Graphics Reference
In-Depth Information
than one column. In addition, the order of the attribute columns can be changed
by clicking and dragging the attribute names to the right or let.
Highlighting groups of auctions:
Ater the attribute/s of interest have been sorted,
groupsof auctions can be selected and their corresponding time series in the let
panels are highlighted. For example, if the attributes table is sorted by the end
day of the auction, it is easy to select all auctions that ended on a weekday from
the table, and see the corresponding time series highlighted, which reveals that
they are the auctions that tend to end with the highest prices (Fig.
.
).
Summary statistics:
he summary statistics tab shows the mean, standard devia-
tion,minimum,max,median,andthequartiles foreachattribute fortheselected
auctions. his is updated interactively when the auctions are filtered with Time-
Boxes, or when users select a subset of auctions manually. For example, while
the median seller rating of all auctions is
, when users apply a TimeBox to
select the auctions that started with a low price, the median seller rating jumps
to
. Moving the TimeBox to select auctions that started with a high price re-
sults in a median seller rating of
, which may imply that starting the auction
with a low starting price is a strategy that tends to be employed by experienced
sellers.
he array of interactive operations described above support data exploration, and
Shmueli et al. (
) describe how these operations can be used for the purpose of
decision-making, through a semi-structured exploration. Exploration can be guided
by a set of hypotheses, and the results can then help the user to find support for and
direct the direct the user towards suitable formal statistical models. In particular,
theyshowhowinsights gainedfromthevisualexploration canimproveseller,bidder,
auction house, and other vendors' understanding of the market, thereby assisting
their decision-making processes.
Forecasting with TimeSearcher
5.5.2
Functional object value forecasting is an area of functional data analysis that has not
received much attention so far. It involves forecasting the value of a curve at a par-
ticular time t (either a particular curve in the data or the average curve), based on
information contained in the functional data and the attribute data. We propose the
following general forecasting procedure:
Select similar items:
For a partial curve (e.g., an ongoing auction that has not
closed), we select the subset of curves that are closest to the curve of interest
in the sense of similar attributes and curve evolutions and dynamics. For the
attribute criterion, this can be achieved either by sorting by attributes and se-
lecting items with similar values for the relevant attributes (e.g., auctions of the
same duration and with the same opening price), or directly by using a filtering
facility that allows theusertospecifylimitsonthevalues ofeachofthe attributes
of interest (this facility is currently not available in the public version of Time-
Searcher). When curve-matching, TimeBoxes can be used to find curves that
have similar structures during the time periods of interest (e.g., auctions with
