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.