The directory of greek blogs Sync.gr run an extended survey on the greek blogosphere, finally answered by 919 bloggers. And we were privileged enough to get early access to the raw data (now available here, thanx to Nick Drandakis); putting our hands on this unique data set was a no brainer.

We finally came up with 258 rules we consider insightful, but you may judge for yourself. The full analysis report is provided here, feel free to read, download or share at will and let us know of your comments!

Creative Survey 2008 was run with a single purpose: To provide insights on the creative professionals in Greece. The survey consisted of 29 questions, answered by 1050 professionals, resulting in a rich data set we were privileged enough to gain access of (a big thank you to the people behind the survey!). And we decided to put our hands on the data and sprinkle some MineKnowledge magic, attempting to extract some deeper insights, next to the useful classical analysis already performed.

So, the original analysis (in greek, 5MB .pdf, you may download it here) resulted in some histograms like these:

Which of the following ones do you use to get informed about your profession? Blogs 49.2%

Amount of productivity bonus earned? Equal or less to 50% of salary, 69.8%

while we came up with some rules you may find insightful, like the following ones:

If job_title=creative_director & occupation=web_design then read_blogs=yes (accuracy 78%)

If employees=21-40 & job_title=creative_director then bonus<50% salary (accuracy 86%)

You may find a dozen of such diagrams, next to almost 250 of such rules in the full analysis report of the Creative Survey (available to view and download below, .5MB .pdf). And we do believe that across the report you’ll discover insights of interest you couldn’t imagine existed, you may let us know your thoughts in the comments.

Plus, in the case that you possess a data set and you would like to take full advantage of its hidden insights and latent knowledge, by now you know what to expect and how to get it.

For our fourth case study, we put our hands on a data set of a survey concerning seismological data. The 10,333 measurements of earthquakes were taken between 1961 and 2002, each one tracking 10 attributes of the incident. Our aim is to indicate the most significant attributes that affect our target value, set as the magnitude of the earthquake. Table 1 gives you the names of these attributes.  

Table 1: Data set at a glance

Table 2 gives you further details on the target attribute. In other words, we attempted to extract patterns out of this data set, that interpret the behavior of each earthquake’s magnitude.

Table 2: Description of the target attribute

Some advanced filtering techniques returned us the following attributes as the most prominent ones, in terms of their information value, given the target.

 Table 3: Attributes of most informational value

With inputs like these, we put into use an extended set of advanced machine learning algorithms to finally bring to light patterns like the ones that follow:

Rule 1: if depth<=10km & year<=1964 then magnitude>6.5R (83% accuracy)

Rule 1 suggests with an accuracy of 83% that for seismic event before 1964, a depth of less than 10 km was expected to result a magnitude of more than 6.5 Richter.

Rule 2: If depth>19km  &  epicentery>=38.26km then Y=1 (89% accuracy)

Rule 2 indicates that when the epicenter of an earthquake is greater than 38.26km and its depth greater than 19 km, then, again, magnitude is expected to be greater than 6.5 Richter.

You may find a more extended description of the above, including a complete set of rules extracted, in the full case study report, which is available below (you may try the full screen view, or even download the .pdf of the document).

Now, if we could return such insights from a limited data set on a major and yet unsolved scientific problem, consider the power of our techniques applied in that data set of yours, yet remaining unused somewhere in your computer or in your intranet.

Still considering it? Learn more about our services here, we’re awaiting for your data, starting from now. 

The third case study we deploy regards a data set of voting records in the US House of Representatives. Given 16 key votes for each one of the 435 members of the House, we focused on extracting patterns explaining the voting behavior, according to the party each one of the representatives belongs to.

The key votes (attributes) are presented in Table 1 that follows.

Table 1: Key votes titles (attibutes)

Table 2 presents further information on the target variable, namely the party each of the Representatives belongs to.

Table 2: Target attribute (part) details

Table 3 provides the votes of most informational value for judging the party of the Representative, as occurred by extensive experimentation with proper filtering algorithms.

Table 3: Key votes (attributes) of importance

With inputs as such, a bunch of advanced machine learning algorithms were run to finally return some critical insights on how voting turned out to be related with the party. Some of the rules extracted are presented below, while you may find an extended set of them in the full analysis report which is available for download at the end of this post.

Rule 01: If crime=yes & religious-groups-in-schools=no & aid-to-Nicaraguan-contras=no then vote=democrat

For example, Rule 01 introduces that Representatives who voted for the crime resolution, but were against to the resolutions about religious groups in schools and the aid to the Nicaraguan contras are expected to be Democrats, with 97% rate of success.

Rule 02: If immigration=yes & adoption-of-the-budget-resolution=no then vote=republican

Rule 02 suggests with a 76% of accuracy that Representatives who voted for the immigration resolution but against the adoption of the budget belong to the republican party.

Rule 03: If physician-fee-freeze=yes & adoption-of-the-budget-resolution=no & El-Salvador-aid=yes then vote=republican

Finally, Rule 03 deploys that those who voted for the physician fee freeze and El Salvador aid regulations, while were against the adoption of the budget resolution, were republicans with an accuracy of 87%.

