Does a historical catch record support a stable latent structure in catch composition across ports, taxa, and years? An assessment of cooperative-level catch statistics from Lake Biwa, Japan, 1962–1972

Historical inland-fisheries records resolved by taxon, location, and year are rare, and whether such records support stable low-dimensional structure is an empirical question. We analyzed annual catch records from Lake Biwa, Japan, covering 44 fisheries cooperatives, 13 taxonomic categories, and 11 years (1962-1972), immediately preceding major coastal modifications associated with the Lake Biwa Comprehensive Development Project. We assessed whether non-negative matrix factorization (NMF), with and without spatial smoothing based on port coordinates, provided a stable representation of port-year catch composition beyond direct summaries. Candidate ranks Q = 1-8 were evaluated using bi-cross-validation, 30 random starts, consensus diagnostics, port-block bootstrap resampling, exclusion analyses, and sensitivity analyses. Bi-cross-validation reached its minimum at Q = 8, with the one-standard-error rule selecting Q = 7, whereas component correspondence across random starts deteriorated sharply at Q ≥ 6. Because these criteria did not agree, Q = 4 was used as an operational reference under a study-specific stability rule rather than as a uniquely selected optimum. At Q = 4, three components dominated by Crucian carp, Freshwater clam, and Landlocked Ayu were reproducible and closely tracked the corresponding observed shares, whereas a mixed fourth component was less stable. Repeating the stability analyses at Q = 5 produced the same qualitative pattern of a reproducible core and less stable additional structure. Component correspondence also depended on data representation. Some observed shares and component scores showed spatial autocorrelation among reporting ports, whereas reconstruction error did not. NMF therefore identified a reproducible core representation of recorded catch composition without supporting a unique and uniformly stable latent structure. These patterns describe recorded catch composition rather than stock abundance and provide a historical reference for future comparison.