Heterogeneous Interactive Snapshot Network for Review-Enhanced Stock Profiling and Recommendation

Heyuan Wang, Tengjiao Wang, Shun Li, Shijie Guan, Jiayi Zheng, Wei ChenIJCAI 2022 · DOI

Abstract

Stock market prediction is a complex task due to its stochastic nature and the significant influence of investor sentiment. Traditional models often treat technical price data and social media sentiment as isolated streams, failing to capture their dynamic interplay. This paper introduces the Heterogeneous Interactive Snapshot Network (HISN), a framework that models daily stock reviews as knowledge-rich heterogeneous graphs and employs a novel "twins-GRU" architecture. This approach enables cross-modal interaction between price and sentiment data over time, leading to a 7.6% improvement in investment returns over state-of-the-art methods.

Introduction

Predicting stock movements requires balancing objective technical indicators with subjective market sentiment. While "text-enhanced" models have attempted to incorporate social media data, they often suffer from semantic isolation—ignoring relationships between entities and topics—and modality separation, where price and text are processed in silos. HISN addresses these gaps by integrating external knowledge into sentiment analysis and allowing continuous interaction between data modalities.

Heterogeneous Interactive Snapshot Network Architecture

Methodology

The HISN architecture is built upon three core components:

  1. Intra-Day Semantic Encoder: For each trading day, the model constructs a Heterogeneous Document Graph. This graph links review texts to latent topics (via LDA) and specific entities (via Wikipedia linking). A Hierarchical Attentive Propagation (HAP) framework aggregates this information, providing a deeper "understanding" of market discourse.
  2. Inter-Day Temporal Encoder (Twins-GRU): To capture the stock's evolution, the twins-GRU allows the hidden states of price and text sequences to influence one another at every time step. This captures how sentiment shifts in response to price volatility and vice versa.
  3. Temporal Aggregation: An attention mechanism weighs historical data based on recency and salience, ensuring that significant market events are prioritized in the final recommendation.

Experimental Results

The model was evaluated on the US S&P 500 and the Ashare&HK (Chinese) markets. HISN consistently outperformed baselines, particularly in turbulent market conditions.

Key Performance Metrics

  • Profitability: On the Ashare&HK dataset, HISN achieved a 7.6% higher Cumulative Investment Return (CIR) and a 10.2% improvement in the Sharpe Ratio compared to the strongest baseline (FAST).
  • Ablation Studies: Removing the "twins-GRU" or the heterogeneous graph components resulted in significant performance drops, validating the importance of cross-modal interaction and structured semantic knowledge.
Performance Analysis across Lookback Periods and Selected Stocks

Conclusion

HISN demonstrates that stock profiling is significantly enhanced when sentiment is anchored to structured knowledge and allowed to interact dynamically with price data. Future work will focus on the "cold start" problem for new IPOs and exploring inter-stock correlations within supply chains to further refine predictive accuracy.

Summary updated

BibTeX
@inproceedings{wang2022heterogeneous,
  author = {Heyuan Wang and Tengjiao Wang and Shun Li and Shijie Guan and Jiayi Zheng and Wei Chen},
  title = {{Heterogeneous Interactive Snapshot Network for Review-Enhanced Stock Profiling and Recommendation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/550},
}
Figure