<div>The widespread and growing usage of machine learning models, especially in highly critical areas such as law, predicate the need for interpretable models. Models that cannot be audited are vulnerable to inheriting biases from the dataset. Even locally interpretable models are vulnerable to adversarial attack. To address this issue a new methodology is proposed to translate any existing machine learning model into a globally interpretable one.</div><div>This methodology, MTRE-PAN, is designed as a hybrid SVM-decision tree model and leverages the interpretability of linear hyperplanes. MTRE-PAN uses this hybrid model to create polygons that act as intermediates for the decision boundary. MTRE-PAN is compared to a previously proposed model, TRE-PAN, on three non-synthetic datasets: Abalone, Census and Diabetes data. TRE-PAN translates a machine learning model to a 2-3 decision tree in</div><div>order to provide global interpretability for the target model. The datasets are each used to train a Neural Network that represents the non-interpretable model. For all target models, the results show that MTRE-PAN generates interpretable decision trees that have a lower</div><div>number of leaves and higher parity compared to TRE-PAN.</div>

No description available.

[pt] As árvores de decisão são estruturas comumente utilizadas em cenários nos quais modelos explicáveis de Aprendizado de Máquina são desejados, por serem visualmente intuitivas. Na literatura existente, a busca por explicabilidade em árvores envolve a minimização de métricas como altura e número de nós. Nesse contexto, definimos uma métrica de explicabilidade, chamada de explanation size, que reflete o número de atributos necessários para explicar a classificação dos exemplos. Apresentamos também um algoritmo, intitulado SER-DT, que obtém uma aproximação O(log n) (ótima se P diferente NP) para a minimização da altura no pior caso ou caso médio, assim como do explanation size no pior caso ou caso médio. Em uma série de experimentos, comparamos a implementação de SER-DT com algoritmos conhecidos da área, como CART e EC2, além de testarmos o impacto de parâmetros e estratégias de poda nesses algoritmos. SER-DT mostrou-se competitivo em acurácia com os algoritmos citados, mas gerou árvores muito mais explicáveis. / [en] Decision trees are commonly used structures in scenarios where explainable Machine Learning models are desired, as they are visually intuitive. In the existing literature, the search for explainability in trees involves minimizing metrics such as depth and number of nodes. In this context, we define an explainability metric, called explanation size, which reflects the number of attributes needed to explain the classification of examples. We also present an algorithm, called SER-DT, which obtains an O(log n) approximation (optimal if P different NP) for the minimization of depth in the worst/average case, as well as of explanation size in the worst/average case. In a series of experiments, we compared the SER-DT implementation with well-known algorithms in the field, such as CART and EC2 in addition to testing the impact of parameters and pruning strategies on these algorithms. SER-DT proved to be competitive in terms of accuracy with the aforementioned algorithms, but generated much more explainable trees.

The most performant Machine Learning (ML) classifiers have been labeled black-boxes due to the complexity of their decision process. eXplainable Artificial Intelligence (XAI) methods aim to alleviate this issue by crafting an interpretable explanation for a models prediction. A drawback of most XAI methods is that they are heuristic with some drawbacks such as non determinism and locality. Formal Explanations (FE) have been proposed as a way to explain the decisions of classifiers by extracting a set of features that guarantee the prediction. In this thesis we explore these guarantees for different use cases: speeding up the inference speed of tree-based Machine Learning classifiers, curriculum learning using said classifiers and also reducing training data. We find that under the right circumstances we can achieve up to 6x speedup by partially compiling the model to a set of rules that are extracted using formal explainability methods. / De mest effektiva maskininlärningsklassificerarna har betecknats som svarta lådor på grund av komplexiteten i deras beslutsprocess. Metoder för förklarbar artificiell intelligens (XAI) syftar till att lindra detta problem genom att skapa en tolkbar förklaring för modellens prediktioner. En nackdel med de flesta XAI-metoder är att de är heuristiska och har vissa nackdelar såsom icke-determinism och lokalitet. Formella förklaringar (FE) har föreslagits som ett sätt att förklara klassificerarnas beslut genom att extrahera en uppsättning funktioner som garanterar prediktionen. I denna avhandling utforskar vi dessa garantier för olika användningsfall: att öka inferenshastigheten för maskininlärningsklassificerare baserade på träd, kurser med hjälp av dessa klassificerare och även minska träningsdata. Vi finner att under rätt omständigheter kan vi uppnå upp till 6 gånger snabbare prestanda genom att delvis kompilera modellen till en uppsättning regler som extraheras med hjälp av formella förklaringsmetoder.

