Habr, hello! On the last day of June, our meeting was held for analysts. It was attended by speakers from Leroy Merlin, Ostrovok.ru and, of course, Avito. We discussed regional A/B tests, managing the delivery of goods in a large online store, predicting profit from new features and data science in delivery.


Under the cut - videos of speeches with timecodes and links to presentations.


ITKarma picture


Regional A/B tests. Why are we needed and how are they arranged - Igor Krasovsky, Avito


Using the example of tasks from Avito, Igor told what regional A/B tests are, when the analyst needs them, what algorithms and mathematics are based on them, and how to measure the accuracy of these algorithms.



00:00 - Presentation of the speaker and topic
00:44 - For what tasks can I use regional A/B tests
04:21 - About the model of regional A/B tests: stages of the test, metric of proximity of the control and test groups
09:51 - Algorithm for selecting a test group and evaluating its accuracy
18:03 - What can be improved in the proposed process: growth points


View Igor’s presentation

Growth model - predict profit from features for prioritization - Pavel Mikhailov, Ostrovok.ru


From the report you will learn how the concept of the growth model works, how it helps to translate product metrics into money and assess the potential profit from new features.


Download an example with an example of a growth model.



00:00 - Presentation of the speaker and topic
01:10 - ICE framework (impact, confidence, ease)
02:26 - What is a growth model and how to measure the influence of features using it
05:55 - How to build a simple growth model from scratch
18:55 - Examples of hypotheses that can be estimated using the growth model
25:22 - How can I improve the base model and why do this at all


View Paul's presentation

The best data-products are born in the fields - Marina Kalabina, Leroy Merlin


Most of the online orders in Leroy Merlin are collected in retail stores.Due to the high speed of turnover of goods, this often led to the fact that the order could not be collected. Then the Data Accelerator team came up with a tool that allowed you to automatically find problematic articles and adjust their number before publishing on the site.



00:00 - Speaker and Topic Presentation
01:37 - Leroyisms - terms for better immersion in the context of the report
02:40 - How to collect orders from the online store, and what problems the collector may have
04:07 - Starting the Data Accelerator unit to make data-driven decisions
04:46 - Guaranteed Stock Product: Its Goals and Implementation Process
13:34 - Results of the implementation of guaranteed flow


See Marina's presentation

How data science Avito helped Delivery - Dima Sergeev, Avito


Several million different products are sold daily on Avito. Not for each of them it is easy to determine whether the seller can put it in a box of 120 × 80 × 50 in size and send it to the buyer. At times, we make such mistakes and offer delivery where it obviously should not be and vice versa. Dima talked about how we cope with this problem and what results were achieved.



00:00 - Presentation of the speaker and topic
01:11 - The history of the appearance of delivery in Avito and the first problems
06:32 - Assessing the extent of the incorrect determination of the ability to deliver goods
11:29 - Classification of goods as a way to solve a problem: data science SWAT is in a hurry to help
17:44 - First successes and side effects
25:47 - Immediate plans


Watch Dima’s presentation

See you at the new mitaps!

.

Source