Flexible constrained sampling with guarantees for pattern mining

Abstract

Pattern sampling has been proposed as a potential solution to the infamous pattern explosion. Instead of enumerating all patterns that satisfy the constraints, individual patterns are sampled proportional to a given quality measure. Several sampling algorithms have been proposed, but each of them has its limitations when it comes to (1) flexibility in terms of quality measures and constraints that can be used, and/or (2) guarantees with respect to sampling accuracy. We therefore present Flexics, the first flexible pattern sampler that supports a broad class of quality measures and constraints, while providing strong guarantees regarding sampling accuracy. To achieve this, we leverage the perspective on pattern mining as a constraint satisfaction problem and build upon the latest advances in sampling solutions in SAT as well as existing pattern mining algorithms. Furthermore, the proposed algorithm is applicable to a variety of pattern languages, which allows us to introduce and tackle the novel task of sampling sets of patterns. We introduce and empirically evaluate two variants of Flexics: (1) a generic variant that addresses the well-known itemset sampling task and the novel pattern set sampling task as well as a wide range of expressive constraints within these tasks, and (2) a specialized variant that exploits existing frequent itemset techniques to achieve substantial speed-ups. Experiments show that Flexics is both accurate and efficient, making it a useful tool for pattern-based data exploration.

Appendix: WeightGen

In this section, we present an extended technical description of the WeightGen algorithm, which closely follows Sections 3 and 4 in Chakraborty et al. (2014), whereas the pseudocode in Algorithm 2 is structured similarly to that of UniGen2, a close cousin of WeightGen (Chakraborty et al. 2015). Lines 1–3 correspond to the estimation phase and Lines 4–8 correspond to the sampling phase. SolveBounded stands for the bounded enumeration oracle.

The parameters of the estimation phase are fixed to particular theoretically motivated values. $\textit{pivot}_{est}$ denotes the maximal weight of a cell at the estimation phase; $\textit{pivot}_{est}=46$ corresponds to estimation error tolerance $\varepsilon _{est}=0.8$ (Line 10). If the total weight of solutions in a given cell exceeds $\textit{pivot}_{est}$, a new random XOR constraint is added in order to eliminate a number of solutions. Repeating the process for a number of iterations increases the confidence of the estimate, e.g., 17 iterations result in $1-\delta _{est}=0.8$ (Line 1). Note that Estimate essentially estimates the total weight of all solutions, from which $N_{\textit{XOR}}$, the initial number of XOR constraints for the sampling phase, is derived (Line 4).

A similar procedure is employed at the sampling phase. It starts with $N_{\textit{XOR}}$ constraints and adds at most three extra constraints. The user-chosen error tolerance parameter $\kappa$ determines the range $\left[ \textit{loThresh},\ \textit{hiThresh}\right]$, within which the total weight of a suitable cell should lie (Line 5). For example, $\kappa =0.9$ corresponds to range $\left[ 6.7,\ 49.4\right]$. If a suitable cell can be obtained, a solution is sampled exactly from all solutions in the cell; otherwise, no sample is returned. Requiring the total cell weight to exceed a particular value ensures the lower bound on the sampling accuracy.

The preceding presentation makes two simplifying assumptions: (1) all weights lie in $\left[ 1/r,\ 1\right]$; (2) adding XOR constraints never results in unsatisfiable subproblems (empty cells). The former is relaxed by multiplying pivots by $\hat{w}_{\textit{max}} = \hat{w}_{\textit{min}} \times \hat{r} < 1$, where $\hat{w}_{\textit{min}}$ is the smallest weight observed so far. The latter is solved by simply restarting an iteration with a newly generated set of constraints. See Chakraborty et al. (2014) for the full explanation, including the precise formulae to compute all parameters.

1.1 Implementation details

Following suggestions of Chakraborty et al. (2015), we implement leapfrogging, a technique that improves the performance of the umbrella sampling procedure and thus benefits both GFlexics and EFlexics. First, after three iterations of the estimation phase, we initialize the following iterations with a number of XOR constraints that is equal to the smallest number returned in the previous iterations (rather than with zero XORs). Second, in the sampling phase, we start with one XOR constraint more than the number suggested by theory. If the cell is too small, we remove one constraint; if it is too large, we proceed adding (at most two) constraints. Both modifications are based on the observation that theoretical parameter values address hypothetical corner cases that rarely occur in practice. Finally, we only run the estimation phase until the initial number of XOR constraints, which only depends on the median of total weight estimates, converges. For example, if the estimation phase is supposed to run for 17 iterations, the convergence can happen as early as after 9 iterations.