- Three Elements of Recommender Systems: Co-Evolution of Users, Information, and Platform
- Role of the Strategy Product Manager
- Reference
Three Elements of Recommender Systems: Co-Evolution of Users, Information, and Platform
In recommender systems, product definition often plays a greater role than algorithms. The reason is that modern recommendation algorithms have matured; the key lies in matching the right scenario. For example:
I. Aligning with Information Scarcity to Richness Phases
- Information Scarcity Phase (Daily uploads < 100)
- Scenarios: Niche professional forums, small community sites.
- Strategies: Basic sorting methods, such as reverse chronological order or popularity sorting.
- Focus: Designing clean layout displays to ensure core content is successfully reached.
- Information Accumulation Phase (Daily uploads 1,000+)
- Scenarios: Vertical platforms (local services, professional knowledge hubs).
- Strategies:
- Build a topic system to achieve structured content presentation.
- Design content guide mechanisms to promote deeper browsing.
- Introduce basic recommendation algorithms, such as topic-based collaborative filtering.
- Information Abundance Phase (Daily uploads 10,000+)
- Scenarios: Large portal sites, e-commerce giants.
- Strategies: Fully implement personalized recommendations, building user-item matching models.
- Focus: Balancing precision and diversity to avoid information cocoons.
II. Aligning with Platform Growth Phases
- Startup Phase
- Characteristics: Limited resources (hardware, algorithms, information volume), lack of user logs.
- Strategies:
- Employ lightweight algorithms (such as TF-IDF, popularity sorting).
- Design structured knowledge systems.
- Prioritize touchpoint efficiency for core content.
- Principle: Algorithms over pure manual operation; simple and effective over complex models.
- Growth Phase
- Focus: Building the recommendation ecosystem.
- Strategies:
- Mature UGC production mechanisms.
- Establish user growth frameworks.
- Optimize content sourcing strategies.
- Goal: Lay the data foundation for personalized recommendations.
- Mature Phase
- Focus: Balancing business value and user experience.
- Core Metrics: User retention, dwell time, conversion rate.
- Strategies:
- Build multi-objective optimization systems.
- Design commercialized recommendation slots.
- Optimize long-term user experience.
III. Aligning with User Cognitive Phases
- User Cold Start
- Core Goal: Build user trust, promote retention.
- Strategies:
- Showcase top-tier high-quality content.
- Leverage basic signals like device type and geographical location.
- Design guided interest selectors.
- Avoid premature monetization.
- Principle: Experience first to cultivate user stickiness.
- Profile Cultivation Phase
- Core Goal: Complete the user profile.
- Strategies:
- Employ Exploration-Exploitation (EE) mechanisms.
- Maintain list diversity.
- Keep commercialization under strict bounds.
- Focus: Balancing recommendation precision with exploration space.
- Profile Stable Phase
- Core Goal: Optimize long-term experience.
- Strategies:
- Design mechanisms to break out of filter bubbles.
- Establish commercialized ad-recommendation blending systems.
- Optimize social recommendation features.
- Principle: Balance commercial value and user experience to achieve sustainable growth.
Role of the Strategy Product Manager
Strategy PMs are responsible for platform value and user experience. Common responsibilities include:
- Defining Use Scenarios
- Such as “You might also like,” “Cross-selling,” “Frequently bought together,” etc. Each requires tailored strategies.
- Clarifying Business Goals
- Shopping cart scenario: fast checkout + auxiliary items.
- Homepage: based on full user profiles.
- Cross-selling: balancing discount thresholds and easy-to-add items.
- Long-term value: retention rate, user path optimization.
- Determining UI Designs
- Requires a combined understanding of algorithms, UI, user behavior, and business models.
- Constructing Model Data
- Handling data differences:
- Feature construction: Can historical price ranges be used directly as a feature? Not quite—gender-based differences across categories are massive. However, brand preferences and price tiers can be migrated.
- Scenario-specific strategies: For viewed items, judge by purchase intent. For purchased items, account for repurchase cycles. For video, separate short vs. long video.
- Leveraging prior knowledge: Prioritize domain rules over raw algorithms for deterministic problems (e.g., using parenting knowledge for maternal and infant product recommendations).
- Handling data differences:
- Algorithmic Experiments
- Defining objective metrics (CTR, CVR, etc.).
- Understanding the role of evaluation metrics (e.g., applying F1 score in different scenarios).
- Structuring input sources and data schemas.
- Leading model iterations.
- Defining objective metrics (CTR, CVR, etc.).
- Diagnosing Problems & Continuous Iteration
- Case studies, online experiments.
- Recommendation System Evaluation:
- Precision: AUC, UAUC, etc.
- Coverage/Diversity: Herfindahl-Hirschman Index (HHI), Gini index.
- Multi-objective balance: balancing content consumption, author ecosystem, and user retention.
- Metric validity: Human evaluation systems.
- System Iteration Ideas:
- Data introduction: Enriching user profiles and tag systems.
- Retrieval optimization: Multi-channel retrieval, efficiency audits.
- Ranking optimization: Multi-objective balancing.
- Rule application: Subjective overrides, real-time rules.
Common Recommendation Strategies
User Experience Strategy
User experience is the critical factor for recommender systems; sound user experience strategies boost user satisfaction and platform stickiness. Strategies focus on diversity, quality, and recency of content by addressing duplicates, filtering low-quality items, and handling spatial/temporal constraints.
