Most of statistics and AI draw insights through modeling discord or variance between sources (i.e., intersource) of information. Increasingly however, research is focusing on uncertainty arising at the level of indivi...
详细信息
In the real-world an effect often arises via multiple causal mechanisms. Conversely, the behaviour of AI systems is commonly driven by correlations which may-or may not-be themselves linked to causal mechanisms in the...
详细信息
ISBN:
(数字)9798350319545
ISBN:
(纸本)9798350319552
In the real-world an effect often arises via multiple causal mechanisms. Conversely, the behaviour of AI systems is commonly driven by correlations which may-or may not-be themselves linked to causal mechanisms in the associated real-world system they are modelling. From an AI and XAI point of view, it is desirable for AI systems to model and communicate primarily, if not exclusively, causal mechanisms between variables, affording strong generalisation performance and effective explanations. Indeed, as we discuss in this paper, it is critical for explanations for a given effect not only to reflect possible causal mechanisms, but to highlight the specific causal mechanisms which led to the effect in the given instance. In this light, we proceed to propose a rule generation framework which generates rules for fuzzy systems that capture possible causal mechanisms between the input variables and the target variable as discovered by data-driven causal discovery algorithms for the given data set. For a given sample, i.e., a specific set of inputs, the obtained fuzzy system provides local explanations which distinguish the locally relevant causal mechanism(s) of its effect from other possible, but not applicable causal mechanisms, and thus avoids both overly simplistic single-cause and exhaustive-potentially misleading explanations. Experiments show that the fuzzy systems obtained by the proposed framework achieve comparable performance compared to classical correlation-based approaches, and provide local explanations which indicate the specific causal mechanism for different effects.
Generating (fuzzy) rule bases from data can provide a rapid pathway to constructing (fuzzy) systems. However, direct rule generation approaches tend to generate very large numbers of rules. One reason for this is that...
Generating (fuzzy) rule bases from data can provide a rapid pathway to constructing (fuzzy) systems. However, direct rule generation approaches tend to generate very large numbers of rules. One reason for this is that such techniques are not designed to differentiate between relationships of variables reflecting a causal link and those where such a link reflects a spurious correlation in the data set. In prior work, we discussed how causal discovery techniques, and specifically the subset resulting of variables within the Markov blanket can be leveraged to focus on the generation of rules for variables with a causal link. In this paper, we broaden this discussion, outlining a road-map to explain how causal discovery and its outputs-causal graphs-can be used towards refining (fuzzy) rule generation techniques. As a next step on this road-map, we present an initial approach which leverages the causal direction captured in the graph to further reduce the set of variables from those captured in the Markov blanket. Initial results show that the approach, combined with a traditional fuzzy rule generation technique such as the Wang-Mendel approach, produces competitive performance and concise rule bases-highlighting a path towards improved fuzzy system interpretability.
The presence of uncertainty, such as datauncertainty (e.g., noise) and model uncertainty (e.g., parameters), directly impacts the efficacy of prediction systems. Recent research is increasingly focusing on the explic...
The presence of uncertainty, such as datauncertainty (e.g., noise) and model uncertainty (e.g., parameters), directly impacts the efficacy of prediction systems. Recent research is increasingly focusing on the explicit modelling of uncertainty within the prediction by deriving an interval-valued, rather than a point-valued prediction. The aim here is to explicitly capture uncertainty within data and model and map it systematically to the prediction outputs, in turn providing usable bounds on the expected prediction. Such bounds can convey crucial insight, such as the expected amount of renewable energy generation expected–enabling the minimal–but sufficient use of fossil fuels and other energy sources. This paper explores the viability of recent advances within non-singleton fuzzy systems to support the efficient and effective prediction of such interval-bounds, while providing an interpretable prediction model. The latter is desirable, as it affords experts not only the ability to validate the model, but allows them to interact with and change the model to adapt to unforeseen circumstances. For example, in a renewal energy context, priming a prediction system for uncommon weather patters can support its effectiveness. The proposed approach integrates both data and model uncertainty into the system recursively, generating systematic prediction intervals. To demonstrate the potential of the approach in integrating these–while supporting interpretability/interactivity, we show a series of synthetic experiments using the Mackey glass time series. The results highlight the capacity of the approach to preserve and recursively model the output uncertainty for multi-step ahead predictions. Further, the experiments demonstrate how the proposed system can provide explanations for given predictions, providing the basis for supporting expert interaction.
