Active learning (AL) is a semi-supervised learning paradigm with human-machine interaction and a limited annotation budget. However, few AL studies have explored distribution inconsistency between the data and the population. In this paper, we consider a basic form of the aforementioned issue,i.e., the training data is non-independently and identically distributed (non-IID) sampled from a class uniformly distributed population. Accordingly, we propose a naïve sample selection plugin, namely generalized active stratified sampling (GASS), to rebalance the sample size of each class during AL iterative process, resulting in a progressive approximation to the population. We generalize statistical stratified sampling to support the uncertainty strata criterion, forming the statistical foundation of GASS. This method, as a plugin, can seamlessly collaborate with popular information-based strategies. GASS shows superior rebalancing capabilities by analyzing the statistical moment and the class imbalanced index under the Probably Approximately Correct (PAC) theory. Furthermore, models derived with GASS have low Rademacher complexity (RC), indicating low generalization error bounds, and GASS also exhibits strong robustness to prediction perturbations. Experiments were conducted on 5 benchmark image datasets, and the results show that GASS significantly boosts the test accuracy by about$2.38\%$/$3.19\%$(paired$t$-test$p=0.01$/0.04) and reduces the empirical RC by about$1.42\%$/$1.94\%$(paired$t$-test$p=0.01$/0.05) on average in class imbalanced/balanced scenarios, respectively. This study establishes a potential benchmark for information-based AL.

Label distribution learning (LDL) employs probabilistic labels to capture the varying degrees of relevance among decision attributes. Existing LDL algorithms usually employ sample correlations to improve their predictive validity. However, they merely employ the superficial features of the sample for correlation analysis, seldom delving into its latent deep features. In this paper, we propose an algorithm to explore the sample topology (ST-LDL) to address this problem. First, we construct a locally weighted directed graph for each target sample. The target sample and its k neighbors are regarded as nodes within the graph. The asymmetrical correlation among these $k+1$ samples is computed separately to determine the weight value of the graph. Then, we utilise the local topology between samples to mine latent information and reconstruct the samples. The local graph is fed into the graph convolutional network to nonlinearly reconstruct the features by mining the latent information. Finally, we designed a new optimization objective function for the reconstructed samples. Experiments are carried out on elven real-world datasets in comparison with seven state-of-the-art algorithms. The results show that our algorithm outperforms several other algorithms, proving the effectiveness of our algorithm.

Co-training is a popular disagreement-based semi-supervised learning method. Learners of different views mutually select reliable unlabeled instances to augment the labeled dataset. Existing co-training style algorithms have cumbersome procedures for selecting confident instances. Furthermore, the pseudo-labels assigned to selected unlabeled instances are not always reliable. In this paper, we propose a safe co-training regression algorithm for multi-view scenarios with two characteristics. An instance selection strategy based on the consistency assumption aims to improve the efficiency of selecting confident unlabeled instances. This strategy makes full use of the information provided by a committee to measure the confidence of unlabeled instances. A safe labeling technique in an ensemble manner is introduced to improve the quality of pseudo-labels. The safe pseudo-labels not only integrate information provided by the committee, but also take into account the part of the receiver. The results over twenty datasets prove the superiority of the proposed algorithm against other state-of-the-art semi-supervised regression algorithms.

Semi-supervised regression attempts to utilize as much unlabeled data as possible and as little labeled data as possible to improve model performance. Methods based on data partitioning can improve regression performance from the perspective of data distribution. However, most partitioning methods only consider the correlation between the data, but not the relationship between the regressor and the data. In this study, a strategy of dividing the data and then regressing is proposed, where data partitioning is based on the relationship between the regression and data. Accordingly, an intuitive and effective algorithm named SRPF, i.e. Semi-supervised regression based on data partitioning and feature mapping, is proposed. First, we divide the labeled dataset into two disjoint subsets based on the relationship between the predicted and actual values of separation regressor. Second, we label these two subsets as distinct classes and use feature mapping to map data features to higher dimensions to better distinguish the data. Third, we construct a partitioner to determine which subset an unlabeled instance belongs to, and then use the regressor on the corresponding data set to make predictions. Finally, during the iterative training process, a self-training method is used to enrich labeled samples. Experiments are conducted on 15 well-known datasets compared to state-of-the-art algorithms. The results show that our method outperforms them in most datasets.

Player behavior modeling is of the utmost importance in game development and player matching. This problem is challenging because behavior is multi-semantic and hard to represent. Existing work often suffers from low generalization ability and high demand for supervisory information. In this paper, we present a behavior representation learning method (BRL) for Reversi players. It learns entirely from unlabeled game records. First, we develop an asymmetric encoder-decoder architecture to learn the mapping between states and actions. The encoder maps game records into a latent subspace for behavior representation. Second, we mask random states to force the encoder to capture the high-level features of the policy. The decoder predicts the corresponding actions according to the latent representation. Coupling these two designs, the semantic relevance of behaviors can be more effectively measured. Experiments were conducted in Reversi, using 14,000 game records of different players to learn behavior representations. Transfer performance in downstream tasks outperforms the supervised method and shows promising scaling ability. This work opens a new way for analyzing and modeling player behavior.

