Publishing Ranking Events

Publish ranking events

The grouped format

Before publishing ranking model events into Fiddler, we need to make sure they are in grouped format (i.e. the listing returned within the same query id—which is usually the group_by argument passed to fdl.ModelInfo—is in the same row with other cells as lists). The first row in the example below indicates there are 3 items returned by query id(srch_id' in the table) 1.

Below is an example of what this might look like.

srch_id	price_usd	review_score	...	prediction	target
101	[134.77,180.74,159.80]	[5.0,2.5,4.5]	[...]	[1.97, 0.84,-0.69]	[1,0,0]
...	...		...	...	...
112	[26.00,51.00,205.11,73.2]	[3.0,4.5,2.0,1.0]	[...]	[10.75,8.41,-0.23,-3.2]	[0,1,0,0]

In the above example, srch_id is the name of our group_by column, and the other columns all contain lists corresponding to the given group.

How can I convert a flat CSV file into this format?

If you're storing your data in a flat CSV file (i.e. each row contains a single item), Fiddler provides a utility function that can be used to convert the flat CSV file into the grouped format specified above.

from fiddler.utils.pandas import convert_flat_csv_data_to_grouped
import pandas as pd

grouped_df = convert_flat_csv_data_to_grouped(input_data=pd.read_csv('path/to/ranking_events.csv'), group_by_col='srch_id')

Call `publish_events_batch`

client.publish_events_batch(
    project_id=PROJECT_ID,
    model_id=MODEL_ID,
    batch_source=grouped_df,
    id_field='event_id',
)

In the above example, the group_by_col argument should refer to the same column that was specified in the group_by argument passed to fdl.ModelInfo.

Update ranking events

Prepare the updating dataframe

We also support updating events for ranking model. You can use publish_events_batch and publish_event APIs with update_event flag to True and keep the grouped format unchanged.

For example, you might want to alter the exisiting target after events are published. You can create a dataframe in the format below where group_by_col,id_col and target_col are required fields. You can either upload the complete group of events within one query_id or the subset contains the changed events.

Complete format

srch_id	event_id	target
101	['001','002','003']	[0,1,0]
...	...	...
112	['367','368','369','370']	[0,0,0,1]

Partial format

srch_id	event_id	target
101	['001','002']	[0,1]
...	...	...
112	['367','370']	[0,1]

Call `publish_events_batch` with `update_event` flag set to True

client.publish_events_batch(
    project_id=PROJECT_ID,
    model_id=MODEL_ID,
    batch_source=grouped_df_update,
    id_field='event_id',
    update_event=True,
)

Or call `publish_event` with `update_event` flag

events_dict = grouped_df_graded.to_dict('index')
for i, group_id in enumerate(events_dict):
    e= events_dict[group_id]
    '''
    first event:
    {'srch_id':101,'event_id':['001','002'],'target':[0,1]}
    '''
    client_v2.publish_event(project_id=project_id, model_id=model_id, event=e, update_event=True, event_id=str(e['event_id']))