Inspection views¶

Beyond a single ranked list, a demo can show layouts, metrics, and plots. This page draws the line between what the framework renders automatically and what you add.

Who provides what¶

Layer	Who supplies it	Examples
Automatic (framework)	the library	session profile, caching, model adapter, side-by-side comparison, seen-filtering, layouts, sidebar widgets from `params`
Recommender adapter	you (the model author)	`get_recommendations()` / `scores()` — how a profile becomes a ranking
Optional content	you (the demo author)	metric tables, score distributions, overlap heatmaps, markdown — via `intro()` / `body()`

Analytical views are helpers you call, not automatic output

sr.evaluate, sr.plot_*, sr.recommendation_overlap_matrix, sr.table, and sr.markdown are functions the library provides, but they render only where you call them in body(). The framework does not compute or show them on its own. All built-in charts are interactive Plotly figures (st.plotly_chart), and any Streamlit-compatible chart can be dropped into body() too.

Layouts¶

Layout	Display
`rows`	One horizontal row of clickable poster cards with side scroll; drives side-by-side comparison
`grid`	Catalog-style clickable poster grid (`n_rows` × `n_cols`) for single-model browsing
`cards`	Swipe deck: one card at a time with Like / Dislike / Skip; refreshes after `swipes_per_refresh` swipes

Compare mode (get_recommendations={...}) always uses rows; the cards deck is single-model.

Metrics¶

Ranking metrics over held-out interactions — hit rate, recall, NDCG, MRR, and catalog coverage:

metrics = sr.evaluate({"Ours": recs_by_user}, test_interactions=test, k=10, all_item_ids=items.item_id)
sr.table(metrics)
sr.plot_metric_comparison(metrics)

Plots¶

sr.dataset_info(items, train)                    # counts, density, distribution tabs
overlap = sr.recommendation_overlap_matrix({"ItemKNN": [1, 2, 3], "EASE": [2, 3, 4]})
sr.plot_overlap_heatmap(overlap)                 # pairwise Jaccard agreement
sr.plot_score_distribution(scores)               # score spread for a model

These connect the aggregate numbers reported in a paper to the concrete recommendations a reader can see. Put them in body() so they render beneath the recommendation interface.