Skip to content

DATA.md

Latest commit

 

History

History
38 lines (25 loc) · 2.32 KB

DATA.md

File metadata and controls

38 lines (25 loc) · 2.32 KB
Raw

Data

The project uses two data sources: a real product catalog and a synthetic clickstream.

Sources

Data Source Size Real or synthetic
Product catalog (product.csv) WANDS from Wayfair 42,994 products Real
Search queries (query.csv) WANDS 480 queries Real
Relevance labels (label.csv) WANDS 233,448 human labels: Exact, Partial, or Irrelevant Real
Clickstream (clickstream.csv) avantiwhenever/WANDS 170,990 events, 12,818 sessions, 5,000 users Synthetic

To download all four files, run ./scripts/download-data.sh. The files go into data/, which Git ignores.

The synthetic clickstream

WANDS has no user behavior. The recommendation strategies need behavior data, so the project generates it.

  • Users. Each of the 5,000 users has 1 to 3 preferred product categories.
  • Sessions. Each user has 1 to 6 sessions. In each session, the user searches one of the 480 real WANDS queries.
  • Events. Each product on the results page gets a view event. Then a cascade model decides on click, add_to_cart, and purchase events.
  • Click probability. The probability depends on the position of the product and on its WANDS relevance grade. The position factor is 1 / position^0.8.
  • Deterministic. With the same seed (default 42), the generator always gives the same file.

CLICKSTREAM.md gives the full method.

Important: this data does not represent real user behavior. Results that use the clickstream show how a strategy behaves on this model of users. They do not predict results with real users.

Effect on the evaluation

The generator uses the WANDS relevance grades to order each session and to make the clicks. Thus, the clickstream metric is partly circular. The evaluation also scores each strategy against the WANDS human labels directly, to give an independent result. See EVALUATION.md.

File format notes

  • The WANDS files have the .csv extension, but they are tab-separated.
  • The WANDS column for the category path is category hierarchy, with a space in the name.
  • WandsProductCsvLoader in rec-support handles both of these points.