Hello everyone, 大家好, Tēnā koutou katoa, my name is Hua. I am really happy to have you guys here. Today, I’m going to talk about artificial intelligence-powered recommendation algorithms, also known as RAs. I will divide my presentation into five parts. This is my overview of what we’ll be covering:
- I am gonna to briefly explain what it is RAs and its history.
- The Advantages
- Challenges, Concerns
- Also, Things the Stakeholders could do for RAs
- and, Conclusion
Now, Let’s get started.
The definition of RAs is a combination of algorithms used by internet platforms to identify and recommend content, products, and services.
The RAs have developed for 30 years can be roughly separated into four stages:
The first stage, before 2000, introduced user and item-based collaborative filtering, which focused on the similarities in computation. These foundational techniques marked the beginning of personalised content.
In the 2000s, Latent Semantic Analysis improved text recommendations by matching items and users in a special space. Then, in 2006, Simon Funk’s matrix factorisation became a big deal during the Netflix Prize competition. This competition had a $1 million US dollar reward and attracted a lot of people involved. They used to improve Netflix’s recommendation algorithm by 10%. Even though the competition was stopped in 2009 due to some controversial issues, it has contributed a lot to the RAs community.
So, as you can see here on the left of the side, all techniques are based on traditional algorithms, computer vision and some supervisored learning models. From the right side, it is going to be real AI-powered.
In 2015, Google introduced wide and deep learning models. In 2016, YouTube started using deep neural networks for recommendations. We also saw innovations like Pinterest’s PinSage for big graph-based recommendations and Alibaba’s deep reinforcement learning for dynamic recommendations.
After 2020, RAs are getting into a prosperous situation where a hundred flowers bloom and a hundred schools of thought. One of the most prominent benchmarks is ChatGPT4 - em. - O. I think many of you already know how powerful they are, right?
OK, now let’s move on to the advantages of RAs:
In terms of End User Experience, it basically includes Customisation, Personalisation and User Engagement. Users can tailor their preferences, and RAs can provide unique content to each individual user. This makes users feel valued and understood, increasing their satisfaction. They are more likely to spend more time on the platform, exploring and interacting with the content. This ongoing engagement is very important for user retention and loyalty.
Regarding platform revenue, it incorporates cross-selling, upselling, targeted marketing, and conversion rates. Everyone knows that in an internet company, as long as you have user traffic, then you can make a lot of money. Netflix’s recommendation system accounts for about 80% of the content watched on the platform. It indicates how significant these systems are in driving user engagement and, consequently, revenue.
However, there are still some challenges and concerns about RAs. First of all, Let’s start with the “Rabbit Hole” effect. This term describes how RAs can lead users to get into a path of increasingly narrow content. For example, YouTube’s algorithm often suggests videos that become progressively more specific. Like me, I was wondering about learning some DevOps techniques. I started with a Docker container. Then, I found a Kubernetes tutorial and decided to learn Kubernetes. As I kept learning, more related tutorials such as Terraform and Ansible kept popping up, and again and again, the endless suggested videos never stopped catching my attention. So this is a problem; It’s not that I don’t want to learn something new just because data science students are always really busy, right? I can not dive into so much; I have other assignments to do.
Next, “Echo Chamber.” This happens when RAs continuously suggest content that aligns with a user’s existing beliefs, reinforcing their opinions without exposure to different perspectives. This can be particularly tough on social media platforms where users only see information that matches their ideologies. This kind of scenario is very dangerous, especially for extremists.
Finally, “Filter Bubble.” This concept is about users getting stuck seeing the same kind of information over and over. The algorithm only displays information they liked before. This means users don’t often get different opinions, which is kind of being isolated. Consequently, the polarisation phenomena might emerge in that situation.
Other issues such as data privacy, security, algorithm fairness, and transparency. I am not talking about these issues today because I found so many classmates have already spoken about them. The reality is that regulations always lag behind algorithm development and implementation.
So we should think about what kind of solutions we could use.
First, platforms should add diversity metrics and user control to their algorithms. For example, YouTube used some metrics to make sure different content creators generate balanced content. Likewise, giving more user control to improve user interfaces and experiences is very important. Platforms like Spotify have easy-to-use interfaces that let users adjust their preferences, which helps with transparency and user authorisation.
Secondly, individuals can improve their digital literacy, which means learning how to navigate online content and understand AI systems. Countries like Finland have media literacy programs in schools to help people become more critical of online information. Moreover, using privacy-focused tools such as the DuckDuckGo browser can also help address this kind of problem.
Finally, students, especially data science students, have a significant responsibility as future stakeholders. We should aim to create fair and unbiased algorithms, develop transparent and trustworthy models, and provide high-quality, reliable data for better recommendations. Suppose we keep focusing on these kinds of areas so that we will have a better digital environment in future.