BSCS 4-3: Elective 4 (Machine Learning) Final Quiz- 6-8-2024

A company has a dataset of customer transactions where only 10% of the transactions are labeled as fraudulent or non-fraudulent. They decide to use a semi-supervised learning approach. Which of the following strategies is most likely to improve the fraud detection model’s performance?

• Discarding the unlabeled data and only using the labeled data to train a supervised learning model

• Using the labeled data to train a supervised learning model and then using that model to label the unlabeled data

• Using the labeled data to initialize the model, then iteratively training the model on both labeled and unlabeled data using techniques like self-training

• Clustering the unlabeled data first and then using the clusters to assign labels

Which of the following is a primary challenge in implementing semi-supervised learning in a real-world scenario?

• Lack of computational power to process large datasets

• Difficulty in defining the model architecture

• Ensuring the unlabeled data is relevant and representative of the problem space

• Lack of effective algorithms for handling labeled data

A biologist has a dataset of images of different species of plants, but only a few images are labeled with their species names. They decide to use semi-supervised learning. Which of the following approaches can be most beneficial?

• Use a semi-supervised learning method that incorporates both labeled and unlabeled images to improve the classification accuracy

• Use an unsupervised learning method to cluster the images and assign species labels based on the clusters

• Train a supervised learning model only on the labeled data and ignore the unlabeled data

• Manually label the unlabeled data to increase the size of the labeled dataset

You are working on a sentiment analysis model for social media comments. Labeling positive and negative comments is easy, but labeling neutral comments is subjective and time-consuming. Which of the following semi-supervised learning approaches might be most effective?

• Training a model on labeled positive and negative comments, then using it to label neutral comments. (Inductive learning)

• Clustering the comments based on word similarity and assigning sentiment labels based on the labeled positive and negative clusters. (Transductive learning with clustering)

• Using self-training, where the model initially learns from labeled data, then iteratively labels the most confident unlabeled comments and adds them to the training set. (Inductive learning with self-training)

• All of the above could be effective depending on the specific data and task

Imagine you have a large dataset of customer images with only a few labeled as "high risk" for credit card fraud. Which of the following statements about using semi-supervised learning for fraud detection is most accurate?

• The unlabeled data will automatically improve fraud detection accuracy without any additional steps

• Semi-supervised learning can be used to identify potential fraudulent transactions based on similarity to the labeled high-risk cases

• The model's performance on unseen fraudulent transactions will be guaranteed to be better than a model trained only on labeled data

• Semi-supervised learning introduces noise into the training data, making it less effective than supervised learning for fraud detection

A company wants to classify product reviews by sentiment (positive, negative, neutral). Labeling all reviews is expensive. They have a small set of labeled reviews and a large set of unlabeled reviews. Which of the following is the biggest challenge they might face when using a semi-supervised learning approach?

• The model might overfit to the labeled data, neglecting the information in the unlabeled data

• The cost of labeling the small set of initial data can be prohibitive

• Semi-supervised learning algorithms are computationally expensive to train

• It is impossible to determine the sentiment of a neutral review without a human label

A tech startup company is developing an AI-driven customer support system to automatically categorize and prioritize support tickets. They have a large volume of historical support tickets but only a small subset has been manually labeled by support agents. The project deadline is tight, and the team needs to deliver a working model within a few weeks. Despite having access to domain experts and a strategy to collect more labeled data, the team faces significant pressure to meet the timeline. What is the primary challenge in this scenario?

• Insufficient quantity of labeled data

• Insufficient domain expertise to label data

• Insufficient time to label and prepare data

• None of the mentioned

When compared to supervised learning, a major advantage of semi-supervised learning is:

• It always guarantees a more accurate model

• It requires less computational resources to train the model

• It is easier to interpret the model's decision-making process

• It can leverage a large amount of unlabeled data to improve model performance

In transductive learning, which of the following statements is true?

• The model generalizes to unseen data by learning from the labeled training set

• The model's goal is to infer labels for the specific unlabeled training data available, without necessarily generalizing to unseen data

• The model uses reinforcement learning to improve its predictions

• The model relies solely on unsupervised learning techniques

Which of the following is a common assumption made in semi-supervised learning that helps in leveraging unlabeled data effectively?

• The underlying structure of the data is consistent between labeled and unlabeled data

• Labeled data is more important than unlabeled data

• Unlabeled data is uniformly distributed

• Unlabeled data can be ignored if it does not match labeled data