Abstract
Explicitly modelling field interactions and correlations in complex document structures has recently gained popularity in neural document embedding and retrieval tasks. Although this requires the specification of bespoke task-dependent models, encouraging empirical results are beginning to emerge. We present the first in-depth analyses of non-linear multi-field interaction (NL-MFI) ranking in the cooking domain in this work. Our results show that field-weighted factorisation machines models provide a statistically significant improvement over baselines in recipe retrieval tasks. Additionally, we show that sparsely capturing subsets of field interactions based on domain knowledge and feature selection heuristics offers significant advantages over baselines and exhaustive alternatives. Although field-interaction aware models are more elaborate from an architectural basis, they are often more data-efficient in optimisation and are better suited for explainability due to mirrored document and model factorisation.
Introduction
In traditional Information Retrieval (IR), multi-field documents—such as a recipe containing a title, a list of ingredients, and a photo—are often "collapsed" into a single embedding or treated as a flat bag of words. This approach ignores the rich, structured relationships between fields. For example, a user searching for "spicy soup" might prioritize the ingredients field over the description field.
Furthermore, fields are often highly correlated; the title "Pepperoni Pizza" inherently implies the ingredient "pepperoni." If a model fails to account for these correlations, it can struggle with redundant information, leading to harder optimization and less accurate ranking. This study investigates whether "field-aware" neural models can better navigate these complexities within the cooking domain.
Methodology
The researchers compared two primary neural architectures designed for multi-field data:
- NRM-F (Neural Ranking Models for Fields): A model that learns query-field interactions but typically ignores field-to-field interactions.
- FwFM (Field-weighted Factorization Machines): A flexible model that explicitly models pairwise interactions between all fields (e.g., title-to-ingredients) as non-linear functions.
Two "interaction selection" strategies were tested: Query-field (modeling only how the search query interacts with recipe fields) and All (modeling every possible pairwise interaction between all fields). The study utilized real-world search log data from Cookpad, using six fields: Query, Title, Ingredients, Description, Country, and Image. Models were trained to predict "clicks" as positive signals.
Analysis of Field Interactions
Analysis of the predictive power of various field interactions revealed several counter-intuitive insights. Notably, the "title-country" interaction emerged as the most significant predictor of a click, highlighting the weight of regional preferences in recipe selection. Conversely, despite the visual nature of recipe browsing, image features showed a relatively low correlation with clicks compared to text fields.
Interestingly, the "query" fields themselves were not the highest-ranked in terms of correlation. This is likely because the dataset only included recipes that had already passed an initial query-based candidate generation phase, meaning the query's influence was already "baked into" the pool of recipes being ranked.
Experimental Results
The researchers used NDCG@20 (Normalized Discounted Cumulative Gain) to measure ranking quality. The FwFM model using query-field interactions provided a statistically significant improvement over all other configurations, including the standard concatenation baseline.
As shown in the performance distribution, the FwFM (query-field) model was not only the highest performer but also the most consistent across different data folds. Surprisingly, the simpler "query-field" configuration often outperformed the "all-field" version, suggesting that modeling every possible interaction can introduce noise that hinders training.
Conclusion
This exploratory study demonstrates that explicitly modeling field interactions, particularly through Field-weighted Factorization Machines, significantly improves recipe retrieval. While the current work focused on the re-ranking stage using offline log data, future research will look toward validating these findings in live A/B tests and exploring more complex embedding techniques like Transformers to capture deeper semantics.
Summary updated