WorksheetsWeb Mining 9
Total questions: 74
Worksheet time: 37mins
Name
Class
Date
1.
In the online advertising setting, what is the role of an ad network?
a)
Select and deliver ads between advertisers and content providers
b)
Write news articles for publishers
c)
Host user email accounts
d)
Provide internet access to users
e)
Translate web pages into multiple languages
2.
Which revenue model is included in the online advertising model?
a)
CPM
b)
Subscription-only pricing
c)
Per-minute billing
d)
Freemium donations
e)
Flat annual licensing
3.
In online advertising, what does CPM stand for?
a)
Cost Per Impression
b)
Cost Per Installation
c)
Clicks Per Minute
d)
Conversion Per Message
e)
Commission Per Merchant
4.
In online advertising, what does CPC stand for?
a)
Cost Per Click
b)
Cost Per Content
c)
Clicks Per Customer
d)
Commission Per Click
e)
Cost Per Campaign
5.
In online advertising, what does CPA stand for?
a)
Cost Per Action
b)
Cost Per Audience
c)
Clicks Per Action
d)
Commission Per Advertiser
e)
Cost Per Ad
6.
Under a CPM model, revenue increases with which factor?
a)
The number of times an ad is shown (impressions)
b)
Only the number of purchases
c)
Only the number of ad clicks
d)
Only the number of unique users
e)
Only the time spent on the page
7.
In the CPC revenue formula, what does CTR represent?
a)
The probability of clicking on an ad
b)
The cost charged per action
c)
The total revenue per impression
d)
The number of advertisers in the auction
e)
The quality score of the landing page
8.
In the CPA revenue formula, what does conversion rate mean?
a)
The probability that a user takes an action after viewing the ad page
b)
The number of impressions per second
c)
The price of an ad impression
d)
The fraction of ads rejected by the network
e)
The number of documents returned by a search engine
9.
In search engine advertising, what triggers showing sponsored ads?
a)
A user's search query
b)
A user's browser window size
c)
A user's screen color theme
d)
The position of the mouse cursor
e)
A user's battery level
10.
In the search engine advertising model, ads are matched primarily to what?
a)
The search query
b)
The user's font settings
c)
The network latency
d)
The page background color
e)
The user's operating system version
11.
Which factor is listed as part of maximizing revenue in search advertising?
a)
Ad page quality
b)
Weather forecast accuracy
c)
Printer driver version
d)
Keyboard layout preference
e)
Screen brightness level
12.
When scoring ads based on content, ads are treated like what?
a)
Text documents
b)
Audio signals
c)
Video streams
d)
Network packets
e)
Hardware devices
13.
In a language-model approach to query-ad matching, what is one scoring idea mentioned?
a)
Estimate the probability of the ad under a query language model
b)
Count pixel differences between ad images
c)
Measure the refresh rate of the browser
d)
Rank ads by file size of the landing page
e)
Sort ads alphabetically by advertiser name
14.
Another language-model scoring idea shown is to estimate which probability?
a)
The probability of the query under an ad language model
b)
The probability of the screen resolution given the query
c)
The probability of a click given the server uptime
d)
The probability of a user given the browser type
e)
The probability of a document given the keyboard layout
15.
Which measure is shown for comparing the ad language model and the query language model?
a)
KL divergence
b)
Euclidean distance
c)
Pearson correlation
d)
Cosine of page angle
e)
Edit distance of URLs
16.
According to the slides, a drawback of the simple content-based model is that it does not use what?
a)
User feedback
b)
Web page titles
c)
Network bandwidth
d)
File compression
e)
Browser cookies
17.
When scoring based on user feedback, what probability is computed for each query and ad page pair?
a)
The probability that the user clicks on the ad page given the query and ad page
b)
The probability that the ad image loads faster than the query
c)
The probability that the query contains exactly two words
d)
The probability that the user changes their keyboard
e)
The probability that the ad page uses HTTPS
18.
