Scanpath prediction in omnidirectional images (ODIs) serves as a critical component for optimizing foveated rendering efficiency and enhancing interactive quality in virtual reality systems. However, existing scanpath prediction methods for ODIs still suffer from fundamental limitations: (1) inadequate modeling and capturing of long-range temporal dependencies in fixation regions, and (2) suboptimal integration of spatial and temporal visual features, ultimately compromising prediction performance. To address these limitations, we propose a novel Dual-Temporal Modulated Diffusion model for Omnidirectional Images Scanpath Prediction, named SalDiff-DTM model, to effectively generate realistic human eye viewing trajectories. Specifically, to effectively model spatial relationships, we propose a novel Dual-Graph Convolutional Network (Dual-GCN) module that simultaneously captures semantic-level and image-level correlations. By integrating both local spatial details and global contextual information across the internal temporal dimension, this module achieves comprehensive and robust modeling of spatial relationships. To further enhance the modeling of temporal dependencies inherent in diverse fixation patterns, we introduce TABiMamba (Temporal-Aware BiLSTM-Mamba), a dedicated module that synergistically combines the contextual sensitivity of BiLSTM with the long-range sequence modeling capabilities of Mamba. This design facilitates deep information flow and context-aware sequential reasoning, thereby enabling high-fidelity capture of intricate temporal correlations. Inspired by the progressive refinement mechanism of diffusion models in various generative tasks, we propose a saliency-guided diffusion module that formulates the prediction problem as a conditional generative process, iteratively yielding accurate and perceptually plausible scanpaths. Extensive experiments demonstrate that SalDiff-DTM significantly outperforms state-of-the-art models, paving the way for future advancements in eye-tracking technologies and cognitive modeling.

Application Programming Interface (API) tutorial is an important API learning resource. To help developers learn APIs, an API tutorial is often split into a number of consecutive units that describe the same topic (i.e. tutorial fragment). We regard a tutorial fragment explaining an API as a relevant fragment of the API. Automatically recommending relevant tutorial fragments can help developers learn how to use an API. However, existing approaches often employ supervised or unsupervised manner to recommend relevant fragments, which suffers from much manual annotation effort or inaccurate recommended results. Furthermore, these approaches only support developers to input exact API names. In practice, developers often do not know which APIs to use so that they are more likely to use natural language to describe API-related questions. In this paper, we propose a novel approach, called Tutorial Fragment Recommendation (TuFraRec), to effectively recommend relevant tutorial fragments for API-related natural language questions, without much manual annotation effort. For an API tutorial, we split it into fragments and extract APIs from each fragment to build API-fragment pairs. Given a question, TuFraRec first generates several clarification APIs that are related to the question. We use clarification APIs and API-fragment pairs to construct candidate API-fragment pairs. Then, we design a semi-supervised metric learning (SML)-based model to find relevant API-fragment pairs from the candidate list, which can work well with a few labeled API-fragment pairs and a large number of unlabeled API-fragment pairs. In this way, the manual effort for labeling the relevance of API-fragment pairs can be reduced. Finally, we sort and recommend relevant API-fragment pairs based on the recommended strategy. We evaluate TuFraRec on 200 API-related natural language questions and two public tutorial datasets (Java and Android). The results demonstrate that on average TuFraRec improves NDCG@5 by 0.06 and 0.09, and improves Mean Reciprocal Rank (MRR) by 0.07 and 0.09 on two tutorial datasets as compared with the state-of-the-art approach.