With the fast development of radio technologies, wireless systems have become more convoluted. This complexity, accompanied by an increase of the number of connections, is translated into a need for more parameters to analyse and decisions to take at each instant. AI comes into play by automating these processes, particularly with Deep Learning techniques, that often show the best accuracy. However, the high performance by these methods also comes with the drawback of behaving like a black box from the view of a human. To this end, eXplainable AI serves as a technique to better understand the decision process of these algorithms. This thesis proposes an eXplainable AI framework to be used on Reinforcement Learning agents, particularly within the use case of antenna resource adaptation for network energy reduction. The framework puts a special emphasis on model adaptation/reduction, therefore focusing on feature importance techniques. The proposed framework presents a pre-model block using Concrete Autoencoders for feature reduction and a post-model block using self-supervised learning to estimate feature importance. Both of these can be used alone or in combination with DeepSHAP, in order to mitigate some of this popular method’s drawbacks. The explanations provided by the pipeline prove useful in order to reduce model complexity without loss of accuracy and to understand the usage of the input features by the AI model. / Med den snabba utvecklingen av radioteknologier har trådlösa system blivit alltmer invecklade. Denna komplexitet, kombinerat med en ökning av antalet anslutningar, innebär att fler parametrar behöver analyseras, och fler beslut behöver fattas vid varje ögonblick. AI kommer in i bilden genom att automatisera dessa processer, särskilt med Deep Learning-tekniker, som ofta uppvisar bäst noggrannhet. Men den höga prestandan med dessa metoder kommer också med nackdelen att tekniken beter sig som en svart låda från en människas synvinkel. Förklarlig AI fungerar därför som en teknik för att bättre förstå beslutet som fattas av dessa algoritmer. Denna avhandling föreslår ett förklarligt AI-ramverk som ska användas inom förstärkningsinlärning, särskilt inom användningsfallet med antenn-resursanpassning för energireduktion i trådlösa nätverk. Det föreslagna ramverket sätter en särskild tonvikt på modellanpassning/modellreduktion. Ramverket innehåller ett förmodellblock som använder Concrete Autoencoders för Feature Reduction och ett post-modellblock som använder självövervakad inlärning för att uppskatta Feature Importance. Båda dessa kan användas ensamt eller i kombination med DeepSHAP, för att lindra några av denna populära metods nackdelar. Feature Importance-uppskattningarna från ramverket visar sig vara användbara för att minska modellkomplexitet utan förlust av noggrannhet och för att förstå användningen av Input Features av AI-modellen.

