Today, I will talk about a recent article that I have participated to as co-author, which has been published in the Applied Intelligence journal a few days ago.
I. Hamimid, F. Nouioua, S. Boutouhami, and P. Fournier-Viger, “Probabilistic periodic frequent pattern mining in uncertain databases,” Applied Intelligence, vol. 56, article 469, 2026. DOI: 10.1007/s10489-026-07388-7
This paper is about algorithms for discovering patterns in data, and more specifically how to find periodic patterns in uncertain transactional data.
To motivation for this paper is as follows. Over the decades, there has been lot of work on frequent pattern mining, which focuses on identifying combinations of items (elements) that occur frequently in transactional data (data represented as records). However, frequency alone does not capture an important property of many datasets: when a pattern occurs. Consider, for example, a collection of transactions representing purchases over time. A group of products might appear frequently, but its occurrences could be scattered irregularly throughout the database. Another group might occur at relatively regular intervals, revealing a recurring behavior. To address this issue a subfield of pattern mining has focused on designing algorithms and techniques for identifying periodic patterns, that is patterns that regularly appear over time.
However, most algorithm for periodic pattern mining consider that the data is certain. But many real-world datasets are inherently uncertain. For instance, sensor observations may be noisy or incomplete, transaction records may only indicate that an item has a certain probability of being present, and data collected from complex environments often cannot be treated as completely deterministic. Uncertain data raises an interesting question for frequent pattern mining: what does it mean for a pattern to be both frequent and periodic when the underlying data itself is uncertain? This paper investigates this problem and proposes a new approach for mining probabilistic periodic frequent patterns in uncertain transactional databases.
In an uncertain transactional database, the presence of an item is not necessarily known with certainty. Instead, an item can be associated with an existential probability representing the probability that the item actually occurs in that transaction. For example, instead of representing a transaction simply as a set of elements such as {A, B, C} we might have probabilities such as {A, B, C} with 90% probability that A occurred, 70% that B occurred, and 70% that C occurred. Consequently, the database does not represent one completely known dataset. Under possible-world semantics, it can be viewed as representing a large collection of possible deterministic databases, each corresponding to one possible realization of the uncertain information.
Existing approaches to periodic frequent pattern mining over uncertain data have explored different ways of dealing with uncertainty. For example, methods based on expected support can be computationally efficient, but they do not directly quantify the probability that a pattern is periodic. Other approaches have considered possible-world semantics, but can rely on simplifying assumptions such as transaction-level uncertainty instead of item-level uncertainty.
The paper addresses this gap by establishing a framework in which the probability of a pattern being periodic is explicitly quantified. The proposed framework models fine-grained item-level existential probabilities under the standard independence assumption used in uncertain itemset mining. This leads to a more meaningful interpretation of a result: rather than simply saying that a pattern has a certain expected support, we can ask how likely it is that the pattern actually satisfies the periodicity requirements across the possible worlds represented by the uncertain database.
In the paper, an algorithm named PPFP-Growth is presented within the proposed framework. PPFP-Growth therefore incorporates a novel decomposition method to process the possible worlds efficiently while computing the probabilistic periodicity of candidate patterns.
We evaluated PPFP-Growth on five benchmark datasets under controlled item-level uncertainty settings and compared its performance with an existing expected-support-based approach, including UPFP-Growth++. The experiments examine the practical computational cost of obtaining probabilistic guarantees for periodicity. The results show that the proposed approach can extract probabilistic periodic frequent patterns while quantifying the additional computational cost associated with explicitly modeling the probability of periodicity. Thus, the work highlights an important trade-off in uncertain pattern mining: obtaining a richer probabilistic interpretation can require additional computation, but it can also provide information that an expected-value approach cannot provide.
Conclusion
This is just a brief overview of this paper. You may see the paper for more details!