This paper explores the parameter generation of the existing ‘Parametrized Model’ (PM) as the state-of-the-art linear interval-valued regression model, highlighting that its strong performance may arise from unexpec...
This paper explores the parameter generation of the existing ‘Parametrized Model’ (PM) as the state-of-the-art linear interval-valued regression model, highlighting that its strong performance may arise from unexpected behavior. Focusing on the approach's core idea of using dynamic reference points from the regressor variables to describe and regress interval-valued data, this paper shows that key parameters in the PM method are not actually restricted, and explores the impact this has on model generation. As part of this analysis, we propose a Restricted Parametrized Model (RPM) which ensures parameter bounds are maintained. We evaluate the approach by conducting linear regression for a series of synthetic interval-valued data sets with different features, discussing its performance against the original PM and other linear interval regression models. The experiments confirm empirically that the PM model violates restrictions on parameter bounds, resulting in the selection of reference points—outside—the regressor intervals–nevertheless producing strongly performing regression models which outperform other models, including RPM, when parameter bounds are violated. We conclude by discussing that the mechanism underpinning the PM regression is different to what is expected, i.e. it commonly selects reference points which are not from the regressor variables, warranting further research to explore whether the PM's parameter bounds can be relaxed, or whether this exposes potential problems, for example for edge cases.
We present JuzzyPy, a Python based fuzzy logic toolkit enabling the creation of type-1, interval type-2, and general type-2 fuzzy logic systems. Fuzzy logic systems are being applied in disciplines across engineering ...
详细信息
ISBN:
(纸本)9781665487696
We present JuzzyPy, a Python based fuzzy logic toolkit enabling the creation of type-1, interval type-2, and general type-2 fuzzy logic systems. Fuzzy logic systems are being applied in disciplines across engineering and sectors such as cyber-security and autonomous systems, where an increasing focus on interpretability and trust is re-emphasising a focus on rule-based systems. This highlights the value of broadly accessible software resources to support both experts and non-experts in efficiently designing fuzzy systems. Python is currently the most popular programming language in the world. JuzzyPy is the first library capable of supporting the development of general type-2 systems in Python, facilitating access to these systems across engineering disciplines. In this paper we delve into the ability of the toolkit and explore its implementation. We hope that this allows for further accessibility in the use of fuzzy logic systems and provides a basis for developers, practitioners and researchers to collaborate on scientific advancement of the field and the tackling of real-world problems.
Machine learning methods are widely used to predict energy consumption, aiming to enhance efficiency and support environmental goals. However, developing these models is traditionally time-consuming and expert-depende...
详细信息
ISBN:
(数字)9781665410205
ISBN:
(纸本)9781665410212
Machine learning methods are widely used to predict energy consumption, aiming to enhance efficiency and support environmental goals. However, developing these models is traditionally time-consuming and expert-dependent. While Automated Machine Learning (AutoML) has emerged as a valuable approach to streamlining machine learning pipelines, including appropriate preprocessing and learning algorithms, it may still require human experts. These experts might be needed to generate new, interpretable features that could significantly improve model performance. This is particularly relevant in complex settings such as the energy domain, where deep learning's automatic feature extraction lacks interpretability. To address this challenge, this exploratory work introduces an automated feature engineering method tailored for energy forecasting problems. It involves generating a comprehensive set of features that can be fed into AutoML, thereby reducing the need for domain knowledge in feature engineering. The proposed method has been validated using eleven datasets from various energy domains, including residential buildings, renewable energy, and regional energy consumption, with state-of-the-art AutoML methods, namely H20, TPOT, AutoGluon, and FLAML. The results demonstrate a noticeable reduction in prediction errors across all the examined datasets.