Label distribution learning (LDL) can more accurately represent the degree of correlation between labels and samples than multi-label learning. However, LDL usually has limited available label data, which is not conducive to training deep learning models. Data augmentation refers to methods for solving limited data problems by introducing unobserved data or latent variables to increase the size and quality of the training dataset. In this paper, we augment the dataset by mapping the features and label distributions of the generated samples to the same subspace, and using the generator to learn the distribution of the original data in this space. First, we use the encoder and generator to extract effective information from sample features and label distributions, respectively. Second, we randomly fuse the existing label distribution to generate a new label distribution, then map it to the subspace and restore the corresponding features through the decoder. Finally, these generated samples are mixed with the original training set to train a model for predicting label distribution. Experimental results on nine real-world datasets show that our proposed algorithm can improve the performance of deep learning models to a certain extent.

Real-world applications always contain complex data objects with multiple semantics. The machine learning tasks generalized from these applications can be formulated as multi-instance multi-label learning (MIML) problems. Each object is represented by a bag of multiple instances and can be associated with multiple labels simultaneously. Extensive studies on the MIML have been conducted during the past few years. However, many of them do not consider local label correlations, which is critical in multi-label scenario. To address this problem, we propose the multi-instance multi-label learning based on parallel attention and local label manifold correlation (MIML-LLMC) algorithm. First, parallel multiple attention mechanisms convert each bag into a number of fusion vectors, one for each label. They are also utilized to find out which instances in the bag trigger respective labels to improve explanability. Second, local label manifold correlation vectors (LLMCs) are constructed based on label manifold generation and clustering. They provide a high level of representation of the label vectors. Third, each fusion vector is concatenated to the LLMC to predict each label independently. With the utilization of attention and LLMCs, MIML-LLMC is able to discover instance-label relations and exploit local label correlations simultaneously. Experiments reveal that our approach is highly competitive to the state-of-the-art MIML algorithms and yields reasonable results in understanding the relations between input patterns and output label semantics.

As data dimensions and volume rapidly increase, attribute reduction using the original data becomes computationally infeasible. Large data frequently contain various redundant attributes and types of noise. This leads to the problems of overfitting and inefficiency in data processing. To address these problems, this paper proposes a general sampling method for attribute reduction by introducing three-way decisions, namely, three-way sampling (3WS), which is the first sampling method that describes the decision boundary accurately while improving the data quality significantly. To improve the effectiveness and efficiency of attribute reduction, we designed a rapid attribute reduction method based on three-way sampling (3WS-RAR). The 3WS-RAR method consists of three main steps: data sampling, attribute reduction, and model effectiveness evaluation. For data sampling, we define the three regions of the 3WS using support vectors to describe the data and use the boundary region as the sampling results. For the attribute reduction, we compute the neighborhood self-information for each attribute while considering the upper and lower approximations. For the effectiveness evaluation, we conducted experiments on 15 relatively large-scale datasets and analysed the influence of parameters. The experimental results reveal that, compared with state-of-the-art attribute reduction models, 3WS-RAR performs better on public benchmark datasets.

Significant place mining in spatiotemporal trajectory data is a key task for mobile pattern mining, useful for supporting location-aware services. State-of-the-art trajectory clustering algorithms utilize a density-based distance measure. However, some major problems with this approach are that (1) results are often inaccurate, especially on data of varying density, (2) the user must fine-tune many thresholds that are unintuitive to set, and (3) boundary points between clusters are often assigned to the wrong locations. Performance is also a major issue as many state-of-the-art algorithms have a very high time complexity. Motivated by these issues, this paper proposes an approach inspired by the data field theory and community detection. It is a graph-based significant place mining algorithm, called GB-SPM, for capturing and characterizing the essence of similarity between nodes. GB-SPM first applies a novel low index neighborhood velocity point filtration method to extract characteristic points. Then, a characteristic point index neighborhood is used to map them to graph nodes. In this way, the original problem is transformed into a community detection problem in complex community networks. Finally, a new edge weight metric is proposed to capture and characterize the nature of similarity between nodes. To evaluate clustering quality, we used the Silhouette (SI) for unannotated data to value inter-cluster separation and intra-cluster homogeneity. To evaluate mining effectiveness, we used Matthew’s correlation coefficient (MCC) for annotated data. Numerous experiments were carried out on real world datasets, and the accuracy and performance of the designed algorithm was compared with the state-of-the-art algorithms. Results show that GB-SPM improves on average SI by 13.9%, MCC by 20.7%, and runtime by 5.15 times.