more work on abstract

Author: pat-alt

dev/submissions/aies/extended_abstract/extended_abstract.pdf

@@ -21,28 +21,32 @@ The remainder of this extended abstract is structured as follows: @sec-main pres
 
 Existing work on CE and AR has largely been limited to the static setting: given some classifier $M: \mathcal{X} \mapsto \mathcal{Y}$ we are interested in finding close [@wachter2017counterfactual], actionable [@ustun2019actionable], plausible [@joshi2019towards, @antoran2020getting, @schut2021generating], sparse [@schut2021generating], diverse [@mothilal2020explaining] and ideally causally founded counterfactual explanations [@karimi2021algorithmic] for some individual $x$. The ability of counterfactual explanations to handle dynamics like data and model shifts remains a largely unexplored research challenge at this point [@verma2020counterfactual]. Only one recent work considers the implications of **exogenous** domain shifts on the validity of recourse [@upadhyay2021towards]. The authors propose a simple minimax objective, that minimizes the counterfactual loss function for a maximal domain and model shift. They show that their approach yields more robust counterfactuals than existing approaches. 
 
-This project investigates **endogenous** domain and model shifts, that is shifts that occur when AR is actually impemented by a proportion of individuals and the classifier is updated in response. @fig-dynamics illustrates this idea for a binary problem involving a probabilistic classifier and a greedy counterfactual generator proposed by @schut2021generating: AR leads to a domain shift, which in turn causes a drastic model shift. As this game of implementing AR and updating the classifier is repeated, individuals who receive and implement algorithmic recourse end up forming a distinct subgroup inside the target class, which may leave them vulnerable to discrimination. Through future experiments we want to investigate if this phenomenon is robust across different benchmark datasets and counterfactual generators. 
+This project investigates **endogenous** domain and model shifts, that is shifts that occur when AR is actually impemented by a proportion of individuals and the classifier is updated in response. @fig-dynamics illustrates this idea for a binary problem involving a probabilistic classifier and a greedy counterfactual generator proposed by @schut2021generating: AR leads to a domain shift, which in turn causes a drastic model shift. As this game of implementing AR and updating the classifier is repeated, the decision boundary moves in the opposite direction of the original training samples in the target class. We consider several aspects of these dynamics as problematic. Firstly, as the decision boundary moves in the direction of the non-target class, counterfactual paths become shorter: in the loan example, individuals that previously would have been denied credit based on their input features are suddenly considered as credit worthy. Average default risk across all borrowers can therefore be expected to increase. Conversely, lenders that anticipate or detect these sorts of dynamics may choose to still deny credit to individuals that have implemented AR, thereby comprimising the validity of AR. 
+
+To the best of my knowledge this is the first work investigating endogenous dynamic in algorithmic recourse. In future experiments I want to investigate how this phenomenon plays out across different benchmark datasets (including German credit [@germancredit1994], Boston Housing [@boston1993] and COMPAS [@compas2016]). Furthermore, I want to asses to what extent the magnitude and direction of domain and model shifts depends on the choice of the counterfactual generator. To this end, I am currently supervising a group of undergraduate students, who are tackling some of these tasks in their final-year research project. 
 
 ![PLACEHOLDER: The dynamics of algorithmic recourse.](www/dynamics.png){#fig-dynamics fig.pos="h" width=250px}
 
