Event

Name: Xiaowen Qi
This dissertation develops a unified research program centered on the multi-objective design of ML systems. The optimization-driven design methods span evolutionary algorithms, architectural co-design, and interpretability-aware training. The proposed methodologies are designed to systematically navigate the trade-off spaces between accuracy and other critical deployment objectives, yielding models that are not only effective but also practical, responsible, and deployable. The research is organized across two principal application domains: (1) smart manufacturing and Industrial Internet of Things (IIoT), where the focus is on efficient spatio-temporal prediction for real-time closed-loop control and energy-aware system deployment; and (2) public safety and criminal justice, where the focus is on integrating heterogeneous data sources, ensuring algorithmic fairness, and providing interpretable predictions.

In the smart manufacturing and IIoT domain, we propose ESPNet, a 3D convolutional neural network for Efficient Spatter Prediction (the origin of the name "ESP") in additive manufacturing (AM). We demonstrate that feedforward spatio-temporal architectures achieve accuracy comparable to recurrent alternatives while being 5.8x faster and less than one-third the model size. Such characteristics help to address the accuracy–latency trade-offs that are critical for real-time closed-loop control in these applications. We also propose the Access Point Placement Genetic Algorithm (AP2GA), a genetic algorithm for energy-efficient access point deployment in industrial IoT environments. AP2GA optimizes communication energy consumption under non-uniform device distributions and unbalanced traffic loads. In the public safety and criminal justice domain, we introduce the Evolutionary algorithm for multi-objective EQuity and RP optimization (E2QRP) to derive optimized Pareto fronts of decision trees. The generated decision trees balance recidivism prediction accuracy against quantitative disparity metrics, and provide criminal justice decision-makers with transparent, interpretable trade-off curves. Additionally, we develop the Temporal + Static + Geographic Dependency (TSGD) model, which integrates time-varying crime histories with static contextual features. TSGD applies parallel-branch processing and a generalized inter-area dependency framework that captures both spatial adjacency and socio-economic similarity.

Building upon the aforementioned foundation, this thesis proposes two research directions for future work. The first direction aims to deepen the efficiency dimension within the context of smart manufacturing deployments. In particular, we introduce a staged 2D+1D decomposition framework that decomposes computationally intensive 3D spatio-temporal convolutions into independently deployable 2D spatial encoders and 1D temporal predictors. This framework not only exhibits high compatibility with existing FPGA acceleration toolchains but also creates a natural cache boundary, thereby enhancing response speed by leveraging spatial feature reuse mechanisms.

The second research direction aims to advance the trustworthiness of public safety systems. We propose a novel white-box spatio-temporal graph neural network (ST-GNN) architecture called Prior-guided Residual Interpretable Spatio-temporal Model (PRISM), which learns the residual beyond known priors and employs disentangled channel message-passing to independently process three semantically distinct graphs: the spatial proximity graph, the contextual similarity graph, and the discovered latent graph. This "white-box'' architecture provides valuable multi-level, and interpretable insights into the underlying drivers of criminal patterns. The model's interpretability is fully embedded within its architectural design itself, rather than relying  on post-hoc retrospective analyses.