Objective 2.4 Apply Problem-Specific Transformations to Data Sets CERTIFIED-DATA-SCIENCE-PRACTITIONER Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   What type of data does GloVe primarily use to generate word embeddings?

   • AA list of word frequencies in each document
   • BA set of labeled documents for supervised classification
   • CA word-word co-occurrence matrix computed from the entire corpus
   • DA sequence of individual sentences used for next-word prediction
   Reveal Solution Discussion Select an answer first
2. 17foundation · easy

   Which of the following best describes word tokenization in natural language processing?

   • ASplitting text into individual tokens such as words or subwords
   • BConverting words into numerical vectors
   • CAssigning a part-of-speech tag to each word
   • DRemoving stop words and punctuation from text
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A startup is building a recommendation system for a niche e-commerce site. They have a large corpus of product descriptions and user reviews. They want to find products that are semantically similar even if they share no exact words (e.g., 'waterproof jacket' and 'rain coat'). Which transformation should they apply to the product descriptions?

   • AApply a bag-of-words model with n-grams (bigrams and trigrams).
   • BApply TF-IDF vectorization to each product description.
   • CUse one-hot encoding of all words and compute cosine similarity between product vectors.
   • DTrain a Word2Vec model on the corpus and average the word vectors for each product description.
   Reveal Solution Discussion Select an answer first
5. 19expert · hard

   A medical imaging team is building a model to detect rare diseases from CT scans. They have a very small labeled dataset (500 scans) and a large unlabeled dataset of CT scans. They want to use an autoencoder to learn latent representations of the scans, then use these representations as features for a classifier. They also want to ensure the representations are compact and capture the most salient features. Which approach should they use?

   • ATrain a Word2Vec model on the image metadata.
   • BTrain a convolutional autoencoder on the unlabeled CT scans, then use the encoder part to generate latent vectors for the labeled scans.
   • CUse a pre-trained CNN (e.g., ResNet) trained on ImageNet to extract features directly from the CT scans.
   • DApply PCA to the raw pixel values of the CT scans.
   Reveal Solution Discussion Select an answer first
6. 20application · medium

   A data science team is building a spam classifier for SMS messages. They have a small dataset of 10,000 messages and need to represent each message as a feature vector. They want to capture the presence of specific keywords (e.g., 'free', 'win') without any semantic relationships. Which vectorization method is most appropriate?

   • AWord2Vec embeddings averaged over the message
   • BTF-IDF vectorization
   • COne-hot encoding of each word in the message
   • DBag-of-words (count vectorizer)
   Reveal Solution Discussion Select an answer first