Table of Contents
Product Trend Calculator
Calculates trend score for products based on their position history.
Maintainer: Silvia
Host
Git
Database
Host: osa_data
DB: meta_db
reads from:
- meta_db_collection (products with given
shop_idandlast_visitnot older than 30 days) - categories (categories with given
shop_idandsessionnot older than 30 days) - history (for relevant products and relevant sessions)
writes into:
- product_trends (unique key: (
picalike_id,sesssion))
Usage
The product trend calculator receives commands and sends stats to the shop conveyor belt through port [tdb]. It follows the feed import in the pipeline.
Relevant Data
The trend score is calculated for all products in meta_db_collection that have last_visit not older than 30 days. Likewise, only sessions not older than 30 days are taken into account for position history and category size.
Weights
Categories
Categories are weighted accordingly to their size in the last relevant session (i.e., the latest session in which the category appears) and the sum of the elements for all categories in their last relevant session.
Example
Session 1
category A: 10 items
category B: 5 items
Session 2
category A: does not appear
category B: 10 items
category C: 20 items
Relevant sessions
category A: session 1 (10 items)
category B: session 2 (10 items)
category C: session 2 (20 items)
Category weights
category A: 10/40 = 0.25
category B: 10/40 = 0.25
category c: 20/40 = 0.5
Sessions
Sessions are weighted according to their distance to the current session measured in weeks as exp(-weeks) and normalized.
Example
Session 0
position: 0.9
weeks since current session: 1.5
time weight: exp(-1.5) = 0.2
result: 0.9*0.2 = 0.18
Session 604800
position: 0.4
weeks since current session: 0.5
time weight: exp(-0.5) = 0.6
result: 0.4*0.6 = 0.24
Session 907200 (current)
position: 0.6
weeks since current session: 0
time weight: exp(0) = 1
result: 0.6*1 = 0.6
Time weighted position
(0.18 + 0.24 + 0.6)/(0.2 + 0.6 + 1) = 1.02/1.8 = 0.57
Score
The product trend score is given by the sum of the category scores divided by the sum of the category weights for all categories in which the product appears in the relevant sessions.
For a given category, the category score is given by the category weight multiplied with the time weighted position.
Example
Category A
category weight: 0.4
time weighted position: 0.5
category score: 0.4 * 0.5 = 0.2
Category B
category weight: 0.2
time weighted position: 0.9
category score: 0.2 * 0.9 = 0.18
Product trend score
sum(category scores)/sum(category weights): (0.2 + 0.18)/(0.4 + 0.2) = 0.38/0.6 = 0.63
Possible problems
Too big chunksize can lead to memory problems.