R&D- People Change: Online Label Shift Adaptation

I was told by my cofounder that this sounds like ChatGPT. Well let's see ChatGPT spit this 🔥.

'sup 🤓. Do you know what sucks with marketing? People change constantly. You figure something out then a few minutes later their mind changes and like I get this with a toddler but are like, all my consumers toddlers? It's a pain in the 🍑 to keep up with shifting trends constantly. I hated this when I was a marketer and since I moved into AI research to deal with this, I figured I would share one tool we use to mirror consumer sentiment in our model. So we're putting out this white paper, an adaptation of our research debuted at The Thirty-Seventh Annual Conference on Neural Information Processing Systems (NeurIPS) 2023 (this is like big boss 💩 in the academic world), that discusses the importance of addressing online label shift (OLS) for:

market researchers
digital marketers
digital merchandisers
brand strategists/identity-ers
digital content managers
ecommerce managers

If you're like a super-nerd then please check out the paper, fascinating stuff.

Introduction

Supervised machine learning algorithms typically assume that data is independent and identically distributed (iid). Aka all data points are equal, but like the great Napoleon (the 🐷 not the 🥖) said: "... some are more equal than ours". (Orwell would be so proud he's getting quoted in such a hyper capitalist setting.) But, real-world environments evolve dynamically. This evolution leads to distribution shifts, where the statistical properties of the data change over time. When these shifts are not addressed, the performance of machine learning models can degrade significantly. That's why what worked last week doesn't work this week, that's why a lot of models can suck. Ours did so we added this in.

One specific type of distribution shift is label shift, where the proportion of different categories or labels changes over time, while the relationship between features and labels remains constant. For example, in digital marketing, the popularity of different product categories may change due to trends or seasonality, but the factors that influence a customer's purchase decision within each category largely remain the same.

This white paper focuses on online label shift (OLS), where label distributions change continually and unpredictably. OLS poses a significant challenge for marketers who need to adapt their strategies in real-time to capture shifting audience preferences.

Understanding Online Label Shift

In technical terms, OLS occurs when the marginal distribution of labels, Q(y), changes over time, while the conditional distribution of features given labels, Q(x|y) (read: Q of x given y is true), remains fixed. In non-technical terms, the popularity of different categories changes, but the characteristics of examples within each category stay the same.

Importance of Addressing OLS:

Improved Accuracy: Adapting to OLS can lead to significant improvements in accuracy compared to ignoring the shift. The paper demonstrates improvements of 1-3% in accuracy by addressing OLS.
Better Label Marginal Estimation: Algorithms designed for OLS can provide more accurate estimates of the true label proportions, allowing for better-informed decision-making without changing the input training data set.
Minimax Optimal Dynamic Regret: OLS algorithms can achieve optimal dynamic regret, which means they perform almost as well as the best possible sequence of models in hindsight.

Implications for aiphrodite.ai's Users

Audience Understanding: OLS methods provide a way to track and understand how audience preferences and behaviors evolve over time. By accurately estimating label marginals, marketers can identify emerging trends and adjust their targeting strategies accordingly.
Personalization: By adapting to shifting label distributions, marketers can deliver more relevant and personalized experiences to their audience. For example, if the proportion of customers interested in sustainable products increases, marketers can adjust their messaging and product recommendations to reflect this shift.
Budget/Resource Allocation: OLS algorithms can help marketers optimize budget allocation by identifying which channel, categories, and/or segments are growing in popularity and which are declining. This information can inform decisions about advertising spend, content creation, and product development.
Real-time Adaptation: OLS methods are designed for the online world, allowing marketers to adapt their strategies in real-time as new data becomes available. This is particularly important in fast-paced digital environments where trends change rapidly.

Advantages of Online Label Shift

No Convexity Assumptions: The algorithms do not require convexity assumptions on the loss functions, making them applicable to a wider range of models, including deep neural networks and decision trees.
Optimal Dynamic Regret: The algorithms guarantee optimal dynamic regret without requiring prior knowledge of the extent of drift in the label distribution.
Computational Efficiency: The algorithms are computationally efficient, making them practical for real-world applications.
Adaptability: The algorithms can be adapted to different types of data and shift scenarios, making them versatile for various digital marketing applications.

Conclusion

Online label shift poses a significant challenge for those understanding shifting audiences, but also presents an opportunity to gain a deeper understanding of such online audiences. By implementing algorithms designed to address OLS, marketers can adapt their strategies in real-time to capture shifting audience preferences, deliver more personalized experiences, and optimize allocation. The novel algorithms discussed in the referenced research paper offer an industry-leading approach to tackling OLS.