Session on "Use Cases and Risks of ML in Capital Markets" | 23rd Dec at 4pm Hi everyone! The AI and Risk Mitigation project is well underway and for the third session, we will be joined by Rachna Maheshwari, Associate Director at CRI… more
The 2023 Monsoon edition is curated by:
- Nischal HP, Vice President of Data Engineering and Data Science at Scoutbee. Nischal curated the MLOps conference which was held online between 23 and 27 July 2021.
- Sumod Mohan, Founder and CEO at AutoInfer. Sumod curated Anthill Inside 2019 edition, held in Bangalore on 23 November.
- AI and Research - covers research, findings, and solutions for challenges on building models in various areas such as fraud detection, forecasting, and analytics. This track delves into the latest methodologies for handling challenges such as large-scale data processing, distributed computing, and optimizing model performance.
- Industrial applications of ML - covers implementation of AI in the industry, with more focus on the AI models, the issues in training, gathering data so, and so forth. ML is being used at scale in industries such as automotive, mechanical, manufacturing, agriculture, and such domains. This track focuses on the challenges in this space, as we see innovation coming out of these industries in the pursuit of using ML on a second-to-second basis.
- AI and Product - covers strategies for building AI products to scale and mitigating challenges. This track provides insights on incorporating AI tools and forecasting techniques to improve model training, developing a working model architecture, and using data in the business context.
There are three phases in the lifecycle of an application - research, application and aftermath of the application.
- Assess capabilities, determining the new frontiers for AI.
- Find a use for the application.
- Learn how to run it, monitor it and update it with time.
The three tracks at the 2023 Monsoon edition of The Fifth Elephant will cover this lifecycle.
The Fifth Elephant 2023 Monsoon edition will be held in-person. Attendance is open to The Fifth Elephant members only. Purchase a membership to attend the conference in-person. If you have questions about participation, post a comment here.
- Data/MLOps engineers who want to learn about state-of-the-art tools and techniques, especially from domains such as automobile, agri-tech and mechanical industries.
- Data scientists who want a deeper understanding of model deployment/governance.
- Architects who are building ML workflows that scale.
- Tech founders who are building products that require AI or ML.
- Product managers, who want to learn about the process of building AI/ML products.
- Directors, VPs and senior tech leadership who are building AI/ML teams.
Sponsorship slots are open for:
- Infrastructure (GPU, CPU and cloud providers) and developer productivity tool makers who want to evangelise their offering to developers and decision-makers.
- Companies seeking tech branding among AI and ML developers.
- Venture Capital (VC) firms and investors who want to scan the landscape of innovations and innovators in AI and who want to source leads for investment in the AI and ML space.
Solving for explainability of fraud detection models
At the TnS(Trust and Safety) team at Swiggy, building powerful fraud detection models that operate at high precision while still capturing maximum fraud has been the uber goal. Our system currently operates at a high level of complexity through various interventions, modelling techniques, and semi-supervised training methods while maintaining robustness.
For the final downstream model, we have always relied on tree-based learners over neural networks. Since are data is primarily tabular in nature, tree-based learners outperformed DNNs significantly on the winning metrics. While tree-based learners are great performers in terms of the final metrics that we’re looking to optimise, it has a few challenges:
1. It inherently restricts us from trying out more complex data structures like images or sequential data, we have tried to integrate such signals through a separate model whose final score is fed into the tree based learner but it significantly adds to complexity of the system.
2. A major press point for Fraud models historically has been a lack of explainability in predictions. We have experimented with LIME and SHAP-based approaches to build an explainable overhead but they’re computationally expensive to run for each record.
While tree-based methods for a deployable model have all these challenges, what works in their favour is that they have historically outperformed DL-based methods by a significant margin. This changes with TabNet, in the original paper(Ref), authors claim that TabNet can match or even outperform tree-based methods while also giving sample-level explainability, which we can also visualise. We explored a tabnet based model for our approach and found it to be on par with tree-based counterpart(xgboost). TabNet also allowed us to compute and store feature level attention within the model logs without any computational overhead.
In the presentation, we’ll be going through the following in depth.
Current pipeline and solution
Challenges in depth
Motivation for TabNet and what it unlocks
Experimental results and conclusion