How Lamoda is working on algorithms that understand the desires of the buyer

Soon, online shopping will be a mix of social media, recommendation platforms, and capsule wardrobe shipments. Oleg Khomyuk, head of the company’s research and development department, told how Lamoda works on this

Who and how in Lamoda works on platform algorithms

At Lamoda, R&D is responsible for implementing most new data-driven projects and monetizing them. The team consists of analysts, developers, data scientists (machine learning engineers) and product managers. The cross-functional team format was chosen for a reason.

Traditionally, in large companies, these specialists work in different departments – analytics, IT, product departments. The speed of implementation of common projects with this approach is usually quite low due to the difficulties in joint planning. The work itself is structured as follows: first, one department is engaged in analytics, then another – development. Each of them has its own tasks and deadlines for their solution.

Our cross-functional team uses flexible approaches, and the activities of different specialists are carried out in parallel. Thanks to this, the Time-To-Market indicator (the time from the start of work on the project to entering the market. — Trends) is lower than the market average. Another advantage of the cross-functional format is the immersion of all team members in the business context and each other’s work.

Project Portfolio

The project portfolio of our department is diverse, although for obvious reasons it is biased towards a digital product. Areas in which we are active:

• catalog and search;
• recommender systems;
• personalization;
• optimization of internal processes.

Catalog, search and recommender systems are visual merchandising tools, the main way a customer chooses a product. Any significant enhancement to the usability of this functionality has a significant impact on business performance. For example, prioritizing products that are popular and attractive to customers in catalog sorting leads to an increase in sales, since it is difficult for the user to view the entire range, and his attention is usually limited to several hundred viewed products. At the same time, recommendations of similar products on the product card can help those who, for some reason, did not like the product being viewed, make their choice.

One of the most successful cases that we had was the introduction of a new search. Its main difference from the previous version is in the linguistic algorithms for understanding the request, which our users have positively perceived. This had a significant impact on sales figures.

All ideas are tested

Before new functionality becomes available to Lamoda users, we conduct A/B testing. It is built according to the classical scheme and using traditional components.

• The first stage – we start the experiment, indicating its dates and the percentage of users who need to enable this or that functionality.
• The second stage — we collect identifiers of users who participate in the experiment, as well as data about their behavior on the site and purchases.
• The third stage – summarize using targeted product and business metrics.

From a business point of view, the better our algorithms understand user queries, including those that make mistakes, the better it will affect our economy. Requests with typos will not lead to a blank page or inaccurate search, the mistakes made will become clear to our algorithms, and the user will see the products he was looking for in the search results. As a result, he can make a purchase and will not leave the site with nothing.

The quality of the new model can be measured by the errata correction quality metrics. For example, you can use the following: “percentage of correctly corrected requests” and “percentage of correctly uncorrected requests”. But this does not directly speak about the usefulness of such an innovation for business. In any case, you need to watch how the target search metrics change in combat conditions. To do this, we run experiments, namely A / B tests. After that, we look at metrics, for example, the share of empty search results and the “click-through rate” of some positions from the top in the test and control groups. If the change is large enough, it will be reflected in global metrics such as average check, revenue, and conversion to purchase. This indicates that the algorithm for correcting typos is effective. The user makes a purchase even if he made a typo in the search query.

Attention to every user

We know something about every Lamoda user. Even if a person visits our site or application for the first time, we see the platform that he uses. Sometimes geolocation and traffic source are available to us. User preferences vary across platforms and regions. Therefore, we immediately understand what a new potential client might like.

We know how to work with a user’s history collected over a year or two. Now we can collect history much faster – literally in a few minutes. After the first minutes of the first session, it is already possible to draw some conclusions about the needs and tastes of a particular person. For example, if a user selected white shoes several times when searching for sneakers, then that is the one that should be offered. We see the prospects for such functionality and plan to implement it.

Now, to improve personalization options, we are focusing more on the characteristics of products with which our visitors had some kind of interaction. Based on this data, we form a certain “behavioral image” of the user, which we then use in our algorithms.

How to change following the behavior of online shoppers

An important part of the development of any product is customer development (testing an idea or prototype of a future product on potential consumers) and in-depth interviews. Our team has product managers who deal with communication with consumers. They conduct in-depth interviews to understand unmet user needs and turn that knowledge into product ideas.

Of the trends that we are seeing now, the following can be distinguished:

• The share of searches from mobile devices is constantly growing. The prevalence of mobile platforms is changing the way users interact with us. For example, traffic on Lamoda over time more and more flows from the catalog to search. This is explained quite simply: it is sometimes easier to set a text query than to use the navigation in the catalog.
• Another trend that we must consider is the desire of users to ask short queries. Therefore, it is necessary to help them to form more meaningful and detailed requests. For example, we can do this with search suggestions.

What’s next

Today, in online shopping, there are only two ways to vote for a product: make a purchase or add the product to favorites. But the user, as a rule, does not have options to show that the product is not liked. Solving this problem is one of the priorities for the future.

Separately, our team is working hard on the introduction of computer vision technologies, logistics optimization algorithms and a personalized feed of recommendations. We strive to build the future of e-commerce based on data analysis and the application of new technologies to create a better service for our customers.

Subscribe also to the Trends Telegram channel and stay up to date with current trends and forecasts about the future of technology, economics, education and innovation.