Explainability Using Bayesian Networks for Bias Detection: FAIRness with FDO

In this paper we aim to provide an implementation of the FAIR Data Points (FDP) spec, that will apply our bias detection algorithm and automatically calculate a FAIRness score (FNS). metrics would be themselves represented as FDOs, could presented via visual dashboard, machine accessible (Mons 2020, Wilkinson et al. 2016). This enable dataset owners monitor level their data. is step forward in making data FAIR, i.e., Findable, Accessible, Interoperable, Reusable; or simply, Fully AI Ready First may discuss context topic with respect Deep Learning (DL) problems. Why are Bayesian Networks (BN, explained below) beneficial for such issues? Explainability – Obtaining directed acyclic graph (DAG) from BN training provides coherent information about independence variables base. generic DL problem, features functions these variables. Thus, one can derive which dominant system. When customers business units interested cause neural net outcome, DAG structure both source importance clarify model. Dimension Reduction — joint distribution associations. The latter play role reducing induce engine: If know random X,Y conditional entropy X Y low, omit since its nearly entire information. We have, therefore, tool statistically exclude redundant Tagging Behavior section less evident those who work domains vision voice. some frameworks, labeling obscure task (to illustrate, consider sentiment problem many categories overlap). tag data, rely on within datasets generate probability. Training BN, when initialize empty DAG, outcomes target parent other nodes. Observing several tested examples, reflect “taggers’ manners”. therefore use DAGs not merely purpose model development learning but mainly taggers policy improve it if needed. conjunction Casual inference Causal Inference highly developed domain analytics. It offers tools resolve questions hand, models commonly do and, real-world raises. There need find framework conjunction. Indeed, frameworks already exist (e.g., GNN). But mechanism merges typical problems causality common. believe flow, described paper, good direction achieving benefits Fairness Bias networks, essence, they reveal columns (or items) modify noise bias, address faults column However, assume have set measure (Purian 2022). networks prominence (as “cause” “effect” data), thus allow us assess overall database. What Networks? motivation using (BN) learn dependencies graphs (DAG), mimic Perrier (2008)). follows probabilistic factorization distribution: node V depends only parents (a r.v independent nodes free node). Real-World Example present way engine tabular python package bnlearn. Since project commercial, variable names were masked; thus, meaningless names. Constructing Our begin by finding optimal DAG. import bnlearn bn = bn.structure_learning.fit(dataframe) now has adjacency matrix found follow: print(DAG[ 'adjmat' ]) outcome form Fig. 1a. Where rows sources (namely arc left elements row) targets (i.e., header receives arcs). drawing obtained get following image: 1b. see rectangle still points arrows itself two discussion Rauber 2021). more variables, I increased number Adding provided new row “False”). following: 1c. So, how construct Now train parameters. Code-wise perform follows: model_mle bn.parameter_learning.fit(DAG, dataframe, methodtype= 'maximumlikelihood' ) change ‘ maximulikelihood ’ bayes beyond. factorized distributions DAG’s structure. given variable: 1d. code create presentation 2. Discussion theoretical concepts usage constructing approximated addition, example end learning: parameters maximum likelihood estimation (MLE) methods, performing inference. metrics, also visualised monitored, taking care FAIRness.