In this week’s workshop, I worked with Facebook Ads and Google Ads to see how these platforms “categorise” me as a user and consumer by examining the advertisement preferences that they assigned to my account. In general, there were not that many differences about how the platforms were seeing me.

While Google categorised me under both professional and leisure themes like jobs, business, productivity, and education, as well as skincare, and fashion advertisements for my account (see IMG 8.1), Facebook’s Ad Preferences section leaned heavily towards lifestyle-related categories (see IMG 8.2).

Interestingly, these categorisations did not fully make sense to me. I spend a significant amount of time engaging with cultural posts, memes, and content related to art and community, yet these interactions seem to be ignored in favor of commercial aspects. For instance, my Facebook feed usually has post suggestions like cultural and academic memes, literature, and opinions. However, the way my Ad Preferences section only focuses on lifestyle categories is overlooking the cultural and creative dimensions of my activity and reduces my interests to what can be monetised through targeted ads. This reveals how platforms prioritise certain data points, those that align with their commercial goals, while sidelining aspects of my identity that do not directly contribute to advertising revenue.

I would say it is hard to change the output since I have actually been interacting differently with the platform, such as engaging more selectively with content that reflects my cultural and creative interests, in hopes of reshaping the data these algorithms process. Yet, nothing really changes with the way the content promoted on my Facebook newsfeed is mostly about clothes, accessories or skincare. Generally, I do not trust social platforms in terms of actually reducing advertisements and promoted content.

The second task that I found interesting while doing was the Sumpter method. For this task, I had to look at the most recent 15 posts from 32 of my Facebook friends and categorise them into 14 categories (see IMG 8.3).

I was fortunate that most of my Vietnamese friends and family are still active on Facebook, so looking at the most recent 15 posts of all of them was not difficult. However, the challenge arose when it came to categorising these posts. The issue is not just the number of categories available, but the nature of the content itself. No matter how many more categories I created in my mind, there would always appear posts that did not quite fit into any one box, or those that could easily fit into multiple categories at the same time. For example, a post that mixes a personal anecdote with political satirical commentary and humor might not easily be categorised as just “Jokes/Memes”.

If I were to iterate on this process, I would likely change or refine the existing categories. The current labels are too broad, and they do not capture the full range of complexity and nuance within the posts. For instance, adding more categories, such as “selfie” “study” or “personal achievements” could help, but even then, the lines would still blur. If the categories were too rigid, they would miss the context of the posts.

Additionally, during the task, I changed the way I interpret posts to include more context. Instead of simply categorising based on surface-level traits, I asked more questions about the post’s intent, its tone, or the larger cultural or emotional context behind it. This allowed me to better capture the essence of each post, but I also had concerns about my own biases and subjectivity.

This process, and the choices I have made, reveal that categorisation is not a straightforward or objective task. The outcome of this iteration suggests that even with well-defined categories, context is essential to accurately capture the meaning behind each post. This is important because it shows that data analysis, particularly in social media and its algorithms, cannot always be reduced to simple categories. There is a complex interplay between the content, context, and intent of posts that algorithms and manual categorisation often fail to capture fully.

This week, I had a chance to work on machine learning, a topic that both fascinates and intimidates me at the same time as a non-STEM student. Why the fascination? From my very personal view, machine learning has been a buzzword to describe the “magic” of Artificial Intelligence, a concept I have been curious about for some time. My interest grew even bigger after reading Dr. Stephanie Dick’s (2019) work on the history of artificial intelligence. I am particularly impressed by the idea that aspects of human intelligence can be broken down and described so precisely and detailed that machines can mimic and “learn” them. However, I am also intimidated by it because it is a vague, complicated, technical, science-y field that I feel like I can never truly understand, or go to the core mechanics of it. This workshop gave me a chance to engage with machine learning hands-on. And, ironically, both spectrums of my feelings towards machine learning are reinforced after this.

We worked with Teachable Machine, a user-friendly and interactive tool for exploring machine learning. The tool offers three main functions: image, pose, and audio. I mainly worked on the image and pose functions as testing the audio function in class felt inconvenient (and I was not entirely comfortable speaking out loud). The image recognition function was both fun and engaging. I collaborated with Chesca and we took turns to appear in front of the camera to see if the machine could differentiate between us. Initially, the machine performed well. It identified each of us with 100% accuracy. However, problems arose when my face was out of frame and only parts of my body, such as my shoulder, visible while Chesca was on screen. The machine still identified me as being fully present (more than 90% me while only 10-20% Chesca), as though my face were still visible.

Later, I individually explored the pose recognition function. This tool detects pose and movement along a horizontal axis and responds to gestures (like heart signs), tilts, and other motions. I experimented with 5 different poses: no pose, tilted head, heart sign, V sign, and half heart sign. The tool did well only when I performed each pose precisely as in previously captured “model” photos. Disappointingly, when I zoomed out farther or zoomed in nearer than the model photos, it started to mess up and identify as another pose (see IMG 7.1). Also, it did a bad job in identifying my hand gestures. For example, when I did the V sign, it incorrectly identified as the finger heart sign (see IMG 7.2).

These experiments helped me realise the strengths and weaknesses of machine learning. I appreciated how accessible and engaging the tool was; and exploring what the machine could do blew my mind a little bit. Yet, its struggles with context and variability really stood out to me. It showed how crucial precision is when training these models. Again, the recurring questions of machine learning and artificial intelligence in general arise. Can these machines truly handle nuances? How much detail needs to be fed into the system to improve its accuracy?

Reference:

Dick, S. (2019). Artificial intelligence. Harvard Data Science Review. https://doi.org/10.1162/99608f92.92fe150c

(I am still working on this week's reflection. Please come back later. Thank you!)

I believe data visualisation is not just a technical skill or a simple way of presenting information; it is a powerful process that shapes how we perceive and understand the world. As discussed in the reading, the ideological work of data visualisation is significant because it is not as neutral as it often seems (Kennedy & Hill, 2017). Data visualisation can be shaped by the subjective choices of the people creating them. These choices, such as what data to use, how to display it, and which aspects to highlight, have a great impact on how the information is interpreted. By presenting data visually, not only reality is presented; it is knowledge that is actively constructed in ways that reflect certain values and power dynamics. This is why visualisations carry so much weight, they can influence public opinion and shape societal understanding of complex issues, even when the full context of how the data is presented is obscured.

The reading also brought up how data visualisation can represent knowledge, and I think this is an important point. The design aspect of data presentation influences what and how we understand it. When creating a visualisation, decisions about scale, colour, and which data points to emphasise or obscure, are considered, all of which affect how the information is received. In my own experience with data visualisation, I have seen how the design of a visualisation can either clarify or confuse, which is why it is so important to approach the process thoughtfully. For example, I have come across numerous powerful data visualisation projects by advocacy sites like Slow Factory and Palestine Youth Movement, which shows the current situation in Palestine, ranging from the number of fatalities to famine and poverty rates. The data presented on their social accounts are carefully curated to emphsise the numbers, figures, words, etc. and helps me understand the situation in a more comprehensive way, and prompts me to reflect on the horror and destruction caused by the ongoing war. This shows that data visualisation is so much more than just an academic or corporate tool. It is an instrument for activism and resistance that is used to push back against the forces that seek to control or distort knowledge.

Reference:

Kennedy, H. and Hill, R.L. 2017. The Pleasure and Pain of Visualizing Data in Times of Data Power. Television & New Media, 18(8), pp. 769-782