Social work practice faces critical challenges in accurately identifying high-risk families for timely intervention due to limited labeled data. This research develops an intelligent system integrating supervised and semi-supervised learning for early warning, coupled with uncertainty-aware resource optimization. We propose a hybrid framework combining Self-training and Co-training leveraging 1,200 labeled and 13,800 unlabeled family cases. The proposed method achieves 85.6% accuracy and 82.4% recall on the high-risk class, with 95% confidence intervals, substantially outperforming supervised-only baselines. A multi-objective allocation model balances effectiveness, cost, and equity while incorporating prediction uncertainty via robust optimization. Sensitivity analysis over 500 bootstrap scenarios demonstrates service quality guarantees. Subgroup fairness audits show consistent recall across demographic groups (max gap 1.2%). The optimized allocation reduces manpower by 22% while improving intervention success from 68.5% to 84.2%. SHAP-based explainability and human oversight support ethical deployment in social services.

To address the “information gap” in social work arising from unstructured data and limited labeled instances, this study proposes a semi-supervised learning framework based on a Teacher-Student BERT architecture. Based on an analysis of 50,000 case records spanning 2019 to 2024, the model incorporates Latent Dirichlet Allocation for topic extraction alongside a multidimensional gap analysis. The proposed method attained a state-of-the-art F1 score of 0.913, markedly surpassing baseline BERT models. Notable findings include a 67.6% increase in mental health-related discourse and the quantification of significant systemic deficiencies, such as inadequate coverage in elderly care (45.8%) and substantial unmet needs in medical assistance (46.4%). Furthermore, to address the profound challenges of class imbalance and the data-hungry nature of transformer models, this study integrated generative artificial intelligence (AI) data augmentation techniques. This approach produced synthetically varied case narratives that preserved the original socio-economic context while expanding lexical and structural diversity, thereby significantly enhancing the classification accuracy of minority classes. Additionally, the use of a Mean Teacher denoising framework and knowledge distillation drastically reduced the computational inference time, rendering the architecture highly suitable for deployment in resource-constrained social welfare environments. Prior to analysis, all case records underwent rigorous automated Personally Identifiable Information (PII) scrubbing utilizing the Microsoft Presidio framework to guarantee data privacy. To ensure reproducibility, the Teacher-Student training codebase, prompt architectures, and anonymized synthetic data samples will be made publicly available upon request. This research effectively transforms administrative textual data into actionable strategic intelligence, offering a scalable and evidence-based tool to enhance resource allocation and inform policy development.

This paper proposes a drift-aware off-policy deterministic actor–critic framework for constrained continuous resource allocation in non-stationary environments. Feasible allocations are ensured by a simplex-parameterized policy using softmax normalization with budget scaling, avoiding projection or Lagrangian tuning. The reward integrates Nash social welfare via mean log-utility, efficiency, fairness, and constraint-violation penalties with adaptive weights. To improve sample efficiency, we adopt prioritized experience replay based on TD error and state novelty. Non-stationarity is detected by KL divergence between recent and historical state-visitation distributions; detected drift triggers buffer refresh and incremental fine-tuning, while Elastic Weight Consolidation mitigates catastrophic forgetting. Experiments across six application-motivated domains (food, medical, housing, education services, employment support, and elderly care) demonstrate improved utilization and welfare with reduced inequality and low decision latency compared with optimization, heuristic, and DRL baselines. Results are reported over multiple runs with mean ± standard deviation and corrected significance tests.

Community engagement is essential to social service delivery, yet traditional community needs assessment remains time-consuming and poorly suited for timely monitoring. This study proposes a semi-supervised learning framework to identify emerging community needs and service gaps from massive, mostly unlabeled, unstructured text. We construct an explicit heterogeneous text graph where each record is a document node linked to keyword and need-category nodes; document–document edges are built using a weighted combination of semantic similarity (BERT cosine), lexical overlap (keyword Jaccard), and temporal proximity. A graph neural network with iterative self-training leverages 3% expert-labeled seed data and the remaining unlabeled corpus to classify records into a 10-category need taxonomy. On 176,602 records, the proposed model achieves F1 = 0.895 and Recall = 0.899, outperforming supervised baselines trained on the same labeled ratio by 23.8% (macro-F1). Post-hoc quarterly aggregation of predictions enables trend monitoring and prioritization of service-gap severity for decision support.