Controlling antennas’ vertical tilt through Remote Electrical Tilt (RET) is an effective method to optimize network performance. Reinforcement Learning (RL) algorithms such as Deep Reinforcement Learning (DRL) have been shown to be successful for RET optimization. One issue with DRL is that DRL models have a black box nature where it is difficult to ’explain’ the decisions made in a human-understandable way. Explanations of a model’s decisions are beneficial for a user not only to understand but also to intervene and modify the RL model. In this work, a state-ofthe-art Explainable Reinforcement Learning (XRL) method is evaluated on the RET optimization problem. More specifically, the chosen XRL method is the Embedded Self-Prediction (ESP) model proposed by Lin, Lam, and Fern [16] which can generate contrastive explanations in terms of why an action is preferred over the other. The ESP model was evaluated on two different RET optimization scenarios. The first scenario is formulated as a single agent RL problem in a ’simple’ environment whereas the second scenario is formulated as a multi agent RL problem with a more complex environment. In both scenarios, the results show little to no difference in performance compared to a baseline Deep Q-Network (DQN) algorithm. Finally, the explanations of the model were validated by comparing them to action outcomes. The conclusions of this work is that the ESP model offers explanations of its behaviour with no performance decrease compared to a baseline DQN and the generated explanations offer value in debugging and understanding the given problem. / Att styra antenners vertikala lutning genom RET är en effektiv metod för att optimera nätverksprestanda. RL-algoritmer som DRL har visat sig vara framgångsrika för REToptimering. Ett problem med DRL är att DRL-modeller är som en svart låda där det är svårt att ’förklara’ de beslut som fattas på ett sätt som är begripligt för människor. Förklaringar av en modells beslut är fördelaktiga för en användare inte bara för att förstå utan också för att ingripa och modifiera RL-modellen. I detta arbete utvärderas en toppmodern XRL-metod på RET-optimeringsproblemet. Mer specifikt är den valda XRL-metoden ESP-modellen som föreslagits av Lin, Lam och Fern [16] som kan generera kontrastiva förklaringar i termer av varför en handling föredras framför den andra. ESP-modellen utvärderades på två olika RET-optimeringsscenarier. Det första scenariot är formulerat som ett problem med en enstaka agent i en ’enkel’ miljö medan det andra scenariot är formulerat som ett problem med flera agenter i en mer komplex miljö. I båda scenarierna visar resultaten liten eller ingen skillnad i prestanda jämfört med en DQN-algoritm. Slutligen validerades modellens förklaringar genom att jämföra dem med handlingsresultat. Slutsatserna av detta arbete är att ESPmodellen erbjuder förklaringar av dess beteende utan prestandaminskning jämfört med en DQN och de genererade förklaringarna ger värde för att felsöka och förstå det givna problemet.

The increasing complexity of Artificial Intelligence (AI) models is accompanied by an increase in difficulty in interpreting model predictions. This thesis work provides insights and understanding of the differences and similarities between explainability methods for AI models. Opening up black-box models is important, especially if AI is applied in sensitive domains such as to, e.g., aid medical professionals. In recent years, the use of Transformer-based artificial neural network architectures such as Bidirectional Encoder Representations from Transformers (BERT) has become common in the field of Natural Language Processing (NLP), showing human-level performance on tasks such as sentiment classification and question answering. In addition, a growing portion of research within eXplainable AI (XAI) has shown success in using explainability methods to output auxiliary explanations at inference time together with predictions made by these complex models. When scoping the different methods, there is a distinction to be made whether the explanations emerge as part of the prediction process or subsequently via a separate model. These two categories of explainability methods are referred to as self-explaining and post-hoc, respectively. The goal of this work is to evaluate, analyze and compare these two categories of methods for assisting BERT models with explanations in the context of sentiment classification. A comparative analysis was therefore conducted in order to investigate quantitative and qualitative differences. To measure the quality of explanations, the Intersection Over Union (IOU) and Precision-Recall Area Under the Curve (PR-AUC) scores were used together with Explainable NLP (ExNLP) datasets, containing human annotated explanations. Apart from discussing benefits, drawbacks and assumptions of the different methods, results of the work indicated that the self-explaining method proved more successful in some instances while the post-hoc method performed better in others. Given the subjective nature of explanation quality, however, this work should be extended in several proposed directions, in order to fully capture the nuances of the