8
$\begingroup$

I'm trying to see if there is a conventional term for this concept to help me in my literature research and writing. When a machine learning model causes an action to be taken in the real world that affects future instances, what is that called?

I'm thinking about something like a recommender system that recommends one given product and doesn't recommend another given product. Then, you've increased the likelihood that someone is going to buy the first product and decreased the likelihood that someone is going to buy the second product. So then those sales numbers will eventually become training instances, creating a sort of feedback loop.

Is there a term for this?

| improve this question | | | | |
$\endgroup$
  • $\begingroup$ Yes, "feedback loop" is how I would always describe this. $\endgroup$ – Sean Owen Apr 3 '15 at 9:28
  • $\begingroup$ I think feedback loop is definitely useful to say, but the literature on performativity and reflexivity seem to address more of the complex interplay that I was trying to discuss. So I think it's useful to say something like, "A model that can act on the concept to change the concept can thus change itself in a sort of feedback loop. Let's call this property (performativity/reflexivity). This is distinct from concept drift, because the model is the cause of the drift." $\endgroup$ – jsmith54 Apr 3 '15 at 15:08
6
$\begingroup$

There are three terms from social science that apply to your situation:

  1. Reflexivity - refers to circular relationships between cause and effect. In particular, you could use the definition of the term adopted by George Soros to refer to reverse causal loop between share prices (i.e. present value of fundamentals) and business fundamentals. In a way, the share price is a "model" of the fundamental business processes. Usually, people assume that causality is one-way, from fundamentals to share price.

  2. Performativity - As used by Donald MacKenzie (e.g. here), many economic models are not "cameras" -- taking pictures of economic reality -- but in fact are "engines" -- an integral part of the construction of economic reality. He has a book of that title: An Engine, Not a Camera.

  3. Self-fulfilling Prophecy - a prediction that directly or indirectly causes itself to become true, by the very terms of the prophecy itself, due to positive feedback between belief and behavior. This is the broadest term, and least specific to the situation you describe.

Of the three terms, I suggest that MacKenzie's "performativity" is the best fit to your situation. He claims, among other things, that the validity of the economic models (e.g. Black-Scholes option pricing) has been improved by its very use by market participants, and therefore how it reflects in options pricing and trading patterns.

| improve this answer | | | | |
$\endgroup$
  • $\begingroup$ All useful terms, but I think agree that performativity is the most applicable, at least in terms of trying to express the concept in writing. Obviously, there is a fair bit of heterogeneity of terminology across fields, though. So, I'm not sure that any of these are really going to help me with any literature research. But that's probably the term that I'm going to try to use. $\endgroup$ – jsmith54 Apr 3 '15 at 2:52
3
$\begingroup$

Though it is not specifically a term, focused on machine learning, but I would refer to such behavior of a statistical model, using a general term side effect (while adding some clarifying adjectives, such as expected or unexpected, desired or undesired, and similar). Modeling outcome or transitive feedback loop outcome might be some of the alternative terms.

| improve this answer | | | | |
$\endgroup$
  • $\begingroup$ @jsmith54: My pleasure. $\endgroup$ – Aleksandr Blekh Apr 3 '15 at 2:57
0
$\begingroup$

I think in the literature the overarching term often used is active learning Active Learning ...in particular Multi-Armed Contextual Bandit

| improve this answer | | | | |
$\endgroup$
  • $\begingroup$ While active learning (AL) seems like a close enough concept, I think that there is a significant difference between AL and situation, described by the OP - the former implies intentional supervising efforts, while the latter implies the lack of such efforts and, even, intentions. Therefore, my preference is either side effect, or feedback loop terms. $\endgroup$ – Aleksandr Blekh Apr 5 '15 at 8:45
  • $\begingroup$ I'd say that active learning, while definitely related, tends to refer to a fairly specific technique intentionally deployed, rather than the unintentional consequence of an arbitrary model type acting on the concept. It strikes me as similar to reinforcement learning. These techniques are likely useful, once you've realized you have a feedback loop in your system, but I agree with Aleksandr that they express a design intention I was implying does not exist in my example. $\endgroup$ – jsmith54 Apr 5 '15 at 18:53
0
$\begingroup$

"I now declare you man and wife", an example of utterance by means of which the speaker performs a particular act, is an instance of a performative act, which changes the thing being modeled and thus changes the concept. Reflexive, iterative, recursive, retroactive, self-generating, are related terms.

| improve this answer | | | | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.