Domain Knowledge

Industry Practice

SNo	Domain	Company	Retrieval Methodology	Class
a	Commerce	Amazon	Item-to-item CF	Item-based
b	UGC	Facebook	Embedding-based retrieval; multiple user-centric sources	User-based
b	UGC	Instagram	Multi-source retrieval; Two-Tower: user and item embeddings	User-item
c	Short video	TikTok	Deep retrieval; user-video interaction	User-item
c	Short video	YouTube Shorts	Deep learning-based candidate generation; user-video	User-item
d	Long video	YouTube	Deep learning-based candidate generation; user-video	User-item
e	Published media	Netflix	Specialized ML personalization models	User-item
e	Published media	Prime Video	Matrix factorization; content-based	User-item
e	Published media	Spotify	Generative retrieval; content + user-item interaction	User-item
f	Inventory retail	Walmart	Hybrid retrieval; traditional + embedding	Item-based
f	Inventory retail	Target	Contextual multi-arm bandits; personalization focus	User-based
g	Travel	Airbnb	Embedding-based retrieval; item2vec-style	Item-based
h	Food delivery	Uber Eats	Two-Tower embeddings; graph learning; user-food	User-item
i	News	Google News	CF on click behavior; user-article	User-item
j	Search engine	Google	Traditional IR; query-document matching; evolving toward user-item based ranking	User-item

Design Consideration

• Item-based is more robust; scalable; and explainable under uncertainty.

• User-item is more powerful and fine-grained when you have dense; high-quality user interaction data.

SNo	Situation	Why Item-Based Wins	Practical Examples
1	Stable; low-churn catalogs	If item properties don’t change rapidly; item-item relationships stay valid for a long time.	Retail (Walmart; Target); long-tail marketplaces.
2	Cold-start users	New or infrequent users have no behavioral data. Item-item similarity avoids the need for personalization.	Food delivery (e.g.; recommend “popular in your area” restaurants).
3	Sparse user history	If users interact very little; you can’t learn good user embeddings. But item similarity can still be robust.	New users on Amazon: recommend similar products without needing deep profiles.
4	Highly diverse user base	When user behavior varies too much (by region; culture; device); item-based models generalize better.	Global apps with fragmented user bases (e.g.; Spotify in different countries).
5	Heavy multi-user devices	If multiple people use the same device/account; user profiles become noisy. Item-based CF is safer and more accurate.	Shared smart TVs; public kiosks; family shopping accounts.
6	Explainability	Item-based recommendations (“users who bought X also bought Y”) are easier to justify to users and auditors.	E-commerce and B2B sales platforms.

Search

Note

• [fennel.ai] Feature Engineering for Personalized Search

Search Advertising

Issue	Why It Matters	Strategic Fixes	Trade-Offs
Relevance vs. Revenue	Showing high-bid but low-relevance ads hurts trust	Hybrid ranking (bid + quality)	Too much relevance filtering lowers revenue
Click Fraud & Ad Spam	Inflated clicks drain budgets	ML-based fraud detection	False positives can hurt advertisers
Ad Auction Manipulation	AI-driven bid shading exploits system	Second-price auctions	Reduced ad revenue
Ad Fatigue & Banner Blindness	Users ignore repetitive ads	Adaptive ad rotation	Frequent ad refreshing increases costs
Query Intent Mismatch	Poor ad matching frustrates users	BERT-based intent detection	Over-restricting ads lowers monetization
Landing Page Experience	High bounce rate = low conversion	Quality Score rules	Strict rules limit advertiser flexibility
Multi-Touch Attribution	Last-click attribution undervalues early ad exposures	Shapley-based attribution	More complexity; slower optimization
Ad Bias & Fairness	Favoring large advertisers hurts competition	Fairness-aware bidding	Less revenue from high bidders

Relevance vs. Revenue Trade-Off

Why It Matters

• Advertisers bid for visibility, but their ads may not always be relevant to the user’s query.

• If high-bid but low-relevance ads are shown, users may lose trust in the search engine.

Strategic Solutions & Trade-Offs

• Quality Score (Google Ads’ Approach) Ranks ads based on a combination of CTR, relevance, and landing page experience, not just bid amount.

