Your Behavior, Modeled
Think about the last time you’ve seen ads or recommendations being made to you while browsing Netflix, Instagram or Amazon. It’s likely that such recommendations are driven by what most people would call “AI” or Artificial Intelligence.
Now think about when the last time the recommendation or ad you got was actually was something you wanted, or if it made you feel like you were really being catered to? I rarely find that these recommendations satisfy what I genuinely desire. Even then, I do feel that these recommendations feel thematically similar, and are optimized so that my attention on the platform is maintained, to keep me on longer to sell me more items or more information. But my experience is from a population of (n=1). However, searching the web, and conversing with friends, I soon realized that my same experiences were experienced by almost everyone around me(well … this could be confirmation bias, or lack site-validation … but I’m not aiming for statistical rigor with this piece).
What I wanted to shed light on in this article is how these recommendations and ad targetting are generated. I also want to propose a few ideas (shared by some notable cogntive scientists) that may improve the nature of technology in the future to encode truly ‘meaningful recommendations’. That being said, I acknowledge I am opening pandora’s box, in that in the future, with these improved recommendation systems, our complete decision making economically could be at the finger-tip of the companies that implement these systems. But I’m willing to risk that.
How do recommendation systems work?
Chances are, you’ve encountered one of two recommendation systems : Collaborative Filtering or Content Filtering. Collaborative filtering takes data from other individuals who share the same interests, and recommend items that others have purchased in-joint with the original item of interest. Content filtering attempts to use descriptive content (like reviews) to then provide recommendations. Similar items with similar ratings/descriptive content will be recommended to the end user. Almost all modern recommendations use a variation of the two, or the combination of the two to provide recommendations and ad targeting. There have been advances, especially in using deep learning methodology (ie: Neural Collaborative Filtering, Deep Matrix Factorization) driving forth these forms of recommendation systems. There are also ‘new-age’ techniques, with advances in reinforcement learning (think the Google’s AlphaGo system that beat the top Go players in the world) that are also being implemented in-joint with collaborative filtering and content filtering. These systems are usually embedded within or are largely similar to collaborative filtering methods. A reinforcement learning system establishes a reward function that the recommendation system can operate by and derive meaning from.
What’s wrong with these methods?
There’s nothing inherently wrong per-se. Simply, there are methods by which we can improve the way we build recommendation systems. And this is the exact problem that cognitive scientists study (psychologists and economists also study this information).
The biggest flaw in the current methods are the assumptions being made. Tom Griffiths, Professor of Psychology and Cognitive Science and Princeton (formerly Berkeley), in his research manifesto, points out the fact that these traditional systems only “predict what you will [purchase/consume] based purely on the similarity of your behavior to the behavior of others”. This obfuscates the fact that there is no active modelling of your personal preference. Ads, for example, are commonly recommended based on reinforcement learning systems where the reward is based on how likely people are to click on webpages/or ads based on recently visited/current webpages. This is exactly why you might see ads for weird things like “sperm-bank donations” when you look up things about sexual fertility (trust me, this happens). Or when you visit a website for a product, and then receive targeted ads for that exact product on Instagram. They are simply trying to maximize click-through rates, based on the assumption that click-throughs=purchases.
What’s a better approach then, if the current approaches aren’t that great?
A lot of the work is still proof-of-concept. With the internet age allowing us to steadily accumulate data about our behavior, we can hope to see more study in the realm of cognitive representations.
Cognitive representations are different from traditional methods, in that their purpose is to encode complex and rich information about preferences, semantic representations, and categorizations we create with respect to objects, feelings, etc. This thus forces us to focus on understanding the behavior of the individual.
Most of the work done to study human behavior still exists within the laboratory settings. Modern cognitive scientists, psychologists, economists all collect data online now. However their work is largely only conducted within the labs they work within, and papers are simply published, and never batted a second look by technologists. This creates a disconnect between those who study behavior and those who study computation and ultimately how we provide recommendations.
What may happen as a result of moving towards learning of cognitive representations?
I’m both optimistic and pessimistic about what will happen. Optimistic, in that we will finally facilitate true meaning through our computational models that are already affecting our daily life. I’m hopeful that this will improve our overall reward system socially, and reward unique preference rather than attempting to conform all behavior into one paradigm, which is likely to happen through methods that reward only one type of behavior (e.g.current use of reinforcement learning for ad targetting), and ones that have high potential to create fallacious representations of behavior (collaborative filtering). On the downside, well, big brother may be able to track and understand your behavior in depth. Who knows, maybe we’ll succumb to the will of our advertising overlords that get us to endly spend money and become fully consumerist as a civilization. I’m just hypothesizing here.
Regardless of the optimism/pessimism, my hopes is that this article can open you to a new field of study that is not commonly discussed, but should be a topic of discussion philosophically. The way we currently attempt to model behavior, through recommendation systems, or message targetting via information filtering, may be sufficient in terms of business purpose, but do not ultimately provide the full value we seek out of these systems. In the future, I hope that we can embed more focus on learning to encode more rich information through cognitive representations; which would only make the value of the models we build for humans more powerful.
I definitely recommend reading Prof. Tom Griffith’s Manifesto to learn more about some of the opinions mentioned in this piece: ( http://cocosci.princeton.edu/tom/papers/ComputationalCognitiveRevolution.pdf)