🤖 AI

Ocean4Rec: Offline LLM-Derived OCEAN Profiles for Request-Time VOD Reranking

Ocean4Rec is an offline LLM-driven reranking framework that pre-computes OCEAN personality profiles for VOD items and users to enable efficient, request-time LLM-free re-ranking, significantly improving NDCG@20 and HR@20 metrics on Samsung Smart TV logs compared to strong industrial baselines.

Original authors: Wonkyun Kim, Sehyun Bae, Kwanki Ahn, Mungyu Bae, Saeun Choi, Soyeon You, Chandra Prabhakar, Sehyun Kim

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Wonkyun Kim, Sehyun Bae, Kwanki Ahn, Mungyu Bae, Saeun Choi, Soyeon You, Chandra Prabhakar, Sehyun Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you run a massive video streaming service, like a digital Netflix for millions of people. You have a team of "recommenders" (algorithms) that suggest movies and shows. Usually, these systems work in two steps: first, they grab a big list of possible movies (the "candidate list"), and second, they sort that list to show the best ones at the top.

The problem is that modern AI (Large Language Models, or LLMs) is really good at understanding why a movie is interesting, but it's also slow and expensive to ask it for advice every single time a user clicks "refresh." It's like hiring a famous food critic to taste every single dish in a buffet before serving it to a customer; the food would get cold, and the bill would be astronomical.

Ocean4Rec is a clever solution from Samsung that gets the best of both worlds: the deep understanding of a smart AI, but the speed of a standard computer. Here is how it works, broken down simply:

1. The "Offline Critic" (The Preparation Phase)

Instead of asking the AI to judge movies in real-time, the team uses the AI offline (when no one is watching).

The Analogy: Imagine a team of expert librarians sitting in a quiet room, reading the summaries and descriptions of thousands of movies. They don't just write a review; they assign each movie a "personality score" based on the OCEAN model (a famous way to describe human personalities):
- Openness (Is it creative and weird?)
- Conscientiousness (Is it structured and disciplined?)
- Extraversion (Is it loud and social?)
- Agreeableness (Is it warm and friendly?)
- Neuroticism (Is it emotional and intense?)
The Result: Every movie in the catalog gets a simple 5-number "personality card" stored in the database. This happens once, not every time you click a button.

2. The "User Taste Profile" (The Calculation Phase)

When you watch a show, the system doesn't ask the AI "What does this user like?" Instead, it looks at the "personality cards" of the shows you've recently clicked.

The Analogy: If you watched three "Open" and "Neurotic" movies recently, the system calculates that your current taste is a mix of "Open" and "Neurotic." It creates a simple 5-number "User Taste Card" for you. It forgets older tastes (like a fading memory) to focus on what you like now.

3. The "Fast Sort" (The Request Phase)

Now, when you open the app and ask for recommendations:

The system grabs a list of 1,000 movies from its standard, fast algorithm.
It looks up your "User Taste Card" and the "Movie Personality Cards" for those 1,000 movies.
It does a quick math check: "Does this movie's personality match the user's current taste?"
It also adds a bonus for Recency (newer movies get a slight boost).
Crucially: It does not call the AI. It just does simple math on the pre-made cards.

Why is this a big deal?

Speed & Stability: Because the AI isn't working during the actual request, the system is incredibly fast and won't crash if the AI server gets busy. It's like having a pre-cooked meal ready to go, rather than cooking from scratch every time a guest arrives.
Better Recommendations: In their tests (using Samsung Smart TV data), this method improved the quality of recommendations significantly compared to just sorting by popularity or recency.
- For one type of algorithm (NCF), it improved the "top 20" recommendations by about 7.6%.
- For another type (LightGCN), it improved them by a huge 61.5%.
Safety: The system is designed to be "fail-safe." If a movie doesn't have a personality card (maybe the description was missing), the system just ignores the AI part and shows the movie based on its standard score. It never breaks the user experience.

The Bottom Line

Ocean4Rec is like hiring a smart consultant to do the heavy thinking before the store opens, so that when customers walk in, the staff can instantly point them to the perfect product without needing to call the consultant back every time. It keeps the recommendations personal and smart, but the system remains fast, cheap, and reliable.

Technical Summary: Ocean4Rec

Problem Statement

Industrial Video-on-Demand (VOD) recommenders require richer content understanding to improve user satisfaction. While Large Language Models (LLMs) offer advanced reasoning capabilities for feature extraction and candidate scoring, integrating them directly into the request-time serving path introduces significant operational challenges. Specifically, per-request LLM inference involves prompt construction, token generation, model invocation, output parsing, and fallback handling. In high-volume, latency-sensitive environments, these operations complicate throughput planning, tail-latency control, capacity isolation, and predictable system operation. The core problem is how to leverage LLM-derived content insights without incurring the latency and stability risks of online LLM calls.

