K M Najib Hayder

For intervention in information disorder on social media, detection alone is insufficient because the volume and speed of circulation make full investigation of all disorder infeasible. This paper develops a conceptual prioritisation framework for information disorder on X (formerly Twitter), using Community Notes as an intervention-focused frontier sampling mechanism to narrative level instead of individual post level. A linguistic risk score measuring volatility of language to proxy for veracity, and a belief proxy score to proxy for vulnerability of users towards believing disordered information were used to define a 9-state risk-belief space, with state transition modelling using a Markov chain under partial observability. The results were used to construct two indices - Credibility Influence Index (CII) for directional drift in belief and the Attention Displacement Index (ADI) for magnitude of linguistic volatily. Monte Carlo bootstrap resampling provides uncertainty estimates. Empirical analysis on six narrative clusters supports a three-way prioritise/observe/deprioritise decision logic, offering a reproducible, uncertainty-aware conceptual triage framework, establishing the feasibility of a robust intervention prioritisation framework.

Fuel poverty statistics in England are reliable at national and regional scales but become increasingly uncertain when disaggregated to local geographies, limiting their use for operational targeting. Crucially, no property-level fuel poverty labels exist in public data, creating a structural challenge for data-driven decision-making. Existing machine learning approaches rely on proxy labels derived from sub- regional estimates, thereby inheriting their uncertainty. This paper proposes a novel, bottom-up framework for property-level fuel poverty risk scoring that avoids direct disaggregation. Instead, aggregate statistics from the Low Income Low Energy Efficiency (LILEE) framework are used as directional priors. These are propagated to individual properties via a geometric mean combination across multiple stratifying dimensions and incorporated into a Bayesian logistic regression model estimated using Hamiltonian Monte Carlo, producing posterior risk probabilities for each property. These probabilities are discretised into a five-class ordinal risk scale.

The resulting labels are then used to train four machine learning models-XGBoost, Random Forest, Deep Neural Network, and Multinomial Logistic Regression on a richer feature set derived from 2.6 million Energy Performance Certificate records linked with income, climate, energy price, and demographic data. This second stage serves a dual purpose: (i) providing convergent validation of the constructed risk measure across models with different inductive biases, and (ii) enabling scalable deployment as a fast, property-level scoring function.

The proposed framework contributes to operational research by addressing decision-making under label uncertainty, combining probabilistic inference with predictive modelling to support targeted interventions. It provides a transparent and reproducible method for prioritising properties for retrofit and support programmes, with direct applicability to schemes such as ECO4 bespoke targeting.