1Z0-184-25 Question 20
Single answerYou are developing a recommendation system on Oracle Cloud Infrastructure to suggest relevant articles to users based on their reading history. You store user interest embeddings in an Autonomous Database instance that supports AI vector search. The goal is to identify the most semantically similar items for each user by focusing on the direction of vectors rather than their magnitude. Which vector distance metric is most suitable in this scenario?
- A
Use Euclidean (L2) distance to measure the magnitude of differences in embeddings
- B
Use Cosine distance to measure the angular similarity between embeddings
- C
Use Manhattan (L1) distance to measure sum of absolute differences between embeddings
- D
Use Hamming distance to measure the number of positions at which the embeddings differ
Show answer and explanation
Correct answer: B
Explanation
When dealing with embeddings that capture semantic or contextual representation of text (or user interests), Cosine distance is frequently recommended. It focuses on the angle between vectors rather than their lengths, making it a strong choice for tasks emphasizing semantic proximity. In Oracle Cloud Infrastructure environments leveraging AI vector search, selecting the right distance metric can greatly impact the accuracy and performance of similarity-based queries. Refer to Oracle Documentation on Autonomous Database Vector Search for best practices and metric recommendations.
- A. Incorrect.
Option 1 (Euclidean distance): While commonly used for many vector-based searches, Euclidean distance emphasizes magnitude and can be less effective for purely semantic comparisons where the angle between vectors matters more than absolute magnitude.
- B. Correct.
Option 2 (Cosine distance): CORRECT. Cosine distance (or Cosine similarity) is well-suited for text or user interest embeddings because it focuses on the orientation (angle) between vectors, effectively capturing semantic similarity where scale can be less important.
- C. Incorrect.
Option 3 (Manhattan distance): Though simpler, Manhattan distance is typically used for specific numerical metrics rather than high-dimensional semantic embeddings, and it does not capture the angular relationship as effectively as Cosine.
- D. Incorrect.
Option 4 (Hamming distance): Hamming distance is specific to binary or categorical data. It is not appropriate for floating-point embeddings representing semantic concepts.