Beyond Bot Labels: Retrospective Consequence-Grounded Ranking of Attributed Cache Harm in E-Commerce Logs

Jisoo Kim, Youngwoo Seo, Seonghoon JeongarXiv 2026 · arXiv

TL;DR

Instead of just hunting for "bots" based on identity, this paper proposes ranking web sessions by the actual damage they do to a cache—specifically, how many "victim" misses they cause by evicting useful data. Their method identifies a tiny fraction of traffic responsible for nearly half of all cache harm, much of which is completely missed by traditional bot labels.

What Problem Does This Paper Address?

Web caches are vital for e-commerce performance, reducing latency and origin server load. However, automated traffic (scrapers, crawlers) often disrupts "cache locality." They may evict objects that human shoppers need, leading to lower cache hit ratios (CHR) and sluggish site performance.

Existing bot defenses typically try to infer automation from User-Agent (UA) strings or navigation patterns. The authors argue this is the wrong target: identity is a poor proxy for consequence. A "good" bot might warm the cache by requesting popular items, while a "human-looking" scraper might be devastating by sweeping through cold, unpopular URIs. This paper shifts the focus from who a client is to what harm they cause to other users.

Methods

The authors introduce a consequence-grounded ranking pipeline that operates retrospectively on server logs:

  1. Attribution: Using a single chronological replay of logs, they attribute every "victim miss" (a miss by a protected human shopper on a previously cached object) to the specific session whose insertion last evicted that object.
  2. Behavioral Modeling: They train a gradient-boosted regressor (HistGBM) to predict this attributed harm using 20 behavioral features (e.g., request volume, URI rarity, and timing). Crucially, they exclude UA strings and identity tags to ensure the model learns harmful behavior rather than just labeling known bots.
  3. Retrospective Bypass Simulation: To evaluate the impact, they simulate "bypassing" the top-ranked harmful sessions—sending their requests directly to the origin without letting them read from or write to the shared cache—and measure the resulting CHR improvement for the remaining human shoppers.
Fig. 1. Pipeline from raw traffic (❶) through preprocessing, attribution, and ranking (❷–❼) to bypass replay and CHR evaluation (❽–❾). Injected traffic enters stress tests only.

Datasets

  • EClog: A primary 183-day trace from a Polish e-commerce store containing ~12.7 million replayed requests.
  • Zanbil: A secondary 5-day trace used to verify that the observed phenomena (like harm concentration) generalize across different stores.
  • Synthetic Personas: Three types of simulated attackers (scrapers and monitors) were interleaved with real traffic for controlled "stress tests" to validate that the attribution method correctly identifies known-harmful actors.

Key Insights and Ideas

  • Harm is highly concentrated: Cache harm follows a steep power-law distribution. As shown in Figure 3, the top 1% of sessions in the EClog dataset were responsible for 48.8% of all attributed cache harm (Gini coefficient of 0.91). This means a very narrow, surgical intervention can have a massive impact.
  • Bot labels are a leaky bucket: Traditional UA-based bot tags missed nearly half (49.1%) of the harm in the top 1% tier. Many high-harm sessions use browser-like UAs to evade detection.
  • Attribution tracks reality: The authors validated their metric using a "leave-one-session-out" check. They found that their single-pass attribution score correlates strongly (Spearman $\rho = 0.80$) with the actual increase in human cache hits when that specific session is removed from the stream.
Fig. 3. Attributed-harm concentration on EClog under 1%-capacity LRU. The top 1% and 10% carry 48.8% and 81% of harm (Gini 0.91). The diagonal denotes equality.
Attributed harm vs. actual victim-hit change

Key Results

  • Efficiency: At a common request-volume target (20%), the proposed ranker achieved 93% of the CHR improvement of a UA-based baseline while selecting 47 times fewer sessions (3.6k vs 167k). This drastically reduces the "audit burden" for security teams.
  • Robustness: The findings were consistent across different cache replacement policies (LRU, ARC, S3-FIFO) and across both the EClog and Zanbil datasets.
  • Victim Safety: The ranker maintained a very low "victim bypass rate" of 0.18%, ensuring that legitimate human shoppers are rarely caught in the crossfire of cache-bypass interventions.

Limitations and Future Work

  • Retrospective Scope: The current method is an offline audit tool. Adapting it for live, "prospective" use would require handling real-time model updates and using only past-looking features.
  • Origin Egress: Bypassing the cache increases the load on the origin server. A production implementation would need "origin-aware budgets" to ensure that protecting the cache doesn't accidentally overwhelm the backend.
  • Simplified Cache Model: The study uses a simplified URI-key application cache. Future work could incorporate more complex HTTP semantics like Time-to-Live (TTL) and Vary headers to better reflect real-world CDN environments.

Summary updated

Abstract

Web caches reduce origin load and latency, but automated traffic can either warm them or evict objects shoppers reuse, lowering the cache hit ratio (CHR). Existing bot defenses generally infer automation from User-Agent (UA) strings or behavior rather than identify sessions that impose cross-session cache harm. We instead formulate the task as ranking sessions by eviction damage and propose consequence-grounded session ranking. A single chronological replay attributes each eviction-induced miss by a protected shopping session (a victim) to the session whose insertion last evicted the object. A behavioral model predicts the target from session behavior alone. Selected sessions bypass the cache but remain origin-served. In a controlled check of 539 non-victim sessions selected across harm levels, attribution order tracks victim-hit changes after removal (Spearman correlation 0.80). On a 12.7-million-request real trace, the top 1% of sessions carry 48.8% of attributed harm. Sessions without bot tags contribute 49.1% of this tier's harm. At the 20% request-volume target, the default conditional CHR difference reaches 93% of the UA baseline, which selects 47 times as many sessions. Six training seeds show little variation. This comparison conditions on the victims that remain, so we also report victim bypass. Local replay and retraining on a second store reproduce harm concentration and a top tier dominated by sessions without bot tags. These results establish attributed cross-session harm as an outcome-based target for retrospective cache audits and selective intervention design.
BibTeX
@misc{kim2026beyond,
  author = {Jisoo Kim and Youngwoo Seo and Seonghoon Jeong},
  title = {{Beyond Bot Labels: Retrospective Consequence-Grounded Ranking of Attributed Cache Harm in E-Commerce Logs}},
  year = {2026},
  eprint = {2611.18027},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18027},
}
Figure