How does Personalized PageRank work for recommendations?

A random walk starts from the target user node. At each step, the walker follows a random outgoing edge with probability 0.85 or teleports back to the source with probability 0.15. After many iterations, nodes with the highest visit frequency are the top recommendations. Items reachable through many similar users naturally rank higher.

What are Node2Vec embeddings and why use them for recommendations?

Node2Vec generates biased random walk sequences from the interaction graph, then trains Word2Vec on those sequences to produce dense vector embeddings for every user and item. Cosine similarity between a user's embedding and item embeddings enables fast approximate nearest-neighbor lookup for candidate generation.

How do you handle cold-start users in a graph recommendation system?

New users have no interaction edges, so graph-based methods cannot personalize. The fallback is popularity-based recommendations: globally top-ranked items by recent interaction count, optionally bucketed by browse category. Once a user accumulates enough interactions (e.g., 5+), the system switches to personalized recommendations.

What is Maximal Marginal Relevance and why use it in recommendations?

MMR balances relevance and diversity when selecting the final recommendation list. At each step it picks the candidate that maximizes a weighted combination of relevance score and dissimilarity to already-selected items. A lambda parameter (~0.7) controls the relevance-diversity trade-off, preventing the list from being dominated by near-duplicate items.

Low Level Design: Graph-Based Recommendation Engine

Graph-Based Recommendation Engine

Model the recommendation problem as a graph: users and items are nodes, interactions are weighted edges. Graph algorithms then surface relevant items for each user.

Interaction Graph Structure

Bipartite graph: Users (U) <--> Items (I)

Edge weights by interaction type:
  view           = 1
  click          = 3
  purchase       = 5
  explicit_rating = raw rating value (1-5)

Higher-weight edges have stronger influence on traversal scores.

Personalized PageRank (PPR)

Start a random walk from the target user node. At each step:

With probability 0.85: follow a random outgoing edge (weighted by interaction weight)
With probability 0.15: teleport back to the source user node

Run N iterations (typically 50-100). Nodes with highest visit frequency = top-K recommendations. PPR naturally handles the graph structure — items reachable through many similar users rank higher.

ppr_scores = personalized_pagerank(
    graph=G,
    source=user_node,
    alpha=0.15,       # teleport probability
    iterations=100,
    top_k=200
)
recommendations = [node for node in ppr_scores if node.type == 'item'][:50]

Node2Vec Embeddings

Generate random walk sequences from the interaction graph, then train Word2Vec on those sequences to produce 128-dimensional embeddings for every user and item node.

walks = node2vec_walks(G, num_walks=10, walk_length=80, p=1, q=0.5)
model = Word2Vec(walks, vector_size=128, window=5, min_count=1)

user_vec  = model.wv['user_42']
item_vecs = [model.wv[f'item_{i}'] for i in all_items]

# Cosine similarity for ANN lookup
candidates = ann_index.query(user_vec, top_k=200)

Collaborative Filtering on Graph

Find users similar to target: those sharing >K item interactions (overlap in bipartite neighborhood)
Collect items those similar users interacted with but target has not
Rank by weighted interaction score from similar users, decay by similarity distance

Offline Pipeline

Daily Spark job:
  1. Load full interaction graph from data warehouse
  2. Run PPR for all active users (parallelized by user partition)
  3. Run Node2Vec training on full graph
  4. Compute top-200 recommendations per user
  5. Write results to Redis sorted sets

Online Serving

-- Pre-computed recs stored in Redis:
ZADD recs:user:42  5.0 item:901  4.8 item:234  4.5 item:567 ...

-- Serving layer reads:
ZREVRANGE recs:user:42 0 49  -- top 50 items

Latency target: <10ms per recommendation request. Redis sorted set lookup is O(log N).

Cold Start

New users have no interaction edges. Fallback strategy: serve popularity-based recommendations (global top items by recent interaction count, bucketed by category). Transition to personalized once the user accumulates >5 interactions.

Diversity: Maximal Marginal Relevance (MMR)

selected = []
candidates = top_200_by_score

while len(selected) < 50:
    best = argmax over candidates of:
        lambda * relevance(item) - (1-lambda) * max_similarity(item, selected)
    selected.append(best)
    candidates.remove(best)
# lambda ~ 0.7 balances relevance vs diversity

MMR prevents the recommendation list from being dominated by near-duplicate items (e.g., ten nearly identical products).