Leveraging Co-Occurrence to Improve Deep Learning Photo-Identification in Social Animals

Photo-identification underpins individual-based inference in numerous ecological studies, but scaling it to decades-long archives remains limited by expert time. Deep learning can accelerate matching, yet most pipelines treat photographs as independent observations and therefore ignore a key aspect of the data collection method: individuals are recorded in structured encounters and often exhibit persistent, non-random associations. We present a model agnostic, encounter-level identification procedure that incorporates social context as a deployable probabilistic component. Given per-image classifier posteriors, we perform log-linear fusion of three information sources: (i) image-based probabilities, (ii) global sighting priors (class frequency), and (iii) an encounter-conditioned context term derived from historical co-occurrence (log lift). The method operates as lightweight post-processing and requires no retraining or architectural changes to the image model. Using a longitudinal photo-identification dataset as a case study (West Coast Transient Bigg's killer whales), we evaluate (a) expert-assisted settings in which a small number of individuals present in an encounter are known without image-level labels, and (b) fully automated settings that initialize context from the model's own high-confidence predictions. On a strict temporal holdout (newest 10%), encounter-context fusion reduces top-1 error by ~14%-25% with expert-assisted seeding; a fully automated variant yields up to ~24% fewer misidentifications once sufficient training history exists, improving Macro-F1 by +0.088 to +0.104, with minimal computational overhead. Placebo and seed-corruption controls confirm that gains depend on meaningful co-occurrence structure and collapse when encounter context is destroyed. By turning encounter structure into a reusable probabilistic component, this work bridges established methods for analyzing animal societies with practical, scalable photo-identification pipelines. The approach is applicable to any system where individuals are repeatedly observed in groups (e.g., cetaceans, primates, ungulates, camera-trap bursts) and provides a transparent mechanism to incorporate social context into automated identification.