Regression methods go for instigating exact prescient models that relate the estimation of an objective or dependent numeric variable to the estimations of an arrangement of independent variables. In the most recent decade or somewhere in the vicinity, most machine learning investigations of regression and also most best in class regression methods are worried about initiating piecewise models. These methods parcel the preparation set and induce a straightforward model in each part. Piecewise models are ordinarily founded on basic consistent and linear models (as in regression and model trees and MARS models ) or polynomials . In this investigation, we assess the ease of use and execution of straightforward models dependent on polynomial equations on standard regression errands. Regardless of the way that piecewise regression models beat basic ones in machine learning literature, we guarantee here that straightforward polynomial equations can be productively induced and have aggressive execution with piecewise models. To approach the regression assignment productively, we create Ciper1, a method for inciting polynomial equations. The method performs heuristic pursuit through the space of hopeful polynomial equations. The inquiry heuristic joins model level of fit to the information with model intricacy. We assess the execution of Ciper on thirteen standard regression informational collections from two open vaults. Experimental assessment incorporates examination with standard regression methods for actuating linear and piecewise linear models, executed inside Weka information mining suite. We thought about various methods as far as prescient error and unpredictability of the induced models. We additionally performed observational inclination difference deterioration of the prescient error on a portion of the informational collections. In regression modeling to depict the connection between variables generally a polynomial regression model is utilized. Polynomials are extremely adaptable and regularly utilized when there is no hypothetical model accessible. To acquire a polynomial regression model, which depicts the relations in information adequately well and does not overfit, commonly the subset selection

Prema Kumari1* Dr. Aswini Kumar2 2

prescient execution of the regression model. Before the subset selection venture, with the end goal to enhance the hopeful model space, a limited arrangement of predefined premise capacities is made and, from that point onward, the subset selection is performed with the premise capacities. The premise capacities ordinarily are characterized as results of the first variables each raised to some request (a positive number). Consequently the objective is to discover a subset that amplifies the prescient execution of the subsequent regression model. With the end goal to discover the subset some sort of hunt must be performed. The most straightforward hunt procedure is the thorough inquiry. Albeit thorough inquiry assurances to locate the best subset, it needs exponential runtime and accordingly is illogical much of the time. Another class, called heuristic pursuit methods, productively navigate the space of subsets, by including and erasing the premise capacities, and utilize an assessment work that coordinates the inquiry into territories of expanded execution. The ordinary models of heuristic pursuit methods are the Forward Selection (otherwise called Sequential Forward Selection, SFS) and the Backward Elimination (otherwise called Sequential Backward Selection, SBS) . SFS begins with a vacant arrangement of chosen premise capacities and iteratively adds the capacity prompting the most elevated execution increment to the arrangement of chosen capacities, until the point that the execution can't be improved any further by including a solitary capacity. SBS begins with the entire capacity set also, iteratively expels the capacity whose expulsion yields the maximal execution increment. The methodology of subset selection expect that the picked settled full arrangement of predefined premise capacities contains a subset which is adequate to portray the objective connection adequately well. Anyway we contend that much of the time the essential arrangement of premise capacities isn't known and should be speculated or picked by involvement (e.g. by determining the maximal request of the subsequent polynomial). Much of the time that means a non-trifling (and long) experimentation process that may create sets of capacities, working with which, in a few issues of moderate dimensionality, may turn out to be computationally excessively requesting notwithstanding for the heuristic pursuit methods (as it will be exhibited in the exact examinations of this investigation). A more advantageous and productive way is let the modeling method itself develop the premise capacities fundamental for making the regression model with sufficient prescient execution. development of premise capacities utilizing heuristic inquiry in the subsequent unending competitor model space. The methodology does not require the client to predefine the arrangement of premise capacities for model creation. We likewise list five of the conceivable refinement administrators, which enable the hunt to discover better models and in addition to do it all the more proficiently, and present an example of the methodology – another regression modeling method called Sequential Floating Forward Polynomial Construction (SFFPC), which is named comparatively to the subset selection method Sequential Floating Forward Selection (SFFS) on which the inquiry system of SFFPC is based. The fairly as of late proposed method Constrained Induction of Polynomial Equations for Regression (CIPER) likewise might be seen as an occasion of the methodology. Anyway it has a few disadvantages with respect to the arrangement of the refinement administrators utilized, which we endeavored to wipe out in our proposed polynomial regression modeling method. To assess the considered methodology in type of our proposed polynomial regression modeling method, SFFPC, we exactly contrast it with two surely understood subset selection methods SFS and SFFS, and in addition to CIPER both on fake and certifiable information. CIPER was produced with regards to inductive databases and imperative based information mining. As of now stated, CIPER utilizes just the first two confusion administrators and, and also SFS, just hunts forward, be that as it may, as a remuneration for the low fanning component, Beam Search procedure is utilized. Here is a diagram of CIPER's settings: Initial express: the state with one capacity that relates to the catch term (this capacity remains in the model consistently and isn't permitted to be altered or erased). Refinement administrators: the first two inconvenience administrators. Pursuit technique: Beam Search. End condition: when no further enhancements are conceivable. Assessment measure: the adjusted two-arrange MDL. In this part, we present ClPER, a machine learning algorithm for discovering polynomial equations. ClPER is a heuristic algorithm that seeks through the space of polynomial equations and discovers one (or a few equations) that fulfill a given arrangement of requirements and have an ideal estimation of the given heuristic capacity. It utilizes shaft pursuit to heuristically look through the space of conceivable equations for ones that fit the information best. ClPER