In the logistic regression approach described, query and ad content are represented as what?
a)
Vectors using a bag-of-words representation
b)
Hand-drawn sketches
c)
Audio waveforms
d)
3D models
e)
Hardware register states
19.
In collaborative filtering for ads, what type of feedback is used as latent user feedback?
a)
Clicks on ad pages
b)
Monitor brightness settings
c)
Printer queue length
d)
Keyboard backlight level
e)
Browser window position
20.
In the collaborative filtering continuation, similarity is computed between which items?
a)
Ads and ads (an ad-ad similarity matrix)
b)
Queries and screen pixels
c)
Users and network routers
d)
Documents and keyboard layouts
e)
Servers and batteries
21.
The query mining section notes that a major search engine can receive queries at what general scale?
a)
Tens of thousands of queries per second
b)
A few queries per hour
c)
Exactly one query per day
d)
Only one query per user per month
e)
No more than ten queries per week
22.
According to the query features slide, what is the average length of a query?
a)
About 2.4 words
b)
About 10 words
c)
About 25 words
d)
About 100 words
e)
About 1,000 words
23.
What does the slide say about the vocabulary distribution of queries compared to website content?
a)
They are different, reflecting different user interests versus available content
b)
They are identical in most cases
c)
Queries always use a larger vocabulary than websites
d)
Websites always use a smaller vocabulary than queries
e)
The vocabulary distribution is irrelevant to search
24.
Which information is included in query logging according to the slides?
a)
Selected documents of the user
b)
The color theme of the browser
c)
The user's screen size history
d)
The user's battery temperature
e)
The user's printer model
25.
Which step is listed under query preprocessing?
a)
Identify query session
b)
Encrypt every query with a secret key
c)
Render each query as an image
d)
Translate every query to Latin
e)
Compress queries into audio
26.
When identifying query sessions, consecutive query pairs can be classified into which class?
a)
Query generalization
b)
Image recoloring
c)
Database sharding
d)
GPU overclocking
e)
Packet routing
27.
Which is a sign of a bot-generated query mentioned in the slides?
a)
Unusually high query rate and recurring query frequency
b)
Queries that contain punctuation
c)
Queries typed with uppercase letters
d)
Queries issued from a mobile phone
e)
Queries that include a brand name
28.
Which operation is part of query standardization in the slides?
a)
Remove stopwords
b)
Increase image contrast
c)
Change screen resolution
d)
Optimize network routing
e)
Encrypt the query log
29.
In the query mining applications, what is labeled as Application 1?
a)
Query suggestion
b)
Spam detection
c)
Video summarization
d)
Speech recognition
e)
Network intrusion detection
30.
According to the slides, which can be used as a basic unit of a language model for queries?
a)
Character
b)
Database table
c)
IP address
d)
Screen pixel
e)
CPU instruction
31.
In an n-gram language model, what does a bigram model estimate?
a)
The probability of a word given the previous word
b)
The probability of a word given the next two words
c)
The probability of a document given a server
d)
The probability of a click given a router
e)
The probability of a query given a keyboard
32.
For query extension, what key assumption about two keywords is stated in the slides?
a)
If one keyword leads to documents containing another keyword, the keywords are likely related
b)
Two keywords are related only if they have the same spelling
c)
Keywords are related only if they appear in the same URL
d)
Keywords are related only if they are synonyms in a dictionary
e)
Keywords are never related across different topics
33.
The query extension model referenced in the slides is attributed to which authors and year?
a)
Hang Cui et al., 2003
b)
Page and Brin, 1998
c)
Manning and Schutze, 1999
d)
Salton, 1971
e)
Ginsberg et al., 2009
34.
In the query extension model, which type of data is explicitly used to support expansion?
a)
User feedback
b)
Screen color settings
c)
Printer logs
d)
CPU temperature traces
e)
Battery discharge curves
35.
