Sione Palu

Sione PaluOct 13, 2026

Automated portfolio optimization is a fundamental quantitative finance task that dynamically allocates capital across multiple assets under non-stationary market conditions. Instead of static holding, it continuously balances return opportunities and risk exposure through systematic position rebalancing, using cross-asset diversification principles to effectively reduce asset-specific risk. The field has evolved through several distinct paradigms: traditional optimization models like Markowitz mean-variance, minimum-variance, risk parity, and cost-aware rebalancing; observation-based reinforcement learning (RL) such as A2C, DDPG and TD3; prediction-augmented models like SARL; and relation-aware frameworks like GPM, StockFormer and RL-TVDT, which use graph neural networks to model cross-asset interdependencies. However, automated portfolio optimization faces persistent challenges from market non-stationarity. Market phase transitions induce sharp distributional shifts in return distributions, volatility levels, and cross-asset correlations. Because existing RL and relation-aware models learn representations solely from local historical windows, they overfit phase-specific short-term patterns and market noise, leading to severe performance drops when market regimes transition. To address these shortcomings, the authors of [1] propose DR-GRL-PO (Distribution-Robust Graph Representation Learning for Portfolio Optimization). It establishes market-phase-invariant cross-asset structural priors across three sequential modules: ◆ 1). MPIGCE (Market-Phase Invariant Graph Contrastive Encoding): Builds Spearman rank correlation graphs and learns invariant representations via temporal graph contrastive pretraining, pseudo-environment inference and adversarial learning with gradient reversal. ◆ 2). DRPC (Distribution-Robust Predictive Coding): Incorporates structural priors to condition dual-scale predictive branches, enforcing invariant ranking consistency to regularize cross-sectional return rankings. ◆ 3). PPL (Portfolio Policy Learning): Fuses structural priors, dual-scale predictive states, and current holdings via cascaded attention into a unified decision state optimized using Soft Actor-Critic (SAC). DR-GRL-PO was evaluated on CSI-300, NASDAQ-100 and Cryptocurrency datasets against traditional baselines, EW (Equal-Weighted), Index B&H, RP (Risk Parity), Cost-Aware, Min-Var, observation-based RL (DDPG, TD3, A2C) and relation-aware architectures (StockFormer, RL-TVDT). Across 5 core metrics, Wealth Ratio (WR), Annual Return (AR), Sharpe Ratio (SR), Maximum Drawdown (MDD) and Omega Ratio (OR). The proposed DR-GRL-PO achieved superior WR, AR and SR across all markets, alongside superior downside-risk control in the crypto space. It outperforms baselines because its distribution-robust priors prevent policies from overfitting local market noise, ensuring reliable decision-making across market phases. The link to the paper [1] is posted in the comments.

Sione PaluOct 8, 2026

Coronary artery (CA) lesion assessment via Coronary Computed Tomography Angiography (CCTA) involves identifying atherosclerotic plaques, characterizing their composition (calcified, non-calcified, mixed), and grading luminal stenosis. It supports CAD diagnosis, risk stratification, and revascularization. However, manual interpretation is labor-intensive and subject to inter-observer variability, motivating automated diagnosis. Deep learning approaches use CNNs on multi-planar reformatted (MPR) or curved planar reformation (CPR) images, 3D patches, and mesh priors. Recurrent CNNs, graph convolutional networks (GCNs), and Transformers have also been explored to capture vessel structure. Despite the advancements of current models, 3 key challenges remain: • (1) maintaining long-range anatomical and hemodynamic dependencies across tortuous vessels, e.g., distinguishing pathological stenosis from normal tapering; • (2) decoupling plaque texture from luminal geometry amid artifacts such as calcium blooming; • (3) scarcity of fine-grained segment annotations and ambiguity in exploiting coarse branch-level reports. To address these limitations, the authors of [1] introduce MG-HGLNet with three modules: • TDE (Topology-Aware Dual-Stream Encoding): Uses bidirectional vessel Mamba (BiV-Mamba) to capture global hemodynamic context with linear complexity, while spatial-aware cross-attention corrects CPR distortions using 3D coordinates. • SSD (Synergistic Spectral-Morphological Decoupling): Separates plaque composition from luminal narrowing using frequency-domain spectral fingerprints via FB-Mamba to guide texture-guided deformable boundary attention (TG-DBA) for precise lumen delineation. • MSO (Mixed-Grained Supervision Optimization): Combines anatomy-aware dynamic prototypes with logical consistency constraints, penalizing illogical “stenosis without plaque” predictions, to learn from coarse branch labels. MG-HGLNet fuses multi-modal 2D planar sequences, 3D volumetric patches, and centerline coordinates into a unified topology-aware representation. Bidirectional Mamba scanning models hemodynamic context along vessel paths, while frequency-domain analysis isolates calcium blooming artifacts and guides deformable attention to delineate blurred luminal boundaries. Against baseline models (Tejero-de Pablos et al., Denzinger et al., Zreik et al., Zhang et al. [2022], Van Herten et al., Ma et al., Le et al., and Zhang et al. [2025]), MG-HGLNet achieved state-of-the-art performance across Accuracy, Precision, Sensitivity, Specificity, and F1-score. It achieved 92.4% accuracy (91.8% F1) for stenosis grading and 91.5% accuracy (90.6% F1) for plaque classification, outperforming all baselines. The gains stem from BiV-Mamba’s bidirectional modeling and SSD’s spectral decoupling, suppressing calcium blooming artifacts. The link to the paper [1] is posted in the comments.