A popularity baseline ranks items by their overall interaction counts and recommends the same top items to every user. It is a simple, nonpersonalized reference: a more complex recommender should justify its added complexity by outperforming it, not merely by outperforming random selection. section 4.4.4 section 4.5.1
To build and test one offline:
- Split historical interactions by time. Count each item's interactions in the training data, then take the top k items. Do not use test-period interactions to determine popularity. section 4.2.1 section 4.4.4
- Give each test user that same ranked list. Compare it with the user's test-period relevance set using the same metric applied to the candidate model. section 4.2.1 section 4.4.4
Listing 4.7 Popularity baseline
def generate_recommendations_popularity(train_df, user_ids, k=10):
"""
Generate popularity-based recommendations.
Returns DataFrame with [user_id, item_id, rank, score]
"""
item_counts = train_df['item_id'].value_counts() #A
top_items = item_counts.head(k) #B
recommendations = []
for user_id in user_ids:
for rank, (item_id, count) in enumerate(top_items.items(), 1):
recommendations.append({ #C
'user_id': user_id,
'item_id': item_id,
'rank': rank,
'score': count
})
return pd.DataFrame(recommendations)
pop_recs = generate_recommendations_popularity(train, test_user_ids, k=10)
pop_metrics = calculate_precision_at_k(pop_recs, ground_truth, k=10)
The implementation counts training-set item IDs, takes the top k, and emits each item’s count as its score in an identical ranked list for every test user. section 4.4.4
Interpreting a comparison
An ALS recommender scores 0.1523 precision@10 against the popularity baseline’s 0.1234: a 23.4% improvement. That suggests ALS has learned something beyond general popularity, but popularity still achieves about 81% of its precision without the added complexity. Whether the gain is worthwhile depends on engineering resources and business goals. section 4.4.4
Treat the result as a starting point, not a deployment decision. Check ranking quality, catalog coverage, and performance across user segments: a good average precision can conceal narrow recommendations or weak results for particular users. Popularity can be especially hard to beat for cold-start users; established users set a higher bar for personalization. If a production recommender already exists, compare against that system too. section 4.10 section 4.5.2


























