In many online platforms—whether you're booking a flight, picking a hotel, or clicking a news article—you are making a discrete choice from a set of alternatives. Accurately predicting these choices is vital for recommendation systems and platform success.
Traditionally, machine learning models predict these choices using an "item-wise" approach: they assign a score to each item in isolation and pick the highest. However, these models usually suffer from a rigid constraint called Independence of Irrelevant Alternatives (IIA). IIA assumes that your preference between a flight to London and a flight to Paris shouldn't change just because a third, expensive flight to Rome is added to the list.
In reality, humans are not perfectly rational. We are heavily influenced by context, decoys, and reference points. While some advanced models try to break IIA, they often fail to scale to millions of data points or lack the theoretical grounding to explain why they work. To solve this, researchers from Harvard have developed a new framework called Set-Dependent Aggregation (SDA), a "Goldilocks" solution that is expressive enough to capture complex human behavior but structured enough to be efficient and theoretically sound.
The SDA Framework: Baking Psychology into Neural Networks
Instead of a single static score function, SDA learns a collection of score functions that are combined based on the specific set of items presented to the user.
The core of their approach incorporates four key principles from behavioral decision theory directly into a permutation-invariant neural network architecture:
- Latent Dimensions: Users evaluate items across multiple hidden features.
- Relative Value ($r$): Value is calculated relative to a set-dependent reference point (e.g., the average price of all flights currently shown on the screen).
- Loss Aversion ($\mu$): Gains and losses are perceived asymmetrically, modeled by an s-shaped function.
- Set-Dependent Weights ($w$): The importance of a specific feature (like "travel time") can change depending on the other options available in the set.
By hard-coding psychological concepts like reference points and loss aversion into the neural architecture, the model achieves targeted flexibility. Unlike "black-box" models (like Deep Sets) that try to be flexible everywhere, SDA maintains a primary score function but allows it to be overruled in specific contexts where human behavior typically deviates from standard logic.
Key Results and Insights
The authors evaluated their model on three massive, real-world choice datasets representing complex, large-scale scenarios: Amadeus (flight itineraries), Expedia (hotel clicks), and Outbrain (news recommendations).
SDA demonstrated significant accuracy gains across the board. On the Amadeus dataset, SDA improved Top-1 accuracy by 8.8 points (a 24.2% increase) over the standard Multinomial Logit (MNL) model. Furthermore, it outperformed much larger and more complex models while remaining compact and easier to train.
Interestingly, the researchers found a "budget of complexity" at play. The dimension of the aggregation (how many score functions are used) acts as a budget for how much the model can violate IIA. As shown in the results above, there are strong diminishing returns: SDA captures roughly 90% of its maximum performance improvement with an embedding dimension of just 4. This suggests that while human choice patterns are complex and context-dependent, they possess a highly learnable, low-dimensional structure.
Future Directions
While SDA represents a major step forward for discrete choice modeling, the current framework is designed strictly for "pick one" scenarios. Future work will need to adapt these principles for sequential choices (picking several items in a specific order) or combinatorial choices (picking a bundle of items). Additionally, while the model is inspired by psychology, mapping its learned "latent dimensions" back to human-understandable attributes remains an exciting frontier for interpretability.
Summary updated