SaFi Bank Space : FS: FireBase Behavioural Data

`featurestore_id`	`entity_type_id`	description	`entity_id`	prefix
`firebase_behaviour_fs`	`onboarding_behaviour_features`		`customerId`	fire_

[ Proposed Ingestion Setup ] [ Future Ingestion Setup ] [ Proposed Features ]

This page documents the features that we will collect from the FireBase application, in order to create risk behavioural indicators at the time of onboarding. This is derived from in-app behaviour at the point of onboarding.

FireBase events are written to BigQuery and the features are extracted from there - See Also: Firebase to BigQuery

Proposed Ingestion Setup

As the data is streamed directly from Firebase into BigQuery, is not possible to ingest this as we would if there were a Kafka topic. Data is streamed from Firebase into the tables below. Each table contains different events related to the application

Propose to only write application behavioural features of customers that have been triggered by the workflow for risk assessment and querying. As such, the application behaviour featurestore, will only contain the behavioural features of individuals that are deemed as “risky”. This is to save computation and writing of data jobs to the online featurestore.

These are the sample tables that the data are streamed to:

As such, in order to create behavioural features for risk assessment, we will need to retrieved these features from the relevant tables in BigQuery. The proposed field names have been laid out in the table below. At present the way to retrieve these set of behavioural features for a given customer would be via a SQL query to BigQuery.

SELECT
  fields
FROM analytics_31354378.events_intraday_
WHERE user_id = "customerId"

Future Ingestion Setup

First Choice: In order to create an architecture that supports real-time ingestion of app events, it seems that the only solution is to stream the FireBase events directly to Pub/Sub. Following which, a DataFlow or Cloud Function can ingest these features into the FeatureStore. This is to be done at the app level.

Google Cloud Pub/Sub: Node.js Client

Non-ideal: Another way is to trigger cloud run actions, using BigQuery ingest events. Meaning to say everytime an app event is ingested into BigQuery, a cloud run is triggered. This is not ideal, as it is dependent on the latency in streaming app events to BigQuery, which is known to have delays. This is more equipped for post-hoc analysis and not real-time event processing of features.

How to trigger Cloud Run actions on BigQuery events

Proposed Features

This documents defines the proposed set of features to be used and how to query - LARK SHEETS
This document defines the various fields that is found in bigquery.analytics_XXXXX.events_.app_events - FireBase BigQuery Definitions

Variable	Description	Field Name in BigQuery	`feature_id`	Type	Is Available?
Customer Id	ENTITY_ID_NAME		-	-
Application Version	Application version used when going through onboarding	`app_info.version`	`fire_app_version`
Clicking on `?` button	Clicking on each "?" Help button - PEP and US Persons Clicking on the "?" Help button to read more about "Terms & Conditions"	event_name = "identity_check_pep_sheet" and "identity_check_us_citizen_sheet" event_name = "privacy_body" (with consentType property to differentiate between T&C, terms of use, privacy policy)	`fire_click_on_help`		- Deployed but event cannot be found by Dkatalis team
Device Model	Device model used when going through onboarding	`device.mobile_model_name`	`fire_device_model`
Event Name	Description of the type of app event	`event_name`	`fire_event_name`
Exit App	Has the client exited/switch to other apps during onboarding	Use firebase_preformance db from `BigQuery.event_name = "_app_in_background"`	`fire_has_exit_app`
Screen Id	Indicator of which screen the user is in	`event_params.firebase_screen`	`fire_screen_id`
Time Per Screen	Time elapsed per each screen	`event_params.engagement_time_msec`	`fire_time_per_screen`
Timestamp	DATETIME field from when user initiates the onboarding process. Need to convert from `timestamp_micros`	`event_timestamp`	`fire_event_timestamp`
Total Keystrokes		`event_name = "screen_analytics"` `event_params.key = "tap"` `event_params.value.string_value = "N"`	`fire_total_keystrokes`
Was the walkthrough linear	Number of times the client pressed back, or gone back to previous screens	`event_name =` `"back_button_listener"`	`fire_pressed_back`
	Auto-Completion/Suggestions of OCR? Auto-fill? Corrections to pre-filled data?	There are 2 event_names: `personal_info_ocr_received` = personal info OCR result received from backend `personal_info_user_sent` = personal info data sent by customer
	Usage of `ctrl + v` to fill data within the app	`event_name = "customer_pasted_from_clipboard_to_text_field"` (old event name is was "customer_pasted_from_clipboard_to_text_f")	`fire_paste_from_clipboard`

SaFi Bank Space : FS: FireBase Behavioural Data

Proposed Ingestion Setup

Future Ingestion Setup

Proposed Features

Attachments: