source: liacs/dbdm/dbdm_5/report.tex@ 385

\documentclass{report}
\usepackage{graphicx}

\title{Databases and Data Mining --- Assignment 5}
\author{Rick van der Zwet}


\begin{document}
\newcommand{\question}[2]{\begin{quotation}\noindent{\Large``}#1. \textit{#2}{\Large''}\end{quotation}}
\maketitle

\question{1a}{Propose an algorithm, in pseudo-code, for the following: The automatic
generation of a concept hierarchy for numerical data based on the equal-width
partitioning rule.}
\begin{verbatim}
input = [] # Input array of all input numbers
num_intervals = %User input of number of intervals needed%
max = maximum(input)
min = minimum(input)
interval_width = (max - min) / num_intervals

output = [] # Output array, where the values of interval k 
            # is stored in value output[k]

for value in input:
  interval = value / interval_width # Find it's correct bin
  output[interval].append(value)    # Put the value inside the bin
endfor
\end{verbatim}

\question{1b}{Propose an algorithm, in pseudo-code, for the following: The automatic
generation of a concept hierarchy for numerical data based on the equal-frequency
partitioning rule.}
\begin{verbatim}
input = [] # Input array of all input numbers
num_intervals = %User input of number of intervals needed%
input_length = length(input) # Number of items in list
interval_width = input_length / num_intervals

sorted_input = sorted(input) # Sort items on value from small to large

output = [] # Output array, where the values of interval k 
            # is stored in value output[k]

interval = 0
counter = 0
for value in sorted_input:
  output[interval].append(value)    # Put the value inside the bin

  # If the width has been 'filled' continue with the next one
  counter++ 
  if counter > interval_width:
    interval++
    counter = 0
  endif
endfor
\end{verbatim}

\question{2}{Suppose that a data warehouse consists of four dimensions,
    date, spectator, location, and game, and two measures, count, and charge, where
    charge is the fare that a spectator pays when watching a game on a given date.
    Spectators may be students, adults, or seniors, with each category having its own
    charge rate.}
\question{2a}{Draw a star schema diagram for the data warehouse.}
See figure~\ref{fig:star}.
% http://it.toolbox.com/blogs/enterprise-solutions/star-schema-modelling-data-warehouse-20803
\begin{figure}[htp]
\centering
\includegraphics[scale=0.6]{star-diagram.eps}
\caption{Star schema diagram 2a data warehose}
\label{fig:star}
\end{figure}

\question{2b}{Starting with the base cuboid [date, spectator, location, game],
what specific OLAP operations should one perform in order to list the total
charge paid by student spectators at GM\_Place 2004?}
% http://en.wikipedia.org/wiki/OLAP_cube#OLAP_operations
You first will need to \emph{slice} on the condition \texttt{game ==
'GM\_Place'}. Secondly you will need to slice on \texttt{date.year == '2004'}.
This will give you all the charges for GM Place in 2004. Next we slice to
\texttt{spectator.type == 'student'}. Lastly we sum all the charges in the
display phase (\texttt{pivot}). 

\question{2c}{Bitmap indexing is useful in data warehousing. Taking this cube
as an example, briefly discuss advantages and problems of using a bitmap index
structure.}
% http://en.wikipedia.org/wiki/Bitmap_index
Bitmap indexing in this case is useful to have a compact representation of for
example the spectator type. As this can only be 4 options a four bit
representation fits the need to store all possible combinations. This required
binary operators to do searches in this set are quick, but their results will
need to be processed before beeing able to represent it.
One other advantage is the fact that the bitmap indexing compresses really well
as patterns are reoccuring.


\question{3}{A data cube, C, has n dimensions, and each dimension
    has exactly p distinct values in the base cuboid. Assume that there are no concept
    hierarchies associated with the dimensions. What is the maximum number of cells
    possible (including both base cells and aggregate cells) in the data cube, C?}
% http://www2.cs.uregina.ca/~dbd/cs831/notes/dcubes/dcubes.html
Take for example 2 dimensions with each 4 distinct values in the base cuboid,
this will a total of $4^2=16$ possibilities. Taking this more general will this be $p^n$.

\question{4}{The Apriori algorithm uses prior knowledge of subset support properties.}
\question{4a}{Prove that all nonempty subsets of a frequent itemset must also be frequent.}
% http://en.wikipedia.org/wiki/Apriori_algorithm
If an itemset $I$ is frequent it means it's occurence classifies a minimal
support level, hence if you take a subset of $I$ ($I_{sub}$). The count will remain the
same -or larger if the subset is also part of an other itemset $J$, which has
distint matches from $I$- as subset $I$. So $I_{sub}$ will also be frequent if
$I$ is frequent.

\question{4b}{Given frequent itemset $l$ and subset $s$ of $l$, prove that the confidence of the rule 
         “$s’ => (l - s’)$” cannot be more than the confidence of the rule “$s => (l - s)$” where $s’$
         is a subset of $s$.}
If $s'$ has an higher confidence than $s$, than a smaller itemset have a higher
conditional propability of also containing a larger itemset. If the repeat this
assumption, you will end up on the left side with a empty set and on the right
the full itemset $l$, which then would have the highest confidence. While this
would not be valid. As rules like 'Customer buying nothing', will also buy
'beer' can not be existing.

