22455 research outputs found
Sort by
Contributions to modeling with set-valued data: benefitting from undecided respondents
Get PDFThis dissertation develops a methodological framework and approaches to benefit from undecided survey participants, particularly undecided voters in pre-election polls. As choices can be seen as processes that - in stages - exclude alternatives until arriving at one final element, we argue that in pre-election polls undecided participants can most suitably be represented by the set of their viable options. This consideration set sampling, in contrast to the conventional neglection of the undecided, could reduce nonresponse and collects new and valuable information. We embed the resulting set-valued data in the framework of random sets, which allows for two different interpretations, and develop modeling methods for either one.
The first interpretation is called ontic and views the set of options as an entity of its own that most accurately represents the position at the time of the poll, thus as a precise representation of something naturally imprecise. With this, new ways of structural analysis emerge as individuals pondering between particular parties can now be examined. We show how the underlying categorical data structure can be preserved in this formalization process for specific models and how popular methods for categorical data analysis can be broadly transferred.
As the set contains the eventual choice, under the second interpretation, the set is seen as a coarse version of an underlying truth, which is called the epistemic view. This imprecise information of something actually precise can then be used to improve predictions or election forecasting. We developed several approaches and a framework of a factorized likelihood to utilize the set-valued information for forecasting. Amongst others, we developed methods addressing the complex uncertainty induced by the undecided, weighting the justifiability of assumptions with the conciseness of the results.
To evaluate and apply our approaches, we conducted a pre-election poll for the German federal election of 2021 in cooperation with the polling institute Civey, for the first time regarding undecided voters in a set-valued manner. This provides us with the unique opportunity to demonstrate the advantages of the new approaches based on a state-of-the-art survey
If interpretability is the answer, what is the question?
Get PDFDue to the ability to model even complex dependencies, machine learning (ML) can be used to tackle a broad range of (high-stakes) prediction problems. The complexity of the resulting models comes at the cost of transparency, meaning that it is difficult to understand the model by inspecting its parameters.
This opacity is considered problematic since it hampers the transfer of knowledge from the model, undermines the agency of individuals affected by algorithmic decisions, and makes it more challenging to expose non-robust or unethical behaviour.
To tackle the opacity of ML models, the field of interpretable machine learning (IML) has emerged. The field is motivated by the idea that if we could understand the model's behaviour -- either by making the model itself interpretable or by inspecting post-hoc explanations -- we could also expose unethical and non-robust behaviour, learn about the data generating process, and restore the agency of affected individuals. IML is not only a highly active area of research, but the developed techniques are also widely applied in both industry and the sciences.
Despite the popularity of IML, the field faces fundamental criticism, questioning whether IML actually helps in tackling the aforementioned problems of ML and even whether it should be a field of research in the first place:
First and foremost, IML is criticised for lacking a clear goal and, thus, a clear definition of what it means for a model to be interpretable. On a similar note, the meaning of existing methods is often unclear, and thus they may be misunderstood or even misused to hide unethical behaviour. Moreover, estimating conditional-sampling-based techniques poses a significant computational challenge.
With the contributions included in this thesis, we tackle these three challenges for IML.
We join a range of work by arguing that the field struggles to define and evaluate "interpretability" because incoherent interpretation goals are conflated. However, the different goals can be disentangled such that coherent requirements can inform the derivation of the respective target estimands. We demonstrate this with the examples of two interpretation contexts: recourse and scientific inference.
To tackle the misinterpretation of IML methods, we suggest deriving formal interpretation rules that link explanations to aspects of the model and data. In our work, we specifically focus on interpreting feature importance. Furthermore, we collect interpretation pitfalls and communicate them to a broader audience.
To efficiently estimate conditional-sampling-based interpretation techniques, we propose two methods that leverage the dependence structure in the data to simplify the estimation problems for Conditional Feature Importance (CFI) and SAGE.
A causal perspective proved to be vital in tackling the challenges: First, since IML problems such as algorithmic recourse are inherently causal; Second, since causality helps to disentangle the different aspects of model and data and, therefore, to distinguish the insights that different methods provide; And third, algorithms developed for causal structure learning can be leveraged for the efficient estimation of conditional-sampling based IML methods.Aufgrund der Fähigkeit, selbst komplexe Abhängigkeiten zu modellieren, kann maschinelles Lernen (ML) zur Lösung eines breiten Spektrums von anspruchsvollen Vorhersageproblemen eingesetzt werden.
Die Komplexität der resultierenden Modelle geht auf Kosten der Interpretierbarkeit, d. h. es ist schwierig, das Modell durch die Untersuchung seiner Parameter zu verstehen.
Diese Undurchsichtigkeit wird als problematisch angesehen, da sie den Wissenstransfer aus dem Modell behindert, sie die Handlungsfähigkeit von Personen, die von algorithmischen Entscheidungen betroffen sind, untergräbt und sie es schwieriger macht, nicht robustes oder unethisches Verhalten aufzudecken.
Um die Undurchsichtigkeit von ML-Modellen anzugehen, hat sich das Feld des interpretierbaren maschinellen Lernens (IML) entwickelt.
Dieses Feld ist von der Idee motiviert, dass wir, wenn wir das Verhalten des Modells verstehen könnten - entweder indem wir das Modell selbst interpretierbar machen oder anhand von post-hoc Erklärungen - auch unethisches und nicht robustes Verhalten aufdecken, über den datengenerierenden Prozess lernen und die Handlungsfähigkeit betroffener Personen wiederherstellen könnten.
