Abhishek Jain

The Feature Management Platform is a self-service platform which enables users to calculate data features to be used in training and prediction for machine learning models The platform processes data event by event and calculates features in streaming mode. Among these features, some require stateful aggregations over time, e.g., “count of distinct logins from one IP” where we have to store state of the given IP over time. The platform’s processing module could be used to store the state and… Show more

The Feature Management Platform is a self-service platform which enables users to calculate data features to be used in training and prediction for machine learning models The platform processes data event by event and calculates features in streaming mode. Among these features, some require stateful aggregations over time, e.g., “count of distinct logins from one IP” where we have to store state of the given IP over time. The platform’s processing module could be used to store the state and do aggregation local to the processing nodes, but it would exhibit the following problems:
* No fault tolerance: If the processing node dies we lose the state.
* Memory constraints of the processing framework.
* Data locality: Distributed processing framework used in the platform means individual nodes can’t update data that are not existing locally.
* Aggregations on different keys for one transaction require “shuffle operations” across the processing framework
* Aggregations on non-primary (e.g. Kafka’s partitioning) keys require a global lock to prevent race conditions on concurrent state updates
* These features might require backfilling (if the aggregation window exceeds the retention on the input stream) which usually happens in batch mode. Hence, the state has to be shared across different processes running on different kinds of input data sources over a variety of time windows.

We solve the above problem by introducing an additional external, distributed cache layer into the platform which can be backfilled, is resilient to repartitioning & failures in the processing system, supports secondary keys w/o reshuffling of data and data structures to easily implement windowed aggregations with configurable TTL (time to live) settings. We decouple the aggregation state into a dedicated, distributed in-memory cache with disk backups. The processing framework stores and retrieves state from the introduced cache layer “leveraging the push down predicate pattern”. Show less

1. Customers purchase Quickbooks in order to more easily manage their financial information, but new First Time Users (FTU) don’t always know where to start and get confused by all the features on the Home Page (HP).

The current method to help new users onboard to the product relies on user’s filling out an optional onboarding question (What would you like to do in Quickbooks?). Based on the responses, a HP widget recommends a rank-ordered list of first actions to take. However, based… Show more

1. Customers purchase Quickbooks in order to more easily manage their financial information, but new First Time Users (FTU) don’t always know where to start and get confused by all the features on the Home Page (HP).

The current method to help new users onboard to the product relies on user’s filling out an optional onboarding question (What would you like to do in Quickbooks?). Based on the responses, a HP widget recommends a rank-ordered list of first actions to take. However, based on the data that we have seen in the past, about 65% of users select none or all of the options, which does not make the onboarding experience very personalized or effective.

2. To improve the recommendations and personalize the onboarding experience in the most relevant way possible for the user, we propose a method to predict user intent (or interest in a particular product feature) based on the user’s information prior to signing up.

The solution (outlined in more detail below) utilizes the following kinds of user information:
Search activity of the user across Intuit partner websites in addition to QuickBooks Online.
Demographics
QB.com web page activity

The above data is processed to feed into Natural Language Processing algorithms such as LDA (Latent Dirichlet Allocation) and Topic Modeling and then analyzed for similarity with the actions available in QuickBooks. The similarity analysis then yields a list of top recommended actions mapped to each user.

In other words, the intent of the user prior to purchasing the product is handed off to make recommendations inside the product once the user starts using it. Show less

This invention leverages a multi-modal feature stream. It is based on a framework to tap into a data source and compute features while pushing the results in a standardized format onto a feature queue. It is treating model outputs as features and supports asynchronous prediction hooks.

This patent filing proposes an architecture for a holistic platform to manage the full lifecycle of data features and attributes for data science and analytics use cases. It is targeted towards Data Workers managing machine learning features to support use cases with real time or offline processing/querying needs. Those workers will be able to implement data transforms of both streaming and batch data sources, add context to data, define and process new and existing features and view metrics on… Show more

This patent filing proposes an architecture for a holistic platform to manage the full lifecycle of data features and attributes for data science and analytics use cases. It is targeted towards Data Workers managing machine learning features to support use cases with real time or offline processing/querying needs. Those workers will be able to implement data transforms of both streaming and batch data sources, add context to data, define and process new and existing features and view metrics on their features. The architecture minimizes dependency on data engineering while reducing the turnaround time to create, modify, deploy, and use features within a consolidated platform. In detail, we envision solving the following user problems.
AI/ML use cases require data features for their models to be computed and made available in a way that it serves both query patterns specific to model training and hosting of smart services, and a variety of special data structures (e.g. one hot encodings). To save cost, effort and duplication, these feature computations and their value output need to be discoverable, shareable and reusable across use cases. A data worker should be able to solely focus on implementing the logic for these transformations without needing the knowledge of the underlying infrastructure, storage solutions, runtime details and query mode (e.g. batch vs streaming). At prediction time, ML use cases need to access aggregates derived from multiple upstream sources the context of which may not be available within the actual product. Finally, the computations should run independently, securely and access controls for feature values need to be in place. Show less