Prema Kumari1* Dr. Aswini Kumar2 2

regression and machine learning. A polynomial over variables can be written in the form: \Where and are constants, and We also defined the length of P as the size of P as the degree of a term as and the degree of P as Verifiably, polynomial models are among the most every now and again utilized observational models for fitting capacities. They are prominent in light of the fact that they have a basic frame; they have surely understood and comprehended properties; they have a moderate adaptability of shapes; and they are computationally simple to utilize. Additionally, every constant capacity characterized on an interim [a,b] can be consistently approximated as nearly as wanted by a polynomial capacity (the Weierstrass estimation hypothesis) made them considerably more attractive. The result of the Weierstrass hypothesis for polynomial regression is that there are numerous polynomials that can give solid match to a given limited dataset. One approach to adapt to this issue is to oblige the space of applicant polynomial equations. ClPER makes utilization of two classes of imperatives for this reason. • Language imperatives are given as a sub/super polynomial of the polynomial we are searching for. They limit the structure of the conceivable polynomial models. Formally, a polynomial P is a sub-polynomial of a polynomial Q if for each term X in P. there exists a term Y in Q. with the end goal that the level of each factor in Y is bigger or rise to than the level of a similar variable in X. For instance, xy2 is a sub-polynomial of x2y4z- • Complexity requirements limit the multifaceted nature of a polynomial. With them we indicate the most extreme length, greatest degree, and greatest number of terms in the polynomial. For instance, one may be keen on equations of degree at most 3 with at most 4 terms.

A refinement administrator executes a capacity that takes as input an equation structure and creates another equation structure by altering the bygone one. The first ClPER refinement administrator builds the length of an equation by one, either by including a first variable (Figure 1). Beginning with the least difficult equation (a consistent), and iteratively applying this refinement administrator, every single polynomial equation can be produced. On account of sub/super polynomial requirements, we begin with an underlying polynomial equation: By applying the refinement administrator all super/sub polynomials of the underlying equation can be produced. Given an articulation x + y, we can refine it in two different ways. First, we can incorporate another linear term yielding x+y + z. Second, we can supplant a current term in the articulation (e.g. x) by increasing it with a variable (e.g, z), yielding another articulation.

