MIM Solutions

We built a model that predicted the probability of buying items depending on the discount received. Initially, to train the model, we used historical data on sales of all items in our client’s store network.
The algorithm developed by our team has been implemented and is currently being used.
In addition, our client organizes many promotions and offers its clients various discounts. In order to optimize the number of ordered products and better planning of marketing campaigns, we have developed a model predicting how a given article will sell in the event of various promotions. As various promotional campaigns often overlap, it was a big challenge to isolate the impact of individual promotional campaigns on sales.
Another challenge was identifying problems on the basis of consumer surveys.
Sometimes customers, after visiting the store, leave comments on various aspects of shopping (prices, promotion, quality of service, etc.). It is not easy for network employees to find out what customers most often complain about – each customer enters a different message.
Therefore, using deep neural networks, we have developed an algorithm that divides comments by topic to facilitate this task.

As part of the activities, we plan to create machine learning models that allow us to determine the optimal hormonal stimulation for a given patient. Stimulation aims to produce the maximum number of eggs, while minimizing the risks associated with stimulation. The model will define a personalized therapy and will take into account, among other things, the patient’s clinical data, stimulation history, and genetic testing.
The recommendation will include i.a. type of stimulation: long / short protocol, stimulation with agonist, antagonist, stimulation with progesterone, etc.
We also plan to recommend optimal doses of the drug. Defining the scope of the recommendation is a key part of the project.
Another decisive element of the system will be functionalities which allow building trust in the algorithms’ results which means determining the certainty of the prediction and indicating the key elements on the basis on which the algorithm made a decision.

ML learning methods are very effective in predicting criminal events. They can be used, among others to predict the place and time where such events may occur and to find spacetime relationships.
Assumptions: The MIM Solutions solution is based on the assumptions similar to the successfully used PredPol system. The PredPol system is the first crime prediction system that achieved significant success in crime prevention. It operates in 11 major cities in the US and the UK. The main advantage of this system is the very small amount of data needed to make forecasts and thus the speed and flexibility of implementation in various centers. The basic version of the system uses solely the information on the place, time and type of event. The aim of the MIM Solutions system is to support police officers in suggesting the place and time of possible future events. The system extracts various crime patterns from historical data (e.g. on Fridays in the evening there are more robberies in this district, on working days break-ins happen most often in this area, etc.) that a human may not notice.
Data used to create the model: The predictive system was based on police data from the decision support system (DSS, in Polish: SWD), which includes the place, time and type of event. The data was processed to obtain the exact coordinates of the place based on the textual description.

This project aims to build the first reliable flight disruption prediction and assessment system that will use the full spectrum of available data.
Other similar systems that are currently being developed rely on limited data and therefore produce results of unacceptable significance. One of the most important aspects of this project is the creation of understandable ML systems,
as a detailed understanding of each claim is required when communicating with the airlines.
Business value:
The project was realized in cooperation with GIVT AG and aimed to:
– help passengers in applying for flight compensation.
– deliver flight reliability predictions, i.e., for a specific flight gives probability that the flight will be delayed.
– alert travelers that the flight might get delayed even before it departs.
For example, our system is able to deliver high quality predictions before the actual flight departure. The figure on the right shows the precision-recall curve for predicting serious flight disruption (i.e., delay by more than 3h) that was done 3h before the flight departure.

Our algorithm made it possible to accurately predict user actions, which is a key component of RTB House technology.
In other words, for a given user and product, we need to predict the likelihood that the user will click on the ad of this product. Our job was to test a variety of models including neural networks, gradient boosting, random forests, factorization machines and various types of regression. We used many techniques, such as k-NN-related functions obtained with LSH (locality-sensitive hashing).
Our LSH regression is available open source in the Practical Approximation Algorithms Library (PAAL)! The solution was created by verifying various hypotheses and testing many different ideas. The cooperation contributed to the rapid development of RTB House and the global success in the online advertising market!

We have created an efficient system that classifies ads in seven different languages from seven different countries. Our solution has replaced much of the work previously done by humans. Moreover, we have designed the system so that adding support for new languages is very easy.
Our system works in two steps:
1. finds a set of good candidate brands for an ad,
2. for each couple – ad and candidate – it predicts whether it is an appropriate match.
This approach allowed us to scale our solution to millions of ads and tens of thousands of brands.
Candidate brands can be found by combining various techniques – extracting text from images, key phrases from ad descriptions, website analysis and logo detection. We identified suitable candidates using state-of-the-art algorithms – XGBoost, LightGBM, deep neural networks, transfer learning and factorization machines. Our solution was presented at the PyData Warsaw 2018 conference.