Daejeon, July 27–31
Dimensionality reduction (DR) is now widely used in Digital Humanities (DH) to visualise high dimensional representations: word- and document-embedding spaces, co-citation networks, image feature manifolds, and other forms of vectorised cultural data (González-Márquez et al., 2022; Noichl, 2021, 2023; Oiva et al., 2024; Rodighiero et al., 2022). Yet in much DH practice, DR remains framed as an expedient means to “see” a dataset, producing a map for exploratory browsing. This paper argues that such a positioning is methodologically insufficient for hermeneutic research. When cultural materials are transformed into feature vectors, they become embedded in latent spaces whose topology encodes relations that can exceed, complicate, or contradict established metadata categories.
DR is the operation that renders these latent relations legible by projecting them into two or three dimensions. In doing so, it does not simply reveal structure; it performs a computational mapping that shapes the very conditions under which structure can be interpreted. DR should therefore be approached as an interpretative act within DH research: a form of spatial transduction through which latent structures become available for scholarly reading.
Indeed, within high-dimensional vector spaces, the value of representation lies less in individual dimensions than in the relational organisation of items. Topological configurations give rise to emergent structures that may be latent in the data and obscured by traditional descriptive models. For these configurations to become available to interpretation, they must be mapped into a space that can be visualised. DR provides this mapping as an operation that re-conditions distance into a domain where relative spatial relations can function as meaningful cues for scholarly reading. This reframing matters because DH uses of DR increasingly appear not only as illustrations but as sites where interpretative claims are made (Noichl, 2021, 2023). Yet such readings often proceed without a systematic account of how the projection was produced, how stable it is, how sensitive it is to parameter choices, and what kinds of structure it is designed to preserve. In response, this paper proposes that DH treat DR outputs as constructed interpretative objects whose validity depends on the transparency and robustness of the mapping procedure.
The paper develops this claim through a technical review and evaluation of four prominent non linear DR algorithms: tSNE (van der Maaten / Hinton, 2008), UMAP (McInnes et al., 2018), TriMap (Amid / Warmuth, 2022), and PaCMAP (Wang et al., 2021). The central proposition is that different DR strategies instantiate distinct mappings, each with specific epistemic affordances and limitations. A hermeneutic approach to DR must therefore move beyond generic recipes and instead cultivate a vocabulary for describing what a projection does: what it preserves, what it exaggerates, what it fragments, what it smooths, and how its parameterisation participates in the production of interpretative claims.
To support this methodological proposal, the paper comparatively evaluates four DR algorithms specifically designed to preserve neighbour relationships, since these are widely used across disciplines (Bodt et al., 2025) and are increasingly adopted in DH since they are expected to be more effective in capturing the non-linear structures of complex and heterogeneous cultural datasets. tSNE prioritises the preservation of local neighbourhoods through a probabilistic similarity model, often at the expense of global structure (van der Maaten / Hinton, 2008). UMAP, by contrast, relies on a graph-based manifold approximation that better retains global organisation and scales more efficiently, while remaining sensitive to parameters such as the number of neighbours and minimum distance (McInnes et al., 2018). TriMap preserves relative distances via sampled triplet constraints, favouring global structure and outlier detection but exhibiting strong dependence on the number of triplets (Amid & Warmuth, 2022). PaCMAP seeks to balance local and global structure by explicitly controlling near, mid-near, and far point relations, resulting in comparatively robust embeddings whose overall spread nonetheless depends on pair selection ratios (Wang et al., 2021).
Because DR behaviour is contingent on the dataset, the evaluation is conducted on four datasets sampled from heterogeneous audiovisual archives (Radio Télévision Suisse- RTS, Prix de Lausanne- PDL, Olympic Games Collection- IOC, Montreux Jazz Festival- MJF). The RTS dataset consists of 60,000 one-second clip sampled from 3,170 representative videos and represented by the average of ResNet frame-level 2,048-d features (Mercier et al., 2019). Next, the PDL dataset contains 27,672 posesuniformly extracted from 1,445 dance recordings using MediaPipe BlazePose (Bazarevsky et al., 2020) and represented as a 99-dimensional vector of 33 flattened keypoints in 3D. Then, the IOC dataset includes 58,313 poses extracted using OpenPifPaf (Kreiss et al., 2022) from Olympic Games footage, with each vector encoding 33 angles between selected 2D keypoints. Lastly, the MJF dataset consists of 39,789 audio tracks from concert recordings, encoded as 87-dimensional feature vectors extracted using Essentia TensorFlow with the MTG-Jamendo genre model (Alonso-Jiménez et al., 2020).
