What logic should be used to label customers (monthly) based on the categories they bought more often in the preceding 4 calendar months? - sql

Question

I have a table that looks like this:
user
type
quantity
order_id
purchase_date
john
travel
10
1
2022-01-10
john
travel
15
2
2022-01-15
john
books
4
3
2022-01-16
john
music
20
4
2022-02-01
john
travel
90
5
2022-02-15
john
clothing
200
6
2022-03-11
john
travel
70
7
2022-04-13
john
clothing
70
8
2022-05-01
john
travel
200
9
2022-06-15
john
tickets
10
10
2022-07-01
john
services
20
11
2022-07-15
john
services
90
12
2022-07-22
john
travel
10
13
2022-07-29
john
services
25
14
2022-08-01
john
clothing
3
15
2022-08-15
john
music
5
16
2022-08-17
john
music
40
18
2022-10-01
john
music
30
19
2022-11-05
john
services
2
20
2022-11-19
where i have many different users, multiple types making purchases daily.
I want to end up with a table of this format
user
label
month
john
travel
2022-01-01
john
travel
2022-02-01
john
clothing
2022-03-01
john
travel-clothing
2022-04-01
john
travel-clothing
2022-05-01
john
travel-clothing
2022-06-01
john
travel
2022-07-01
john
travel
2022-08-01
john
services
2022-10-01
john
music
2022-11-01
where the label would record the most popular type (based on % of quantity sold) for each user in a timeframe of the last 4 months (including the current month). So for instance, for March 2022 john ordered 200/339 clothing (Jan to and including Mar) so his label is clothing. But for months where two types are almost even I'd want to use a double label like for April (185 travel 200 clothing out of 409). In terms of rules this is not set in stone yet but it's something like, if two types are around even (e.g. >40%) then use both types in the label column; if three types are around even (e.g. around 30% each) use three types as label; if one label is 40% but the rest is made up of many small % keep the first label; and of course where one is clearly a majority use that. One other tricky bit is that there might be missing months for a user.
I think regarding the rules I need to just compare the % of each type, but I don't know how to retrieve the type as label afterwards. In general, I don't have the SQL/BigQuery logic very clearly in my head. I have done somethings but nothing that comes close to the target table.
Broken down in steps, I think I need 3 things:
group by user, type, month and get the partial and total count (I have done this)
then retrieve the counts for the past 4 months (have done something but it's not exactly accurate yet)
compare the ratios and make the label column
I'm not very clear on the sql/bigquery logic here, so please advise me on the correct steps to achieve the above. I'm working on bigquery but sql logic will also help

Answer 1

Consider below approach. It looks a little bit messy and has a room to optimize but hope you get some idea or a direction to address your problem.
WITH aggregation AS (
SELECT user, type, DATE_TRUNC(purchase_date, MONTH) AS month, month_no,
SUM(quantity) AS net_qty,
SUM(SUM(quantity)) OVER w1 AS rolling_qty
FROM sample_table, UNNEST([EXTRACT(YEAR FROM purchase_date) * 12 + EXTRACT(MONTH FROM purchase_date)]) month_no
GROUP BY 1, 2, 3, 4
WINDOW w1 AS (
PARTITION BY user ORDER BY month_no RANGE BETWEEN 3 PRECEDING AND CURRENT ROW
)
),
rolling AS (
SELECT user, month, ARRAY_AGG(STRUCT(type, net_qty)) OVER w2 AS agg, rolling_qty
FROM aggregation
QUALIFY ROW_NUMBER() OVER (PARTITION BY user, month) = 1
WINDOW w2 AS (PARTITION BY user ORDER BY month_no RANGE BETWEEN 3 PRECEDING AND CURRENT ROW)
)
SELECT user, month, ARRAY_TO_STRING(ARRAY(
SELECT type FROM (
SELECT type, SUM(net_qty) / SUM(SUM(net_qty)) OVER () AS pct,
FROM r.agg GROUP BY 1
) QUALIFY IFNULL(FIRST_VALUE(pct) OVER (ORDER BY pct DESC) - pct, 0) < 0.10 -- set threshold to 0.1
), '-') AS label
FROM rolling r
ORDER BY month;
Query results