Causality is one of the most fundamental properties of our universe. As humans, we take advantage of this property all the time. When we take action, we do so in hope of achieving some outcome.
But we don’t just hope for any old outcome…
No Sir…
We hope for achieving a very specific outcome which we hope will bring us closer to our goal. We often even try to quantify and measure the outcome to track our progress and make corrections when necessary.
This describes a typical pattern of our behavior: we take action, observe its result, and decide on the next action. However, no action happens in isolation from the others and by measuring the outcome only based on our last action we discard the significance of all our previous actions. Moreover, the measurement of the outcome may be costly or difficult to obtain and, as a result, impossible to obtain on demand, let alone continuously. This situation happens in a multitude of real-life situations. In healthcare, we apply series of medical actions (medical procedures, drugs intake, dietary regulations, lifestyle changes, etc.) to improve the health of each patient, which is periodically measured by some indicators (e.g., blood work, heart rate, tumor size). In marketing, we apply series of advertising campains via different media (TV, email, regular mail, personal visits, etc.) to make our products recognizable or simply make people like or buy them. But in both of these cases, we only analyze one action at a time when it comes to assessing its impact on the given outcome. While I perfectly understand and value this approach, I also claim that it misses out on potentially a lot of information about the compound effect of series of actions and their influence on one another.
Designing dedicated experiments to incorporate all of these options is obviously impossible since there are too many combinations to consider. However, in todays world each action and measurement leaves behind a digital footprint in the data we so eagerly gather.
And this is where we come in.
Our goal is to create a method capable of mining such data to automatically extract patterns of actions or events leading to certain outcomes. We call them state-changing sequential patterns, where by state we understand the measured outcome and by sequential pattern — a sequence of actions leading to a given change in state. This report presents a continuation of our previous research on this matter, published as a work-in-progress paper [1], where we introduced the concept of state-changing sequential patterns. In this report, together with my graduate student Mariusz Popiół, we intend to develop this idea further by proposing methods for finding such patterns.
Before we begin, let me tell you a little secret. Although I’ve already given you two prominent examples where this problem occurs (healthcare and marketing), these are not what I had in mind when I came up with this idea. My initial inspiration and the ultimate goal of this research is to answer the following question. What makes people change their minds?
Let me explain.
It is rather obvious that you are not going to convince an avid flat-earther about the roundness of our planet by presenting him with a single argument. He’s probably heard it before! Yet, people are changing their minds about their most deeply held convictions every day, some of which are hugely consequential! People change their minds about: who they are going to vote for, should they vaccinate their kids or not, what to eat, is climate change real, etc.. This even concerns something as deep and rooted in us as our religious convictions!
So, how is it that people change their minds? I think that it’s uncontroversial to say that — in most cases — it’s a process. A process consisting of a series of events which eventually tip the scales. Not all of these events work. Some may be more effective than others. Some may work in the opposite directions. Some may need time to kick in.
I believe that state-changing sequential patterns may be our shot at figuring this out, or at least be a part of the solution. And today, thanks to tons of data from social media, we are in the best position ever to try to do it.
So, let’s try!
In our previous research [1], we introduced the broad concept of state-changing sequential patterns with a simple proof-of-concept solution. The solution actually discarded the value of state and was solely focused on the direction of its change. Moreover, the time between the events and state measurements was also neglected.
In this study, we intend to take these as well as some other factors into consideration. Firstly, we want to treat state as-is, regardless if it’s categorical, ordinal, or continuous. This way, we can find patterns of events not only based on their frequency but also their impact on state. Secondly, we want to factor in the time of the events and state measurements to be able to detect delayed responses to some of the events as well as spontaneous changes in state. Thirdly, we aim to better identify the moments in which the shifts in state occur since we use them to split the sequences into smaller sequences where the value of state is monotone. It seems trivial — just split whenever the state changes its monotonicity, right? Well, not so fast! A relatively small change in state does not necessarily have to indicate anything significant. Virtually all measurements naturally fluctuate without any apparent impact.
The addition of these constraints makes our previous solution completely unusable, as it actually redefines the problem a little bit, so we have to start from scratch. Luckily, they are also completely independent, so we can work on each of them separately. Let’s start with time.
[1] M. Piernik, Solomiewicz Joanna, and A. Jachnik, “Assessing the effectiveness of sequences of treatments using sequential patterns,” in Artificial intelligence in medicine, 2019, vol. 11526, pp. 131–135.