When you search for information online, you are usually looking for a specific answer contained within a single document. Traditional information retrieval (IR) systems are beautifully optimized for this task: they match query keywords against a massive index of pages and rank them by immediate relevance.
However, a different challenge arises when you want to find an authoritative source rather than a single piece of content—such as subscribing to a specialized blog, tracking an industry influencer, or identifying a central portal for a hobby. In these scenarios, standard keyword matching breaks down. A generalist news site might publish a single viral post about cycling that ranks highly for keyword relevance, but a user looking for a dedicated community hub wants a source that covers cycling deeply and continuously.
To solve this, a distinct paradigm known as source distillation or feed retrieval has emerged. Instead of evaluating individual documents in isolation, this approach aggregates content across time to identify streams that exhibit a sustained, recurring interest in a specific topic.
The Architectural Split: Macro-Documents vs. Voting Models
To evaluate a source, an IR system must first decide how to represent it. The literature generally splits into two architectural philosophies for aggregating evidence across multiple posts.
The first approach is the macro-document model. As explored by Elsas et al. in their work on Retrieval and Feedback Models for Blog Distillation (TREC 2007), this strategy concatenates an entire feed's historical output into one massive document. By treating the source as a single entity, standard document retrieval and query expansion techniques (such as leveraging external structures like Wikipedia) can be applied directly to capture the source's overall thematic footprint, as further demonstrated by Arguello et al. in Document and Query Expansion Models for Blog Distillation (TREC 2008).
The second approach is the voting model. Rather than merging text, this strategy indexes individual posts independently. When a query is issued, the system finds relevant posts and treats each hit as a "vote" for its parent source. Hannah et al. demonstrated the efficacy of this approach in University of Glasgow at TREC 2007: Experiments in Blog and Enterprise Tracks with Terrier (TREC 2007), showing that combining post-level proximity features with structural voting significantly outperforms raw document matching.
To make this computationally viable, Weerkamp et al. introduced a two-stage framework in Blog feed search with a post index (Inf. Retr. 2011). A fast "Posting" model first filters candidate documents, and a detailed "Blogger" model then aggregates the scores. This multi-stage approach achieves high precision while reducing computational overhead by over 70%. Additionally, Na et al. showed in KLE at TREC 2008 Blog Track: Blog Post and Feed Retrieval (trec 2008) that evaluating passage-level relevance within these posts helps filter out the boilerplate noise inherent to decentralized publishing.
Macro-Document Approach:
[Post 1] + [Post 2] + [Post 3] ---> [ One Giant Document ] ---> Standard IR Index
Voting / Two-Stage Approach:
Query ---> [Post Index] ---> Hits: Post 1 (Source A), Post 2 (Source A), Post 3 (Source B)
---> Aggregation / Voting Model ---> Rank: Source A > Source B
The Virality Problem: Why Simple Aggregation Fails
Whether you use macro-documents or voting models, simple summation introduces a critical flaw: it is highly susceptible to transient spikes. If a major news event occurs, a generalist blogger might post twenty times about it in a single week. A simple count or text aggregation would mistake this sudden burst for long-term expertise, misidentifying the generalist as a niche authority.
To separate true authority from transient engagement, systems must evaluate the temporal distribution of the content. Early web track experiments by Shakery in Relevance Propagation for Topic Distillation UIUC TREC 2003 Web Track Experiments (TREC 2003) recognized the need to identify central portals rather than just content-similar pages. In dynamic streams, this intuition translates to measuring thematic persistence.
As Ernsting et al. discovered in Language Modeling Approaches to Blog Postand Feed Finding (TREC 2007), the frequency with which a creator publishes on a specific topic over an extended timeline is a far more reliable indicator of feed authority than the recency or immediate volume of the posts. He et al. reinforced this in University of Glasgow at TREC 2008: Experiments in Blog, Enterprise, and Relevance Feedback Tracks with Terrier (TREC 2008), proving that a consistent pattern of activity across time serves as a powerful filter against noisy, short-term data bursts.
Quantifying Consistency via Temporal Entropy
To operationalize thematic persistence, researchers turned to information theory—specifically, Shannon entropy. Entropy allows algorithms to mathematically weigh a topic's immediate intensity against its long-term dispersity across time.
In ICTNET at Blog Track TREC 2010 (TREC 2010), Xu et al. introduced an entropy-based "Recurring-Degree" metric. Instead of merely counting relevant posts, the metric divides the timeline into discrete intervals (e.g., weeks or months) and measures how evenly the relevant content is distributed across those buckets. A source with a high volume of posts concentrated in a single week yields low entropy (high predictability, indicating a transient spike), whereas a source that consistently publishes a few relevant posts every month yields high entropy across intervals, signaling a sustained, recurring interest.
Transient Spike (Low Temporal Dispersity / Low Value Source):
Intervals: [Week 1: 20 posts] [Week 2: 0 posts] [Week 3: 0 posts] [Week 4: 0 posts]
Result: Low recurring degree.
Sustained Interest (High Temporal Dispersity / High Value Source):
Intervals: [Week 1: 2 posts] [Week 2: 3 posts] [Week 3: 2 posts] [Week 4: 2 posts]
Result: High recurring degree.
This significance-dispersity trade-off is highly effective at separating structural campaigns from flash-in-the-pan news. Oghaz et al. formalized this in Probabilistic Model of Narratives Over Topical Trends in Social Media: A Discrete Time Model (arxiv 2020) by treating time as a discrete category to extract persistent narratives from chaotic social media streams.
The same principle applies to video platforms. In Deriving latent social impulses to determine longevous videos (www 2014), Hu et al. bypassed raw view counts—which are easily skewed by viral events—by isolating discrete interest triggers from their natural temporal decay. This allowed them to mathematically score a video's true "longevity."
Coherence and Predictability
Beyond tracking time, source distillation requires analyzing the internal focus of the creator's history. A source that writes about everything from cooking to quantum physics might occasionally hit a topic consistently, but it lacks the specialized authority of a dedicated channel.
To capture this, He et al. developed a topical coherence measure in An effective coherence measure to determine topical consistency in user-generated content (Int. J. Document Anal. Recognit. 2009). Their framework calculates how tightly a source's posts cluster together in a vector space compared to a general background collection. If the posts are tightly bound, the source demonstrates high topical consistency.
This historical consistency is highly predictive of future behavior. In Predicting Author Blog Channels with High Value Future Posts for Monitoring (aaai 2011), Wu et al. showed that authors who maintain a narrow, consistent niche are significantly easier to model and predict than generalists. By leveraging these historical consistency profiles, their system can anticipate which specialized channels are most likely to publish high-value content on a specific topic in the near future, before the conversation even spreads.
Ultimately, these models demonstrate that surface-level engagement metrics are poor proxies for authority. As surveyed by Ciceri et al. in Assessing Online Media Content Trustworthiness, Relevance and Influence: an Introductory Survey (KDWeb 2015), low-quality information propagates through networks just as fast as high-quality data. By shifting the focus away from transient metrics like likes or shares, and toward structural signals like content persistence and temporal consistency, source distillation provides a robust blueprint for navigating noisy, decentralized media landscapes.
Key Foundational Work
Language Modeling Approaches to Blog Postand Feed Finding (2007) — contributed 2 concept elements to this direction.
Retrieval and Feedback Models for Blog Distillation (2007) — contributed 1 concept elements to this direction.
University of Glasgow at TREC 2007: Experiments in Blog and Enterprise Tracks with Terrier (2007) — contributed 1 concept elements to this direction.