Development of solutions using machine learning and data analysis

Objectives: to improve their forecasts of purchases by product categories in order to reduce the costs of illiquid, but at the same time fall into seasonal waves of increased demand. This task required immersion in the business processes of procurement, inventory balances, assortment and retail sales. Having plunged into business with our heads, we learned how to track the moments of the category deficit and its overabundance, adding this data to the time series with information about sales in all categories, we compiled our own algorithm for modeling the optimal volume of purchases. As a result, sales increased, and warehouses began to overflow less often.

Task: the local TV channel wanted to reduce the burden on consultants and coordinators, to whom most of the clients turned. They noticed that most of the questions are typical and the coordinators answer them with the same messages. They wanted to make themselves an online assistant and turned to us. We have collected all the detailed answers, made a Word to Vector model out of them and trained the annoying model from Spotify to select the most appropriate answer for the question. It was important to implement the functionality so that the client could call a person at any time or make a purchase directly from the chat, so we added a classifier that distinguished the client's requirements related to the purchase of an advertising placement from questions and from the requirements to call a person. As a result, the average response rate to the question became close to the usual offline request for help in the store, and consultants began to deal more with new, complex and atypical questions, and the company now saves all questions and answers for machine learning.

Task: to increase the speed of finding fraudulent ads. Employees of the anti-fraud department noticed that they often set up filters in a specific way for each district of the city, taking into account the height, novelty of the building, proximity to the subway, etc. all to find atypical ads. This procedure takes a lot of time and requires highly qualified employees, but the popularity of the service has increased in recent years and the department needed search automation. We analyzed the data to detect anomalies and built a model that combines a tree ensemble and regression, which itself finds potential outliers and questionable price values for the proposed conditions. It turned out that the model is good to observe changes in real estate prices. Over time, we finalized it and it began to participate in pricing. Two birds with one stone!

The online store solved the problem of increasing demand for an expanded assortment. He asked us to improve the recommendation system for his users so that customers would be more inclined to buy new products for themselves, and not only from the niches of previous purchases. We collected data on customers and their purchases for 2 years, formed a metric that counts the increase in unique purchases in the user's basket over the test period and began to build a recommendation system. The most effective approach turned out to be an ensemble of a user-based approach to recommendations coupled with KNN on the product names in the receipt for each client. The forecast announced a 49% hit of recommendations in the client's future products, but this is only in theory, because recommendations affect the purchase itself. An A/B test was conducted, which showed a statistically significant result of the growth of total sales, among which the increase in purchases falling on new products for users was 30.5%

Task: to reduce the cost of sending SMS with discounts, since most of the customers attracted in this way ignored them. Uplift modeling was carried out on CatBoost classifiers, which indicated potential 10-30% of customers whose mailing lists potentially increased the company's profit. When it came to evaluating the result of selective mailing, it turned out that the conversion to the purchase of discount goods has improved twice! And the cost of mailing has almost tripled.