Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DSFNet: Learning Disentangled Scenario Factorization for Multi-Scenario Route Ranking

Jiahao Yu, Yihai Duan, Longfei Xu, Chao Chen, Shuliang Liu, Kaikui Liu, Fan Yang, Xiangxiang Chu, Ning Guo

TL;DR

A novel method, Disentangled Scenario Factorization Network (DSFNet), which flexibly composes scenario-dependent parameters based on a high-capacity multi-factor-scenario-branch structure and a novel regularization is proposed to induce the disentanglement of factor scenarios.

Abstract

Multi-scenario route ranking (MSRR) is crucial in many industrial mapping systems. However, the industrial community mainly adopts interactive interfaces to encourage users to select pre-defined scenarios, which may hinder the downstream ranking performance. In addition, in the academic community, the multi-scenario ranking works only come from other fields, and there are no works specifically focusing on route data due to lacking a publicly available MSRR dataset. Moreover, all the existing multi-scenario works still fail to address the three specific challenges of MSRR simultaneously, i.e. explosion of scenario number, high entanglement, and high-capacity demand. Different from the prior, to address MSRR, our key idea is to factorize the complicated scenario in route ranking into several disentangled factor scenario patterns. Accordingly, we propose a novel method, Disentangled Scenario Factorization Network (DSFNet), which flexibly composes scenario-dependent parameters based on a high-capacity multi-factor-scenario-branch structure. Then, a novel regularization is proposed to induce the disentanglement of factor scenarios. Furthermore, two extra novel techniques, i.e. scenario-aware batch normalization and scenario-aware feature filtering, are developed to improve the network awareness of scenario representation. Additionally, to facilitate MSRR research in the academic community, we propose MSDR, the first large-scale publicly available annotated industrial Multi-Scenario Driving Route dataset. Comprehensive experimental results demonstrate the superiority of our DSFNet, which has been successfully deployed in AMap to serve the major online traffic.

DSFNet: Learning Disentangled Scenario Factorization for Multi-Scenario Route Ranking

TL;DR

A novel method, Disentangled Scenario Factorization Network (DSFNet), which flexibly composes scenario-dependent parameters based on a high-capacity multi-factor-scenario-branch structure and a novel regularization is proposed to induce the disentanglement of factor scenarios.

Abstract

Multi-scenario route ranking (MSRR) is crucial in many industrial mapping systems. However, the industrial community mainly adopts interactive interfaces to encourage users to select pre-defined scenarios, which may hinder the downstream ranking performance. In addition, in the academic community, the multi-scenario ranking works only come from other fields, and there are no works specifically focusing on route data due to lacking a publicly available MSRR dataset. Moreover, all the existing multi-scenario works still fail to address the three specific challenges of MSRR simultaneously, i.e. explosion of scenario number, high entanglement, and high-capacity demand. Different from the prior, to address MSRR, our key idea is to factorize the complicated scenario in route ranking into several disentangled factor scenario patterns. Accordingly, we propose a novel method, Disentangled Scenario Factorization Network (DSFNet), which flexibly composes scenario-dependent parameters based on a high-capacity multi-factor-scenario-branch structure. Then, a novel regularization is proposed to induce the disentanglement of factor scenarios. Furthermore, two extra novel techniques, i.e. scenario-aware batch normalization and scenario-aware feature filtering, are developed to improve the network awareness of scenario representation. Additionally, to facilitate MSRR research in the academic community, we propose MSDR, the first large-scale publicly available annotated industrial Multi-Scenario Driving Route dataset. Comprehensive experimental results demonstrate the superiority of our DSFNet, which has been successfully deployed in AMap to serve the major online traffic.
Paper Structure (30 sections, 1 theorem, 14 equations, 5 figures, 6 tables)

This paper contains 30 sections, 1 theorem, 14 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multi-Scenario Ranking
  4. DSFNet vs. MoE & MuSeNet
  5. Method
  6. Scenario Factorization Network
  7. Disentangled Factorization
  8. Overview.
  9. Neuron centroid repulsion.
  10. Contrastive neuron clustering.
  11. Disentangling regularization.
  12. Enhancing Awareness of Scenario
  13. Scenario-aware batch normalization.
  14. Scenario-aware feature filtering.
  15. The Proposed Dataset
  16. ...and 15 more sections

Key Result

Lemma 3.1

When $d\geq N-1$, $\min \max_{i\neq j} {\bar{\mathbf{w}}^{(i)}}\cdot\bar{\mathbf{w}}^{(j)} \Leftrightarrow \forall_{i\neq j}\,\bm{\theta}^{(i,j)} \\ = \arccos(-\frac{1}{N-1}) \Leftrightarrow \min \max_{i\neq j} \Vert \bm{\theta}^{(i,j)}-\arccos(-\frac{1}{N-1})\Vert^2$.

Figures (5)

  • Figure 1: Comparison for the schemes of multi-scenario ranking method. (a) Maintain as many network branches as scenario number and one scenario-specific branch is selected in a forward propagation. (b) Directly generate the dynamic scenario-specific parameters of the network based on the scenario feature $S$. (c) Our method by composing dynamic parameters in a multi-factor-scenario-branch style.
  • Figure 2: The illustration of DSFNet with four FSLs (i.e.$N=4$). In each layer, scenario factorization structure composes the dynamic network parameter $\mathbf{W}(\mathbf{s})$ via gating the four shared factor parameters $\{ \tilde{\mathbf{W}}_i \}_{i=1}^4$ from FSLs with the gates $\alpha(\mathbf{s})$. The disentangling regularization constrains the four factor parameters by $\mathcal{L}_{NCR}$ keeping the neuron centroids $\{ \bar{\mathbf{w}}^{(i)} \}_{i=1}^4$ as far away from each other as possible and $\mathcal{L}_{CNC}$ pushing the directions of the neurons in each $\tilde{\mathbf{W}}_i$ to cluster appropriately (\ref{['subsec:DR']}). SABN and SAFF are integrated to enhance the network awareness of $\mathbf{s}$ (\ref{['subsec:SA']}). Bias is omitted for figure simplicity.
  • Figure 3: The number of samples (#Samples) and proportion of the positives (Pos%) of the 72 scenarios, which are further divided into four subsets (by four background colors), and the statistics of the subsets are exhibited in the table.
  • Figure 4: Visualization results. a) Convergence of our disentangling regularization $\mathcal{L}_{DR}$. b) Sensitivity of two key hyperparameters (i.e.$N$ and $\delta_\theta$), where $\delta_\theta=\frac{1}{\kappa} \arccos(-\frac{1}{6})$ and $\kappa$ varies linearly within each pair of adjacent graduations.
  • Figure 5: Interpretability experimental results, showing the route features each FSL focuses on. The explanations of the feature names are exhibited in \ref{['tab:routefeat']}.

Theorems & Definitions (1)

  • Lemma 3.1