explainability methods. / Parallellt med den ökande komplexiteten hos modeller med artificiell intelligens (AI) följer en ökad svårighet att tolka förutsägelser som modellerna gör. Detta examensarbete fokuserar på skillnader och likheter mellan förklaringsmetoder för AI-modeller. Att skapa mer transparens kring modellerna är viktigt, speciellt om AI ska appliceras i känsliga områden som t.ex. inom hälso- och sjukvård. Under de senaste åren har användningen av Transformer-baserade artificiella neurala nätverk som Bidirectional Encoder Representations from Transformers (BERT) blivit vanligt inom Natural Language Processing (NLP). Resultaten som modellerna når på uppgifter såsom sentimentklassificering och svar på frågor är på en mänsklig nivå. En växande del av forskningen inom eXplainable AI (XAI) har dessutom kunnat visa stora framsteg inom användandet av förklaringsmetoder, för att bistå förutsägelserna som dessa komplexa modeller gör med förklaringar. I kategoriseringar av metoderna särskiljs det ofta mellan huruvida förklaringarna uppstår som en del av förutsägelsen, tillsammans med modellen eller om de skapas efteråt via en separat modell. Dessa två kategorier av förklaringsmetoder kallas självförklarande och post-hoc. Målet med detta arbete är att utvärdera, analysera och jämföra dessa två kategorier av metoder som används för att hjälpa BERT-modeller med förklaringar i samband med sentimentklassificering av text. En jämförande analys genomfördes därför för att undersöka kvantitativa och kvalitativa skillnader. För att mäta kvaliteten på förklaringar användes Intersection Over Union (IOU) och Precision-Recall Area Under the Curve (PR-AUC) tillsammans med dataset skräddarsydda för just Explainable NLP (ExNLP) innehållande mänskligt annoterade förklaringar. Förutom att diskutera fördelar, nackdelar och antaganden med de olika metoderna, pekade resultaten på att den självförklarande metoden presterade bättre i vissa fall medan post-hoc-metoden presterade bättre i andra. Med tanke på hur kvaliteten av förklaringar till stor del handlar om en subjektiv bedömning bör dock detta arbete utvidgas i flera riktningar – föreslagna i detta arbete – för att fånga alla nyanser av förklaringsmetoderna.

The deployment of human-augmented machine learning (ML) systems has become a recommended organizational best practice. ML systems use algorithms that rely on training data labeled by human annotators. However, human involvement in reviewing and labeling consumers' voice data to train speech recognition systems for Amazon Alexa, Microsoft Cortana, and the like has raised privacy concerns among consumers and privacy advocates. We use the enhanced APCO model as the theoretical lens to investigate how the disclosure of human involvement during the supervised machine learning process affects consumers' privacy decision making. In a scenario-based experiment with 499 participants, we present various company privacy policies to participants to examine their trust and privacy considerations, then ask them to share reasons why they would or would not opt in to share their voice data to train a companies' voice recognition software. We find that the perception of human involvement in the ML training process significantly influences participants' privacy-related concerns, which thereby mediate their decisions to share their voice data. Furthermore, we manipulate four factors of a privacy policy to operationalize various cognitive biases actively present in the minds of consumers and find that default trust and salience biases significantly affect participants' privacy decision making. Our results provide a deeper contextualized understanding of privacy-related concerns that may arise in human-augmented ML system configurations and highlight the managerial importance of considering the role of human involvement in supervised machine learning settings. Importantly, we introduce perceived human involvement as a new construct to the information privacy discourse. Although ubiquitous data collection and increased privacy breaches have elevated the reported concerns of consumers, consumers' behaviors do not always match their stated privacy concerns. Researchers refer to this as the privacy paradox, and decades of information privacy research have identified a myriad of explanations why this paradox occurs. Yet the underlying crux of the explanations presumes privacy concern to be the appropriate proxy to measure privacy attitude and compare with actual privacy behavior. Often, privacy concerns are situational and can be elicited through the setup of boundary conditions and the framing of different privacy scenarios. Drawing on the cognitive model of empowerment and interest, we propose a multidimensional privacy interest construct that captures consumers' situational and dispositional attitudes toward privacy, which can serve as a more robust measure in conditions leading to the privacy paradox. We define privacy interest as a consumer's general feeling toward reengaging particular behaviors that increase their information privacy. This construct comprises four dimensions—impact, awareness, meaningfulness, and competence—and is conceptualized as a consumer's