Federated multitask optimization (FMTO) allows private collaboration for distributed black-box optimization, but often suffers from negative transfer under strong task heterogeneity. We propose FMDLES, a personalized FMTO framework that couples sequence-based relatedness estimation with inverse-guide transfer and local refinement. Each client trains an auxiliary inverse mapping guide that predicts decision vectors from target objective values, providing coarse basin-level proposals without sharing landscapes or solutions. To support collaboration across heterogeneous tasks without exchanging solver-specific parameters, we introduce a solver-parameter-agnostic relatedness signature based on improvement-rate sequences that privately capture optimization dynamics. The server performs personalized Top-K aggregation using this sequence similarity to produce client-specific inverse guides, which are then exploited locally via NN-guided candidate generation followed by local surrogate-assisted differential evolution (LSA-DE). Experiments on Arxiv2017, CEC2022, and a heterogeneous kinematic arm control task show competitive and often improved performance over representative FMTO baselines across the studied heterogeneity settings.

Open Relation Extraction (OpenRE) aims at clustering relation instances to extract relation types. By learning relation patterns between named entities, it clusters semantically equivalent patterns into a unified relation cluster. Existing clustering-based OpenRE methods only consider the information of the instance itself, ignoring knowledge of any relations between instances. Therefore, a Cross-Attention Fusion based Graph Convolution Auto-Encoder (CAGCE) method for Open Relation Extraction is proposed. The Auto-Encoder learns the semantic information of the sentence instance itself, and the Graph Convolution Network learns the relational similarity information between sentences. Then, the two heterogeneous representations are crossed and fused layer-by-layer through a cross-attention fusion mechanism. Finally, the fused features are used for clustering to form the relation types. A comparison with baseline models using the FewRel and NYT-FB datasets shows the effectiveness and superiority of the proposed method.

Due to the availability of huge number of Web Services (WSs), calculating similarity between services is still a challenge. Usually, the similarity is measured in different dimensions, i.e. input/output parameters, QoS, correlations between preconditions and effects, etc. In this paper, we present an approach to measuring the degree of similarity between semantic web services considering multiple dimensions as a whole. It can integrate the results obtained from different matching algorithms in a generic and customizable manner. So that, the user may determine the relative importance of each dimension and obtain a single rank for candidates services matching a given service.