Our Case Studies
AI supervision and observability is core part of every Artificial Intelligence Project. Uncover ML Cube Platform’s Real-world Applications in Our Case Studies!
Application
Sales Forecast
In a multinational company, strategic planning is a very complex task given the number of customers, the number of shops, and the variability between countries. Sales forecasting is one of the most important and valuable tools for making informed and functional decisions. We consider a setting in which a complex hierarchical supervised learning model has been developed to predict sales, learning from customers behaviour.
The implemented model is employed as a simulation, to test different business scenarios (cost of devices, marketing expenditure, trial period length) and identify the best strategy.
SALES FORECAST
Model Maintenance Challenges
Customers behaviour may change over time due to multiple factors: this may impact simulations and lead to wrong decisions. Being the implemented model a hierarchical one, the even the drift of a single model can have a large effect. Since the output of a model is the input of the following one, errors propagates along the models chain, influencing the final sales prediction.
SALES FORECAST
ML cube Platform Benefits
The ML cube Platform allows to monitor all the different models of the chain at the same time, raising an alarm when there is a drift in one of them. This allows to re-train only the models that are under-performing, reducing the computational burden. When a drift is present in input data, the Platform can detect it even before target data are available, promptly notifying users.
Application
Automated Trading
Automated trading systems are steadily increasing their impact on financial markets, however, generating profitable strategies using solely historical data remains an ambitious goal. The possibility of developing an algorithm that can plan for an uncertain future without relying on economic or financial assumptions makes Reinforcement Learning an attractive alternative to classical approaches.
We consider a setting in which an RL model is trained with historical financial data in order to execute automated trades.
AUTOMATED TRADING
Model Maintenance Challenges
The effectiveness of such strategies is strongly conditioned to market stationarity. This hypothesis is challenged by the regime switches frequently experienced by practitioners. Therefore, it is paramount to identify drifts in the market as soon as possible. Moreover, when specific models are available for different regimes, it may be difficult to select the best model to employ.
AUTOMATED TRADING
ML cube Platform Benefits
The ML cube Platform allows to monitor both the data used to compute predictions, and the prediction quality itself. Whenever a drift is detected in the data, the platform sends an alert to inform the user that the market regime might have changed. This may help a trader to promptly act to understand what is going on, even before starting to lose money. In case different models are available, the expert learning feature itself automatically selects the best running model, limiting the incurred losses. Finally, in case none of the available models performs in a satisfactory way, a new trading strategy may be trained by weighting past data according to the suggestion provided by the ML cube Platform.
Application
E-commerce
Pricing products within an e-commerce can be a complex problem. This is because an e-commerce store often has a large number of products in its catalogue and defining a price for each one is cumbersome and not always possible due to exogenous factors. We consider a setting in which an online pricing algorithm based on Multi-Armed Bandit is in production. At each timestep the algorithm chooses among a set of possible prices, quickly learning the optimal solution, based on purchase data.
E-COMMERCE
Model Maintenance Challenges
While the model in production is effective in proposing the optimal prices in stationary conditions, adapting to market changes may be difficult. A periodical re-training using the most recent window of data may be sufficient to deal with a slow market change. However, when drifts are abrupt, this kind of datasets risks to present drifts within it.
E-COMMERCE
ML cube Platform Benefits
Thanks to the ML cube Platform it is possible to re-train the model only when drifts are detected, and thus, to train the model with data compatible with current data distribution. This allows not only to obtain better performance, but also to avoid unnecessary training when conditions are stationary.
Application
PV System Maintenance
Photovoltaic systems are becoming increasingly popular and more widely used as a renewable energy source for the transition in electricity production. Solar panels can suffer a considerable number of failures due to constant exposure to the elements, resulting in corrosion of the devices’ electrical components. In addition to stopping electricity production, a failure can also be dangerous by short circuits resulting in fires or other electrical problems.
We consider a setting in which machine learning models have been trained to identify the presence of faults, their nature and intensity, by means of electrical and atmospheric data. Fault diagnosis allows the maintenance technician to know what repair measures need to be taken in advance, thus considerably reducing time and costs.
PV SYSTEM MAINTENANCE
Model Maintenance Challenges
Photovoltaic panel are subject to physical degradation of hardware components through years, being exposed to various atmospheric agents. This phenomenon reduces their power production, and influences the fault diagnosis system performance, since it changes the panel nominal behaviour. This change is incremental, thus, it is particularly hard to detect. Moreover, being atmospheric data subject to yearly seasonality, it is important for the dataset to include several years in order to learn the correct pattern at different moments of the year.
PV SYSTEM MAINTENANCE
ML cube Platform Benefits
By continuously monitoring model performance and input data, the ML cube Platform can detect this kind of drifts, notifying the user when a certain confidence threshold has been passed. Furthermore, it can propose a training dataset which contains data with different levels of degradation, by weighting them in the proper way. This feature allows to avoid discarding past data after a degradation, which is paramount when seasonality is present.