Multi-task news engagement prediction via bidirectional LSTM and adaptive attention

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Current news engagement behavior prediction models are limited by unidirectional sequential dependencies and insufficient bidirectional contextual information capture, resulting in insufficient fusion of long- and short-term preferences and difficulty adapting to rapidly fluctuating user interests and breaking news scenarios. To improve prediction accuracy and interpretability, this article constructs a prediction model based on a bidirectional multi-head adaptive attention mechanism. The model first utilizes a bidirectional LSTM (Long Short-Term Memory) to encode the context of a user’s historical news sequence, balancing long-term stable preferences with short-term dynamic interests. Subsequently, a multi-head adaptive attention mechanism is applied to parallelize the fine-grained interactions between the user sequence and candidate news in eight independent subspaces. An adaptive weighting mechanism is constructed, dynamically adjusting the contribution of each attention head based on the temporal intensity of the sequence and the diversity of the news, improving sensitivity to breaking news and interest transitions. Finally, multi-task learning is used to simultaneously predict click-through rate (CTR) and dwell time, achieving synergistic gains and sharing useful latent features, enabling refined modeling of user interest evolution and content preferences. Experimental results show that the model in this paper achieves an AUC of 0.786 ± 0.002 on the MIND dataset, which is statistically significant (p < 0.01) compared with the best baseline model LSTUR (0.781 ± 0.003); the MRR reaches 0.352 ± 0.003, nDCG@5 is 0.263 ± 0.004, and nDCG@10 is 0.314 ± 0.003, all of which are better than all the comparison models. Consistent performance advantages are also achieved on the Adressa dataset. The research concludes that the proposed method effectively addresses the shortcomings of existing models in capturing dynamic interests and integrating long- and short-term preferences, and provides a reference for building personalized news recommendation systems with time-series awareness and adaptive capabilities.