How JUMP Boosts Session Click‑Through and Dwell Time with a Triple‑Layer RNN

Abstract

JUMP uses a novel three‑layer RNN structure to encode a user session, employing a fast‑slow layer to alleviate short‑session issues and an attention layer to mitigate session noise. Extensive experiments show that JUMP outperforms state‑of‑the‑art methods in both click‑through rate and dwell‑time prediction.

Dwell Time Estimation

The paper adopts survival‑analysis ideas, treating the time a user stays on content as the occurrence time of a “leave” event. By mapping observed dwell times O to a simple distribution f (e.g., Gaussian or Gamma) and using maximum‑likelihood estimation, the model approximates the dwell‑time distribution, which after log‑transformation follows a normal distribution.

Model Learning

Each session is represented as a sequence of item‑click and dwell‑time pairs. The joint likelihood factorizes into a click‑through component and a dwell‑time component. JUMP employs a three‑layer RNN: an Embedding layer, a Fast‑Slow layer, and an Attention layer.

Attention Layer

The attention mechanism assigns weights a to the outputs of the fast‑slow layer, suppressing noisy interactions.

Fast‑Slow Layer

The fast‑slow layer processes each input through a unit containing a fast sub‑cell and a slow sub‑cell; the slow sub‑cell captures long‑term dependencies while the fast sub‑cell focuses on recent information.

Embedding Layer

The bottom layer maps each (item, dwell‑time) tuple to a vector and applies batch normalization.

Experiments

JUMP is compared with mainstream session‑based prediction models (GRU, IGRU, NARM, DTGRU, RMTP, ATRP, NSR) on RecSys15, CIKM16, and Reddit datasets. Results show JUMP achieves the highest Recall, MRR, and NDCG for click‑through prediction and superior accuracy for dwell‑time estimation.

Conclusion

The proposed JUMP algorithm jointly predicts click‑through rate and dwell time using survival analysis for time modeling, a three‑layer RNN with attention to improve robustness, and a fast‑slow structure to enhance learning from short sessions. Extensive experiments demonstrate significant improvements over existing methods on multiple public datasets.