The goal of missing value imputation is to replace the missing values in the dataset with specific values. In particular, this preprocessing step plays a crucial role in knowledge discovery and data mining, as most data analysis methods assume complete data and cannot be applied directly to datasets with missing entries. Among various approaches, the neural network-based missing value imputation method has recently gained significant attention due to its superior prediction accuracy based on its excellent capability to fit the given training dataset. Specifically, these approaches conventionally begin by applying a naïve missing value imputation to fill all missing entries in the dataset and then train the network on the completed data. Thus, the performance of missing value imputation can be limited because the neural network is trained on an unreliable dataset filled with roughly guessed values. Instead, we may consider an alternative strategy to use only the features without missing values or carefully imputed features obtained during the imputation process, which can be regarded as an asymmetric process because it progressively adds the newly imputed features into the training dataset. In this study, we propose an effective neural network-based imputation method that incrementally constructs a cumulative feature set during training. The experimental results on 25 publicly available datasets showed that the proposed method outperforms conventional methods significantly.

Unsupervised feature selection seeks to identify informative, nonredundant features without using class labels. Motivated by Fano’s inequality, we use the entropy of the selected feature subset as a task-agnostic heuristic for its information capacity. Because multivariate joint entropy is difficult to estimate and exact evaluation of all pairwise joint entropies is computationally costly, we propose a landmark-based block approximation inspired by the Nyström method. Features with high marginal entropy are selected as landmarks. Joint entropy is computed exactly for pairs involving at least one landmark, while non-landmark pairs are approximated in constant time per pair using the geometric mean of their marginal entropies. We establish deterministic bounds for the surrogate and incorporate the resulting hybrid estimator into an incremental feature selection strategy. Experiments on 25 public benchmark datasets show that the proposed method achieves the lowest average rank across eight evaluation metrics. It exhibits significant advantages over most baselines on information-preservation metrics while maintaining competitive classification and clustering performance.

Crack segmentation plays a vital role in ensuring structural safety, yet practical deployment on resource-limited platforms demands models that balance accuracy with efficiency. While high-accuracy models often rely on computationally heavy designs to expand their receptive fields, recent lightweight approaches typically delay this expansion to the deepest, low-resolution layers to maintain efficiency. This design choice leaves long-range context underutilized, where fine-grained evidence is most intact. In this paper, we propose an evidence-preserving receptive-field expansion network, which integrates a multi-scale dilated block to efficiently capture long-range context from the earliest stages and an input-guided gate that leverages grayscale conversion, average pooling, and gradient extraction to highlight crack evidence directly from raw inputs. Experiments on six benchmark datasets demonstrate that the proposed network achieves consistently higher accuracy under lightweight constraints. Each of the three proposed variants—Base, Small, and Tiny—outperforms its corresponding baselines with larger parameter counts, surpassing a total of 13 models. For example, the Base variant reduces parameters by 66% compared to the second-best CrackFormer II and floating-point operations by 53% on the Ceramic dataset, while still delivering superior accuracy. Pareto analyses further confirm that the proposed model establishes a superior accuracy–efficiency trade-off across parameters and floating-point operations.