\question{5}{An optimization in frequent item set mining is mining closed patterns, or mining max
    patterns instead. Describe the main differences of mining closed patterns and mining
    max patterns.}
% http://www.dataminingarticles.com/closed-maximal-itemsets.html
By definition the itemset $X$ is closed means $X$ is frequent and there exists
no super-pattern $X /in Y$ wth the same support as $X$. While $X$ is max if $X$
is frequent and there exists no \textit{frequent} super-pattern $X /in Y$.
Maximum frequent itemsets has the downside that you don't not know the actual
support of those sub-itemsets. The closed frequent itemsets helps purging a lot
of itemsets that are not needed.

\question{6}{The price of each item in a store is nonnegative. For
each of the following cases, identify the kinds of constraint they represent (e.g.
anti-monotonic, monotonic, succinct) and briefly discuss how to mine such association
rules efficiently:}
% www.cs.sfu.ca/CC/741/jpei/slides/ConstrainedFrequentPatterns.pdf
\question{6a}{Containing one free item and other items the sum of whose prices is at least \$190.}
This is \emph{monotonic}, as adding items will never lower the sum of prices.
Mining this will be bestly done by first finding the free items and next order
the other items based on price, starting with the most expensive ones first.
The soon you get to \$190 can al it's supersets to matching as well, next you
remove the last item and try to match up again. Continue this till you removed
the most expensive item and tried to match up again. Next \emph{JOIN} with all
the free items and the list is complete.

\question{6b}{Where the average price of all the items is between \$120 and \$520.}
This is convertible \emph{anti-monotonic} and convertible \emph{monotonic} if
you look at one contrain.  \emph{anti-monotonic} If itemset $S$ violates the
sub-constraint $avg(S.Price) \le \$520$. so does every itemset with $S$ in
prefix item value descending order.  \emph{monotonic} If itemset $S$ constraint
$avg(S.Price) \ge \$120$ so does every itemset having S as prefix item values
descending order.

Satifing both conditions how-ever does make it neither \emph{anti-monotonic}
nor \emph{monotonic}. A fast way to generate this set is to use the algoritm
used in 6a but modify the number constraint on-the-fly. Like average of 3 items
between a centrain range (120 -- 520) It the same of checking whether the sum
of 3 items is between ($3*120$ -- $3*520$).

\question{7}{Suppose a city has installed hundreds of surveillance cameras at strategic locations in
busy streets. All the captured video (each stream being 640 pixels x 480 pixels, 24-
bits (RGB) per pixel, 25fps) is streamed to a central observation post in real time.
Describe an efficient system that would allow real time data mining and continuously
querying these video streams for abnormal events, such as explosions, people in
panic, etc.. Also discuss the computational costs and the memory requirements of
your system.}

Every camera generates $640 *  480 * 24 bits * 25fps = 180.000 kbit/sec \approx
21.9 MByte/sec$. I would first use a compression algoritm to send-the-data over
the wire by sending the changes only and not the full frame every time. Secondly 
applying abnomality detections using heuristics. Keeping the detected
abnormalities available un-compressed and uncut and save all other data in a
much lower resolution (1fps), with a compression ratio of 4:1 this would make
storage stream of approx 19GB/camera/day. Processing power required depends on
the algoritms used, but a 500Mhz CPU/camera should fit the bill when simple
algoritms are used.

\question{8}{A flight data warehouse for a travel agent consists of six
dimensions: traveler, departure (city), departure\_time, arrival (city), arrival\_time,
and flight; and two measures count, and avg\_fare, where avg\_fare stores the concrete
fare at the lowest level but the average fare at the other levels. Suppose the cube is
fully materialized. Starting with the base cuboid [traveler, departure, departure\_time,
arrival, arrival\_time, flight], what specific OLAP operations (e.g roll-up flight to
airline, etc.) should one perform in order to list the average fare per month for each
business traveler who flies American Airlines (AA) from Los Angeles (LA) in the
year 2007?}

Note bit tricky as we use the departure\_time to determine the year and
arrival\_time to determine the month of the flight. When these values are not
set to the same day we might run into inconsistencies.


\begin{verbatim}
* Roll-up departure\_time to 'Year'
* Roll-up flight to 'Airline'
* Dice departure = 'LA' and departure_time = '2007' and flight = 'American Airlines'
* Roll-up traveler to 'Consumer/Business'
* Roll-up arrival_time to 'Month'
* Slice traveler = 'Business'
\end{verbatim}

\question{9}{In graph mining what would be the advantage of the described apriori-based approach
over the pattern growth based approach (see lecture slides) and vice versa.}
The advantage comes in the simplicity of the algotitm when it get to
computations needed and it's setup, but this does not make it effient. 

On the other hand apriori uses the costly process of candidate generation and
testing, making it more 'heavy' then pattern grow based. and also avoid costly
database scans.

% http://en.wikipedia.org/wiki/Association_rule_learning

\bibliography{report}

\end{document}