Overview
The Longest Filled Common Subsequence (LFCS) problem, categorized as NP-hard, was investigated. This problem has documented applications in bioinformatics, specifically in areas such as gene mutation prediction and genomic data reconstruction. Current evaluation practices for existing LFCS approaches, which include exact, metaheuristic, and approximation algorithms, have primarily focused on smaller instances. This limits the insights available regarding their scalability.
This work introduces a novel benchmark dataset characterized by significantly larger problem instances than those previously utilized. The authors contend that existing datasets lack the necessary discriminative power to adequately assess algorithmic performance at scale. To address the computational challenge posed by larger instances, an adaptive Construct, Merge, Solve, Adapt (CMSA) framework was developed. This framework is designed to iteratively generate and refine promising subproblems, utilizing feedback from prior iterations.
The research also presents an engineering contribution: a novel application of LFCS to the identification of songs from degraded audio excerpts. This application employs real-world energy-profile instances derived from popular music. Furthermore, an empirical explainability analysis was conducted to identify critical feature combinations influencing algorithm performance, shedding light on problem features linked to the success or failure of various approaches across different instance types.
Research Context
The Longest Filled Common Subsequence (LFCS) problem is recognized as NP-hard. Its relevance extends to bioinformatics, where it is applicable to tasks such as gene mutation prediction and genomic data reconstruction. Prior methods for solving LFCS, encompassing exact, metaheuristic, and approximation algorithms, have typically been evaluated using small problem instances. This established evaluation paradigm offered limited insight into the scalability of these algorithms when confronted with larger, more complex data structures.
Approach
The research methodology involved several key components. First, a new benchmark dataset was introduced, comprising significantly larger problem instances compared to those used in previous studies. This dataset was designed to overcome the limitations of existing datasets, which were deemed insufficient for discriminating algorithm performance at scale.
To efficiently solve these large instances, an adaptive Construct, Merge, Solve, Adapt (CMSA) framework was employed. The CMSA framework operates by iteratively generating promising subproblems through a component-based construction process. These subproblems are subsequently refined using feedback derived from prior iterations. Subproblems within the CMSA framework are solved using an external black-box solver.
For the application component, the research utilized real-world energy-profile instances. These instances were derived from popular music and were central to the proposed novel application of LFCS for song identification from degraded audio excerpts. An empirical explainability analysis was also performed. This analysis aimed to identify combinations of features critical to algorithm performance, thereby revealing which problem characteristics correlated with the success or failure of different approaches across various instance types.
Findings
Extensive experiments were conducted on both standard and the newly introduced set of large-sized instances. The adaptive CMSA framework demonstrated state-of-the-art performance, achieving superior results compared to four other known approaches. Specifically, among 1,510 problem instances for which optimal solutions were known, the adaptive CMSA approach successfully matched the proven optimum on 1,486 instances. This corresponds to an accuracy rate of 98.4% for these specific instances. These results indicate a substantial improvement in scalability when dealing with large problem instances.
Why This Matters
The development of a new benchmark dataset for larger instances of the Longest Filled Common Subsequence (LFCS) problem provides a more robust tool for evaluating algorithm scalability, addressing a limitation of prior datasets. The demonstrated state-of-the-art performance of the adaptive CMSA framework on these larger instances suggests improved practical applicability for NP-hard problems in bioinformatics. The novel application of LFCS to song identification from degraded audio excerpts introduces a new potential domain for this computational problem, using real-world data derived from popular music.
Potential Applications
- Song identification from degraded audio excerpts, using real-world energy-profile instances derived from popular music.
- Bioinformatics applications, including gene mutation prediction and genomic data reconstruction, leveraging improved scalability for larger instances.