🎥 Recording
Welcome from ML Organisation panel and highlight on what's going on with ML in Adevinta

🎥 Recording
First Kaggle Triple Grandmaster Abhishek Thakur will share his learnings, experiences and vision on the future of ML.

🎥 Recording
In this talk we will present how we improved the performance of the most impactful product provided by Personalisation called Related Items, which presents users with similar ads to the one they are currently looking at. We currently generate recommendations for this product by a combination of two batch algorithms that run in an hourly fashion: a graph-based Collaborative Filtering (CF) algorithm and a text-matching Content-Based (CB) approach. This offline architecture has 2 main limitations: (1) low scalability / high computational costs incurred by precomputing all recommendations every hour, even though only 10-20% of them are actually served; and (2) the item cold-start problem, not being able to return recommendations for classified ads published in the last hour.
We addressed these limitations by moving towards an online hybrid recommendation architecture on top of Elasticsearch, powered by a neural recommender model called LightFM. This model learns item embeddings from user interactions which then are used to retrieve related-items recommendations in an online manner via efficient approximate nearest neighbour search. We will also talk about the novel streaming content ingestion pipeline built on top of Yotta’s self-serve services, which is key to be able to have access to near real time ad content data and generate recommendations also for new ads, which allows us to reach almost 100% coverage. The most recent A/B test on Segundamano.mx has shown a significant uplift in several user-centric online metrics, which validates the effectiveness of the new architecture not only in terms of cost reduction and coverage, but also in user engagement impact.

🎥 Recording
Connecting buyers and sellers in a safe and secure environment is one of the biggest challenges in online marketplaces. Probabilistic models built upon user-item databases address the challenge, but often encounter issues such as lack of stability and robustness. These issues are magnified in fraud scenarios where datasets are highly imbalanced, noisy and malicious users deliberately adapt their behaviors to avoid detection.
In this context, we leveraged the power of existing open sources machine learning libraries H2O and Catboost and designed a pipeline to collect, process and predict the likelihood of a private seller’s listing data to be fraudulent. We found that the stacked ensemble model provides the best performance (F1=0.73) when compared to other commonly used models in the field. Further, our models are benchmarked on a public Kaggle Dataset, TalkingData AdTracking Fraud Detection Challenge where we compared them to other studies and highlighted their generalizability and effectiveness at handling online fraud.

🎥 Recording
Our team (PnR) provides recommendations for various markets, and in this talk we want to give an overview of how our solutions evolved, what were the driving factors, how do they work and what are the advantages and disadvantages of different approaches we saw during our journey.

🎥 Recording
In this talk we will share our way of working and dive into our personalised homepage recommendations. The first half will focus on our ML development lifecycle, the tech stack, the challenges we faced and how we overcame some of them. In the second half we will show how we increased revenue and connections on Marktplaats by adding a personalised ranking layer on top of multiple existing recommender systems. We will discuss the details of its two core components, one for producing relevant content and one for adding a taste of inspiration. We will also provide an overview of the experiments we did in the past year to share our key learnings.

Quiz Challengue, Who knows more about ML?

🎥 Recording
Welcome from Organisation comitee and introduction to Analytics Day

🎥 Recording
Prof. Dr. Arif Wider is a full professor of software engineering at HTW Berlin and a fellow technology consultant (part-time) with ThoughtWorks Germany, where he served as Head of Data & AI before moving back to academia.

🎥 Recording
It was mid of 2019 when the distributed Data mesh concept developed by Zhamak Dehghani arrived to BIC, and there we were able to identify some of the problems we were suffering while building analytics data for central teams.
From this moment on, we started slowly to shift central teams analytics data development to this new paradigm. It's been a long way to get there, as it is not only about engineering or data, it has a big impact on Data culture.
In this session we are going to get and introduction to Data mesh, how BIC is enabling it and some learnings and fails from the trip.

🎥 Recording
An A-Z time series forecasting methodology including all steps from data preparation, exploratory analysis, model fitting/validation/selection (currently in the form of an R script, building a Shiny app is WiP). The idea is to help end-users with no coding skills, to generate robust forecasting and achieve standardisation and transparency of forecasting methods across the organisation. Tested on Vibrancy, SEO and Paid Digital use cases with promising results.

🎥 Recording
Taking from the experience of Q3 and Q4 at Subito, integrating the Houston tool and starting the upskilling programme we will show how challenging but rewarding is to move an entire organisation towards a more data driven product process and creating an experimentation culture.

🎥 Recording
The goal of self-service analytics is to enable everyone in Adevinta to leverage available data in making business decisions - not an easy task considering thousands of Adevintans are making decisions every day!
This presentation will give an overview of what we’ve learned so far about self-service analytics, share best practices and showcase some of our most successful tools.

TBC