Accurately predicting stock prices remains a challenging task due to the volatile and complex nature of financial markets. In this study, we propose a novel hybrid deep learning framework that integrates a large language model (LLM), a Linear Transformer (LT), and a Convolutional Neural Network (CNN) to enhance stock price prediction using solely historical market data. The framework leverages the LLM as a professional financial analyst to perform daily technical analysis. The technical indicators, including moving averages (MAs), relative strength index (RSI), and Bollinger Bands (BBs), are calculated directly from historical stock data. These indicators are then analyzed by the LLM, generating descriptive textual summaries. The textual summaries are further transformed into vector representations using FinBERT, a pre-trained financial language model, to enhance the dataset with contextual insights. The FinBERT embeddings are integrated with features from two additional branches: the Linear Transformer branch, which captures long-term dependencies in time-series stock data through a linearized self-attention mechanism, and the CNN branch, which extracts spatial features from visual representations of stock chart data. The combined features from these three modalities are then processed by a Feedforward Neural Network (FNN) for final stock price prediction. Experimental results on the S&P 500 dataset demonstrate that the proposed framework significantly improves stock prediction accuracy by effectively capturing temporal, spatial, and contextual dependencies in the data. This multimodal approach highlights the importance of integrating advanced technical analysis with deep learning architectures for enhanced financial forecasting.

Deep learning on graphs and especially graph convolutional networks (GCNs) have shown superior performance in collaborative filtering. Most GCNs have a message-passing architecture that enables nodes aggregate information from neighbours continuously through multiple layers. This also leads to a common issue called over-smoothing: as the number of layers increases, the node embeddings become similar and the performance of tasks such as link prediction degrades severely. In this work, we propose motif-based graph attention network for web service recommendation (MGSR) that alleviates the over-smoothing issue by incorporate network motifs in layer propagation. Extensive experiments show that our model outperforms state-of-the-art approaches.

Graph motifs capture crucial structural patterns within user–item interaction graphs and offer meaningful semantics that are typically underutilised in conventional graph-based collaborative filtering. Existing contrastive learning methods in recommender systems rely primarily on simple perturbations, which limits their ability to leverage deeper motif-based structural information. To address this gap, we propose a novel motif-guided contrastive learning framework for recommender systems. To exploit complementary structural biases inherent to user–item interactions, our approach explicitly incorporates three distinct motifs into the construction of the contrastive view. By contrasting views that capture the structural differences among different motifs, our model learns meaningful group-level relationships and potentially suppresses noise arising from sparse or isolated interactions. Extensive experiments on four real-world recommendation datasets validate that our motif-based contrastive approach achieves the best overall performance, outperforming state-of-the-art baselines on three of the four benchmarks with statistically significant margins on the more skewed datasets, while remaining competitive on the fourth, which demonstrates notable robustness and improved accuracy.

The prediction of stock market price trends has always been a challenging issue, attracting widespread attention from both economists and computer scientists. Recently, integrating stock prices with news data has been shown to be an effective strategy for enhancing the accuracy of the prediction task. Yet, many current methods fail to fully leverage the intricate inter-stock relationships inherent in stock news.Applying deep learning, especially Graph Convolutional Networks (GCNs), to predict stock trends has demonstrated advanced performance. This method employs a message-passing architecture, enabling nodes to progressively aggregate information from neighboring nodes across multiple layers. In this paper, we propose a novel approach: the Motif-based Graph Convolutional Network for Stock Prediction (MGCN-SP). This model mitigates the over-smoothing problem by incorporating network motifs into the layer propagation process. Specifically, we first generate a motif graph by correlating stocks with stock news. Then, we encode the stock price information and stock news into features using the scaled dot product attention adapted from the transformer architecture. After that, we apply the motif-based graph convolutional network. This framework is designed to jointly refine the embeddings of both stock news and stock time series data using a transformer encoder and to estimate the likelihood of target movements. Finally, we conducted extensive implementation in the U.S. stock market, and the experimental results demonstrate that our method outperforms some state-of-the-art approaches.