Research:Understanding Engagement with Images in Wikipedia/Second Round of Analysis

Despite the crucial role of Wikipedia as a central hub for knowledge sharing and learning, little is known about its visual content and the way readers interact with it while browsing the encyclopedia. In this study, we aim to fill this gap by providing a characterization of reader engagement with images on Wikipedia. Specifically, we quantify reader engagement with images, explore the main drivers of such engagement and ask whether images may support the reader's need for additional information when navigating Wikipedia.

Data & Methods

Data

Traffic data

For the goals of this study, we make use of the server logs collected for four weeks in March 2021 for English Wikipedia. From the logs, we extracted three types of reader actions:

• image views: corresponding to image visualizations in the Media Viewer, displayed anytime a reader clicks on an image;
• page loads: logged every time a reader visits any Wikipedia page (in our study, we only consider articles, i.e. Wikipedia pages in the main namespace);
• page previews: logged whenever a reader hovers over a link to an article while browsing Wikipedia via desktop.

Image and article data

At the time of our data collection, English Wikipedia consisted of 6.2 million articles containing 5 million unique images. It is worth noting that only 44% of the articles were illustrated by at least one image. We characterized the images in our dataset by a rich set of features inspired by the literature around the cognitive perception of images in instructional or web environments, such as:

• features from the image context: page topic, length, popularity, readability, length of image caption, image placement (in terms of (i) the image text offset from the beginning of the page and (ii) its positioning in one template—infobox, inline, gallery), and image resolution;
• features from the image content extracted by running a set of computer vision-based classifiers: image quality, presence of faces, and presence of outdoor settings.

Methods

We quantify reader engagement with images by means of three main metrics:

• the global click-through rate: defined as the fraction of reading sessions with at least one click on an image, and measuring the overall reader engagement with images;
• the image-specific click-through rate: defined for each image as the ratio of clicks to impressions;
• the conversion rate: defined for each wikilink as the fraction of sessions that clicked on a preview after hovering on it.

Results

How much do readers engage with images?

On average, we find the global click-through rate to be 3.5% across English Wikipedia, or, in other words, one in 29 page loads results in a click on at least one image. Notably, this is one order of magnitude higher than engagement with citations on Wikipedia . We also observe that engagement is higher for desktop (5%) than for mobile web (2.6%) readers, with behaviors that vary also across the days of the week: on desktop devices especially, readers tend to click more on images during the weekdays (Monday to Friday) rather than on weekends.

What factors drive reader engagement?

Coefficients of the logistic regression model predicting the image-specific click-through rate from the image features. (A) Coefficients of the model trained on the article topics. (B) Coefficients of the model trained on the other variables.

To address this question, we perform a set of multivariate analyses:

• starting with an exploratory and correlation analysis, we find that engagement tends to be lower for longer and popular articles. Regarding the image size, we observe no clear effect on engagement. When considering the image position instead, we find the images in galleries to have a higher image-specific click-through rate than those in the infobox or within the text. Finally, we observe high engagement among images depicting outdoor sceneries, but low engagement when showing faces.
• to understand the predictive power of the image features collected, we train a logistic regression classifier using the image-click through rate as target (split at the median into high vs. low values). We observe that clicks on images are more often related to topics such as transportation, visual arts, geography, and military. On the contrary, clicks on images are less likely in education, sports, and entertainment articles. Regarding visual content, we observe a strong positive effect of outdoor settings. Also, images in galleries are more likely to be clicked than images in other parts of the text. Noteworthy, the presence of faces has a negative impact in predicting a high level of click-through rate. To dive deeper into this latter result, we apply a clustering algorithm to the image features. From the analysis of the clusters, we observe that images within unpopular biographies show a higher click-through rate than images in popular biographies.

Are images useful to fulfill part of the reader’s information need when navigating the website?

Conversion rate of previews with vs. without images across different levels of page popularity.

When browsing an article, readers can hover over wikilinks to display a short preview of them. Such previews contain a concise textual description of the target article and may contain an image or not. We ask whether such an illustration may be useful to the reader to fulfill its information needs during its navigation. To address this question, we compare the conversion rate of previews with vs. without images. The results show that it is more likely to click on non-illustrated previews than on illustrated ones and that this pattern is consistent across different popularity levels.