Layers of Data: How Context Helps Make Better Decisions

In 2012, Harvard Business Review declared the role of Data Scientist to be “the sexiest job of the 21st century.” Glassdoor added fuel to the fire in 2016 by listing data scientist at number one on their list of the best jobs of the year. Their number one job in 2019? You guessed it: data scientist. So what is it about these “high-ranking professionals with the training and curiosity to make discoveries in the world of big data” that is just so damn sexy? For starters, there’s more data than ever, and most of that data has been created extremely recently. In 2017, Data management platform Domo estimated that 90% of all data had been created within the prior two years — to the tune of 2.5 quintillion bytes of data per day. It’s no wonder that companies are interested in discovering how best to use this unprecedented volume of data to their benefit. However, voluminous data isn’t necessarily relevant data. This incongruity has led to companies scrambling over the past decade to implement programs to interpret this data and generate actionable insights. As Duke economist Dan Ariely stated in 2013:

“Big data is like teenage sex: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it…”

Data scientists, therefore, have become highly valuable for their ability to make sense of this sea of information. By drawing connections between different data points and assessing the bigger picture of why that data might matter, data scientists are able to help companies surface meaningful insights that can lead to better decisions.

Why Data is Meaningless Without Context

On its surface, data is simply a collection of numbers, words, etc. It is only when data is presented with a context that it becomes meaningful. Take, for example, this Google trends graph:

Without any context, it’s completely meaningless. There is certainly something cyclical to the frequency, but it’s not entirely clear what or why that frequency is occurring, or why it matters. Now, if I told you that the trend graph was for the search term “pie,” you could begin to start drawing meaningful conclusions about the data: namely, something every year is causing a spike in searches for the term “pie.” If I continue to provide contextual clues, the data becomes more and more relevant. For example, allow me to further clarify that the search is limited to the United States, and the spikes seem to be focused on one major and one minor spike:

Knowing that searches for “pie” in the United States significantly peak in November every year, we can begin to make inferences about what the data means. At this point, you might be saying to yourself “clearly, searches for pie are spiking significantly around Thanksgiving.” However, this is where context becomes critically important. If you were born and raised in the US, chances are that you have the context to know that Thanksgiving is in November, and that pie is often served at Thanksgiving, therefore the spike in pie searches directly correlates to people planning their Thanksgiving celebrations. However, if you were not raised in the US, you might lack the contextual reference to understand why pie searches peak in November in the US.

So what about that minor spike in March? What you may have guessed at this point is that the spike in March is likely a result of the lesser-known holiday of Pi Day that occurs every March 14th (3.14…get it?). Once again, this data is only insightful to those that possess the proper context to provide relevance. If you’ve never heard of Pi Day, the March data spike might not provide you with much insight.

With the proper context, decisions can be made based upon the data, but also based upon what the data means. In our pie example, the CMO of a hypothetical pie manufacturer would be armed with information about traffic habits (data), as well as topical interests related to those habits (context). Taken together, this could support an ad campaign where the data informs when and how to reach consumers, and the context informs what they might find relevant and interesting.

It bears mentioning that this example is an oversimplification, and likely would require additional data and context through both quantitative and qualitative means. Regardless, hopefully it’s becoming clear that those sexy data scientists can help improve decision making by researching trends and contextual clues. What may be less clear is how they could possibly begin to do this with 2.5 quintillion bytes of daily data. This is where technology enters the equation.

How Computers Can Help Navigate Contextual Clues

Before we get into silicon and algorithms, let’s take a moment to think about what it is that makes Data Scientists able to do their job so well. Broadly speaking, they all have access to the most powerful computer in existence: the human brain. Dr. Bobby Kasthuri, a neuroscientist who studies attempts to “map the brain at a scale that’s never been mapped,” estimates that a single human brain has something along the lines of one quadrillion connections linking over a hundred billion neurons. It is this extreme complexity that leads to consciousness and cognition. Effectively, this complexity and power is what makes humans human. This separation between biological processes and the high-level behavioral traits they influence is ultimately the difference between neuroscience and cognitive sciences. This all begs the question: if we could make a computer as complex as the human brain, would it be able to “think.”

As of the publishing of this paper, the level of complexity of the human brain is orders of magnitude greater than the most powerful computers that are available to us, and as such, true computer cognition is still a long way from reality. If we look at the neuroscience/cognitive science relationship, there are distinct parallels between electrical engineering and computer science. The theory behind how computers behave (computer science) is limited to the practical constraints of the hardware (electrical engineering). Computers, therefore, require input and instructions from humans in order to behave a certain way. In other words, they require data (input) and context (instructions) in order to generate some valuable output. Data Scientists can use their experience, insight, research, and critical thinking to help computers “learn” which context is relevant for a given data set. In turn, the computers can rapidly and independently process this data within the given context. This is effectively how machine learning works. Once trained, the computers can take new input and determine if that input matches the established context, and take action accordingly, offering a new output that is judged against the established contextual understanding. This loop is essentially what we describe as artificial intelligence.

Despite having these rudimentary tools available to us, computers are still a long way from thinking. As Oren Etzioni states in How To Know if Artificial Intelligence is About to Destroy Civilization:

“Machines possess only a narrow sliver of humans’ rich and versatile learning abilities. To say that machines learn is like saying that baby penguins know how to fish. The reality is, adult penguins swim, capture fish, digest it, regurgitate into their beaks, and place morsels into their children’s mouths. AI is likewise being spoon-fed by human scientists and engineers.