Methodology: Ocean4Rec

Ocean4Rec proposes a reranking layer that decouples LLM reasoning from the online serving path. The system operates on a "materialize offline, score online" principle, utilizing the OCEAN (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) personality taxonomy as a schema-constrained content representation.

1. Offline Item Profiling

Process: An LLM (Google Gemini 2.5 Pro) is used as an offline annotator to process VOD metadata (title, plot, description, genre).
Output: Each eligible item is mapped to a five-dimensional integer vector representing OCEAN scores, accompanied by a confidence score and a short reasoning string.
Validation: The pipeline validates generated records for format and range. Items with missing metadata or invalid LLM outputs receive low-confidence neutral fallback profiles, ensuring they remain eligible for the base recommender.

2. User Taste Profiling

Construction: User profiles are not static; they are dynamic vectors derived from recent implicit interactions (clicks, deep-links) with profiled items.
Aggregation: A user's OCEAN vector ( $z_u$ ) is a time-decayed weighted average of the OCEAN vectors of consumed items ( $q_i$ ).
Decay: Interactions are weighted using an exponential half-life decay function ( $w_{time}(a) = 0.5^{a/h}$ ), where $a$ is the interaction age in days and $h$ is the half-life (90 days in the reported implementation). This captures the temporal dynamics of user preferences.

3. Request-Time Reranking

At serving time, the system retrieves a scored candidate list (Top 1000) from a base generator (e.g., NCF or LightGCN). No LLM is invoked. The reranking score $S(u, i)$ is a linear combination of three components:
$S(u, i) = \alpha B(u, i) + \beta P(u, i) + \gamma R(i)$

$B(u, i)$ : The base recommender score (min-max normalized).
$P(u, i)$ : The OCEAN compatibility term, calculated as the Pearson correlation between the user vector $z_u$ and item vector $q_i$ , clipped and transformed to $[0, 1]$ .
$R(i)$ : A catalog recency term based on the content release age.
Fallback: If a user or item lacks a valid OCEAN profile, the OCEAN weight ( $\beta$ ) is reassigned to the base score ( $\alpha$ ), ensuring the system degrades gracefully to the base ordering.

Key Contributions

The paper makes three specific contributions to the RecSys Industry Track:

Offline LLM Materialization Pattern: It describes a pattern for integrating LLM-derived OCEAN content profiles into an industrial reranker without requiring request-time LLM calls, thereby preserving the predictability of the serving path.
Same-Space Representation: It defines a unified representation where items are profiled from metadata and users are represented by the time-decayed aggregation of consumed item profiles, enabling direct compatibility scoring without learned alignment.
Temporal-Holdout Evaluation: It reports same-candidate, temporal-holdout offline evaluations on anonymized Samsung Smart TV VOD logs using NCF and LightGCN generators, scoping all empirical claims to these specific settings.

Results

Evaluations were conducted on anonymized logs from a Samsung Smart TV VOD environment, using a temporal holdout (training data up to March 31, 2026; evaluation on April 1–27, 2026).

Performance vs. Baselines:
- NCF Generator: Ocean4Rec improved NDCG@20 by 7.6% over the stronger "Base+Recency" baseline. HR@20 improvements were inconclusive.
- LightGCN Generator: Ocean4Rec improved NDCG@20 by 61.5% over "Base+Recency" and HR@20 by 67.3%.
Component Analysis: The "Base+OCEAN" component alone showed mixed results (improving HR but sometimes lowering NDCG/MRR compared to Base), indicating that OCEAN is most effective when paired with both base relevance and recency signals.
Secondary Diagnostics: Comparisons with a metadata-only implementation of Persona4Rec on public datasets (Amazon-Books, Yelp) showed competitive results, though the paper notes these are sanity checks rather than direct head-to-head baselines due to architectural differences (e.g., lack of base-score fusion in the diagnostic).

Significance and Claims

The paper positions Ocean4Rec not as a universal solution, but as a bounded auxiliary content-taste feature that preserves the deployability advantages of a request-time-LLM-free serving path.

Operational Stability: The primary significance is architectural: it removes the variability of prompt construction, token generation, and model invocation from the critical serving loop, ensuring predictable latency and throughput.
Interpretability: Unlike dense embeddings, the five-dimensional OCEAN space provides operators with an inspectable feature set to diagnose why items are promoted or demoted.
Privacy and Governance: The system strictly separates user behavior data from LLM inputs (the LLM sees only item metadata). User OCEAN vectors are treated as internal recommendation features, not psychological labels, with specific governance controls recommended for profile retention and exposure monitoring.
Modest Scope: The authors explicitly state that the results are offline point estimates. They do not claim online A/B test impact, measured latency savings, or business outcome improvements (e.g., retention), noting that total system cost depends on factors like catalog size and batch cadence which were not measured. The work serves as evidence for the viability of offline LLM materialization in industrial settings.