Hierarchical fuzzy systems (HFSs) are claimed to be an excellent approach to reducing the number of rules in Fuzzy logic systems (FLSs). Further, HFSs have also been shown to have the potential to reduce complexity an...
详细信息
ISBN:
(数字)9798350319545
ISBN:
(纸本)9798350319552
Hierarchical fuzzy systems (HFSs) are claimed to be an excellent approach to reducing the number of rules in Fuzzy logic systems (FLSs). Further, HFSs have also been shown to have the potential to reduce complexity and improve interpretability for FLSs. However, designing an interpretable HFS is a challenging task. This is due to the HFSs' structures, which have multiple subsystems, layers and topologies. This paper put forward an approach to present a design guidelines framework to build interpretable HFSs. The framework consists of five key guidelines for interpretable HFSs. It is demonstrated using a real-world example to design interpretable HFS for the Iris classification problem. This study contributes to providing insight into a pathway for designing interpretable HFSs, as can be used in practice.
The growing use of intervals in fields like survey analysis necessitates effective aggregation methods that can summarize and represent such uncertain data representations. The Interval Agreement Approach (IAA) addres...
详细信息
ISBN:
(数字)9798350319545
ISBN:
(纸本)9798350319552
The growing use of intervals in fields like survey analysis necessitates effective aggregation methods that can summarize and represent such uncertain data representations. The Interval Agreement Approach (IAA) addresses this by aggregating interval responses into Fuzzy Sets (FSs), capturing both intra- and inter-participant agreement, while minimizing information loss. While offering a powerful modeling tool, the IAA does not natively offer a measure of central tendency, which is itself an interval of particular utility in real-world applications. In contrast, the Interval Weighted Average (IWA) has been used for directly measuring the central tendency of intervals. While straightforward and effective, it is not designed, nor able to, summarize interval data in terms of their agreement, as the IAA does. To bridge this gap, this paper introduces Interval Agreement Weighted Average (IAWA), which is specifically designed to reflect both the central tendency and agreement. This is achieved by first modeling interval agreement as FSs using the IAA, and then transforming these FSs into intervals using the IWA. We demonstrate the approach by conducting sensitivity analyses to explore the behavior of the proposed approach in detail. Our findings suggest that the IAWA is a highly effective measure of central tendency. Additionally, it also partially inherits IAA's ability to reflect the agreement of sets of intervals. We conclude by highlighting the potential and growth of the use of intervals in information elicitation, within, and beyond survey research, underpinning a new degree of understanding of both intra- and inter-source uncertainty.
LIME, or ‘Local Interpretability Model-agnostic Explanations’ is a well-known post-hoc explanation technique for the interpretation of black-box models. While very useful, recent studies show that LIME suffers from ...
LIME, or ‘Local Interpretability Model-agnostic Explanations’ is a well-known post-hoc explanation technique for the interpretation of black-box models. While very useful, recent studies show that LIME suffers from stability problems: explanations provided for the same process can be different, making it difficult to trust their reliability. This paper investigates the stability of LIME when explaining multivariate time series classification problems. We demonstrate that due to the temporal dependency in time series data, the traditional artificial neighbour generation methods used in LIME have a higher risk of creating out-of-distribution inputs. We disucss how this behavior is one of the reasons resulting in unstable explanations. In addition, LIME weights neighbours based on user-defined hyperparameters which are problem-dependent and hard to tune, and we show how unsuitable hyperparameters can contribute to the generation of unstable explanations. As a preliminary step towards addressing these issues, we propose to employ a generative approach with an adaptive weighting method in the LIME framework. Specifically, we adopt a generative model based on variational autoencoder to create within-distribution neighbours, reducing the out-of-distribution problem, while the adaptive weight method eliminates the need for user-defined hyperparameters. Experiments on real-world datasets demonstrate the effectiveness of the proposed method in providing more stable explanations using the LIME framework.
暂无评论