HySent-Trans: A Hybrid Deep Learning Framework for Sentiment Analysis Stock Market Prediction and Job Recommendation

Abstract

 

Introduction In the era of pervasive digital communication social media platforms like Twitter Reddit and LinkedIn generate vast unstructured data streams rich in public opinions sentiments and behavioral insights Traditional sentiment analysis struggles with linguistic complexities such as sarcasm slang and domain specific jargon limiting its efficacy in real time applications like stock market prediction and job recommendations HySent Trans addresses these gaps through a novel hybrid deep learning framework integrating CNN BiLSTM GRU and Transformer architectures with FinBERT embeddings and BERTopic modeling for multi source cross domain analysis Evaluated on 250000 plus posts and financial datasets it achieves 92 4 percent sentiment accuracy RMSE 1 83 for stocks and 89 6 percent Top 5 job recommendation precision outperforming baselines like CNN GRU

Objectives This research proposes HySent Trans a hybrid deep learning framework integrating CNN BiLSTM Transformer for sentiment analysis stock prediction and job recommendation using multi source social media and financial data

Methods The HySent Trans framework integrates multi source data from Twitter Reddit LinkedIn and financial time series OHLCV for sentiment analysis stock prediction and job recommendation Raw text undergoes preprocessing via lowercasing stop word removal lemmatization and sentiment scoring with VADER TextBlob and FinBERT followed by time alignment with market and job data Feature engineering generates FinBERT embeddings 768 dim BERTopic topic modeling 512 dim numeric normalization via MinMaxScaler and lag based temporal features like moving averages The hybrid architecture stacks CNN layers for local pattern extraction BiLSTM for sequential modeling of trends and Transformer encoders for long range dependencies culminating in a fusion layer for multimodal integration Models train with Adam optimizer on 80 20 splits across 5 to 30 day windows achieving 92 4 percent accuracy 0 905 F1 score and 1 83 RMSE with ablation studies confirming 4 7 percent gains from hybrid fusion

Results HySent Trans a hybrid CNN BiLSTM Transformer framework excels in sentiment analysis with 92 4 percent accuracy and F1 score 0 905 outperforming CNN GRU 87 2 percent and RNN 85 1 percent on 250K social media posts from Twitter Reddit and LinkedIn Stock predictions yield RMSE 1 83 better than baselines 2 3 to 2 6 using 5 year OHLCV data Top 5 job recommendations achieve 89 6 percent accuracy surpassing collaborative filtering 81 2 percent Ablation shows 4 7 percent accuracy gain from hybrid fusion and 67 percent retraining reduction via cross domain transfer enabling scalable real time finance and HR decisions

Conclusions HySent Trans advances sentiment analysis stock prediction and job recommendation through its hybrid CNN BiLSTM Transformer architecture achieving 92 4 percent accuracy RMSE 1 83 and 89 6 percent Top 5 job accuracy outperforming baselines by 8 to 12 percent Its cross domain transfer learning reduces retraining by 67 percent enabling scalable real time applications in finance and HR Future multimodal extensions promise broader impact.