assessment of contextual factors affecting their privacy perceptions and their global predisposition to respond to those factors. Importantly, interest was originally included in the privacy calculus but is largely absent in privacy studies and theoretical conceptualizations. Following MacKenzie et al. (2011), we developed and empirically validated a privacy interest scale. This study contributes to privacy research and practice by reconceptualizing a construct in the original privacy calculus theory and offering a renewed theoretical lens through which to view consumers' privacy attitudes and behaviors. / Doctor of Philosophy / The deployment of human-augmented machine learning (ML) systems has become a recommended organizational best practice. ML systems use algorithms that rely on training data labeled by human annotators. However, human involvement in reviewing and labeling consumers' voice data to train speech recognition systems for Amazon Alexa, Microsoft Cortana, and the like has raised privacy concerns among consumers and privacy advocates. We investigate how the disclosure of human involvement during the supervised machine learning process affects consumers' privacy decision making and find that the perception of human involvement in the ML training process significantly influences participants' privacy-related concerns. This thereby influences their decisions to share their voice data. Our results highlight the importance of understanding consumers' willingness to contribute their data to generate complete and diverse data sets to help companies reduce algorithmic biases and systematic unfairness in the decisions and outputs rendered by ML systems. Although ubiquitous data collection and increased privacy breaches have elevated the reported concerns of consumers, consumers' behaviors do not always match their stated privacy concerns. This is referred to as the privacy paradox, and decades of information privacy research have identified a myriad of explanations why this paradox occurs. Yet the underlying crux of the explanations presumes privacy concern to be the appropriate proxy to measure privacy attitude and compare with actual privacy behavior. We propose privacy interest as an alternative to privacy concern and assert that it can serve as a more robust measure in conditions leading to the privacy paradox. We define privacy interest as a consumer's general feeling toward reengaging particular behaviors that increase their information privacy. We found that privacy interest was more effective than privacy concern in predicting consumers' mobilization behaviors, such as publicly complaining about privacy issues to companies and third-party organizations, requesting to remove their information from company databases, and reducing their self-disclosure behaviors. By contrast, privacy concern was more effective than privacy interest in predicting consumers' behaviors to misrepresent their identity. By developing and empirically validating the privacy interest scale, we offer interest in privacy as a renewed theoretical lens through which to view consumers' privacy attitudes and behaviors.

Artificial Intelligence (AI) has emerged as a transformative force in wireless communications, driving innovation to address the complex challenges faced by communication systems. In this context, the optimization of limited radio resources plays a crucial role, and one important aspect is the Modulation and Coding Scheme (MCS) selection. AI solutions for MCS selection have been predominantly characterized as black-box models, which suffer from limited explainability and consequently hinder trust in these algorithms. Moreover, the majority of existing research primarily emphasizes enhancing explainability without concurrently improving the model’s performance which makes performance and explainability a trade-off. This work aims to address these issues by employing eXplainable AI (XAI), particularly counterfactual and causal analysis, to increase the explainability and trustworthiness of black-box models. We propose CounterFactual Retrain (CF-Retrain), the first method that utilizes counterfactual explanations to improve model performance and make the process of performance enhancement more explainable. Additionally, we conduct a causal analysis and compare the results with those obtained from an analysis based on the SHapley Additive exPlanations (SHAP) value feature importance. This comparison leads to the proposal of novel hypotheses and insights for model optimization in future research. Our results show that employing CF-Retrain can reduce the Mean Absolute Error (MAE) of the black-box model by 4% while utilizing only 14% of the training data. Moreover, increasing the amount of training data yields even more pronounced improvements in MAE, providing a certain level of explainability. This performance enhancement is comparable to or even superior to using a more complex model. Furthermore, by introducing causal analysis to the mainstream SHAP value feature importance, we provide a novel hypothesis and explanation of feature importance based on causal analysis. This approach can serve as an evaluation criterion for assessing the model’s