Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Neuro-Symbolic Framework for Sequence Classification with Relational and Temporal Knowledge

Luca Salvatore Lorello, Marco Lippi, Stefano Melacci

TL;DR

This work tackles sequence classification when background knowledge spans relational and temporal dimensions. It introduces LTLZinc, a benchmarking framework that generates multi-channel, temporally constrained sequences and provides datasets to evaluate neural-only and neuro-symbolic pipelines. A multi-stage architecture combines neural image classification with symbolic constraint reasoning and temporal state prediction, enabling end-to-end evaluation through $LTL_f$-based temporal logic. Key findings show that while symbolic-symbolic configurations often outperform neural setups, neuro-symbolic methods face training instabilities and the relational-temporal coupling remains challenging, underscoring the need for tighter integration and robust time-aware reasoning in real-world tasks.

Abstract

One of the goals of neuro-symbolic artificial intelligence is to exploit background knowledge to improve the performance of learning tasks. However, most of the existing frameworks focus on the simplified scenario where knowledge does not change over time and does not cover the temporal dimension. In this work we consider the much more challenging problem of knowledge-driven sequence classification where different portions of knowledge must be employed at different timesteps, and temporal relations are available. Our experimental evaluation compares multi-stage neuro-symbolic and neural-only architectures, and it is conducted on a newly-introduced benchmarking framework. Results demonstrate the challenging nature of this novel setting, and also highlight under-explored shortcomings of neuro-symbolic methods, representing a precious reference for future research.

A Neuro-Symbolic Framework for Sequence Classification with Relational and Temporal Knowledge

TL;DR

This work tackles sequence classification when background knowledge spans relational and temporal dimensions. It introduces LTLZinc, a benchmarking framework that generates multi-channel, temporally constrained sequences and provides datasets to evaluate neural-only and neuro-symbolic pipelines. A multi-stage architecture combines neural image classification with symbolic constraint reasoning and temporal state prediction, enabling end-to-end evaluation through -based temporal logic. Key findings show that while symbolic-symbolic configurations often outperform neural setups, neuro-symbolic methods face training instabilities and the relational-temporal coupling remains challenging, underscoring the need for tighter integration and robust time-aware reasoning in real-world tasks.

Abstract

One of the goals of neuro-symbolic artificial intelligence is to exploit background knowledge to improve the performance of learning tasks. However, most of the existing frameworks focus on the simplified scenario where knowledge does not change over time and does not cover the temporal dimension. In this work we consider the much more challenging problem of knowledge-driven sequence classification where different portions of knowledge must be employed at different timesteps, and temporal relations are available. Our experimental evaluation compares multi-stage neuro-symbolic and neural-only architectures, and it is conducted on a newly-introduced benchmarking framework. Results demonstrate the challenging nature of this novel setting, and also highlight under-explored shortcomings of neuro-symbolic methods, representing a precious reference for future research.
Paper Structure (24 sections, 2 equations, 3 figures, 1 table)

This paper contains 24 sections, 2 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Linear temporal logic over finite traces
  4. Neuro-symbolic integration of logic formulas
  5. Related Works
  6. Neuro-symbolic temporal reasoning
  7. Neuro-symbolic finite state automata
  8. Relationships to this paper
  9. Sequence Classification with Relational and Temporal Knowledge
  10. Problem definition
  11. The LTLZinc Framework
  12. Methodology
  13. (ic) Image classification
  14. (cc) Constraint classification
  15. (nsp) Next state prediction
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Stages of our architecture. We implemented cc and nsp in multiple ways, that we listed by means of the green (symbolic) and reddish (neural) dashed-borders sub-blocks.
  • Figure 2: Q2. cc- nsp accuracy trade-off for different families of architectures (i.e., Neural/Symbolic cc, Neural/Symbolic nsp, where Neural nsp is indicated either by MLP or GRU, to compare them as well). The horizontal dashed line indicates the performance of a deterministic baseline of nsp always choosing the successor state most represented in the training set.
  • Figure 3: Q3. Accuracies for Task 4 with oracular predictors. Oracle types (perfect, flip, confidence) are described in Section \ref{['sec:exp']}.