IML ist nicht nur ein sehr aktiver Forschungsbereich, sondern die entwickelten Techniken werden auch weitgehend in der Industrie und den Wissenschaften angewendet.
Trotz der Popularität von IML ist das Feld mit fundamentaler Kritik konfrontiert, die in Frage stellt, ob IML tatsächlich dabei hilft, die oben genannten Probleme von ML anzugehen, und ob es überhaupt ein Forschungsgebiet sein sollte:
In erster Linie wird an IML kritisiert, dass es an einem klaren Ziel und damit an einer klaren Definition dessen fehlt, was es für ein Modell bedeutet, interpretierbar zu sein. Weiterhin ist die Bedeutung bestehender Methoden oft unklar, so dass sie missverstanden oder sogar missbraucht werden können, um unethisches Verhalten zu verbergen. Letztlich stellt die Schätzung von auf bedingten Stichproben basierenden Verfahren eine erhebliche rechnerische Herausforderung dar.
In dieser Arbeit befassen wir uns mit diesen drei grundlegenden Herausforderungen von IML.
Wir schließen uns der Argumentation an, dass es schwierig ist, "Interpretierbarkeit" zu definieren und zu bewerten, weil inkohärente Interpretationsziele miteinander vermengt werden. Die verschiedenen Ziele lassen sich jedoch entflechten, sodass kohärente Anforderungen die Ableitung der jeweiligen Zielgrößen informieren. Wir demonstrieren dies am Beispiel von zwei Interpretationskontexten: algorithmischer Regress
und wissenschaftliche Inferenz.
Um der Fehlinterpretation von IML-Methoden zu begegnen, schlagen wir vor, formale Interpretationsregeln abzuleiten, die Erklärungen mit Aspekten des Modells und der Daten verknüpfen. In unserer Arbeit konzentrieren wir uns speziell auf die Interpretation von sogenannten Feature Importance Methoden. Darüber hinaus tragen wir wichtige Interpretationsfallen zusammen und kommunizieren sie an ein breiteres Publikum.
Zur effizienten Schätzung auf bedingten Stichproben basierender Interpretationstechniken schlagen wir zwei Methoden vor, die die Abhängigkeitsstruktur in den Daten nutzen, um die Schätzprobleme für Conditional Feature Importance (CFI) und SAGE zu vereinfachen.
Eine kausale Perspektive erwies sich als entscheidend für die Bewältigung der Herausforderungen: Erstens, weil IML-Probleme wie der algorithmische Regress inhärent kausal sind; zweitens, weil Kausalität hilft, die verschiedenen Aspekte von Modell und Daten zu entflechten und somit die Erkenntnisse, die verschiedene Methoden liefern, zu unterscheiden; und drittens können wir Algorithmen, die für das Lernen kausaler Struktur entwickelt wurden, für die effiziente Schätzung von auf bindingten Verteilungen basierenden IML-Methoden verwenden
Navigating pyoverdine and beyond: the role of tripartite efflux pumps in Pseudomonas putida KT2440
Get PDFEinfluss der Patientenlagerung auf den intraoperativen Blutverlust in der orthognathen Chirurgie - eine randomisierte, kontrollierte Studie
Get PDF
Language attitudes and social identity
Get PDFThis study investigates two constructs of fundamental importance in sociolinguistics and social psychology i.e. language attitudes and social identity. They are investigated through a mixed-method study conducted with Russian-Speaking immigrants in Israel and Germany. Migration provides an especially fitting context for the study of attitudes towards language and of the positioning of the self in society in that it involves the crossing of different kinds of boundaries: geographic, linguistic, cultural, social, emotional etc. Specifically, post-Soviet migration to Israel and Germany poses a compelling case-study of identity negotiation processes: being highly regulated by the two states, immigration policy largely relies on the same categories of ethnonational identity forcibly attributed to the so-called nationalities in the former Soviet Union, amongst which Soviet Germans and Soviet Jews.
The qualitative data corpus consists of 56 sociolinguistic interviews with young-adult immigrants conducted in Israel and in Germany. The informants are first to 1.5 generation immigrants from countries of the former Soviet Union to Israel and Germany. Data for quantitative analysis stem from an online survey carried out with 761 Russian-speaking participants from countries of the former Soviet Union based in Israel and Germany.
Through the lens of language attitudes emerges a complex picture of post-Soviet migration to Israel and Germany. While Russian-speaking immigrants to Germany report a relatively broad access to German language programs and a secure command of German, their degree of identification with the core values of German society is rather low. On the other hand, Russian-speaking immigrants to Israel report issues at learning Hebrew and a low level of interactions in Hebrew, which can be put in relation with segregation issues and ideological tensions in Israeli society. At the same time, Russian-speaking immigrants to Israel can be observed to identify with Israeli culture, thus overtly or covertly reappropriating their contested Jewish identity.
This study advances a conceptual model of the interactions between language, migration and identity.
From the analysis emerge significant theoretical and methodological implications for future sociolinguistic research on language attitudes. While they are often regarded as correlates to language competence, this study illustrates that language attitudes allow a grounded analysis of the way in which speakers situate themselves and others in society. Moreover, this study shows that a triangulation of interdisciplinary - sociolinguistic, psychological, sociological and other - methods is necessary to attain the due degree of analytical depth when dealing with language attitudes and identity-related phenomena
