Machine learning is only a minor part of a typical artificial intelligence project. Nevertheless, it is this part that marks each project and generates value. The higher the accuracy of the forecasting algorithms, the greater the value generated by their implementation.
Finding the most accurate model is a particularly time-consuming task and requires teams of experienced data scientists. However, even the data scientist profession is not immune to the trend towards automating tasks through algorithms, and we can currently observe the emergence of new technologies called "AutoML" (automated machine learning).
With Verteego, we apply this trend to the field of forecasts.
The basic element of Verteego is Synapp®. It contains training templates, a parameter file and interfaces to connect to datasets, other applications or other Synapps®. Each new predictive application starts with the creation of one or more Synapps®.
Choosing the right algorithmic model is the key to achieving high accuracy in the forecasting phase. There are hundreds of models on the market, most of them belonging to open source libraries. Verteego offers in its native version the most frequently used algorithms (decision trees, regressions, time series, neural networks, etc.) and makes it easy to "plug in" other libraries. During model training, Verteego automatically compares the different models activated through parameterization files according to prioritization criteria that have been defined by the user.
Model hyperparameters (user selectable and so called as opposed to parameters not selected by the user) play a key role in forecasts performance. Thus, the same model can produce totally divergent results depending on the set of selected hyperparameters. The user can "guide" the Synapp® in the choice of hyperparameters by indicating specific ranges in the configuration file. Nevertheless, Verteego natively integrates the most powerful hyperparameter selection methods to achieve high accuracy results, even without manual selection.
One of the most time-consuming tasks in the work of the data scientist is the correct selection of the explanatory variables of the model (also called features). It is important to include as many explanatory variables as possible, as these increase the accuracy of the forecasts, without adding superfluous variables that could create unnecessary "noise" and deteriorate the quality of the forecasts. Verteego uses the best techniques for evaluating the relevance of the best performing variables to relieve the user of this sometimes complex task.
One of the main strengths of Verteego is its ease of use. We wanted to design a technology with which it becomes possible to create and put into production a machine learning application in less than 10 minutes and we think we have succeeded in our bet. Verteego is prepared in such a way that it is possible to launch a first learning phase without any prior manual parameterization. The installation of Verteego automatically includes a documented REST API and requires no additional time.
However, it was also important for us to provide experienced users with a high level of technological depth as well as freedom in parameterization. As a result, once the first Synapp® is deployed, it can evolve significantly through changes made to the configuration file and the integration of new datasources. As accuracy increases, the application remains ready for production.
Verteego is not a solution for data preparation (like Talend, Trifacta, Dataiku and others). Nevertheless, sometimes it can be convenient to modify input data "on the fly", without having to fully regenerate the underlying datasets.
For this purpose, Verteego allows, very intuitively, to set up preprocessing rules through the settings file. For example, it is possible to generate additional variables, calculated from other variables, to define rules for replacing certain values, to exclude outliers according to certain well-defined criteria, etc.
In some cases, there may be anomalies in the forecasting results. This happens in particular when the input data is not of optimal quality. It then becomes necessary to correct the results by implementing business rules of various types (e.g. correction of outliers, replacement of null values, etc.). Verteego makes it easy to define these rules through its settings file.
Your data sets can be very heterogeneous. Depending on the input data sets used in the training, one or the other algorithm may be more efficient. However, it is technically complex to use different algorithms in combination .
With Verteego, this constraint is a thing of the past, so you no longer have to choose between different approaches. Depending on the type of data, Verteego will combine the best modeling approaches for each subset of data to achieve the best overall accuracy.
Explanability and transparency are essential to make machine learning acceptable for use within an organization. Decisions or recommendations provided by machine learning algorithms must be fully understandable to those who wish to use them.
Therefore, the training of the models as well as the generation of the predictions must show perfect transparency on the origin of the results, the weighting of the variables used, the selected algorithms and their hyperparameters.
The predictions generated by Verteego are intended to generate decision making in the organization, either fully automated or driven by humans following the recommendations issued by artificial intelligence. It is then essential that the origin of real decisions is traceable over time.
