Avoiding Over-Personalization with Rule-Guided Knowledge Graph Adaptation for LLM Recommendations

📝 Paper Summary

User modeling Conversational personalization

A neuro-symbolic framework mitigates filter bubbles in recommender systems by selectively restructuring the user's Personalized Knowledge Graph at inference time to suppress over-reinforced feature associations without model retraining.

Core Problem

Over-personalization in recommender systems leads to Personalized Information Environments (PIEs), where algorithms reinforce existing biases and narrow content exposure (filter bubbles).

Why it matters:

Users become trapped in content silos that reduce discovery and agency, repeatedly seeing similar items based on past interactions
Existing solutions often require expensive retraining of large models or rely on opaque embeddings that lack interpretability and user control
Without explicit mechanisms to break these cycles, users struggle to diversify their recommendations even when they desire novelty

Concrete Example: A user who frequently rates tomato-based Italian dishes highly creates a strong 'Italian + Tomato' bias. When asking for a new Italian dish, standard systems recommend more 'Tomato Pasta' or 'Lasagna' (In-PIE), whereas the user might prefer 'Pesto Pasta' (Out-PIE)—relevant to Italian cuisine but breaking the tomato dominance.

Key Novelty

Rule-Guided Knowledge Graph Adaptation

Treat the user's profile as a structured Personalized Knowledge Graph (PKG) that can be symbolically edited at inference time
Detect specific 'PIE-inducing' feature pairs (e.g., Italian + Tomato) where user preference is statistically over-amplified
Apply symbolic rules (Soft, Hard, or Removal) to down-weight these specific edges in the graph before passing it to the LLM, steering generation toward diverse but relevant items

Architecture

Overview of the recommendation pipeline: PIE detection, PKG Adaptation, and LLM Inference.

Evaluation Highlights

Soft adaptation increases Out-PIE (novel, relevant) recommendations from 25.2% to 32.4% compared to a global adaptation baseline
Reduces Invalid recommendations (irrelevant content) from 49.0% to 46.0% while breaking filter bubbles, unlike prompt-based methods which degrade relevance
Prompt-based natural language instructions fail to avoid PIEs effectively, yielding only 19.3% Out-PIE results and a high 62.1% Invalid rate

Breakthrough Assessment

7/10

Offers a lightweight, interpretable alternative to retraining for filter bubble mitigation. While the absolute gains are moderate, the symbolic inference-time control is a valuable architectural shift.

⚙️ Technical Details

Problem Definition

Setting: LLM-based recommendation using Personalized Knowledge Graphs (PKGs) as context

Inputs: User query q and a user-specific Personalized Knowledge Graph (PKG)

Outputs: Recommendation item o that satisfies q while avoiding over-personalization (PIE)

Pipeline Flow

PIE Detection (Identifies biased feature pairs)
Adaptation Strategy (Modifies PKG structure)
Prompt Construction (Injects adapted PKG)
LLM Inference (Generates recommendation)

System Modules

PIE Detector

Identify feature pairs (F_given, F_bias) that cause over-personalization based on rating deviations

Model or implementation: Statistical heuristic (feature-pair bias score q_bias)

PKG Adapter

Apply symbolic rules to modify the PKG structure/ratings to suppress detected biases

Model or implementation: Rule-based logic (Soft/Hard/Removal)

Prompt Generator

Linearize the adapted PKG into a text prompt for the LLM

Model or implementation: Template-based formatter

Recommender LLM

Generate the final recommendation using the adapted context

Model or implementation: Qwen3-0.6B (fine-tuned)

Novel Architectural Elements

Inference-time modification of the user profile (PKG) itself rather than the model weights or decoding strategy
Feedback-driven tuning loop that learns a user-specific 'adaptProportion' to balance novelty and relevance

Modeling

Base Model: Qwen3-0.6B

Training Method: KTO (Kahneman-Tversky Optimization)

Adaptation: Full fine-tuning (implied, no PEFT mentioned)

Trainable Parameters: Not explicitly reported in the paper

Training Data:

Customized version of Food.com Recipes and Interactions dataset
Data points are (prompt, completion, label) triplets

Key Hyperparameters:

learning_rate: 0.05 (for adaptation proportion tuning, not model training)
adaptProportion: Learned per-user or global

Compute: Not reported in the paper

Comparison to Prior Work

vs. Prompt-based: Manipulates the input knowledge structure (PKG) directly rather than relying on the model's instruction following, yielding higher validity
vs. Retraining methods: Operates purely at inference time with lightweight symbolic rules, avoiding expensive model updates
vs. PEFT/LoRA [not cited in paper]: Modifies the *data* (context) fed to the model rather than the *parameters* (adapters) of the model

Limitations

Evaluation limited to a single domain (Food/Recipes) and relatively small model (0.6B parameters)
Requires explicit maintenance of a structured Personalized Knowledge Graph (PKG) for every user
Feedback loop for learning adaptProportion requires synthetic simulation, which may not perfectly match real user reactions

Reproducibility

Code: https://github.com/brains-group/KGAdaptation

Publicly available: Code, dataset construction scripts, and result generation pipelines at https://github.com/brains-group/KGAdaptation. Missing: Final trained model weights not explicitly linked.

📊 Experiments & Results

Evaluation Setup

Recipe recommendation using Food.com dataset

Benchmarks:

Food.com Recipes (Constraint-based recommendation)

Metrics:

Out-PIE rate (Diversity/Novelty)
In-PIE rate (Over-personalization)
Invalid rate (Relevance failure)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Comparison of different adaptation strategies showing Soft Adaptation strikes the best balance.
Food.com	Out-PIE rate	25.2	32.4	+7.2
Food.com	Invalid rate	49.0	46.0	-3.0
Food.com	Out-PIE rate	19.3	32.4	+13.1
Food.com	Invalid rate	62.1	46.0	-16.1

Main Takeaways

Symbolic manipulation of the user profile (PKG) is more effective than natural language prompting for breaking filter bubbles.
Personalized tuning of the adaptation strength (adaptProportion) significantly outperforms using a global fixed value for all users.
Soft adaptation (inverting ratings) preserves relevance better than Hard adaptation (setting ratings to zero) or Removal (deleting edges), likely because it retains some structural information about the user's cuisine preferences while flipping the specific feature valence.

📚 Prerequisite Knowledge

Prerequisites

Knowledge Graphs (KGs) and triples
Recommender Systems concepts (filter bubbles, personalization)
Large Language Models (LLMs) and prompting

Key Terms

PKG: Personalized Knowledge Graph—a structured representation of a specific user's preferences, storing items they liked and their features (e.g., ingredients, tags)

PIE: Personalized Information Environment—an algorithmically reinforced filter bubble where recommendations are overly narrowed by past preferences

neuro-symbolic: Combining neural networks (like LLMs) with symbolic logic or structured knowledge (like KGs) to leverage strengths of both

feature co-occurrence: When two features (e.g., 'Italian' and 'Tomato') frequently appear together in items a user likes

KTO: Kahneman-Tversky Optimization—a method for aligning language models to human utility functions using binary (good/bad) feedback

triples: The fundamental unit of a Knowledge Graph, consisting of (Subject, Predicate, Object), e.g., (Pizza, hasIngredient, Tomato)

Out-PIE: A recommendation that matches the user's query and broad interest but avoids the specific over-represented feature bias

In-PIE: A recommendation that reinforces the existing over-represented bias (e.g., recommending another tomato dish to a tomato-heavy profile)