LFPeers supports analysts (epidemiologists and decision-makers) in a "learning from peers" analysis of multimodal time-oriented data. The central idea is to identify data objects (e.g., countries) that behaved similarly to a reference object before a user-defined time point t0, and then explore how these peers diverged after t0, correlating observed variations with external factors to explain cause-and-effect relationships.
The input is a multimodal collection of data objects (countries) described by multiple temporal attributes (e.g., confirmed cases, stringency index, transit station mobility, reproduction rate) and static attributes (e.g., GDP, population density, median age). The analyst selects a reference object of interest (e.g., the USA) and begins familiarising themselves with its temporal development by inspecting bundled line charts, a world map, and a table of attribute values.
The analyst defines the temporal search parameters: a focal temporal attribute, a time range ending at t0 (separating the search phase from the exploration phase), a time series descriptor (equidistant sampling, statistical measures, PAA, PIP), and a distance metric (Euclidean or Dynamic Time Warping). LFPeers supports up to four parallel similarity concepts — each with independent attribute, time range, descriptor, and metric choices — whose results are intersected to form the final peer group. The system computes feature vectors for all objects in each similarity concept and ranks them by distance to the reference.
The system presents search results through coordinated views: a similarity-sorted table with sparklines and colour-coded static attributes, bundled line charts with the reference (blue), peers (orange), and context (grey), a t-SNE scatter plot of feature vectors with similarity-preserving colours, and a world map. The analyst inspects these views to identify the peer group — selecting the top-k nearest neighbours from the table, brushing clusters in the scatter plot, or clicking countries on the map. They assess whether the peer group is coherent (similar colours in projection, geographic plausibility, visual alignment of line chart bundles). If not, they refine by adjusting similarity criteria, adding/removing parallel concepts, or manually including/excluding individual objects. This cycle repeats until a satisfactory peer group is formed.
In the exploration phase (after t0), the analyst inspects a grid of bundled line charts showing how all temporal attributes develop for the peer group. They identify "fan-out" patterns — attributes where peers that were previously similar start diverging — which signal potentially interesting causal factors. The analyst selects a target attribute showing notable divergence and defines a time point t1 at which to measure peer state.
At the selected time point t1, the system computes correlations between the target attribute values and all other temporal/static attributes across the peer group. Results are ranked by correlation strength and visualised in scatter plots. The analyst inspects correlation details to identify explaining factors — attributes that were similar before t0 but became differently correlated after t0 are prime candidates for causal explanations. The analyst synthesises findings: which policies or factors distinguish successful from unsuccessful outcomes, what can be learned from peers for future decision-making.