Attention Span For Personalisation

📝 Abstract

A click on an item is arguably the most widely used feature in recommender systems. However, a click is one out of 174 events a browser can trigger. This paper presents a framework to effectively collect and store data from event streams. A set of mining methods is provided to extract user engagement features such as: attention span, scrolling depth and visible impressions. In this work, we present an experiment where recommendations based on attention span drove 340% higher click-through-rate than clicks.

💡 Analysis

A click on an item is arguably the most widely used feature in recommender systems. However, a click is one out of 174 events a browser can trigger. This paper presents a framework to effectively collect and store data from event streams. A set of mining methods is provided to extract user engagement features such as: attention span, scrolling depth and visible impressions. In this work, we present an experiment where recommendations based on attention span drove 340% higher click-through-rate than clicks.

📄 Content

1 Attention Span For Personalisation Joan Figuerola Hurtado
joan@specifiedby.com j.figuerolahurtado@napier.ac.uk

ABSTRACT A click on an item is arguably the most widely used feature in recommender systems. However, a click is one out of 174 events a browser can trigger. This paper presents a framework to effectively collect and store data from event streams. A set of mining methods is provided to extract user engagement features such as: attention span, scrolling depth and visible impressions. In this work, we present an experiment where recommendations based on attention span drove 340% higher click-through-rate than clicks. Keywords Recommender systems, feature mining, data collection.

1. INTRODUCTION Feature mining aims to extract features from data streams [1]. These features are then fed into a machine learning (ML) algorithm to complete a certain task. Feature mining is a key problem in ML. The design of those features determines the success of a ML algorithm. Since recommender systems (RS) extensively use ML algorithms, feature mining is also a key problem.
   A good feature should be informative, invariant to noise or a given set of transformations, and fast to compute [1]. There are several features for RS such as clicks, ratings, or purchases. A click on an item is arguably the most widely used feature in RS [2,3,4,5]. It is informative in a sense that it shows users’ preferences. Furthermore, it is fast to compute, invariant to noise, easy to collect and often non- sparse. However, a click is just one out of 174 events modern browsers can trigger [6]. The data generated by these events is underexplored in feature mining for RS. Consequently, collecting event-stream data for feature mining becomes a challenging task when there are several events. A server could easily be brought down and a client could become non- responsive if there was an event listener for each event triggered by a web browser. Besides that, 50% of the Internet’s traffic is generated by crawlers [10], thus it’s important to detect non-human traffic in order to improve data quality and reduce the amount of data to be stored. As a matter of fact, a data collection framework can provide a structured and easy way to capture interactions between entities as well as facilitating the research of the next generation of features for RS. Some recent research has been focused on attention span. “Attention span is the amount of concentrated time one can spend on a task without becoming distracted. Most educators such as psychologists agree that the ability to focus attention on a task is crucial for the achievement of one’s goals.” [11] In the domain of webpage ranking, [12,13,14] introduced attention span to improve the relevance of search results and personalise them. So the same search query returns different results depending on the user. YouTube also uses the amount of time a user spent watching a video to make personalised recommendations. The motivation to use attention time was to better surface the videos that viewers actually watch, over those that they click on and then abandon [8]. [7,8] Demonstrated that attention span can slightly improve news recommendations. In that context, attention span refers to the time a user has spent interacting with media content. However, the event collectors presented in [7] do not take into account several scenarios where a user’s focus might have faded away, thus giving a false signal of user’s attention. Besides that, attention span can not only be applied to news media, it can also be applied to other domains. Attention span could achieve better results by improving the data collectors and the methods to mine it. In addition, visible impressions [9,15] and scrolling depth [16] are also becoming relevant when trying to better understand how users behave online. This paper provides a novel framework to effectively collect data from event streams. The framework has a low impact on the client and the server. In addition, it is also able to detect and filter out non- human traffic. Furthermore, we present methodology to mine: attention span, scrolling depth and visible impressions. Finally, an experiment is shown where recommendations based on attention span achieved a significant increase in click-through-rates (CTR) compared recommendations based on clicks.
2. METHODOLOGY 2.1 Data Collection An event is an action from a user to an item. For example, a user clicks on a product in an ecommerce site. User is an entity, click is an action and product is a target entity. An event is the atomic unit for data collection and its structure is defined as follows:
   • entityId: identifier of the entity. • entityType: type of the entity. • targetEntityId: identifier of the target entity. • targetEntityType: type of the target entity. • type: event type. • timestamp: time when the event happened. • ip: ip of the client

This content is AI-processed based on ArXiv data.