Verteego keeps in memory all historical models as well as the sets of predictions generated. This allows users to go back on past decisions, understand the parameters of the models used, compare them with each other and possibly reuse them.
Verteego was designed to accelerate the implementation of applications that increase the performance of business teams. As our pilot customers have been using our technology, we have identified recurring business issues that we have decided to integrate natively into our solution.
Verteego contains a dozen Synapps® models already configured and ready to use. These templates allow you to launch applications without any configuration and to understand the incoming data in order to answer complex business questions in no time at all.
Some examples of pre-configured Synapps® in Verteego : sales forecasting, traffic forecasting, markdown optimization, price optimization, etc.
In many cases of use, you will need to plan for data with little or no history (for example, planning new fashion collections, product launches, opening new points of sale, etc.).
Verteego technology incorporates features specifically designed to effectively address this common problem and will automatically detect correlations between variables that best describe your data without history to automatically identify the most accurate prediction strategy.
In forecasting demand, it is essential to be able to differentiate between out-of-stock and low sales in order to obtain accurate forecasts.
Verteego contains a ready-to-use out-of-stock detection processor that uses machine learning to finely understand the correlations in your data in order to estimate whether a value is an out-of-stock or zero or low sales. The annotations you make to your data can then be used as a new input variable to make your prediction models more accurate.
In some cases, it may be useful to generate multiple predictions with slightly different inputs. For example, to predict sales quantities under different price, promotion, or merchandising scenarios.
As this is a particularly common use case, we have developed a dedicated model that allows you to create a Synapp® in just a few clicks.
Whether you are looking to forecast your annual revenue, quantities sold in a future fashion collection or hourly traffic in your points of sale or online store, whether you want to generate a single template for all your data or millions of separate templates, one for each possible combination of items and points of sale, Verteego has the right solution for you.
Training and prediction resolutions can easily be specified in the configuration files without having to adapt anything else to the underlying infrastructure.
Putting machine learning applications into production and maintaining them represents an important part of the time spent on a typical project. Verteego has been designed to highly automate the time spent on these tasks.
Verteego was built to work in virtual environments, hosted in the cloud or locally. If you don't want to worry about running Verteego in your own environments, our SaaS solution manages everything for you. Whichever runtime solution you choose, your license will always include all code updates, enhancements and features.
Verteego is independent of the cloud technology you wish to use. Our technology is containerized and can be deployed without difficulty and in its native version at major cloud providers (Google Cloud, Amazon Web Services, Microsoft Azure, etc.).
If you prefer to use Verteego in SaaS, the solution will run in a secure and dedicated GCP environment.
Verteego has been designed to make no difference between a proof-of-concept and a production application. You can test many configurations quickly and inexpensively, and when you are ready for deployment, simply increase the data volume or the scope of your project, because the execution environment will be the same.
Verteego is designed to improve the performance of your organization without having to change any processes. You simply benefit from more accurate forecasts that improve your overall performance.
Verteego comes with a documented high performance API that allows you to make your existing systems interact with our technology.
Verteego is designed to be fast in all situations. It will use all available resources to reduce the time needed for deployment, model learning and forecasting. And when that work is done, resource consumption slows back down to near zero so you only pay for what you use.
The Verteego license was inspired by the pricing models of the major cloud providers. You only pay for what you consume. Our financial charges mainly come from the computing power consumed during the training of the machine learning models. The consumption of predictions and data storage represents only a minor part of the budget to be allocated.
Verteego platform's parallelization technology (based on Kubeflow) allows the use of virtually unlimited computing resources in the cloud to create its learning models. Thanks to the parallelization capabilities employed natively by our technology, Verteego will be able to create accurate models in a decent timeframe, regardless of the volume and complexity of your data.
Data sources can be internal or external, raw or calculated data, hosted locally or in the cloud.
By simply modifying the configuration files, Verteego can connect to many different data sources.
Verteego can easily use data from different sources by configuring how to combine them.
The confidentiality of your data is a top priority for us. In its SaaS version, we rely on the best practices for securing infrastructures as well as the tools made available through the Google Cloud Platform. Upon request, we would of course be happy to provide you with a detailed description of our security policy and programming practices.