Accept we measure the unpredictability of the polynomial equation as its length. The refine¬ment administrator expands the multifaceted nature of the equation by one. either by including another linear term or by adding a variable to a current term. First, a subjective linear (first degree) term can be added to the present equation. Unique consideration is taken that the recently presented term is not quite the same as every one of the terms in the present equation. Second, we can build the multifaceted nature by adding a variable to one of the terms. Once more, care ought to be taken that the subsequent term is unique in relation to the various terms in the present equation. Note that the refinements of a given polynomial are super-polynomials of it. They are insignificant refinements as in they increment its intricacy by one unit.

Prema Kumari1* Dr. Aswini Kumar2 2

To evaluate equations, we calculate different measures of the degree of fit of an equation to a given dataset. Two measures commonly used for regression are the mean squared error (MSE) and the multiple correlation coefficient R. Various other types of prediction error measures are often used. These include mean absolute error, maximum absolute error, and root mean square error. Most of these are well-known from statistics. In the machine learning literature, the measure RE. defined as, where is the variance of the dependent variable, is often used to evaluate the performance of regression approaches. The normalization with the variance allows for comparisons of performance across different datasets. The original ClPER implementation uses the AdHoc heuristic, defined as follows where P is the polynomial equation being assessed, len(P) is its length, MSE(P) is its mean squared error, and m is the quantity of preparing models. The second term of the AdHoc heuristic capacity estimates the level of attack of an offered equation to the information and the first term presents a punishment for the multifaceted nature of the equation. With this punishment, the AdHoc heuristic capacity presents an inclination toward less complex equations.

ClPER looks through the space of conceivable equations by utilizing a bar seek algorithm. Anytime, it keeps up an arrangement of b most ideal equations (the pillar) that fulfill the forced imperatives. The output of ClPER comprises of the last substance of its shaft, i.e., the best polynomial equations. The best dimension layout of the ClPER algorithm is appeared in Table 1. First, the shaft is introduced either with the least complex polynomial equation P = C (where C is steady), or with a client determined negligible polynomial. In each inquiry emphasis, an arrangement of new, more mind boggling polynomials is created from the polynomials in the shaft by utilizing a refinement administrator. The coefficients previously the terms in a polynomial are fitted by utilizing the method of minimum squares. For every one of the produced polynomials, the estimation of the AdHoc heuristic is ascertained. Toward the finish of the cycle, the equations with littlest heuristic qualities are held in the bar. likewise stop if the substance of the shaft is unaltered in the last emphasis. Such a circumstance happens when each polynomial created in the last cycle has a more regrettable heuristic score gauge than the polynomials as of now in the pillar.

A few enhancements for fitting the coefficients of the produced polynomial structure can be presented. The information are spoken to as a grid X. where the quantity of lines is the quantity of cases, and the quantity of sections is the quantity of terms (r) in addition to one (the first segment is loaded up with ones). The minimum squares gauge for the coefficients of the equation is where y is the vector of values we are trying to predict2. In Equation 19, the multiplication is computationally expensive because of the large number of rows. Let and be terms in equation A. and terms in equation B, such that Then the appropriate elements in the matrices and are equal. We store all generated elements from the matrices We reuse them for calculating the matrices of the subsequently generated polynomials. This optimization considerably lowers the computational cost of ClPER at the expense of some memory.

The experimental assessment by Todorovski et al. demonstrates that ClPER outflanks linear regression and stepwise linear regression on a large portion of the trial datasets considered. The stepwise regression methods gain exactness with expanding the maximal

Prema Kumari1* Dr. Aswini Kumar2 2

significantly more mind boggling models and tend to over-fit the preparation information. The consequences of stepwise regression demonstrate that further enhancement is conceivable in the event that we increment the degree further: in any case, this is obstinate for vast datasets. Stepwise regression will in general create an ever increasing number of complex models as d increments, and the execution of stepwise polynomial regression is extremely delicate to the estimation of d. choosing the ideal d esteem is a nontrivial issue, since it can vary starting with one dataset then onto the next. For viable reasons, the selection would be guided by computational multifaceted nature issues (the quantity of precomputed higher degree terms). The general exactness of ClPER is practically identical to the precision of regression trees. The relative exactness enhancement is higher for littler datasets. The last give inadequate statistical support to various fractional models got from parts of the dataset (as in regression trees and model trees), yet adequate support for a solitary equation over the whole dataset (as in ClPER).

