Near-identical but distinct results should be grouped, not listed side by side

What it is

Ten results say almost the same thing and remain ten objects one can still choose among: hardback and paperback of one book, colourways of one model, a wire report and a column on the same event. They must not be fused into one representative, and they must not occupy ten nearly identical snippet rows. Grouping similar-but-not-same results gathers them into a group that takes one row’s slot, while the members stay distinct options. Hearst and Pedersen’s Scatter/Gather, and later search-result clustering, target this “alike, but still a choice” set.

Side-by-side listing fails as visual false duplication: people think dedup never ran and skip the block. Fusing into one row fails by deleting a real choice. Grouping takes the middle: fewer rows, options intact.

Why it happens

Highly similar snippets share scent. When several rows in a row barely change, people start a “already seen” skip, and later rows that carry the difference that matters (in stock, shorter, official) slide by as copies. Grouping marks “this cluster” as one comparison unit, so differences are expected inside the group rather than as the next row happening to differ. Working memory moves from “another similar piece” to “there are several versions here.”

Ungrouped juxtaposition also distorts the feel of ranking: near-tied similar documents form a territorial block and push truly different objects off the first screen. Grouping occupies space by cluster, diversity happens between clusters, and the monopoly that similarity creates is broken.

Studying it

Contrast fully expanded juxtaposition, fusion into one row, and grouping-by-similarity, and watch whether people still find the one different option inside the group, and whether they skip the whole block.

• Paradigms: one result set, three UIs; the task names an attribute that exists only inside the similar cluster (“the paperback,” “the colour that is in stock”). Cluster quality uses standard search-result clustering metrics (purity, agreement with human topics); the interface effect uses the task.
• Independent variables: grouping on or off, similarity threshold for a cluster, how many members are shown by default.
• Dependent variables: discovery rate of the in-group target, rate of skipping the whole cluster as duplication, number of distinct-topic clusters on the first screen.
• Methodological note: if the task target happens to be the cluster representative, grouping and fusion look equally good. Put the target on a non-representative member. If automatic clusters do not match user topics, the study is measuring clustering failure, not failure of grouping as a presentation.

Where it stops holding

When members are fully interchangeable for the current task, grouping is empty ceremony and fusion is cleaner. Forced grouping at too low a similarity ties unrelated objects and reads as a claim that they are equivalent. Results that already have a spatial or temporal axis (images, maps, timelines) fight a second grouping by textual similarity. Expert comparison (two contracts side by side) needs juxtaposition; grouping blocks the comparison.

Applying it

• Group results that are alike but not identical. One group, one row; the heading states what is shared. Do not run ten near-paraphrase snippets as ten rows.
• The threshold can be looser than near-duplicate detection, but still answer “do these still need a separate choice?” If they do not, leave them to identity merge.
• Default to one representative plus “N similar results,” not a fully expanded group.
• Verify with a query whose top ten holds six near-paraphrase news items and one truly different document. After grouping, the first screen should still show the different one, and the six news items should not each take a row. If the six still sit in parallel, or the different one is bundled into the news group, grouping has not aimed at “alike, but still a choice.”

Related

• Within the group: G3.15.1 Multiple sources of the same entity should collapse to one representative · G3.15.3 Grouping hides within-group differences and needs an expand path · G3.15.4 Dedup rules must be intelligible so known items do not silently vanish · G3.15.5 Group by the dimension the user is deciding on, not by a database field
• Adjacent: G3.05 Result ranking · G3.06 Result snippets · G1.02 Organization systems
• Search terms: search result clustering · Scatter/Gather · similar results