In the extension model notation, what does P(Dk | wi in the query) represent?
a)
The probability that document Dk is selected when keyword wi appears in the query
b)
The probability that wi is misspelled in the document
c)
The probability that Dk is indexed after wi
d)
The probability that the user edits wi into uppercase
e)
The probability that Dk is deleted from the web
36.
In the extension model notation, what does the frequency f(wi, Dk) in query sessions count?
a)
How often keyword wi appears and document Dk is selected in the same query session
b)
How many times the browser is refreshed while viewing Dk
c)
How many advertisers bid on keyword wi
d)
How often a server times out on Dk
e)
How many times wi is translated into another language
37.
In the query extension procedure, what is done after scoring terms in related documents?
a)
Use the top scoring terms to construct an expanded query
b)
Delete the query log immediately
c)
Randomly shuffle query words
d)
Replace all terms with stopwords
e)
Disable user feedback collection
38.
In the disease warning application, what is the core idea connecting queries to illness?
a)
People searching for disease information is proportional to how many people are sick
b)
People search less when they are sick
c)
Queries are unrelated to real-world events
d)
Only medical professionals issue disease queries
e)
Disease can be detected only from page layout
39.
If an advertiser only wants to pay when users click, which revenue model matches that goal?
a)
CPC
b)
CPM
c)
CPA
d)
Ad exchange
e)
Display setting
40.
A publisher wants revenue even if users never click. Based on the slides, which pricing model best fits this preference?
a)
CPM
b)
CPC
c)
CPA
d)
Collaborative filtering
e)
Query generalization
41.
Which model requires both a click and a subsequent action to generate revenue, according to the revenue formulations in the slides?
a)
CPA
b)
CPM
c)
CPC
d)
Vector space model
e)
Stemming
42.
Two ads have the same CPC price and the same number of impressions. Which ad yields higher expected revenue under CPC?
a)
The ad with higher click-through rate (CTR)
b)
The ad with lower click-through rate (CTR)
c)
The ad with lower conversion rate
d)
The ad with higher impression cost (CPM)
e)
Both ads always yield the same revenue
43.
Under a CPA model with the same number of impressions and the same CTR, what would increase expected revenue the most?
a)
Higher conversion rate
b)
Lower conversion rate
c)
Lower ad page quality
d)
Fewer impressions
e)
Removing stopwords
44.
When selecting ads to maximize revenue in sponsored search, which combination of factors is explicitly listed in the slides?
a)
Match the query, advertising costs, and ad page quality
b)
Browser font size, screen resolution, and keyboard layout
c)
Network bandwidth, CPU speed, and disk space
d)
Video length, audio quality, and subtitle language
e)
Time of day, device color, and battery level
45.
If ads are treated like text and matched by similarity to a query, which two methods are named as approaches in the slides?
a)
Vector space model and language model
b)
Neural rendering and ray tracing
c)
Packet switching and circuit switching
d)
Hashing and encryption
e)
Clustering and page segmentation
46.
The slides show multiple language-model-based scoring directions for query-ad matching. Which pair best reflects those directions?
a)
Score ad given the query model, and score query given the ad model
b)
Score user given the browser, and score browser given the user
c)
Score document length given the server, and score server given the document
d)
Score pixels given the screen, and score screen given the pixels
e)
Score time given the clock, and score clock given the time
47.
Why is the simple content-based model described as suitable for short popular queries?
a)
It is simple and works well when many similar queries are frequent
b)
It always uses rich user feedback on rare queries
c)
It requires no query text at all
d)
It guarantees real-time processing for the long tail
e)
It eliminates the need for ad selection
48.
A query is rare and belongs to the long tail. Based on the slides, what is a likely weakness of the simple content-based model?
a)
It hardly handles rare queries well
b)
It becomes more accurate on rare queries by default
c)
It automatically learns from clicks without training
d)
It ignores query text and only uses costs
e)
It only fails on popular queries
49.
Compared with content-only scoring, what additional signal does the user-feedback approach explicitly incorporate?
a)
Estimated probability of a click for a query and ad page pair
b)
The font size of the query text
c)
The background color of the results page
d)
The number of tabs open in the browser
e)
The user's screen orientation
50.