performance. / Artificiell intelligens (AI) har dykt upp som en transformativ kraft inom trådlös kommunikation, vilket driver innovation för att möta de komplexa utmaningar som kommunikationssystem står inför. I detta sammanhang spelar optimeringen av begränsade radioresurser en avgörande roll, och en viktig aspekt är valet av Modulation and Coding Scheme (MCS). AI-lösningar för val av modulering och kodningsschema har övervägande karaktäriserats som black-box-modeller, som lider av begränsad tolkningsbarhet och följaktligen hindrar förtroendet för dessa algoritmer. Dessutom betonar majoriteten av befintlig forskning i första hand att förbättra förklaringsbarheten utan att samtidigt förbättra modellens prestanda, vilket gör prestanda och tolkningsbarhet till en kompromiss. Detta arbete syftar till att ta itu med dessa problem genom att använda XAI, särskilt kontrafaktisk och kausal analys, för att öka tolkningsbarheten och pålitligheten hos svarta-box-modeller. Vi föreslår CF-Retrain, den första metoden som använder kontrafaktiska förklaringar för att förbättra modellens prestanda och göra processen med prestandaförbättring mer tolkningsbar. Dessutom gör vi en orsaksanalys och jämför resultaten med de som erhålls från en analys baserad på värdeegenskapens betydelse. Denna jämförelse leder till förslaget av nya hypoteser och insikter för modelloptimering i framtida forskning. Våra resultat visar att användning av CF-Retrain kan minska det genomsnittliga absoluta felet för black-box-modellen med 4% samtidigt som man använder endast 14% av träningsdata. Dessutom ger en ökning av mängden träningsdata ännu mer uttalade förbättringar av Mean Absolute Error (MAE), vilket ger en viss grad av tolkningsbarhet. Denna prestandaförbättring är jämförbar med eller till och med överlägsen att använda en mer komplex modell. Dessutom, genom att introducera kausal analys till de vanliga Shapley-tillsatsförklaringarna värdesätter egenskapens betydelse, ger vi en ny hypotes och tolkning av egenskapens betydelse baserad på kausalanalys. Detta tillvägagångssätt kan fungera som ett utvärderingskriterium för att bedöma modellens prestanda.

Deep Learning models are ubiquitous in machine learning. They offer state-of- the-art performance on tasks ranging from natural language processing to image classification. The drawback of these complex models is their black box nature. It is difficult for the end-user to understand how a model arrives at its prediction from the input. This is especially pertinent in domains such as medicine, where being able to trust a model is paramount. In this thesis, ways of explaining a model predicting paroxysmal atrial fibrillation from sinus electrocardiogram (ECG) data are explored. Building on the concept of feature attributions, the problem is approached from three distinct perspectives: time, scale, and frequency. Specifically, one method based on the Integrated Gradients framework and one method based on Shapley values are used. By perturbing the data, retraining the model, and evaluating the retrained model on the perturbed data, the degree of correspondence between the attributions and the meaningful information in the data is evaluated. Results indicate that the attributions in scale and frequency are somewhat consistent with the meaningful information in the data, while the attributions in time are not. The conclusion drawn from the results is that the task of predicting atrial fibrillation for the model in question becomes easier as the level of scale is increased slightly, and that high-frequency information is either not meaningful for the task of predicting atrial fibrillation, or that if it is, the model is unable to learn from it. / Djupinlärningsmodeller förekommer på många håll inom maskininlärning. De erbjuder bästa möjliga prestanda i olika domäner såsom datorlingvistik och bildklassificering. Nackdelen med dessa komplexa modeller är deras “svart låda”-egenskaper. Det är svårt för användaren att förstå hur en modell kommer fram till sin prediktion utifrån indatan. Detta är särskilt relevant i domäner såsom sjukvård, där tillit till modellen är avgörande. I denna uppsats utforskas sätt att förklara en modell som predikterar paroxysmalt förmaksflimmer från elektrokardiogram (EKG) som uppvisar normal sinusrytm. Med utgångspunkt i feature attribution (särdragsattribution) angrips problemet från tre olika perspektiv: tid, skala och frekvens. I synnerhet används en metod baserad på Integrated Gradients och en metod baserad på Shapley-värden. Genom att perturbera datan, träna om modellen, och utvärdera den omtränader modellen på den perturberade datan utvärderas graden av överensstämmelse mellan attributionerna och den meningsfulla informationen i datan. Resultaten visar att attributioner i skala- och frekvensdomänerna delvis stämmer överens med den meningsfulla informationen i datan, medan attributionerna i tidsdomänen inte gör det. Slutsatsen som dras utifrån resultaten är att uppgiften att prediktera förmaksflimmer blir enklare när skalnivån ökas något, samt att högre frekvenser antingen inte är betydelsefullt för att prediktera förmaksflimmer, eller att om det är det, så saknar modellen förmågan att lära sig detta.