Google Professional Machine Learning Engineer Question 11
Single answerGoogle Cloud PlatformYou are working for a retail company that wants to forecast daily sales for its products using historical data. The dataset contains multiple time-series for different products, and there are missing values in the historical records. Which machine learning approach would be the most suitable for building an accurate predictive model in this scenario?
- A
Use a boosted trees model with time-based features such as day of the week, month, and lagged sales values.
- B
Apply matrix factorization to decompose the dataset into user and item latent features for recommendations.
- C
Train an autoencoder to impute missing values and use it directly for predicting future sales.
- D
Use a linear regression model with engineered time-series features to predict future sales.
Show answer and explanation
Correct answer: A
Explanation
The best approach for forecasting daily sales with time-series data is to use a boosted trees model with relevant time-based features. Boosted trees are robust to missing data, can capture non-linear relationships, and perform well on structured datasets. While other techniques like matrix factorization or autoencoders have their use cases, they are not ideal for this specific scenario.
- A. Correct.
Boosted trees models can handle missing values and are highly effective for structured data, such as time-series with engineered features like lagged sales values. This makes them a strong choice for accurate forecasting.
- B. Incorrect.
Matrix factorization is better suited for recommendation systems rather than time-series forecasting. It works by finding latent structures in user-item interactions, which does not directly apply to this scenario.
- C. Incorrect.
Autoencoders are effective for dimensionality reduction or anomaly detection, but they are not inherently designed for time-series forecasting. While they can help impute missing values, they are not a complete solution for this task.
- D. Incorrect.
Linear regression can be used for time-series forecasting with engineered features, but it may struggle with complex patterns or non-linear relationships in the data, which are better captured by boosted trees.