Abstract
This paper introduces a unified Theory of Information Matching (TIM) that models relevance as a bidirectional, logical relationship between the properties of an information need (a query or user) and an information item (a document or product). By mathematically mapping what a query seeks in a document and what a document seeks in a query, the authors derive a flexible probabilistic ranking framework applicable to both ad-hoc text retrieval and collaborative filtering.
Introduction: The Asymmetry Problem
At its core, Information Retrieval (IR) involves matching two sets of features: a user's query and a database of documents. However, traditional probabilistic models, such as BM25 or standard Language Models, are fundamentally asymmetric. They typically fix one side of the equation—optimizing the document for a fixed query or vice versa.
This asymmetry makes it difficult to simultaneously incorporate rich, heterogeneous features from both sides, such as a document's PageRank alongside a user's historical click-through data. In collaborative filtering, models often rely on computationally heavy dimensionality reduction or explicit similarity metrics. The Theory of Information Matching (TIM) aims to bridge these gaps by proposing a single, unified probabilistic theory that natively ingests all available information from both the query and the document simultaneously.
A Unified Theory of Bidirectional Matching
The central hypothesis of TIM is that relevance is a two-way street. Both queries and documents are defined by sets of properties. Relevance occurs when the item possesses the properties sought by the query, and the query possesses the properties sought by the item.
To formalize this, the authors introduce two mapping matrices: $Y$, representing which item properties a specific need is looking for, and $Z$, representing which need properties a specific item is looking for. Using Bayesian transformations and independence assumptions, they derive a ranking function that scores relevance based on the joint probabilities of these properties.
Applications: Text Retrieval and Collaborative Filtering
The general theory is applied to two specific domains:
- Ad-Hoc Text Retrieval: Properties are defined based on vocabulary terms. A 2-Poisson mixture model handles term frequencies, with parameters estimated via Expectation-Maximization (EM) and Gibbs sampling.
- Collaborative Filtering (CF): The theory models user "preference" and item "appeal." This allows relevance to propagate through a network of users and items without requiring explicit user-user similarity calculations or latent vector spaces (like SVD).
Key Insights and Feature Integration
A significant byproduct of the ad-hoc retrieval derivation is that it naturally recovers the Inverse Document Frequency (IDF) scoring function. Unlike the classic Robertson-Spärck Jones model, TIM derives IDF without needing the assumption that the entire collection is a "non-relevant" set.
Furthermore, by treating relevance as a dual "seeking" process, the model creates a mathematical scaffolding that naturally ingests arbitrary features—such as geolocation, authority scores, or user history—from both sides of the search equation.
Experimental Evaluation
The Unified Model (UM) was tested on title-only queries against strong IR baselines, including BM25 and Language Models with Jelinek-Mercer (LM-JM) and Dirichlet smoothing. The evaluation used seven standard TREC document collections (e.g., FT-8, FBIS-8, LA-8, WT10G).
- Superior Accuracy: The UM achieved the highest Mean Average Precision (MAP) and Mean Reciprocal Rank (MRR) across almost all tested datasets.
- Significant Gains: On the FT-8 dataset, UM achieved a MAP of 0.347 (vs. BM25's 0.323) and an MRR of 0.724 (vs. BM25's 0.649). These improvements were statistically significant on several datasets.
- Robustness: The model achieved these results on short queries without relying on explicit relevance feedback, proving the effectiveness of the term-based parameter estimation.
Conclusion and Future Work
While the current mathematical derivations rely on simplifying assumptions to remain tractable, the TIM framework provides a robust foundation for multi-feature matching. Future work includes extending the rank function to modern web search by incorporating click-through data and applying the theory to multimedia retrieval, job search, and targeted advertising.
Summary updated