• Hybrid Ranking Model (Revenue + User Engagement) Balances ad revenue vs. user satisfaction.

Trade-Offs

• Prioritizing high-relevance, low-bid ads reduces short-term revenue.

• Prioritizing high-bid, low-relevance ads hurts user trust & long-term retention.

Click Spam & Ad Fraud

Why It Matters

• Bots & malicious actors inflate clicks to waste competitor ad budgets (click fraud).

• Some advertisers run low-quality, misleading ads to generate fake engagement.

Strategic Solutions & Trade-Offs

• Click Fraud Detection (Googles Invalid Click Detection) Uses IP tracking, anomaly detection, and ML models to filter fraudulent clicks.

• Post-Click Analysis (User Behavior Analysis) Detects bots based on engagement (bounce rate, session length, interactions).

Trade-Offs

• False Positives May block legitimate traffic, harming advertisers.

• False Negatives Fraudulent clicks still get monetized, increasing costs for real advertisers.

Ad Auction Manipulation & Bid Shading

Why It Matters

• Sophisticated advertisers use AI-driven bidding strategies to game real-time auctions.

• Bid shading techniques lower ad costs while maintaining high visibility.

Strategic Solutions & Trade-Offs

• Second-Price Auctions (Vickrey Auctions) Advertisers only pay the second-highest bid price, reducing manipulation.

• Multi-Objective Bidding Models Balances advertiser cost efficiency and search engine revenue.

Trade-Offs

• Too much bid control reduces revenue Search engines may earn less per click.

• Aggressive bid adjustments can reduce advertiser trust If advertisers feel theyre losing transparency, they may pull budgets.

Ad Fatigue & Banner Blindness

Why It Matters

• Users ignore repetitive ads after multiple exposures, reducing CTR over time.

• If ads look too much like organic results, users may feel deceived.

Strategic Solutions & Trade-Offs

• Adaptive Ad Rotation (Google Ads Optimize for Best Performing Mode) Dynamically swaps low-performing ads with higher-engagement creatives.

• Ad Labeling Transparency Clearer Sponsored tags improve user trust but reduce click rates.

Trade-Offs

• Refreshing ads too frequently raises advertiser costs.

• Too much ad transparency leads to lower revenue per impression.

Query Intent Mismatch

Why It Matters

• Search queries are often ambiguous, and poor ad matching leads to bad user experience.

• Example Searching for Apple Should the search engine show Apple iPhones (commercial intent) or apple fruit (informational intent)?

Strategic Solutions & Trade-Offs

• Intent Classification Models (BERT, T5-based Models) Classify queries into commercial vs. informational intent.

• Negative Keyword Targeting (Google Ads’ Negative Keywords) Advertisers block unrelated queries from triggering their ads.

Trade-Offs

• Restricting ads based on intent can lower revenue.

• Allowing broad ad targeting risks user dissatisfaction.

Landing Page Experience & Conversion Rate Optimization

Why It Matters

• Even if an ad gets high CTR, if the landing page is misleading or slow, users bounce without converting.

• Google penalizes low-quality landing pages via Quality Score reductions.

Strategic Solutions & Trade-Offs

• Landing Page Quality Audits (Googles Ad Quality Guidelines) Checks for page speed, relevance, mobile-friendliness.

• Post-Click Engagement Monitoring Uses bounce rate, time-on-site, conversion tracking to refine ranking.

Trade-Offs

• Strict landing page rules limit advertiser flexibility.

• Relaxed rules allow low-quality ads, reducing long-term trust.

Multi-Touch Attribution & Ad Budget Allocation

Why It Matters

• Users may see an ad but not convert immediately Traditional last-click attribution ignores earlier touchpoints.

• Advertisers struggle to allocate budgets across search, display, social, and video ads.

Strategic Solutions & Trade-Offs

• Multi-Touch Attribution Models (Shapley Value, Markov Chains) Assigns fair credit to different ad exposures.

• Cross-Channel Conversion Tracking Tracks user journeys across search & display ads.

Trade-Offs

• More complex attribution models require longer training times.

• Over-attributing upper-funnel ads can inflate costs without clear ROI.

Fairness & Ad Bias Issues

Why It Matters

• Some ad auctions are biased against small advertisers, favoring large ad budgets.