In this area, we present a few enhancements of the first ClPER algorithm. We first present the enhanced refinement administrator, a noteworthy enhancement over the former one. We next present the enhanced MDL heuristic, trailed by a heuristic dependent on a cross-approval gauge of forecast error. At long last, we portray the treatment of clear cut properties.

Adding a term to a linear (in the parameters) equation dependably diminishes its error (in any event on preparing information). Be that as it may, supplanting a term with a more mind boggling adaptation of it (duplicated by a variable) doesn't really diminish the error of the equation. On the off chance that we add y to x, yielding x+y, we will lessen the error of the equation. Notwithstanding, in the event that we supplant x with xy, the substitution may really expand the error of the equation. This has persuaded us to alter the refinement administrator in ClPER. Other than the two kinds of refinements considered in the first form of ClPER. we present a third one. We take a term in the equation, influence a duplicate, to increase the duplicate with another variable and add the item back to the equation. For instance, with the new administrator x + y can be refined to x + y + xy by replicating the term x, term xy to the equation. The old refinement administrator dependably builds the multifaceted nature of an equation by one. As outlined in Figure 2, the new refinement administrator can expand the unpredictability of an equation impressively. Along these lines, we present an additional disentanglement venture in ClPER. For each equation in the shaft, we take a stab at expelling every one of its terms: if this yields an equation with a superior heuristic esteem, when contrasted with the first equation, we include the recently framed rearranged equation to the pillar. The additional rearrangements step is the last sort of refinements that we use inside the new refinement administrator. Note that the new refinement administrator has now four kinds of refinements: the first two, i.e., including a solitary variable term (e.g.,) and multiplying a term (e.g.,), the third refinement type that adds a multiplied term (e.g.,) and the last type of simplification refinement (e.g.,).

The fanning component of the new refinement administrator relies upon the quantity of variables V and the quantity of terms in the present equation r. The upper bound of the expanding factor is since there are at most conceivable refinements that expansion the quantity of terms r and at generally V. r conceivable refinements that expansion just the degree yet not the quantity of terms. The fractional multifaceted nature of the additional rearrangements step is just the quantity of terms of the equation O(r). The improvement step is rehashed until there are better equations created with it. In theory, if this progression is rehashed r times, we can

Prema Kumari1* Dr. Aswini Kumar2 2

In this area, we have figured the spreading component of the old refinement administrator as O(V.r). As should be obvious, the old and the new refinement administrators have comparative spreading factors i.e., O(V.r) and separately.

Encoding the Polynomial Structure - In order to encode the structure of a polynomial, we follow the refined MDL1 approach . We first partition the space of candidate models into subgroups of models with equal complexity c. A particular model can be then encoded using (log stands for the binary logarithm) bits, where denotes the number of models in the class In the case of polynomials, we partition the space of candidate polynomial structures into classes at several levels. At the highest level, we group together the candidate polynomials with the same length and the same size m. Recall that for a polynomial the size m is defined as the number of terms m, and the lengthis defined as (note also that ). We refer to these classes as For example G(l,l), contains polynomial structures with a single term, which has to be linear (length 1). Similarly, G(l,2) contains structures with a single term of second degree, while G(2,4) contains structures with two terms, and the degrees of the terms can be up to four (since the length is 4). At the second level, we partition each in subclasses with fixed term degrees All polynomials in this subclass have m terms with degrees Note that For example, G(2,4) can be broken into two subclasses of and G'(2,2). The first subclass contains polynomial structures with one linear term and one third-degree term, while the second subclass contains polynomial structures with two second-degree terms. Now we have to calculate how many sub-classes there are in a single class and also calculate how many polynomial structures there are in each class. The number of structures in each , can be easily calculated using a procedure roughly depicted in Figure 3. Given the degree of the first term , we have to choose variables from the set , where variables can appear in the selection more than once. Thus, the number of elements from a set of n elements. This number equals . Continuing the same reasoning for all m terms, we find the number of possible structures in to be . However, if there are several values that are equal, we will encounter the same term many times, which means that the above formula over-estimates the number of possible structures. The remedy is to divide the number with the factorial of repetitions observed in the tuple. For example, when dealing with the case we have to divide the above product with since a fifth degree term appears twice and a second degree term appears three times Note also that each multiplicative term decreases by 1 for each degree value repetition (see Figure 3).