In contrast to machine learning, human learning maps a personal motivation (“I want to drive to be independent of my parents”) to a strategic learning plan (“Take driver’s ed and practice on weekends”). A human formulates specific learning targets (“Get better at parallel parking”), collects and labels data (“The angle was wrong this time”), and incorporates external feedback and background knowledge (“The instructor explained how to use the side mirrors”). Humans identify, frame, and shape learning problems. None of these human abilities is even remotely replicated by machines. Machines can perform superhuman statistical calculations, but that is merely the last mile of learning.”

How We can Use Artificial Intelligence to Make Better Decisions

Clearly, the capacity for technology to “think” is still a far cry from our own. However, AI still offers significant benefits when we think of it as one of several tools available to improve how we make decisions.

If computers lack the ability to truly “think” and generate strategic decisions, should we still use the term artificial “intelligence”? I would argue yes, provided we consider AI by the following definition: AI is a computer system that relies on human input in order to process possible outputs given the context of the data. Without human intelligence, artificial intelligence is meaningless.

Thinking about this definition of artificial intelligence, it becomes clear that the value of any output will be limited by the accuracy of the data and the relevance of the context. Additionally, the scope of the context will be the largest driving factor that “colors” any results. If we are to accept that artificial intelligence functions at least similarly to human intelligence, albeit significantly less powerfully, then it stands to reason that we should first think about how we use our own capacity for intelligence to inform decisions. Consider this quote from Bob Suh, Founder and CEO of machine learning firm OnCorps and former chief technology strategist at Accenture:

People with high EQs (emotional quotient) have distinct advantages. They learn other people’s patterns, know how to convey empathy, and carefully weigh responses to choose the best one based on how they anticipate someone will react. EQ helps leaders make better decisions and produce better outcomes.

Considering this definition of emotional intelligence, patterns are derived from responses (data) being carefully weighed (context). All of this falls in line with our prior definition of machine learning. The primary addition is in Suh’s description of high EQ individuals knowing “how to convey empathy.” This is where human understanding and AI output can act together for better decision making.

The most commonly accepted definition of empathy (and the one that Suh is referencing in relation to high EQ individuals) is “the psychological identification with or vicarious experiencing of the attitudes of another.” Less commonly discussed is the secondary definition of empathy, which is “the imaginative ascribing to an object, as a natural object or work of art, feelings or attitudes present in oneself.” If we think of “a decision to be made” as an object to which we can ascribe our own feelings or attitudes, then we can see how decision making is a series of actions where we apply our own emotional context towards a final outcome. This is where the complexities of the human brain supersede the capabilities of computers, and allow us to make more complex decisions than computers could on their own.

Avoiding the Dangers of Contextual Bias

In the examples we mention above, it is the variability of the individuals’ emotions and experiences that ultimately allow for complex decision making. Each person will bring a different, slightly nuanced interpretation to both input instruction and output assessment. Therefore, it is critical to consider how the context used to train AI can lead to biased results. While there are clearly certain biases that an individual may bring to how the interpret and use any AI output, this section will specifically be looking at how input bias can lead to significant, yet often hidden issues. Why? Because if you are tasked with making a decision based solely on the output, you may lack any context for what shaped the input, which could lead you towards drastically incorrect interpretations of the data.

One of the most blatant examples of a biased context is historical biases. These occur when the data set is limited to a small contextual subset due to cultural, legal and/or systemic reasons that may no longer be valid. Take, for example, the widely reported issues related to Amazon’s attempt to create an AI algorithm to vet new hires and find the best candidates for development positions with the company. Once implemented, the results skewed significantly towards a preference to recommend male candidates over female candidates. The reason? Amazon built the system using a decade of applications as the context for success/failure. The problem was that an overwhelming number of candidates and hires in technical roles had been male. While it is no secret that historically (and presently) tech sector jobs are dominated by men, that historical context must be accounted for in any data models in order to prevent biased hiring decisions.

Another example of biasing results relates to the earlier discussion regarding human experience and perspective acting to provide context to a data set. There is a phenomenon related to the discrepancy between what humans say they are feeling/thinking, and how their brain is actually responding. As we discussed earlier, we can separate the processing (neuroscience) from the output (cognition). In many cases, our life experiences lead us to (often subconsciously) alter our outputs to meet what we believe to be acceptable (socially, situationally, or otherwise). Emotions are, by their nature, subjective. This, of course, can affect how we train data models, which will add bias to any AI outputs.

If we recognize that contextual subjectivity will inherently bias data training, we must then examine what potential impact this could have on real-world decisions that are based upon these subjective models. One of the earliest and most thorough investigations into the impact of contextual bias came from ProPublica in May of 2016. They conducted a thorough analysis of “risk assessment” algorithms that were being used in courtrooms across the United States to determine the risk of recidivism for individuals coming through the judicial system. Ultimately, they found that the algorithms were significantly more likely to identify black defendants as being of higher risk for recidivism than white defendants, which led to significantly higher sentences for those individuals. The problem was that this did not correlate with actual rates of repeat offense. White defendants were chronically under-scored compared to their actual rates, while black defendants were chronically over-scored compared to their actual rates. In reality, recidivism rates between races is far closer than the scoring would indicate. While the specific design of the algorithm in question is proprietary, and therefore we are unable to conclusively determine the exact reasons for its failures, it brings to light the dire consequences of what can happen when we “get it wrong.” People lost years of their lives because it was not considered that contextual biases might incorrectly skew outputs.

If there is room for such significant, potentially life altering consequences to contextual bias impacting AI outputs, it begs the questions: should we be using AI at all? Is it still too rudimentary compared to the human brain to realistically provide us with anything meaningful? In short, no, we should not abandon the potential benefits that AI presents. However, we must constantly be aware of how the technology applies context to data, how that context adds bias, and what that means for any outputs going forward. If we recognize what AI is good for: namely, acting as but one of many tools we can use to supplement our decision making processes, then we can begin to harness the potentially significant benefits within our personal and business lives.