CONTENTS

Abstract

To map its input into a feature vector, JUMP adopts a novel three-layered RNN structure which includes a fast-slow layer for very short sessions and an attention layer for noisy sessions.

Introduction

The main contributions are summarized as follows.

  • Dwell time is modeled by survival analysis, which has sound underlying mathematical principles. Based on this, a joint predictor is designed for both user click and dwell time.

  • A novel three-layered RNN structure is constructed to map the session input to feature vector. Instead of forcing it to encode all information into one fixed-length vector, we use an attention layer to allow the network to refer back to the input sequence, which improves the robustness against noise.

  • We also use redesigned fast-slow cells to enhance the performance for very short sessions. In the meanwhile, Recurrent Highway sub-cells are embedded in the fastslow cell to keep gradient flow stable.

Architecture

avatar