The full report, which you may download here (.pdf), contributes more than 30 such rules and patterns, fully representative of our services. And, please take into account that, focusing on the party of each representative was one of the options, rules and patterns extracted could also focus on the expected votes on a specific resolution, taking the voter’s party into consideration, it goes further.

Consider also the power of such an analysis when the given data set regards, for example, voters and their characteristics. If you are used to think some bars and pies, relating one attribute with the vote, as razor-sharp survey analysis, get ready to be impressed with the extent and value of the results and insights you’ll gain from a mineknowledge analysis.

Still considering it? Learn more about our services here, we’re awaiting for your data, starting from now.

The second case study we present regards the analysis of a survey conducted to find out factors that shape the purchase decision of electrical switches and sockets. The survey, prepared for a well-known manufacturer and distributor of electrical switches and sockets, resulted in a capable data set of answers; however we considered its default statistical analysis as poor and we attempted to put our hands on it. What follows is the results of our own analysis, demonstrative of the value of our approach, which you may find of interest.

The initial data set is comprised of 253 questions (attributes, all of structured textual nature) and answers from 309 customers’ answers (instances). The question titles are clustered and presented in the following Table 1.

Table 1: Question titles (attributes)

Table 2 presents the characteristics of the attribute that served as the target to our analysis, namely the decision to buy or not the products of the under study brand.

Table 2: Target attribute details

Advanced filtering and prioritization algorithms, fine tuned for the data set under focus, essentially contributed a ranking of the available attributes, in regards to their information value on predicting the target attribute price. The 10 most important attributes are presented in Table 3.

Table 3: Questions of most importance

With valuable inputs like these, we put into use an extended set of advanced machine learning algorithms to finally come up with the underlying structure and patterns that cannot but remain unseen under a typical statistical analysis. We present and discuss a few out of these patterns hereafter, while you may find a more extended selection out of these results in the full report, which is available or download below.

Rule 1: if known_brand_brand= absolutely_agree then brandbougth_brand=yes (88% success)

Rule 1 comes up with the very significant -especially under a marketing perspective- pattern that a customer who is well informed about the under focus brand, is expected to select its products for purchase, with a significant certainty of 88%.

Rule 2:  if most important_char_ endurance= yes & advertising_in_magazines = yes then brandbought_brand= yes (85% success)

Rule 2 reveals that a customer who considers endurance as a significant characteristic and has viewed at least one of the brand’s advertisements in magazines will buy the brand’s products. This rule came up with a success rate of 85% across the data set.

Rule 3: if most important_char_quality=yes & important_char_endurance=yes  & important_char_security=yes then brandbought_brand= yes (74% success)

Rule 3 suggests with a certainty of 74% that a customer who considers quality, endurance and security as important factors that shape her purchase decision, is expected to select the under focus brand, versus the other ones available. Yes, the brand seems to be very strong, at least that is what the data set and the data mining analysis performed indicates.

The full report presents 31 such rules and you may download it here (.pdf). And we do believe that such insights, remaining latent in most of your data sets, consist a solid potential for you to upgrade your inputs to decision making. So, why omit such insights, when you can just mine your knowledge? Send us your data sets right now, and just wait for the very treasures of your data to be revealed!

The first case study to present refers to a survey conducted on customers of the automotive industry. As the feature under focus served the frequency of internet usage, regarding topics related to the industry. 

The initial data set is comprised of 37 questions (attributes, all of structured textual nature) and answers from 319 customers (instances). The questions’ titles are presented in the following Table 1.

Table 1: Question titles (attributes)

Table 2 gives further information on the target variable, namely the internet usage and the possible answers to this question.

Table 2: Target attribute details

Extensive experimentation with some advanced filtering techniques resulted in introducing the most significant questions, in regards to their informational value according to the target question. Table 3 gives the 10 attributes of most importance.

Table 3: Questions of most importance

While the latter Table clearly provides some critical insights on the factors that correlate with the web surfing patterns of customers, the real payoff comes when performing data mining on these inputs. Extensive experimentation was conducted with a bunch of sophisticated machine learning algorithms, which were fine tuned to finally result in extracting rules and patterns like the ones that follow.

Rule 01: If car_owner=yes and carvalue=35-50k then internet_use=frequently

For example, Rule 01 introduces that  a typical owner of a car valued between 35 to 50 thousand euros is expected to use the internet frequently for searching relevant to cars information.

Rule 02: If 16v=no and age=18-25 and spoiler=yes then internet_use=never

On the other hand, Rule 2 suggests, with a 80% certainty, that a young user (aged 18-25) who does understand about spoilers but don’t know about 16v engine features, is not expected to search for relevant to cars information on the internet.

The full report, available for download here (.pdf), finally contributes more than 50 such patterns about the data set under focus, and you may consider it as indicative of our services.

So, if you are interested in gaining such valuable insights of your own data sets, you may learn more about our services here, while we are awaiting your data, starting from now.