In the logistic regression approach, what is the role of the weight parameters learned from user feedback?
a)
They map query and ad vectors to a click probability
b)
They compress images for faster loading
c)
They translate queries into another language
d)
They assign each query a random ad
e)
They remove stopwords from documents
51.
In the collaborative filtering approach, why would kNN be useful according to the slides?
a)
To find similar ads based on query-based representations
b)
To convert queries to lower case
c)
To compute CPM revenue directly
d)
To restore Vietnamese accents
e)
To identify bot queries from rate alone
52.
If you have an ad-ad similarity matrix, what kind of inference does it support in the collaborative filtering setup described?
a)
Estimating relevance of an ad to a query using similar ads
b)
Encrypting the query log for storage
c)
Rendering the ad as an image
d)
Detecting long tail queries by word count
e)
Replacing the search engine index
53.
Which item belongs to query logging rather than query preprocessing in the lecture slides?
a)
List of relevant documents
b)
Stemming
c)
Remove stopwords
d)
Filter bot query
e)
Convert to lower case
54.
A reasonable preprocessing pipeline in the slides includes identifying sessions, filtering bots, and standardizing queries. Which order best matches that pipeline?
a)
Identify session, filter bot queries, then standardize
b)
Standardize, filter bots, then identify session
c)
Filter bots, standardize, then identify session
d)
Standardize only; the other steps are not used
e)
Identify session only; the other steps are not used
55.
Two consecutive queries have the same content but the user changes the search scope. How does the slides' session classification describe this case?
a)
Same content but different search scope
b)
Query detailing
c)
New query content
d)
Bot query duplication
e)
Vietnamese accent restoration
56.
Which pair of steps is listed specifically for Vietnamese query standardization?
a)
Restore accent and tokenize
b)
Remove stopwords and tokenize
c)
Convert to upper case and stem
d)
Encrypt and compress
e)
Cluster and classify
57.
Why does the language model approach for query suggestion require a large query dataset, according to the slides?
a)
Because it learns probabilities from query data and needs many examples
b)
Because it must store every web page on the internet
c)
Because it only works with images, not text
d)
Because it depends on ad exchanges for training
e)
Because it avoids using any historical data
58.
If you move from a unigram model to a bigram model as described, what extra context is incorporated?
a)
The previous word when predicting the next word
b)
The next two words when predicting the previous word
c)
The screen resolution of the user
d)
The advertiser's monthly budget
e)
The layout of the results page
59.
Why might query expansion based only on textual content fail to meet user needs, as stated in the slides?
a)
It may not capture what users actually want without user feedback
b)
It always reduces the number of relevant documents
c)
It prevents session identification
d)
It makes bot filtering impossible
e)
It removes the need for ranking
60.
In the query extension procedure, what must happen before constructing the expanded query?
a)
Find related documents and score candidate terms from those documents
b)
Delete all stopwords from the web index
c)
Compute CPM and CPC revenue for each keyword
d)
Replace every term with a synonym dictionary entry
e)
Sort ads by advertiser name
61.
A campaign is optimized for completed sign-ups, and the advertiser wants to pay only when a sign-up happens. Which model best aligns with that objective?
a)
CPA
b)
CPC
c)
CPM
d)
Ad exchange
e)
Content match setting
62.
An ad is shown many times but receives zero clicks. Which revenue model can still generate revenue in this situation (assuming the ad is displayed)?
a)
CPM
b)
CPC
c)
CPA
d)
Click probability model
e)
Query session model
63.
An ad has a high click-through rate but a very low conversion rate after the click. Based on the revenue dependencies shown in the slides, which pricing model would be least favorable for the publisher's expected revenue?
a)
CPA
b)
CPC
c)
CPM
d)
Display advertising
e)
Vector space model
64.
