I Like C4.5 (j48 at Weka software)

@Henrique, that is for decision tree analysis. Where a node is allowed to have only two successors. The structure I am referring to is not restricted in terms of number of successors or predecessors. 

I have no idea; I don't understand your question. I have no idea what you mean by "efficiency." I don't know what you mean by "knowledge structure" but the example of ITA on Wikipedia is about scaling items (especially their "extremity"). I don't know why (or how) anyone would apply C4.5 or any method of classification to items of a survey.

But I know a lot about scaling items. Educational and social scientists have been studying methods for scaling questionnaires for over 100 years and we have some excellent solutions. Classical test theory (CTT) or item response theory (IRT) would be good choices for scaling items.

IRT is a family of related robust statistical models about how observed item responses are linked to the latent construct(s) being measured. There are a large number of IRT models. For example, if the 7-item German nationality scale in the Wikipedia article had been analyzed with the IRT 2PL model, here's what you would have gotten: (1) statistical and heuristic tests of model-data fit for each item; (2) inteval extremity scale values for each item (as opposed to the extremely rough "pseudo ordering" obtained by ITA); (3) an indication of the association of each item with the underlying construct of German nationality. Even the simpler CTT would give you #2 and #3.

Also, the ITA analysis required dichotomization. Using IRT we could have applied the graded response model (GRM), the generalized graded response model (GGRM), or the generalized partial credit model (GPCM) to the Likert responses without dichotomization. In this case, we could have chosen to model the Likert responses either as if (a) all the thresholds were constant across items (i.e., the difference between, say, "Very important" and "Important" were the same across all items) or (b) allowed these thresholds to vary by item (i.e., "Very important" and "Important" etc. were allowed to mean different things across the items). And we would gotten a fourth result: (4) the relative extremity/thresholds of these response points. CTT also works fine with Likert data. Clearly, psychometric scaling methods are more applicable to Likert data than is ITA.

The JSS paper on the ITA software (and the Wikipedia article) uses an example of a West German nationality scale.  In the attached spreadsheet, I did a quick CTT analysis of that scale using the same dataset. I think the CTT analysis is clearer than the IITA or CITA analyses. The ordering is clearer and the corrected item-total correlation (CITC) shows that items 5 and 6 (the most and least important items) are the weakest. The other items have CITC values between 0.5 and 0.7, these two had values near 0.40. Not shown in the attached item-level data, the scale has a coefficient alpha estimate of reliability of 0.80, which is quite adequate, particularly for a seven-item scale.  CTT was being used over a century ago and was formalized in the 1960's. The statistics are extremely simple to compute.  It is incomprehensible to me why someone would use ITA over CTT in analyzing a scale, unless it was because they were unaware of CTT.

ITA bears some resemblance to a technology that was popularized in the 1950's called "scaleogram analysis." It's ancient history now, but I think it lost out to methods like IRT and Rasch analysis because it assumed a deterministic model, whereas IRT, CTT, etc. assume a stochastic model (sometimes responses are not as they should be: people mark things incorrectly; bright students make mistakes; poor students make  lucky guesses; etc.).  I think ITA shares this deterministic approach.  I think that means that it cannot make a proper accounting for measurement error (but that's speculation).

BTW, the only reasonable way to analyze data is to apply both empirical and theoretical methods. But for situations where you truly have no idea about the structure of the data, I'd recommend factor analysis, but it only works on ordinal data.

Educational psychologists are also working on models for inferring knowledge states. For example, can we analyze algebra item responses to see what misunderstandings a student has, so we can correct them? Turns out it's hard to do this.  For example, see: https://www.ets.org/Media/Research/pdf/RR-92-57.pdf

Finally, if you want to produce a way to predict one variable from responses to a questionnaire (or, really, any other data) there are many methods. Multiple regression predicts a continuous dependent variable (DV), logistic regression predicts a binary DV, linear discriminant functions (LDF) "classifies" for a nominal DV. I did a paper in the 2000's comparing C4.5, neural networks, and LDF for predicting RIASEC job codes from personality survey data. In cross-validation, LDF won because both C4.5 and ANN's overfit the data.

But if you'd like, there are a number of methods that fall under the class of "empirical keying" that optimize a composite for predicting a ordinal DV.