Continuous Automated Machine Learning

Machine Learning product deployment is difficult and takes time. As business changes, redeploying working ML models becomes even more difficult.

HAKsys's Continuous Automated ML deployes fast.
It requires very little maintenance and intervention in production to produce high accuracy results continuously.

HAKsys does this by supporting the entire data and analytics pipeline continuously by its:
- its ability to integrate to different data sources easily
- automated data cleaning, transformation and selection procedures
- advanced modeling enabled by automated configuration of hundreds of different machine learning models
- continuous machine learning pipeline to support continuous business
- high performance and scalability on real time data streams
- augmented intelligence enabled by understandable models that can learn from data and humans, continuously

Continuous Automated ML for Retail

Restaurant Demand Prediction

Challenge:

Food preparation is performed in stages. Meat is thawed, cooked and then served with the other ingredients. The cooked meat and the prepared food have a very short lifetime and they are discarded after a certain amount of time. On the other hand, having too little prepared meat causes longer customer queues.
The restaurant manager tries to track stages of food preparation and amount of food prepared, however, demand keeps changing dynamically and depends on many variables.

Result:

Continuously trained tazi models predict demand for the next 15 minutes for 10 food items and in 4 major locations. Heterogeneous features on many aspects of customer behaviour, such as the demand and price for previous time slots and locations, actual and predicted weather conditions, holidays, #cashiers available, customer satisfaction feedback, marketing campaigns are used by the models, Mean absolute error of 4% is obtained.

Clothing Retail Demand Prediction

Challenge:

Whether each type and color of clothing will be sold at a particular store needs to be determined so that enough items can be shipped. Currently experienced allocators decide on what to send to each location, however their accuracy might be as low as 55%. This means not only the extra cost of returning the items that are sold, but also lost opportunity for items that could be sold. Demand for each item at each location keeps changing dynamically and depends on many variables related to the item, other items in the store, weather, customer segment, price and marketing campaigns.

Result:

On 4 different category of clothing whether a certain item would be sold for the next 14 days is predicted. tazi accuracy of demand prediction is first 3% better and then with more data 9% better than the allocator. Considering that each percent improvement results in 20 million USD in revenues, tazi benefit is remarkable.

Flight Delay Prediction

Challenge:

The airline has more delay probability than other higher cost airlines and this causes customer complaints. In order to be able to take action for delay, the company wants to predict which flight will be delayed at least 24 hours in advance. Delay prediction after the gate is closed also needs to be predicted so that customers and the crew can be informed beforehand. Blackbox ML models are not acceptable since they do not help with mitigaring actions to prevent delay.

Result:

Delay prediction 24 hour in advance with explanation showed that the number of elderly or child passengers, checked in baggage and weather all contribute to delay. Company is in the process of planning actions, such as allocating more crew, when delay is expected with high probability.