What I Learned From My AI Engineer Interview: Recommendation Systems, TF‑IDF, Word2Vec & SVM Explained

I am a Java engineer who, after years away from advanced mathematics, decided to self‑study AI and focus on the most popular area of algorithm engineering: recommendation systems.

After reading "Machine Learning" and "Statistical Learning Methods" and participating in Tianchi beginner contests, I landed several interview invitations.

Interviewer: What are the common approaches for recommendation systems? Me: Content‑based recommendation, collaborative filtering, matrix factorization, and latent semantic models.

Interviewer: Explain how content‑based recommendation works. Me: Represent items as text vectors (e.g., keyword extraction, bag‑of‑words, TF‑IDF) and compute similarity distances.

Interviewer: What types of collaborative filtering exist and what is their basic principle? Candidate: User‑based and item‑based collaborative filtering; both are nearest‑neighbor methods that recommend items based on similar users or similar items.

Interviewer: Which method do you use most and why? Me: Item‑based collaborative filtering, because item similarity is more stable and provides interpretable explanations.

Interviewer: How would you handle the cold‑start problem when new data arrives? Me: Use content‑based recommendations to mitigate cold‑start, or collect additional user/item information.

Interviewer: What similarity measures can be used after converting text to TF‑IDF vectors? Candidate: Cosine distance, Pearson similarity, Jaccard similarity, etc.

Interviewer: Do you understand the word2vec model structure? Explain hierarchical softmax and negative sampling. Candidate: Word2vec is a shallow classifier (e.g., CBOW predicts a target word from surrounding context). Hierarchical softmax builds a Huffman tree to turn the large‑vocabulary softmax into a series of binary decisions; negative sampling randomly selects a small set of negative words to approximate the full softmax.

Interviewer: Write the principle formula of an SVM. Candidate: (Explanation of SVM margin maximization and decision function omitted for brevity.)

The interview highlighted my gaps in mathematical foundations, prompting me to study machine‑learning mathematics and seek structured courses to bridge the knowledge gap.