• Discriminatory ad targeting (e.g., gender/race bias in job/housing ads) can lead to regulatory penalties.

Strategic Solutions & Trade-Offs

• Fairness-Constrained Bidding (Googles Fairness-Aware Ad Auctions) Adjusts auction weights to prevent dominance by large advertisers.

• Bias Detection in Ad Targeting (Auditing Models for Discriminatory Targeting) Ensures fair exposure of diverse ads.

Trade-Offs

• Too much fairness correction may reduce revenue from high-bidding advertisers.

• Too little correction risks regulatory lawsuits (e.g., Facebooks 2019 lawsuit for discriminatory ad targeting).

Music

Challenge	Why Its Important	Trade-Offs
Personalization vs. Serendipity	Users want relevant music but also expect some new discoveries.	Too much personalization Feels repetitive. Too much exploration Feels random.
Repetition & Content Fatigue	Users get frustrated if the same songs appear too often.	Strict anti-repetition May exclude user favorites. Loose constraints Risk of overplaying certain songs.
Context & Mood Adaptation	Users listen to music differently based on mood; time; activity (workout; relaxation).	Explicit mood tagging is effective but requires manual input. Implicit context detection risks wrong assumptions.
Balancing Popular & Niche Tracks	Highly popular songs dominate engagement; making it hard for lesser-known songs to gain exposure.	Boosting niche tracks improves diversity; but may lower engagement metrics.
Cold-Start for New Songs & Artists	Newly released songs struggle to get exposure due to lack of engagement signals.	Over-boosting new music can lead to reduced user satisfaction.
Playlist Length & Engagement Optimization	Users may not finish long playlists; leading to low engagement metrics.	Shorter playlists increase completion rate; but longer ones improve session duration.

Playlist Generation & Curation in Music Recommendation Systems

Types of Playlists & Their Challenges

Playlist Type	Example	Key Challenges
Personalized Playlists	Spotifys Discover Weekly; YouTube Musics Your Mix	Ensuring balance between familiar & new tracks.
Mood/Activity-Based Playlists	Workout Mix; Chill Vibes; Focus Music	Detecting mood & intent dynamically.
Trending & Algorithmic Playlists	Spotifys Top 50; Apple Musics Charts	Avoiding popularity bias while staying relevant.
Collaborative & Social Playlists	Spotify Blend; Apple Musics Shared Playlists	Handling conflicting preferences in shared lists.
Genre/Artist-Centric Playlists	Best of 90s Rock; Jazz Classics	Ensuring diversity within a theme.

Solutions to Key Playlist Challenges

Challenge	Solution	Trade-Off
Over-Personalization (Echo Chamber)	Inject 5-20% exploration (Multi-Armed Bandits)	Too much exploration may decrease CTR
Repetition & Content Fatigue	Anti-repetition rules (e.g.; same song cannot appear in back-to-back sessions)	May prevent users from hearing favorite tracks
Cold-Start for New Songs	Boost underexposed songs using metadata (tempo; genre)	Over-promoting new songs may harm engagement
Context-Aware Playlists	Use real-time signals (e.g.; running mode detects movement; adjusts tempo)	Misinterpreted context may cause poor recommendations
Playlist Completion Rate	Optimize for average session length (shorter playlists for casual users; longer for engaged users)	Shorter playlists may reduce playtime per session

Common Problems

Cold-Start Problem for New Artists & Songs

• Why It Matters:

  • New artists and newly released tracks struggle to get exposure since they have no engagement history.

• Strategic Solutions & Trade-Offs:

  • Metadata-Based Recommendations (Genre, BPM, lyrics embeddings) Useful for early exposure but lacks engagement feedback.

  • Collaborative Boosting (Linking new artists to known artists) Improves visibility but risks inaccurate pairing.

  • User-Driven Exploration (Playlists like Fresh Finds) Promotes new songs but may not reach mainstream listeners.

• Example:

  • Spotifys Fresh Finds is a human-curated playlist designed for emerging artists.

Popularity Bias & Lack of Exposure for Niche Artists

• Why It Matters:

  • Big-label artists dominate recommendations, making it hard for new/independent musicians to gain visibility.

  • Overemphasis on top charts and algorithmic repetition reinforces the same mainstream music.

