Polars groupby + value_counts

I need some help with polars:

I have a dataframe with a categorical values column

┌───────────────────┬──────────────┬────────┐
│ session_id        ┆ elapsed_time ┆ fqid   │
│ ---               ┆ ---          ┆ ---    │
│ i64               ┆ i32          ┆ cat    │
╞═══════════════════╪══════════════╪════════╡
│ 20090312431273200 ┆ 0            ┆ intro  │
│ 20090312431273200 ┆ 1323         ┆ gramps │
│ 20090312431273200 ┆ 831          ┆ gramps │
│ 20090312431273200 ┆ 1147         ┆ gramps │
│ …                 ┆ …            ┆ …      │
│ 20090312431273200 ┆ 5197         ┆ teddy  │
│ 20090312431273200 ┆ 6180         ┆ teddy  │
│ 20090312431273200 ┆ 7014         ┆ teddy  │
│ 20090312431273200 ┆ 7946         ┆ teddy  │
└───────────────────┴──────────────┴────────┘

And I want to transform the fqid-column to look like this:

┌───────────────────┬─────────────┬────────────┬────────────┐
│ session_id        ┆ fqid_gramps ┆ fqid_intro ┆ fqid_teddy │
│ ---               ┆ ---         ┆ ---        ┆ ---        │
│ i64               ┆ i32         ┆ i32        ┆ i32        │
╞═══════════════════╪═════════════╪════════════╪════════════╡
│ 20090312431273200 ┆ 1           ┆ 1          ┆ 4          │
└───────────────────┴─────────────┴────────────┴────────────┘

That is, I would like to:

• Groupby over session_id,
• Make a value_counts() over fqid,
• Rename columns so that it would be 'fqid_' + category,
• Turn them into columns (transpose),
• Add them to the result.

Technically, I could achieve this without groupby by using something like

column_values = train['fqid'].value_counts().with_columns(pl.concat_str(pl.lit('fqid' + '_').alias('fqid'), pl.col('fqid').cast(pl.Utf8))).transpose()
column_values = column_values.rename(column_values.head(1).to_dicts().pop()).slice(1)

But when I am trying to make an aggregating function from this replacing train['fqid'] with pl.col('fqid') and making a groupby('session_id').aggregate(func('fqid')) it gives me nothing but errors like AttributeError: 'Expr' object has no attribute 'with_columns'.

Could you kindly suggest a proper way of making this operation?

Starting from

train=pl.from_repr(
     """┌───────────────────┬──────────────┬────────┐
        │ session_id        ┆ elapsed_time ┆ fqid   │
        │ ---               ┆ ---          ┆ ---    │
        │ i64               ┆ i32          ┆ cat    │
        ╞═══════════════════╪══════════════╪════════╡
        │ 20090312431273200 ┆ 0            ┆ intro  │
        │ 20090312431273200 ┆ 1323         ┆ gramps │
        │ 20090312431273200 ┆ 831          ┆ gramps │
        │ 20090312431273200 ┆ 1147         ┆ gramps │
        │ 20090312431273200 ┆ 5197         ┆ teddy  │
        │ 20090312431273200 ┆ 6180         ┆ teddy  │
        │ 20090312431273200 ┆ 7014         ┆ teddy  │
        │ 20090312431273200 ┆ 7946         ┆ teddy  │
        └───────────────────┴──────────────┴────────┘""")

we can do

(
    train
        .groupby(['session_id','fqid']) 
        .count()  #same as .agg(pl.count())
        .pivot('count','session_id','fqid','first')
        .select('session_id', pl.exclude('session_id').prefix('fqid_'))
)

shape: (1, 4)
┌───────────────────┬────────────┬─────────────┬────────────┐
│ session_id        ┆ fqid_teddy ┆ fqid_gramps ┆ fqid_intro │
│ ---               ┆ ---        ┆ ---         ┆ ---        │
│ i64               ┆ u32        ┆ u32         ┆ u32        │
╞═══════════════════╪════════════╪═════════════╪════════════╡
│ 20090312431273200 ┆ 4          ┆ 3           ┆ 1          │
└───────────────────┴────────────┴─────────────┴────────────┘

Since the output of what you want is a count rather than an average or aggregation of the fqid column itself, we include it in the groupby. Next, we do a pivot to make the results wide. The output of pivot doesn't keep the original column name so we have to add that back manually. We do that in a select by first taking the session_id and then adding to that every column except session_id with the prefix 'fqid_' to get the final desired result.

Incidentally, I'm not using value_counts because it returns a list of structs so we can't do, for example, train.select(pl.col('fqid').value_counts().over('session_id'))