10h11 participated on December 18th in the Hackathon Lyreco Frictionless Supplies and was voted favorite. We then wanted to explain to you the solution we have developed.
Lyreco’s objectives were as follows :
- Optimize inventory management ;
- Avoid shortages of supplies ;
- Automate orders.
Thanks to the history of orders placed on the Lyreco Webshop, we have implemented a technological engine that allows us to predict from one week to the next the quantities ordered of each product reference for each of Lyreco’s customers. The prediction period can be adjusted according to the frequency of customer orders. This predictive engine is based on the Machine Learning XGBoost algorithm (eXtreme Gradient Boosting). It feeds and learns from new orders placed by each customer in order to be more and more efficient in terms of prediction.
To evaluate the performance of our model, we use the RMSE (Root Mean Square Error) to obtain the average deviation in the prediction of the quantities ordered for a customer’s product. This deviation only makes sense if it is compared to the minimum and maximum quantities ordered during the model’s test period. The RMSE is therefore not a percentage! Indeed, regression models (quantity prediction) are not evaluated using percentages, unlike classification models (class prediction). For more information on these 2 prediction families, you can refer to article.
Showcasing knowledge through use
In order to automate orders and thus save customers time, we have designed a chatbot directly integrated into the Lyreco WebShop. Thus, as soon as a customer connects to the Webshop, he automatically receives an order proposal, via the chatbot, that he can validate, modify or refuse. The proposed order is the result of the predictive engine. For example, if the customer is used to ordering pens of reference A every week, the model (via the chatbot) will propose to the customer to order a certain quantity of reference A according to the orders placed by this customer. This solution allows the customer to spend less time ordering the supplies they need.
We could also imagine a system of recommendations for products not ordered by a customer. For example, and to simplify the approach, let’s imagine that 80% of Lyreco customers who order reference A also order reference B. Then, we could recommend to the 20% of the other customers who order this reference A to order this reference B via the chatbot. The objective is not to invade the customer with order suggestions but on the contrary to offer him relevant products likely to interest him.
Thanks to the NEMO dashboard we have designed, Lyreco can know the results of the predictive engine and see in real time the chatbot conversations of each of its customers. In addition, NEMO allows you to know the customers who are most loyal to the order proposals made to them via chatbot. This can enable Lyreco to anticipate orders from loyal customers and thus optimize its inventory management, for example.
Vimeo link : https://vimeo.com/318002218
10h11 adapts its products according to the needs of its customers, and of its customers’ customers. Features can be added, modified or deleted to the NEMO dashboard.
 https://www.analyticsvidhya.com/blog/2018/09/an-end-to-end-guide-to-understand-the-math-behind-xgboost/  https://machinelearningmastery.com/classification-versus-regression-in-machine-learning/