Questions tagged [influence-function]

a measure of how strongly the model parameters or predictions depend on a training instance.

An influence function tells you the effect of a change in one observation on an estimator. It's' useful in studying model robustness and calculating variance-covariance matrices for certain types of estimators, especially when more straightforward methods become hard to implement.

Influence functions are basically an analytical tool that can be used to assess the effect (or "influence") of removing an observation on the value of a statistic without having to re-calculate that statistic. They can also be used to create asymptotic variance estimates.

The influence function could be very useful to understand and debug deep learning models.

9 questions
7
votes
0 answers

Is there a measure of the robustness of a statistic?

I got a question today when talking about mean and median, IQR and variance. Is there a numerical measure of the robustness of a statistic? I must confess that I had never thought about that before, so the question took me aback. I have not found…
3
votes
1 answer

How does an influence function-based estimator estimate a target functional for an unknown distribution?

How exactly does a "1-step" influence function-based estimator estimate a target functional (like average treatment effect) for an unknown distribution? As described in Aaron Fisher and Edward H. Kennedy's tutorial (2019), the aim is to find an…
RobertF
  • 4,380
  • 6
  • 29
  • 46
2
votes
1 answer

Influence function used in partykit for binary classification

What is the influence function used for binary classification in the R package partkit, specifically for the conditional tree (ctree). I could not find any details in the R package documentation. In the vigente I found this paragraph about…
Kozolovska
  • 1,027
  • 6
  • 11
2
votes
1 answer

How is the asymptotic justification of the "linearization by influence function method" for surveys established?

The survey R package recently adopted the "linearization by influence function" method of estimating covariances between domain estimates. The central paper justifying this method is Deville (1999). I'm trying to understand the main asymptotic claim…
2
votes
0 answers

How to calculate the Influence function for the half variance

So we know that the influence function $IF$ for a functional $v$ at a point $y$ is roughly defined as: $$IF(v,F,y)=lim_{e\rightarrow 0} \frac{v(Y,G_y)-v(Y,F_y)}{e}$$ where $$G_y(y)=1(Y>y)*e + (1-e)F_y(y)$$ and $F_y$ is the CDF of y. The question is,…
Fcold
  • 644
  • 2
  • 10
2
votes
0 answers

How to find influence function of $\lambda=\log(\mu)$ such that $\mu=E(X)$?

The original question is that $X$ is a random variable that $E(X)=\mu$. We are interested in statistical functionals $\theta=\int\log(x)dF(x)$ and $ \lambda=\log(\mu)$. The first part of the question asks us to derive the influence function for…
JoZ
  • 579
  • 3
  • 12
0
votes
0 answers

Optimal new feature selection that reduces uncertainty on a fit

I have a dataset that contains a couple of hundred results from a computational task related to plasma physics. The computation time of the task is, due to the high complexity of the underlying physics, in the order of weeks. I am tasked to come up…
0
votes
0 answers

Recentered Influence Function

I am trying to build upon a paper which transforms the dependent variable into a recentered influence function by quantile and quantile ratios. However, I notice that when I do this in stata using rifvar, the vast majority of the resulting…
NDG
  • 61
  • 4
0
votes
0 answers

Robustness of latent variable models

Influence functions are a tool to study robustness. They tell us the effect of perturbing one datapoint on the trained parameters. E.g. by taking $x_i \mapsto (1+\epsilon)x_i$. How can this be used on latent variable models? If a model pairs each…
900edges
  • 339
  • 1
  • 9