The slides list several disadvantages of the simple content-based model. Which pair of disadvantages most directly conflicts with handling huge query volume and rare queries?
a)
Hardly process in real-time and hardly handle rare queries
b)
Simple model and suitable for popular queries
c)
Uses user feedback and handles long tail well
d)
High conversion rate and low CTR
e)
Requires query logging and removes stopwords
65.
The lecture contrasts content-based scoring with scoring based on user feedback. For long-tail queries, which reason best explains why user feedback is valuable in the framework described?
a)
It can estimate click likelihood for specific query and ad page pairs beyond text similarity alone
b)
It removes the need to match ads to queries
c)
It guarantees that all queries are short and popular
d)
It replaces the need for an ad network
e)
It converts CPM into CPA automatically
66.
The query features slides note that many users click the first result and mostly use only the first two results. If you learn Pr(click | query, ad page) from click logs, what risk follows most directly from this behavior?
a)
Clicks may reflect position bias rather than true relevance alone
b)
Clicks become independent of the query text
c)
Clicks become identical for all ads
d)
Clicks no longer depend on the ad page
e)
Clicks can only be used for CPM pricing
67.
A new ad page has never been clicked before. Based on the two user-feedback approaches described (logistic regression with content vectors vs collaborative filtering using click-based interactions), which approach is more likely to still produce a score using only what is explicitly available about the ad?
a)
Logistic regression with query and ad content vectors
b)
Collaborative filtering using only the interaction matrix
c)
KL divergence between language models
d)
Bot filtering based on query rate
e)
Query session identification
68.
In the collaborative filtering setup described, ads are represented by queries to compute similarity and kNN is used. Which situation would most directly reduce the reliability of ad-ad similarity?
a)
Very sparse click feedback across many query and ad pairs
b)
Lowercasing all queries
c)
Removing stopwords from queries
d)
Restoring accents in Vietnamese
e)
Using query suggestion as an application
69.
A user submits 'best gaming laptop' and then immediately submits 'laptop'. Using the session categories listed in the slides, how should the second query be classified relative to the first?
a)
Query generalization
b)
Query detailing
c)
New query content
d)
Bot query duplication
e)
Same content but different search scope
70.
Bot filtering in the slides uses signals like unusually high query rate and recurring query frequency. Which human behavior could most plausibly be misclassified as bot activity if you rely only on those signals?
a)
A user repeatedly refining a query during an urgent task
b)
A user converting a query to lower case
c)
A user removing stopwords once
d)
A user clicking the first result once
e)
A user issuing a single rare long-tail query
71.
The query extension assumption says two keywords are likely related if one leads to documents containing the other. Which scenario most challenges this assumption while still fitting the slide description?
a)
A popular document contains many unrelated keywords, causing co-occurrence without true relatedness
b)
Two keywords have different capitalization
c)
One keyword is in Vietnamese and one is in English
d)
A query contains stopwords
e)
A query has only one word
72.
In the query extension procedure, you select the top n highest scoring terms to form an expanded query. What is the most direct trade-off implied by this step?
a)
Choosing more terms may improve coverage but can also introduce less relevant terms
b)
Choosing more terms always reduces the need for user feedback
c)
Choosing fewer terms always increases the query volume
d)
Choosing more terms converts CPC revenue into CPM revenue
e)
Choosing fewer terms prevents session identification
73.
The disease warning application assumes searches for disease information are proportional to sickness. Which external factor would most directly violate this proportionality while keeping the mechanism described unchanged?
a)
News coverage causing many healthy people to search for the disease
b)
Stopword removal during standardization
c)
Using a unigram model for suggestion
d)
Filtering bot queries from logs
e)
Using kNN in collaborative filtering
74.
The n-gram formulas shown add 1 to observed counts. What outcome does this most directly prevent when estimating probabilities from counts?
a)
Assigning zero probability to events that were not observed in the training data
b)
Assigning equal probability to all words regardless of data
c)
Making probabilities depend on screen resolution
d)
Making probabilities depend on ad page quality
e)
Making probabilities independent of the query dataset size
100 %