- Duplicate Optimization
- Complete duplicates: Identifying plagiarism, copies.
- Detail page duplicates: De-duplication rules.
- List page/Title/Similarity duplicates: Scattering/diversity rules.
- Content Quality
- Low-quality content identification: User negative feedback + manual annotation.
- Down-ranking strategies.
- Sourcing and boosting top-tier content.
- Spatial and Temporal Constraints
- Recency: Processing short, medium, and long-term fresh content.
- Locality: Localization and regional distribution strategies.
Cold Start Strategy
The cold start problem is a classic challenge. It refers to the situation where the system lacks sufficient user history or item interaction data, making it difficult to generate accurate recommendations.
- User Cold Start: Leveraging device information, registration data, and basic context to quickly establish an initial profile.
- Explicit interest gathering: Guiding users to select initial interest categories.
- Implicit feature inference: Inferring user cohort characteristics from IP address, device type, etc.
- Lookalike mapping: Matching new users to similar existing cohorts.
- Popular content recommendation: Showcasing the platform’s high-quality trending items.
- Interest exploration: Designing diverse recommendations to probe interest boundaries.
- Immediate feedback loop: Giving higher weights to the user’s initial clicks to speed up profile building.
- Item Cold Start: Analyzing inherent attributes like title, cover, and description.
- Text extraction: Using NLP to parse topics and keywords.
- Multimodal analysis: Incorporating image and video features.
- Similarity mapping: Linking new items with similar trending items.
- Long-tail coverage: Setting up exposure guarantee mechanisms for new items.
- Traffic test: Routing new items to a small sliver of traffic to measure performance.
- Creator trust transfer: Leveraging the creator’s historical record to estimate initial quality.
Ecosystem Strategy
Ecosystem strategy is the core support framework, focusing on the positive loop between creators, content, and users.
A healthy ecosystem balances all parties’ interests, fostering quality content creation, improving user experience, and driving sustainable growth. Effective ecosystem strategies create a positive loop: “high-quality creators produce premium content, premium content attracts more users, and users provide feedback and incentives to creators.”
- Creator / Publishing Ecosystem
- Tools and guidelines: Sponsoring content creation tools, publishing guides, and best practices.
- Creator quality index: Rating creators by originality, health, and professionalism.
- Platform tone guidelines: Steering creation directions via recommendation biases.
- Creator incentives: Revenue sharing, creator funds, and honorary tier systems.
- Governance: Fostering new creators, anti-spam, and creator training.
- Content Ecosystem
- Quality assurance: Audits, user reports, and automated low-quality filters.
- Multi-format support: Diversifying media (text, images, video, audio) and genres (knowledge, entertainment, news).
- Long-tail support: Providing exposure opportunities for high-quality niche content.
- Distribution balance: Controlling ratios of trending vs. fresh content.
- Safety & Copyright: Filtering sensitive content, copyright protection.
- User Ecosystem
- User segmentation: Tailored recommendations based on user engagement levels.
- Low-engagement users:
- Definition: New, low-activity, or returned users.
- Active churn-risk users: Users whose activity levels are dropping.
- Strategy: Content recommendation adjustments and ad exposure protection.
- Personalization strategies:
- Interest exploration.
- Diversity enhancement.
- Long-term interest tracking.
- Historical user cohort shifts analysis.
- Off-platform users:
- Surveys and competitive analysis.
- Sharing link optimizations.
- Trending event tracking.
- Push optimization:
- Multi-touch attribution: Accurately attributing app opens to push events.
- Activation path optimization: Tuning clicks and app startup pipelines.
- Low-engagement users:
- Community atmosphere: Fostering healthy interactions, comment moderation.
- User feedback: Weight designs for likes, favorites, and shares.
- Understanding how trending content propagates and identifying key nodes.
- User retention: Personalized push, interest cultivation, and social loops.
- User journey: Guiding new users to become power users.
- Diagnosing dropping activity.
- DAU, day-1 retention, and stay time attribution.
- Tracking historical shifts in user cohorts.
- User segmentation: Tailored recommendations based on user engagement levels.
Layout and Interaction Design
Interaction design directly affects user experience and recommendation effectiveness. Smart designs reduce choice fatigue, improve data gathering, and naturally guide exploration:
- Reducing Choice Fatigue
- Example: Netflix’s autoplay previews, enabling users to preview content without clicking.
- Example: Spotify’s “Daily Mix” playlists, offering personalized music with a single click.
- Improving Data Sourcing Quality
- Example: YouTube’s “Not interested” and “Don’t recommend channel” buttons, capturing precise negative feedback.
- Example: Amazon’s separation of star ratings and written reviews, lowering feedback friction.
- Guiding User Exploration
- Example: TikTok’s “Discover” page, highlighting new content categories using visual layouts.
- Example: Pinterest’s relevant content grids, naturally steering users into related interests.
- Enhancing Recommendation Transparency
- Example: LinkedIn’s “Because you followed X” recommendation explanations.
- Example: Facebook’s “Why am I seeing this ad?” disclosures.
- Optimizing Instant Feedback Loops
- Example: Instagram’s double-tap to like, simplifying feedback.
- Example: Medium’s progressive reading bar, implicitly gathering content consumption speed.
These designs improve user experience and feed high-quality training logs back to recommendation models, forming a positive optimization loop.