Machine learning of product recommendation and the matching algorithm rely on large quantities of data. There are two key stages to the algorithm:
- Recognition of the best-match product based on user input. In this, our algorithm processes several data sources (user inputs, knowledge on the user, product popularity in the area, etc.).
- Finding the best alternatives for the selected product. The idea behind it is to suggest the best price-to-quality ratio. This is done through our algorithm reading the blueprint of the product and then comparing it with other products in our database. This is the next step compared to the current specification-based alternative product algorithms on the market.
We are currently developing two areas of predictive modelling:
- Current product value estimation
We are focusing on correct estimation of the product price. This requires state-of-the-art machine learning methods: sophisticated gradient boosting methods, random forests, deep NNs, etc.
- Future product value prediction
In this predictive model, we are focusing on predicting the price of the products in the course of time. This is based on several types of machine learning methods. We approach the problem as a time series prediction and evaluate different methods and distance functions suitable for the task at hand. We have therefore started with ensembles of predictive clustering trees for time series prediction (coupled with quantitative distances, dynamic time warping distance, and cosine similarity). We will also use other standard methods used for time series prediction (forecasting).
We initially intended to use one of the already existing chatbot frameworks to use for chatbot functionality for our discovery features. After testing them, however, we started a branch of chatbot framework development within our company. This was done because all the frameworks on the market had some limitations that could affect our end product functionalities. Our first chatbot will mainly be used to guide you through the process of product discovery, but the framework itself is developed in a way that can easily be trained in additional interactions.