20+ products. Millions of active customers. Insane amount of data and complex domain. Come join me in this talk to know the journey we at Gojek took to predict which of our products a user is most likely to use next.

A major problem we faced, as a company, was targeting our customers with promos and vouchers that were relevant to them. We developed a generalized model that takes into account the transaction history of users and gives a ranked list of our services that they are most likely to use next. From here on, we are able to determine the vouchers that we can target these customers with.

In this talk, I will be talking about how we used recommendation engines to solve this problem, the challenges we faced during the time and the impact it had on our conversion rates. I will also be talking about the different iterations we went through and how our problem statement evolved as we were solving the problem.

Recommendation systems (RecSys) are the core engine for any personalized experience on eCommerce and online media websites. Most of the companies leverage RecSys to increase user interaction, to enrich shopping potential and to generate upsell & cross-sell opportunities. Amazon uses recommendations as a targeted marketing tool throughout its website that contributes 35% of its total revenue generation [1]. Netflix users watch ~75% of the recommended content and artwork [2]. Spotify employs a recommendation system to update personal playlists every week so that users won’t miss newly released music by artists they like. This has helped Spotify to increase its number of monthly users from 75 million to 100 million at a time [3]. YouTube's personalized recommendation helps users to find relevant videos quickly and easily which account for around 60% of video clicks from the homepage [4].

In general, RecSys generates recommendations based on user browsing history and preferences, past purchases and item metadata. It turns out most existing recommendation systems are based on three paradigms: collaborative filtering (CF) and its variants, content-based recommendation engines, and hybrid recommendation engines that combine content-based and CF or exploit more information about users in content-based recommendation. However, they suffer from limitations like rapidly changing user data, user preferences, static recommendations, grey sheep, cold start and malicious user.

Classical RecSys algorithm like content-based recommendation performs great on item to item similarities but will only recommend items related to one category and may not recommend anything in other categories as the user never viewed those items before. Collaborative filtering solves this problem by exploiting the user's behavior and preferences over the items in recommending items to the new users. However, collaborative filtering suffers from a few drawbacks like cold start, popularity bias, and sparsity. The classical recommendation models consider the recommendation as a static process. We can solve the static recommendation on rapidly changing user data by RL. RL based RecSys captures the user’s temporal intentions and responds promptly. However, as the user action and items matrix size increases, it becomes difficult to provide recommendations using RL. Deep RL based solutions like actor-critic and deep Q-networks overcome all the aforementioned drawbacks.

Present systems suffer from two limitations, firstly considering the recommendation as a static procedure and ignoring the dynamic interactive nature between users and the recommender systems. Also, most of the works focus on the immediate feedback of recommended items and neglecting the long-term rewards based on reinforcement learning. We propose a recommendation system that uses the Q-learning method. We use ε-greedy policy combined with Q learning, a powerful method of reinforcement learning that handles those issues proficiently and gives the customer more chance to explore new pages or new products that are not so popular. Usually while implementing Reinforcement Learning (RL) to real-world problems both the state space and the action space are very vast. Therefore, to address the aforementioned challenges, we propose the multiple/distributed Q table approaches which can deal with large state-action space and that aides in actualizing the Q learning algorithm in the recommendation and huge state-action space.

References:

1. "https://www.mckinsey.com/industries/retail/our-insights/how-retailers-can-keep-up-with-consumers":https://www.mckinsey.com/industries/retail/our-insights/how-retailers-can-keep-up-with-consumers
2. "https://medium.com/netflix-techblog/netflix-recommendations-beyond-the-5-stars-part-1-55838468f429":https://medium.com/netflix-techblog/netflix-recommendations-beyond-the-5-stars-part-1-55838468f429
3. "https://www.bloomberg.com/news/articles/2016-09-21/spotify-is-perfecting-the-art-of-the-playlist":https://www.bloomberg.com/news/articles/2016-09-21/spotify-is-perfecting-the-art-of-the-playlist
4. "https://dl.acm.org/citation.cfm?id=1864770":https://dl.acm.org/citation.cfm?id=1864770
5. "Deep Reinforcement Learning based Recommendation with Explicit User-Item Interactions Modelling": https://arxiv.org/pdf/1810.12027.pdf
6. "Deep Reinforcement Learning for Page-wise Recommendations": https://arxiv.org/pdf/1805.02343.pdf
7. "Deep Reinforcement Learning for List-wise Recommendations": https://arxiv.org/pdf/1801.00209.pdf
8. "Deep Reinforcement Learning Based RecSys Using Distributed Q Table": http://www.ieomsociety.org/ieom2020/papers/274.pdf

Machine Learning and AI are changing or would say have changed the way how businesses used to behave. However, the Data Science community is still lacking good practices for organizing their projects and effectively collaborating and experimenting quickly to reduce “time to market”.

During this session, we will learn about one such open-source tool “DVC”
which can help you in helping ML models shareable and reproducible.
It is designed to handle large files, data sets, machine learning models, metrics as well as code

Sample this – two cities in India; Mumbai and Pune, though only 80kms apart have a distinctly varied spoken dialect. Even stranger is the fact that their sign languages are also distinct, having some very varied signs for the same objects/expressions/phrases. While regional diversification in spoken languages and scripts are well known and widely documented, apparently, this has percolated in sign language as well, essentially resulting in multiple sign languages across the country. To help overcome these inconsistencies and to standardize sign language in India, I am collaborating with the Centre for Research and Development of Deaf & Mute (an NGO in Pune) and Google. Adopting a two-pronged approach: a) I have developed an Indian Sign Language Recognition System (ISLAR) which utilizes Artificial Intelligence to accurately identify signs and translate them into text/vocals in real-time, and b) have proposed standardization of sign languages across India to the Government of India and the Indian Sign Language Research and Training Centre.

As previously mentioned, the initiative aims to develop a lightweight machine-learning model, for 14 million speech/hearing impaired Indians, that is suitable for Indian conditions along with the flexibility to incorporate multiple signs for the same gesture. More importantly, unlike other implementations, which utilize additional external hardware, this approach, which utilizes a common surgical glove and a ubiquitous camera smartphone, has the potential of hardware-related savings at an all-India level. ISLAR received great attention from the open-source community with Google inviting me to their India and global headquarters in Bangalore and California, respectively, to interact with and share my work with the TensorFlow team.

DS, ML, and AI have moved very far from just running the models only at your local machine. Nowadays models are running in production and helping the business at decision making, which in turn increased the expectations for continuously running the models and making the changes online, but remember running this at a large scale is no easy task.
During this session, we will learn about one such approach with Docker, Kubernetes and Kubeflow which can help us not only in developing but also deploying models at scale, allow us to use distributed setup and Hyperparameter tuning

Reinforcement learning helped Google's "AlphaGo" beat the world's best Go player. Have you wondered if you too can train a program to play a simple game?

Reinforcement learning is a simple yet powerful technique that is driving many applications, from recommender systems to autonomous vehicles. It is best suited to handle situations where the behavior of the system cannot be described in simple rules. For example, a trained reinforcement learning agent can understand the scene on the road and drive the car like a human. In supply chain management, RL agents can make decisions on inventory ordering.

In this talk, I will demonstrate how to train a RL agent to a) cross a maze and b) play a game of Tic-Tac-Toe against an intelligent opponent c) act as a warehouse manager and learn inventory ordering; with the help of plain python code. As you participate in this talk, you will master the basics of reinforcement learning and acquire the skills to train your own dragon.

Beyond predictions, some ML models provide rules to identify actionable sub-populations in support-confidence-lift paradigm. Along with making the models interpretable, rules make it easy for stakeholders to decide on the plan of action. We discuss rule-based models in production, rule-based ensembles, anchors using R package: tidyrules.