The aim of this choice is methodological rather than domain-specific: these datasets exhibit different feature types, densities, and degrees of non-linearity, allowing the comparison to foreground a central point for DH: DR methods do not have context-free behaviour. The same algorithm can generate dramatically different structures depending on the representational regime that produced the feature vectors and the distributional properties of the dataset, as shown in Figure 1. Consequently, algorithm selection cannot be separated from representational critique.
The evaluation adopts a set of quantitative metrics drawn from recent DR literature (Becht et al., 2019; Huang et al., 2022; Kobak / Berens, 2019; Wang et al., 2021; Zhang et al., 2021; Bodt et al., 2025), with a focus on those that operationalise local and global structure preservation. For local structure, the Neighbourhood Kept Ratio (NKR) is used to compare the overlap between each point’s nearest neighbours in the original and projected spaces. For global structure, three complementary measures are employed: a Global Score (GS) benchmarking cluster-level arrangement against a PCA reference; Random Triplet Accuracy (RTA), capturing the preservation of relative ordering across random point triplets; and Distance Spearman Correlation (DSC), measuring rank correlation between pairwise distances in high- and low-dimensional spaces. Figure 2 reports these evaluation metrics for the four DR algorithms and four datasets tested.
In addition, the paper considers practical dimensions that bear directly on hermeneutic reliability: computational efficiency, shown in Figure 3, and, crucially, stability across runs, assessed by computing the mean and the standard deviation of Spearman rank correlations between pairwise distance matrices obtained from 10 runs on the same sample of 10,000 points, as indicated in Figure 4.
The resulting comparison produces two findings that are particularly instructive for DH hermeneutics. First, qualitative inspection indicates a clear behavioural distinction between tSNE and the other methods: tSNE tends to generate more fragmented cluster structures, whereas UMAP, TriMap, and PaCMAP more readily reveal coarser global formations that appear consistent across methods and datasets. Second—and more methodologically challenging—quantitative results do not neatly align with standard expectations. While tSNE dominates local preservation (as anticipated), it also performs surprisingly well on certain global metrics (GS and RTA), with other methods outperforming it primarily on DSC. Conversely, TriMap and PaCMAP—often presented in the literature as strong candidates for balancing global and local structure—produce subpar scores on several metrics in these AV datasets, except for distance-rank correlation. These discrepancies point to potential mismatches between algorithmic objectives, metric assumptions, and dataset properties. They also invite a caution that is directly relevant to DH: metric-driven selection can be misleading if metrics are treated as neutral arbiters rather than as formalised proxies grounded in specific assumptions about what “structure” is.
From a hermeneutic standpoint, this leads to a key methodological conclusion: DR should be incorporated into DH research as a comparative and iterative practice rather than a one-shot technical step. Because projections are parameter-sensitive and, in many cases, stochastic, interpretative readings require attention to robustness: the extent to which observed patterns persist across runs, across plausible parameter settings, and across alternative mapping strategies. Parameter tuning is not an unfortunate inconvenience but part of the interpretative labour through which the analyst actively configures a legible space. The question is not “which algorithm is best?” but rather “which mapping is appropriate for which interpretative task, under which constraints, and with what degree of reproducibility?” In this framing, DR becomes a method for articulating and testing hypotheses about latent structure: a way to explore whether apparent clusters correspond to meaningful semantic groupings, whether gradients align with historically grounded continuities, or whether outliers index genuinely anomalous artefacts versus pitfalls of the projection. By reframing DR as computational mapping, the paper contributes a methodological vocabulary for DH scholarship that seeks to make interpretative use of latent spaces. DR maps are neither transparent windows nor purely aesthetic surfaces; they are situated constructions whose validity depends on the explicit articulation of representational choices, algorithmic configurations, and evaluative criteria. The paper therefore advocates for a mode of reporting that treats DR projections as part of an argumentative chain: specifying algorithm choice, hyper-parameters, initialisation, stability checks, and the metrics used to justify claims about local and global structure, while also acknowledging where metrics fail to capture what matters for interpretation.
In sum, the paper argues that dimensionality reduction occupies a pivotal role in contemporary DH research precisely because it mediates between imperceptible latent structures and interpretable spatial forms. Its outputs should be read as interpretative objects shaped by computational decisions, and its practice should be integrated into hermeneutic workflows through comparative evaluation, robustness testing, and reflexive reporting. By combining a technical review of four DR algorithms with a dataset-sensitive evaluation and a hermeneutic reframing of projection as mapping, the paper offers both practical guidance and a conceptual stance: DR is not merely how we visualise vector spaces; it is one of the ways by which we make them available to interpretation, and thus one of the ways computational methods participate in the production of knowledge in the Digital Humanities.