Create competitive
ML solutions in days.
Not months.
Customer churn prediction
Predict credit default
Predict winner in F1
Modern businesses rely on relational data structures. Think of it as the digital backbone of any industry.
Unfortunately, preparing relational data for machine learning requires massive amounts of manual work.
getML automates that.
Trusted by thousands of data scientists, fortune 500 companies & public institutions.
Boost your teams' productivity
getML is a tool for data scientists
getML combines unique feature learning algorithms with state-of-the-art AutoML. Using getML, you can build predictive analytics models on the most complex relational data schemata without a single line of SQL (or equivalent) code for data transformation.
Discover relational learningSpeed up your data science process
Custom ML solutions
Our cross-industry experience in machine learning enables us to build productive applications that are tailor-made for your needs. Starting from data engineering, over feature and machine learning to cloud deployment, we offer to conceptualize, code and maintain custom ML applications.
How we can support youAn expensive problem
Machine Learning
needs features
Building features is an expensive process - it takes up to 90% of your project's time and ties up scarce expert capacities.
It is the #1 bottleneck of every AI/ML journey on relational data.
Popular applications that
rely on features:
- Customer Churn
- Customer activation
- Predictive maintenance
- Sales forecasts
- Resource planning
- Up- and Cross-Selling
- Product return predictions
Real-world data is stored in relational databases.
In a relational database, the relevant information for your prediction models is spread across multiple tables. This general, non-flat data structure applies to any domain or industry.
Read our technical intro to relational data with code examples
But relational data is not compatible with ML.
Every state-of-the-art ML algorithm requires its inputs to adhere to a very specific format: A single, flat table. Prediction-relevant information in a relational database can span multiple tables and does not fulfuill this requirement.
Background article: Researchers have forgotten about relational dataMaking relational data compatible with ML-algorithms is a manual process
It requires domain experts, data scientists and a lot of code to build flat feature tables from relational data.
Domain Experts have day-to-day experience in the business domain. For predicting loan default, these might be the people whose job it is to handle loan applications. Their experience is valuable for defining meaningful features.
Data Scientists are responsible for the feature construction process. Their job is to translate domain knowledge into code that transforms raw data into a format that ML algorithms can understand.
Code, a lot of code. Good machine learning systems require hundreds of features. Each feature can contain hundred lines of code. Features are difficult to code, maintain and can change over time.
Effects of manual feature engineering on ML projects
- A feature’s quality is hard to evaulate, bad features lead to bad predictions and lost profits
- Features can change over time and you need to redo feature engineering
- Relying on the experts’ intuition, Data Scientist don’t know if they have missed something
Effects of manual feature engineering on ML projects
Andrew NG, Co-founder of deeplearning.ai
time-consuming, requires expert
knowledge. Applied machine learning
is basically feature engineering.”
The missing piece
An algorithm that learns features
A Machine Learning model can only be as good as its features. Excellent features are the result of a large-scale search.
The success of this search depends on three key properties:
- The number of features evaluated
- The depth of each feature
- The manual effort required
Here is why feature learning delivers better results, faster.
Here is why feature learning delivers
better results, faster.
Domain-knowledge driven
Manual Feature Engineering
Why you want to avoid manual feature engineering
- Months of manual work
- Requires hard-to-get knowledge
- Reduced predictive performance
- Months of manual work
- Requires hard-to-get knowledge
- Reduced predictive performance
A lot of code and lengthy meetings with domain experts for a small number of features.
Statistical Optimization
Feature Learning
Generates flat feature tables from relational data. All it needs:
- Understanding of data model
- Reduced predictive performance
- Python skills
- Understanding of data model
- Defined prediction target
Let algorithms learn from billions of features instead of coding them by hand.
relational data structures.
Deep learning, has enabled automated feature engineering for images and sound data. getML's feature learning algorithms automate feature engineering on relational data and classical time series.
Data scientists building predictive analytics models using getML's feature learning algorithms can reduce their workload by up to 90%.
A whole lot faster
Business Owner. Project leads can evaluate new AI business cases in days not months by enabling their teams to automate the predictive analytics process with getML. getML's feature learning algorithms deliver accurate models that enable a solid assessment of the maximum potential profit from day one.
Data Scientist. The workload of data scientists is reduced by up to 90% by using getML for feature learning. At the same time, data scientists can stay focused on statistics, making prediction models less dependent on domain experts, even discovering new feature logic.
Return on Investment. Increase your ROI with getML: Its end-to-end automation allows you to validate new business cases faster. Stop rewriting hundreds of SQL, pandas or R/data.table scripts in your production environment. Automatic retraining of your feature logic helps keep your models profitable by adapting features to ever changing patterns in your data.
Beating popular frameworks
Benchmarking getMLs algorithms against popular frameworks from AlterYX, Facebook & Blue Yonder.
Predictive performance. In predicting traffic volume, getML's relational learning algorithms outperform Prophet's classical time series approach by ~14% and tsfresh's brute force approaches to feature engineering by ~26%.
Feature depth is key to predicting baseball players saleries. getML’s Relboost feature learning algorithms beats featuretools propositionalization approach by five percentage points (in terms of the R-squared).
Speed. All of getML’s algorithms rely on our custom-built C++-native in-memory database engine. In benchmarks against propositionalization algorithms tsfresh & featuretools, we find that getML is between 34x to 179x faster, depending on the dataset.