 ## Related and Future Work {#sec-related}
 
 ### Benchmarking CE in Julia
 
-Alongside my research I have developed open-source implementations related to explainable AI. [CounterfactualExplanations.jl](https://www.paltmeyer.com/CounterfactualExplanations.jl/stable/) is a Julia package that can be used to generate counterfactual explanations for models developed and trained not only in Julia, but also in other popular programming languages like Python and R. I have recently submitted the package along with a companion paper as a proposal for a main talk at [JuliaCon](https://juliacon.org/2022/). [BayesLaplace.jl](https://www.paltmeyer.com/BayesLaplace.jl/dev/) is a small Julia package that can be used to recover Bayesian representations of deep neural networks through Laplace approximation in a post-hoc manner. It is inspired by a recent paper [@daxberger2021laplace] and has also been submitted to JuliaCon. Finally, [deepvars](https://github.com/pat-alt/deepvars) is an R package that implements an approach towards vector autoregression that leverages deep learning. This was originally my master's thesis and later presented at the NeurIPS 2021 MLECON workshop. I have also published several blog posts on explainable AI and probabilisitic ML in an effort to make my research accessible to a broad audience. 
+Until recently there existed only one open-source library that provides a unifying approach to generate and benchmark counterfactual explanations for models built and trained in Python [@pawelczyk2021carla]. To address this limitation I have developed [CounterfactualExplanations.jl](https://www.paltmeyer.com/CounterfactualExplanations.jl/stable/): a Julia package that can be used to generate counterfactual explanations for models developed and trained not only in Julia, but also in other popular programming languages like Python and R. The package and companion paper are currently pending acceptance for a main talk at [JuliaCon](https://juliacon.org/2022/). 
+
+### Probabilistic Methods for Realistic CE
 
-### Probabilistic methods for realistic counterfactual explanations
+To ensure that the generated explanations are realistic it is important to understand which input-output pairs are likely and which are not. To quantify their joint likelihood, previous work has either relied on generative models or restricted the analysis to probabilistic models that incorporate uncertainty in their predictions. While the former approach is more versatile since it is applicable to both deterministic and probabilistic models, the latter is computationally much more efficient. The approach proposed by @schut2021generating and used to generate the examples in @fig-dynamics falls into the latter category. In future work I want to explore how recent advances in post-hoc uncertainty quantification, most notably Laplace Redux [@daxberger2021laplace], can be leveraged to generate realistic and unambiguous counterfactual explanations for any model.^[For some initial work on this see my Julia implementation of Laplace Redux: [BayesLaplace.jl](https://www.paltmeyer.com/BayesLaplace.jl/dev/).]
 
-Probabilistic machine learning can be leveraged in this context and more generally facilitates inference and interpretability. It is also closely related to Bayesian statistics, which has played an important role in both finance and economics for many years. 
+### XAI for Time Series Models
 
-To ensure that the generated explanations are realistic it is important to understand which input-output pairs are likely and which are not. To quantify their joint likelihood, previous work has either relied on generative models or restricted the analysis to probabilistic models that incorporate uncertainty in their predictions. While the former approach is more versatile since it is applicable to both deterministic and probabilistic models, the latter is computationally much more efficient. In my work I want to explore how recent advances in post-hoc uncertainty quantification can be leveraged to generate realistic and unambiguous counterfactual explanations for any model.
+Datasets in Finance and Economics often involve autoregressive time series, which is modelled differently than static, cross-sectional data. One popular approach to macroeconomic forecasting, for example, is Vector Autoregression (VAR), which involves an interpretable system of linear equations. Recent work has shown that machine learning can be leveraged in this context to improve forecasting performance, albeit at the sacrifice of explainability [@altmeyer2021deep].^[For an open-source implementation of our approach see [deepvars](https://github.com/pat-alt/deepvars): an R package Deep Vector Autoregression]. Through my PhD research I hope to ultimately adress that shortfall through explainable AI and probabilistic methods.
 
-### CE for time series
+#### Counterfactual Explanations
 
-Data sets in finance and economics typically involve time series data. Therefore, I am naturally interested in the application of explainable AI to sequential data, an area which has so far not been explored extensively. In the future, I want to work on counterfactual explanations for time series models. I am also interested in seeing if and how Laplace approximation can be used for Bayesian deep learning with time series data. I hope that the findings from both of these projects can ultimately be used to build complex but interpretable time series models for classification and forecasting in finance and economics. 
+I would like to investigate if and how CE can be used to leverage machine learning for time series forecasting without sacrificing explainability. In the context of conventional VAR counterfactual outcomes are actually well-studied [@kilian2017structural], which makes macroeconomic forecasting and interesting case study. More generally though, work on applying CE to black-box models for time series data has so far been limited and therefore warrants attention. 
 
-### Explainable black-box models for time series
+#### Probabilistic Methods
 
-For example, I would like to leverage effortless Bayesian deep learning to make our proposed Deep Vector Autoregression model explainable. 
+I am also interested in seeing if and how Laplace Redux can be applied to reccurent neural networks. 
 
 ## References