Learn how we use machine learning to focus on marketplace users with promoting campaigns based on their specific interests.

An introduction to the Adevinta marketplaces project

The concept is straightforward: at Adevinta, we would like to focus on audiences with promoting campaigns based on their specific interests. Nevertheless, in an effort to do that, we’d like to discover our users’ interests. As these interests could also be seasonal, we discuss with them as . We assume that we are able to find out about our users’ hobbies based on their previous browsing behaviour. The answer is intuitive: we start by segmenting users based on their past browsing behaviour, then goal and enrich their experiences depending on which segment they’re in.

Nevertheless, a single user can have multiple hobbies, the hobby could also be seasonal, or the interest could also be deep within the user’s behaviour timeline. Due to this fact, segmenting users for promoting campaigns may not lead to sales. It can also be ineffective in increasing traffic (visits, clicks, or conversions) to our marketplaces (the platforms where buyers can sell their goods or services to a select group of shoppers).

Due to this fact, we redefined the issue to binary classification, where we simply classify whether the user is a possible buyer or not. With this approach, we are able to find potential buyers in each category and increase sales by targeting promoting specific to them.

What we set out to attain and why

Our marketplaces have a big user base. Some users are needs-based, one-time users, whilst others are more regular as a result of their habits or season. HomeNGarden users, for instance, usually tend to visit our marketplaces within the spring and summer to enhance the looks of their gardens when the weather allows them to spend more time there. One other example is summer sports enthusiasts, who’re prone to visit marketplaces during summer or before, to buy summer sports equipment. Consequently, understanding specific user interests in our marketplaces may very well be useful to achieve the correct customer, at the correct time, with the correct product. We call this the

Why do we’d like to achieve the correct users with the correct product at the correct time? An important reasons are:

• To extend user activity in our marketplaces and drive more business
• To motivate customers to persist with their routines
• To search out recent potential customers

Find out how to hack that hobby

We began working on this seed idea as a team during a hackathon in March 2022. The difficult task was defining the issue. We considered segmenting user behaviours, but that didn’t align with our goals because a single user can have multiple hobbies (overlapping interests), and user interests may change based on the season. We were also unable to learn the segments as a result of an absence of definitive segment labels. Without clear evidence of whether this solution could generate more traffic and sales, we decided to deal with understanding the acquisition intention of individual customers based on their behaviour.

We began by specializing in a single hobby, HomeNGarden, and defined a baseline.

Baseline: customers who purchased within the previous nine months will make one other purchase in the approaching months.

The baseline results are adequate, with a precision (percentage of correct users caught by the model) of 75% and a recall (percentage of total right users who did the model capture) of 44%.

The initial machine learning (linear approach) models couldn’t outperform the baseline metrics, so we continued with carrying out A/B testing on the e-mail campaign using the baseline results. The e-mail campaign was successful. This demonstrated that targeting potential buyers will increase the reach of our marketplaces, so we reiterated the identical model approach.

Dataset preparation

We used data that was available including: listing views, searches, searches saved by them, bids, favourites, unfavourites, and messages, if buyers interacted with sellers.

We would have liked historical purchase information from users for labelling, but peer2peer payments are typically only available for a subset of things in a listing, leading to limited label data to discover patterns from. So, we defined and labelled connection2point (C2P) as a purchase order. Below is an outline of C2P, with a visible representation in Figure 6. Consequently, we were in a position to generate enough label data for our model.

C2P: A buyer messages a seller, and the vendor responds, then the customer messages back, indicating that each the customer and the vendor made a connection.

Defining the issue for the model to know

The dataset is a chronological collection of user events, including purchases, but we wanted to work out tips on how to tell the model which specific behaviour indicates purchase intent. The plan was to watch behaviour for a set period and see if it repeats itself. But how precisely and for a way long should we examine the behaviour? We would have liked a mechanism to separate user behaviour into sessions with, and without buys.

We experimented with various static time partitions to watch user behaviour and feed it right into a learning model with assorted session-based features. Nevertheless, these initial partitions were too large to pinpoint potential buyer behaviour.

Within the second-hand e-commerce market, the vast majority of users don’t subscribe. Due to this fact, there is no such thing as a established contract between users and the marketplace. Users are typically only energetic for a brief period before deciding whether or not to buy an item, so we selected a more dynamic approach to partitioning the behaviour into smaller partitions. We then observed the behaviour in smaller sessions.

We collected user interactions, sorted them by timestamp, and cut them into dynamic sessions. Most of our sessions are half-hour or less, following e-commerce industry standards. The next two rules describe how we partitioned the user behaviour dynamically:

• When a p2p (purchase) event occurs
• When there is no such thing as a activity from the user for greater than half-hour

The feature engineering stage

The engineering of features is crucial because it means we are able to use data collected from our platforms to make assumptions about our users. For the aim of this test, we identified and defined the attributes of user behaviour that will represent purchase intent. Based on the research and domain knowledge, we identified the feature set. We discuss a couple of behavioural attributes and explain why and the way they indicate purchase intent.

The variety of item views within the session: The upper the variety of items viewed, the upper the interest within the product or category.

The whole period of time staying within the session: The longer a customer spends, the more interested the user is in a product or category. A positive relationship exists between dwell time and user interest within the items during a session.

The period of time spent on each product in a session: The upper the extent of interest, the longer the user spends on an item.

The best and lowest popularity of viewed items in a session: Popularity features indicate how popular a product/listing is in a given session. The more popular the product/listing, the more likely it’s to be purchased. It shows that the user has a selected item in mind and is viewing the listing multiple times.

Whether the user is solely browsing aimlessly (window shopping) or in the event that they are on the lookout for something specific in a specific category: A user who browses in a specific genre is more prone to buy a product from that category.

If the user browses in patterns that align with specific seasons: We derive characteristics that represent seasonality because it might be an excellent indicator when combined with purchasing behaviour.

The modelling stage

We selected a straightforward feed-forward neural network (Multi-Layer Perceptron) since it is sweet at detecting non-linear relationships and interactions between features.

The model outperforms the baseline, with a precision of 97% and a recall of 70%.

We scheduled the model in near real-time and derived the likelihood of user intent to purchase a product in each category. Currently, potential buyers receive listing offers based on their interests or hobbies but we would like to make use of the outcomes of this model further to customize users’ homepages for higher personalisation, along with targeted campaigns.

In conclusion — successful!

Our approach worked well with the machine learning method providing stronger results than the rules-based method. With a model that’s generic and simply adaptable to other categories or hobbies, we are able to alter the feature and model setup to search out potential customers for specific campaigns. For instance, if we desired to run a campaign to advertise sports gear to customers, we could goal users who’ve higher levels of purchase intent within the sports category. Also, we imagine that this product has the potential to scale across our different marketplaces.

What comes next?

More improvements are on the best way as we don’t consider the order of events in our current approach. For instance, when a user views the listing, spends significant time on the listing, favourites, and bids on the listing, this sequence of behaviour may indicate a powerful indication of intent to purchase. Using sequence behaviour and sequence approaches will enable us to enhance the relevance and strength of our model. Once the present A/B test results can be found, we are going to keep improving our model to be even more practical in identifying users with intent to purchase across our marketplaces.

Do you might have any comments on our methodology? Or tips about further improvements? Please get in contact.