Abstract
This paper presents a novel multidisciplinary framework bridging social science (SS) and computer science (CS) to investigate socio-technical issues systematically. By integrating qualitative analysis with machine learning (ML), the framework provides a transparent, ethically-informed, and bias-aware methodology for identifying and quantifying objects of inquiry. A core feature is the explicit mapping of bias entry points and mitigation strategies, enhancing the robustness and explainability of research results. The framework was validated through an investigation of online harms encountered by minoritised ethnic (ME) communities accessing digitalised social housing services in the UK, identifying key vulnerabilities such as discrimination, digital poverty, and language barriers.
Introduction
As essential services like social housing, health, and energy transition to "digital by default" models, there is a significant risk of exacerbating existing societal inequalities. Investigating these "online harms" is inherently complex, often hindered by the silos between social scientists, who provide deep human context, and computer scientists, who offer scalable analytical tools. Furthermore, the reliance on "big data" often introduces hidden biases, while many ML models remain opaque "black boxes."
This paper addresses these challenges by proposing a transparent, reproducible pipeline where social science rigor—such as inductive coding—directly informs ML models. This synergy ensures that automated analysis remains grounded in human experience while providing the scale necessary to influence policy and service design.
The Multidisciplinary Framework
The proposed framework is structured into three distinct phases: Data Collection, Data Analysis, and Verification. This sequence ensures that findings are not only discovered through rigorous qualitative methods but are also scalable and cross-referenced against broader datasets.
- Data Collection (Part 1): Focuses on defining the object of inquiry and shared vocabulary. It utilizes inductive coding (IC), where themes emerge naturally from interview transcripts rather than being forced into pre-existing theoretical frameworks.
- Data Analysis (Part 2): Integrates SS and CS techniques. The authors utilize Anchored Topic Modeling (CorEx), which allows researchers to guide the model with "seed words" (e.g., "housing," "health"). This ensures the machine-generated topics align with research goals while discovering latent patterns. Additionally, Sentiment Analysis is performed using a fine-tuned GPT-2 Large model, coupled with SHAP (SHapley Additive exPlanations) to provide explainable AI (XAI) insights.
- Verification (Part 3): Findings from primary qualitative data are compared against large-scale public or alternate datasets to ensure generalizability and robustness.
Mapping and Mitigating Bias
A defining characteristic of this framework is the visualization of the "Bias Loop." The research cycle explicitly identifies where human and algorithmic biases can enter the process, marked by "lightning bolt" icons in the framework diagrams.
Personal bias may be introduced during transcription, translation, or recruitment, while algorithmic bias often stems from biased training data or model selection. To counter this, the framework incorporates "mitigation points" (inverted triangles), where researchers apply reflexivity, triangulation, and multidisciplinary collaboration to minimize these effects.
Case Study: Digital Vulnerabilities in Social Housing
To test the framework, the authors investigated the vulnerabilities of minoritised ethnic (ME) communities in the UK social housing sector. The study utilized 100 semi-structured interviews (the PRIME dataset) and was verified against the EVENS Dataset (Evidence for Equality National Survey), which contains over 4,600 data points.
Through the application of Latent Class Analysis (LCA), the researchers demonstrated that ethnic groups are not a monolith. Different communities face distinct hurdles; for instance, Bangladeshi groups presented unique clustering patterns, while Black African communities showed the highest relevance to social housing vulnerabilities.
Key Results
- Vulnerability Identification: The framework successfully isolated four primary vulnerabilities: discrimination, digital poverty, lack of digital literacy, and lack of English proficiency.
- Model Performance and Explainability: The sentiment analysis model achieved 72.88% overall accuracy. XAI analysis revealed that the word "difficult" was the strongest predictor of negative sentiment across both language and digital literacy domains, highlighting a universal barrier to service access.
- Disproportionate Impact: The analysis confirmed that Black African communities are significantly more likely to experience these vulnerabilities when interacting with digitalised social housing services.
Conclusion
By bridging the gap between social science and computer science, this framework offers a path toward more ethical and transparent socio-technical research. While the study was limited by convenience sampling and a relatively small sub-sample for deep-dive sentiment analysis, it provides a scalable model for future investigations. Future work will involve co-design workshops with stakeholders to translate these findings into practical, harm-reducing tools for minoritised communities.
Summary updated