Synerise Monad: Apply science to behavioral data. Automatically.
We are presenting MONAD - an award winning AI solution for behavioral modeling and behavior predictions.
The current research focus in AI community is on image and text processing, whereas companies in e-commerce, retail, travel, telco or financial industries have an abundance of interconnected data reflecting behaviors of their clients eg. what they bought, what they clicked, what transactions they made, how they interacted with a mobile app, mobile game or the call-center.
Until now, transforming raw behavioral data and handcrafting features to fit specific models and paradigms was cumbersome, slow and ineffective. Due to limitations of manual feature engineering, important behavioral cues were often lost. Feature engineering was often the bottleneck for large ML projects.
Thanks to MONAD this is no longer the case.
Cutting-edge research brought to behavioral event data for the first time.
Monad is extending the idea of self-supervised learning, known from the latest Large Language Models for texts (e.g. GPT, BERT), or Diffusion Models (e.g. DALL-E, StableDiffusion) for images, to the realm of rich, behavioral event-level data. The concept of a self-supervised foundational model is extremely powerful, but before MONAD it was only applied to texts and images. MONAD takes a step forward and allows every organization to train a behavioral model of all their customers and fine-tune it to various business applications.
Process your whole datalakes irrespective of size and complexity.
Under the hood, MONAD analyzes the full spectrum of historic behaviors of the whole population of customers, with the goal of predicting their future behaviors, without any supervision. To this end, it utilizes ultrafast proprietary graph representation learning algorithms, combined with recent advances in high dimensional density estimation. MONAD understands complex interconnected relations between multiple large tables storing various types of behavioral event data, such as clicks, purchases, card transactions, mobile app interactions, and others. It can easily process them together and with simpler user/product attribute tables.
MONAD’s outputs, called Universal Behavioral Profiles, can serve as an input to an unlimited number of supervised models like propensity prediction, churn prediction, recommendations, anomaly detection, customer scoring and others. Just like large language models can be fine-tuned to any application, MONAD is not restricted to a predefined set of use cases – the sky is the limit.
Experience record speed and industry leader quality at the same time.
Thanks to extremely efficient algorithms and optimizations, MONAD is blazingly fast. The whole training pipeline for 10M+ customers and a year worth of their events can complete within a couple hours on a single commodity server, without significant investments in infrastructure and compute. Subsequent fine-tuning for supervised applications is even faster, and requires very little labeled data.
MONAD’s effectivness was battle-tested by its commercial users but it also took the podium of Stanford KDDCup, Rakuten challenge, Twitter RecSys Challenge and Booking.com Data Challenge, competing with companies like Nvidia, Baidu, DeeepMind, Intel, PARC, Rakuten as well as top universities in the world.
Experiment faster. Cut human effort and time spent on modeling.
The aim of MONAD is to help data scientists quickly bootstrap their work by removing the bottlenecks in ML model production. MONAD is flexible and simple to use for business teams and more tech-savvy data science teams alike. It is proven to significantly decrease development time of machine learning models, effectvely allowing to deploy and test models in production in a matter of days.
Start experimenting quickly with popular cloud storage providers.
Getting started with MONAD is very straightforward. As long as raw, event-level data is stored in one of the supported databases like Google BigQuery or Snowflake, after a simple configuration, MONAD is ready to go and deliver value immediately.
Standard process of building AI application/models eliminated by Monad:
- Identify entities, metadata, attributes
- Find good joins
- Filter data
- Transform data
- Evaluate cardinalities
- Create embeddings
- Create feature transforms
- Find good parameters
- Identify useful features
- Create feature aggregates
- Calculate vectors
- Store vectors
- Create target variables
- Train models & tune models
- Analyze features, add more features
- Scale & manage infrastructure, add servers
- Serve models & monitor model health
How to query the future with Monad – selected problems to solve:
Use domain expertise where it matters.
With Monad, the time and effort usually spent on handcrafting feature creation, data pipelines, and infrastructure management can now be better utilized to choose business objectives, quickly test new hypotheses and refine modeling targets. All standard use cases like scoring customers predicting churn, identifying purchase propensities and product affinities, personalizing content and recommendations, detecting fraud and anomalies are available out-of-the-box. Users can also easily add custom features derived from deep domain expertise.
In addition to a collection of built-in models, Monad can be used to predict any targets supplied by the user. Models can be modified, customized, and tuned for special purposes. Helpful by-products such as embeddings, entity similarities, and optimal feature transforms are exposed for power users to utilize. Monitoring changes and feature drift over time enables teams to adapt and stay ahead of ever-changing business realities and customer demands.
Let business learn from the AI, instead of teaching it.
Understanding why things happen enables us to consciously shape the future. For every model, key events and features affecting predictions can be analyzed – both in aggregate, as well as at an individual prediction level.
- Financial Services
- Media & Publishing
What You need to start?
Set of raw event-level data linked to entities being modeled (for example: customers, devices, employees, cars, players, credit cards, anonymous IDs etc.).
Example of core built-in AI applications :
- Real-time LTV calcs
Concept used by leaders
"We are currently using EMDE for generating candidates to facilitate downstream recommendation systems. It generates recommendations using density-based rich customer representation. It allows us to trace customer look-alikes (‘People Like You’) to find similar users with similar cuisine/taste preferences as well as price affinity. We used Cleora for customer-restaurants graph data […] And to our delight, the embedding generation was superfast (i.e <5 minutes). For context, do remember that GraphSAGE took ~20hours for the same data in the NCR region. Cleora + EMDE gives us a generalised framework for recommendations […] We are exploring ways to use it in other applications such as search ranking, dish recommendations, etc. “ – Zomato.com Data Science team
Zomato is an multinational restaurant aggregator and food delivery company founded 2008 owned inter alia by Uber and AliPay. Zomato provides information, menus and user-reviews of restaurants as well as food delivery options from partner restaurants in select cities. The service is available in 24 countries and in more than 10,000 cities. In financial year 2021, the average monthly active users for Zomato were 32.1 million users.
Designed for hyperscale. Banking Sector. Case study
Case 1: compute propensity towards products aggregated into classes. Classes represent similar financial products (e.g. investment or insurance products). The goal is to compute probability scores of purchasing at least 1 product from each group, for each customer. The predictions should reflect probability of purchase within a given timeframe from now - e.g. 2 months.
Case 2: compute propensity towards purchase via a given channel. There are multiple existing sales channels, which are grouped into digital and non-digital channels. The goal is to predict the probability of each user buying at least 1 product via the digital/non-digital channel within a set timeframe.
All models should allow to understand which factors (such as sociodemographic, seasonal or interaction-level features) motivate user decisions of purchase/lack of purchase.
Dataset and technical information
Synerise Monad delivers state-of-the-art models with individual propensities for 45 product/channel combinations for all 35M customers with full interpretability insights within 3 hours of pretraining and 3 hours of inference.