 10-20x less likely to occur than if a human did the resume scan themselves.