• Strategic Solutions & Trade-Offs:

  • Fairness-Aware Re-Ranking (Exposing lesser-known artists) Promotes diversity but may reduce engagement.

  • User Preference-Based Exploration (Blending familiar & new artists) Increases discovery but harder to balance.

  • Contextual Boosting (Surfacing niche content in certain playlists) Encourages exploration but risks user dissatisfaction.

• Spotifys Fix:

  • Discover Weekly and Release Radar to highlight emerging artists.

Balancing Exploration vs. Personalization in Playlists

• Why It Matters:

  • Users want to hear familiar songs but also expect discovery of new tracks.

  • Too much exploration reduces engagement, too little keeps users stuck in their existing preferences.

• Strategic Solutions & Trade-Offs:

  • Reinforcement Learning-Based Ranking (Balancing Novelty & Familiarity) Dynamically adjusts exploration but requires more data.

  • Hybrid Personalized Playlists (50% known, 50% new) Encourages discovery but still risks disengagement.

  • Diversity Re-Ranking Models (Ensuring mix of different artist popularity levels) Enhances engagement but increases complexity.

• Spotifys Fix:

  • Discover Weekly mixes familiar artists with newly recommended artists.

Repetition & Content Fatigue (Avoiding Overplayed Songs)

• Why It Matters:

  • Users dislike hearing the same songs too frequently in personalized playlists.

  • Music recommendation systems tend to reinforce top tracks due to high past engagement.

• Strategic Solutions & Trade-Offs:

  • Play-Session Awareness (Avoiding recently played tracks) Prevents fatigue but risks reducing personalization strength.

  • Diversified Playlist Generation (Embedding Clustering) Encourages discovery but may introduce unrelated tracks.

  • Temporal Diversity Constraints (Recommender-aware time gaps) Reduces overexposure but adds complexity to ranking models.

• Spotify & Apple Musics Fix:

  • Autogenerated playlists (e.g., Daily Mix, Radio) have anti-repetition constraints.

Context-Aware Recommendations (Music for Different Situations)

• Why It Matters:

  • Music preferences vary by context (workout, driving, studying, relaxing), but most recommenders treat all listening the same.

• Strategic Solutions & Trade-Offs:

  • User-Controlled Context Tags (Spotifys Mood Playlists, YouTube Musics Activity Mode) More control but adds friction.

  • Implicit Context Detection (Using location, time, device, previous context switches) Improves automation but risks privacy concerns.

  • Adaptive Playlist Generation (Real-time context-aware re-ranking) Better real-world usability but increases computational costs.

• Industry Example:

  • Spotifys Made for You mixes genres based on past listening sessions.

Short-Term vs. Long-Term Personalization

• Why It Matters:

  • Users music preferences change over time, but most recommendation models overly rely on recent activity.

  • Recommending only recently played songs can overfit short-term moods and ignore long-term preferences.

• Strategic Solutions & Trade-Offs:

  • Session-Based Personalization (Short-Term Context Models) Captures mood-based preferences but can overfit recent choices.

  • Hybrid Long-Term + Short-Term Embeddings (Contrastive Learning on Listening History) Balances nostalgia & discovery but computationally expensive.

  • Decay-Based Weighting on Past Behavior Helps phase out stale preferences but requires careful tuning.

• Spotifys Approach:

  • Balances On Repeat (long-term) and Discover Weekly (exploration).

Multi-Modal Recommendation (Lyrics, Podcasts, Audio Similarity)

• Why It Matters:

  • Music discovery can be driven by lyrics, themes, artist backstories, and spoken content (podcasts).

  • Traditional recommendation models focus only on collaborative filtering (listening history).

• Strategic Solutions & Trade-Offs:

  • Lyrics-Based Embeddings (Thematic music recommendations) Enhances meaning-based recommendations but requires NLP processing.

  • Cross-Domain Music-Podcast Recommendation (Shared interests) Improves discovery but harder to rank relevance.

  • Audio Similarity-Based Retrieval (Matching based on timbre, rhythm) Better for organic discovery but requires deep learning models.

• Industry Example:

  • YouTube Music cross-recommends music & podcasts based on topics.

Social Media

Video

E-Commerce