Seqret: Mining Rule Sets from Event Sequences

TL;DR

The paper tackles the problem of mining conditional dependencies in event sequences by formulating sequential rules $X \rightarrow Y$ within a Minimum Description Length framework. It introduces Seqret, a greedy approach with two variants (Seqret-Candidates and Seqret-Mine) that discover succinct, non-redundant rule sets by balancing model cost and data likelihood, and by using gap-aware rule windows and significance-guided rule generation. Extensive experiments on synthetic and real-world datasets demonstrate Seqret’s ability to recover ground truth, resist noise, and produce interpretable rules that surpass state-of-the-art baselines in compression and usefulness. The work provides practical algorithms and open-source resources, with potential extensions toward causal modeling and deeper structural analysis of event sequences.

Abstract

Summarizing event sequences is a key aspect of data mining. Most existing methods neglect conditional dependencies and focus on discovering sequential patterns only. In this paper, we study the problem of discovering both conditional and unconditional dependencies from event sequence data. We do so by discovering rules of the form $X \rightarrow Y$ where $X$ and $Y$ are sequential patterns. Rules like these are simple to understand and provide a clear description of the relation between the antecedent and the consequent. To discover succinct and non-redundant sets of rules we formalize the problem in terms of the Minimum Description Length principle. As the search space is enormous and does not exhibit helpful structure, we propose the Seqret method to discover high-quality rule sets in practice. Through extensive empirical evaluation we show that unlike the state of the art, Seqret ably recovers the ground truth on synthetic datasets and finds useful rules from real datasets.

Figures (11)

• Figure 1: Toy example of a rule $ab \rightarrow cd$ in an event sequence. Each occurrence of head $ab$ triggers the rule. The first is followed by tail $cd$ and hence a 'hit' whereas the second is not and hence a 'miss'. $\mathit{supp}(ab \rightarrow cd)\xspace = 1$ and $\mathit{conf}(ab \rightarrow cd)\xspace = 0.5$.
• Figure 2: Toy example showing an encoding of sequence $S$ using rule set $R$. The encoding consists of three code streams. $C_t$ encodes if a triggered rule hits or misses. $C_d$ encodes the delay between the trigger and the rule tail. $C_g$ encodes the gaps in the tails. Together, they form a cover $C$ of $D$ given $R$.
• Figure 3:
• Figure 4:
• Figure 5: