Questions & explanations
1. In a Bayesian network with nodes A, B, C where A causes B and B causes C, are A and C independent if B is known?
Yes, they are independent given B. Because the path from A to C goes through B. B is in the middle and is known. In a chain A->B->C, knowing B blocks the path. So A and C become conditionally independent given B. For example, if A is 'raining', B is 'sprinkler is off', and C is 'grass wet', then if you know the sprinkler is off, the rain directly affects wetness, and knowing rain gives no extra info about wetness beyond the sprinkler status? Actually careful: In this chain, if B is known, A and C are independent. So if you know sprinkler state, rain and wetness are independent (since rain only affects wetness via sprinkler? That's not realistic but in the network structure it holds.) More typical: A=age, B=income, C=car price. If you know income, age and car price are independent.
2. How do companies collect internal data for BI?
Companies collect internal data through their everyday operations. Point-of-sale systems record each sale. Customer relationship management (CRM) software stores interactions with customers. Enterprise resource planning (ERP) systems track inventory, orders, and finances. Employees enter data into these systems, or sensors automatically log machine outputs. For example, when a cashier scans a product, the system saves the time, price, and product code. Later, BI extracts this data from the company's databases. Collecting the data accurately and consistently is important. Often, companies use software to integrate all these sources into one place.
3. Compare structured and unstructured data in BI data gathering.
Structured data is organized in rows and columns, like a sales table with columns for date, product, and price. It is easy to store in databases and analyze with simple queries. Unstructured data has no fixed format — emails, videos, social media posts, and images. It is harder to process because it needs special tools like text analysis or image recognition. For BI, structured data is used for standard reporting, while unstructured data adds deeper context. For example, structured data shows a drop in sales; analysis of unstructured customer reviews might reveal why. Both types are valuable but require different methods to collect and analyze.
4. Why do companies use a separate data warehouse instead of a regular database?
Regular databases are made for lots of small, fast transactions like processing orders. They work well but are not built for big analytical queries that scan millions of rows. Data warehouses are optimized for reading large amounts of data and doing complex calculations. They also combine data from multiple departments into one place, which a regular database often does not do. For example, a warehouse can merge sales and inventory data from different systems. This makes it faster and easier to run BI reports. Also, using a separate warehouse keeps the daily transaction systems running smoothly because heavy queries do not slow them down.
5. What role does correlation between assets play in portfolio optimization?
Correlation measures how two assets move together. If they are positively correlated, they tend to go up or down at the same time. If they are negatively correlated, one goes up when the other goes down. Portfolio optimization uses correlation to reduce risk. For example, if you own stocks and bonds, they often have low or negative correlation. When stocks fall, bonds may rise, balancing the portfolio. This lowers overall risk without reducing expected return much. The optimization model finds the combination of assets with low correlations to each other. The goal is a portfolio where losses in one asset are offset by gains in another.
6. What is preposterior analysis in decision making?
Preposterior analysis means thinking about how useful a new piece of information could be before you actually get it. You imagine what you would decide after seeing each possible outcome from the information. Then you compare the best choice with and without that information. This helps you decide if it is worth paying for the information. For example, a company might ask: 'Should we spend money on a market study before launching a new product?' They use preposterior analysis to see if the study would change their decision enough to justify its cost. The value of clairvoyance is the maximum you would pay for perfect information.
7. Compare the optimization of energy distribution vs consumption.
Optimizing energy distribution focuses on moving power efficiently from plants to users. This includes managing transformers and lines to reduce losses. For example, a utility might reconfigure the grid to balance load. Optimizing consumption focuses on how customers use energy. This includes demand response and efficiency programs. Both aim to reduce waste and cost. Distribution optimization is about the network, while consumption optimization is about end-user behavior. They work together: better distribution reduces losses, and better consumption reduces total demand. Prescriptive analytics can optimize both simultaneously.
8. Compare mass advertising with targeted digital advertising in terms of optimization.
Mass advertising, like TV or billboard ads, reaches many people at once but cannot be easily personalized. It is hard to measure exactly who buys because of the ad. Targeted digital advertising, like social media or search ads, shows ads only to people who match certain traits, like age or interests. Digital ads allow quick testing: you can change the budget, message, or audience daily. Mass advertising is less flexible and needs more spending to test. For optimization, digital advertising gives more data and control. However, mass advertising can build brand awareness broadly. The best mix depends on the product and market.
9. How would you use d-separation to check if two variables are independent given a set of evidence?
First, draw the Bayesian network. Then consider all paths between the two variables. For each path, check if it is blocked by the evidence set. A path is blocked if there is a node on the path that is in the evidence set and the node is in a chain or fork structure, or if there is a collider that is not in the evidence set and none of its descendants are in the evidence set. If every path is blocked, the variables are conditionally independent given the evidence. If at least one path is not blocked, they may be dependent. This check helps in reasoning about probabilistic relationships without doing complex calculations.
10. Compare early decision support systems with today's BI tools.
Early decision support systems (DSS) ran on single computers and used simple tables. They needed experts to write commands and could take hours to run a query. Today's BI tools are cloud-based, work on any device, and update data in seconds. They offer colorful dashboards, maps, and graphs that anyone can click on. DSS focused on specific problems, while BI covers the whole business. Modern BI also uses artificial intelligence to suggest insights. The biggest change is ease of use — today, a store manager can open a dashboard on a tablet and see sales instantly. Early DSS were powerful but limited to technical staff.
11. What is one main difference between GDPR and CCPA in how they define personal data?
GDPR (General Data Protection Regulation) defines personal data very broadly as any information relating to an identified or identifiable person. CCPA (California Consumer Privacy Act) defines personal information more narrowly, focusing on data that identifies, relates to, or could reasonably be linked to a consumer or household. For example, IP addresses are personal data under GDPR, but under CCPA they are only personal information if they can be linked to a specific device or person. The key difference is the breadth: GDPR covers almost any identifier, while CCPA requires a link to a consumer or household.
12. Suppose a decision tree has two decision nodes in a row. Describe how you would roll back such a tree.
You start from the rightmost decision node. For that node, you look at its branches (each leading to chance or terminal nodes). Compute the expected value for each branch by rolling back any chance nodes further right. Then choose the branch with the highest expected value. Replace that decision node with the value you chose. Then move left to the next decision node. Now you treat it the same way: consider its branches, compute expected values using the values you already calculated to the right, and pick the best. Continue until you reach the root decision node. This gives the optimal sequence of decisions.