Having the number of equation structures in each class, we now turn to the problem of calculating the number of classes within each G(m J). The size of G grows according to the recursive formula The first additive term corresponds to the cases when the classes contain linear terms (there is an with value 1), while the second corresponds to the cases when all terms in the classes have a degree at least 2 (all ). In the first case, when removing the linear term, we obtain polynomials with m — 1 terms and length . In the second case, we can remove one variable from each of the terms, which leads to polynomials with the same number of terms (m) and length Take for example G(2,4), mentioned above: G(l,3)

Prema Kumari1* Dr. Aswini Kumar2 2

up to degree two, and G(2,2) contains structures with two terms, up to degree 2. Figure 4 depicts the relationship between the G and G' classes of polynomial structures.

The recursive formula always leads to one of the two simple G classes, which contain a single subclass. The first is with a single subclass where The second simple class is with a single subclass where 1 is repeated times. In this case, Finally, note that the recursive formula above can lead to the illegal situation with in such cases Now, having this partitioning and the number of polynomials in each partition, we can decompose the code for each candidate polynomial into four components. First, we have to encode its length and for this we need bits (the second double logarithm term is necessary, since we do not know the magnitude of / in advance). Second, we encode the number of terms m. for which we need bits (remember that ). Third, we can identify a particular class within the class using bits. Finally, we identify the specific polynomial structure within using bits. Putting these four components together gives us the final formula: for the number of bits necessary to encode the polynomial structure. for any binary attribute X. X = X. for an arbitrary degree d. Taking this into account, the number of structures in each subclass has to be adjusted. Note also that we need an additional bit to encode the fact that a polynomial structure contains a binary attribute. From here on. we assume that a polynomial structure contains one or more binary attributes. Consequently, there are polynomials in the class that are equiv alent because of the binary attributes. Thus this number should be recalculated. First, we need one bit of information to identify if the polynomial has binary attributes. If it does not, then the number is calculated just like before. Let the number of binary attributes be If the polynomial has binary attributes, then for every monomial, we need to encode the number of binary attributes that that monomial contains. For this we need bits of information (as the number of binary attributes is smaller then the degree of the monomial). Let thei-th monomial contain binary attributes. The number of monomials that have degree. The number of monomials that have degree for which every variable is contained only once, is . The number of monomials that have degree and have binary attributes is The number of possible structures in is If there are several values that are equal, we will encounter the same term many times. The remedy is to divide the number with the factorial of repetitions observed in the tuple. The Complete Scheme - The complexity of the polynomial structure, explained above, plus the stochastic complexity of the linear regression model3 gives the total complexity of the model: A representation of the stochastic multifaceted nature of the linear model and the unpredictability of the polynomial structure is given in Table 2. For this models we utilize the auto-cost dataset. This informational index has a solitary target and 33 traits, which incorporate curb Weight, width, length and alternate properties that show up in Table 2. On account of a multi-target model4, the structure of the polynomial does not change, the unpredictability

Prema Kumari1* Dr. Aswini Kumar2 2

whole the stochastic complexities of the linear models for every one of the objectives with the multifaceted nature of the structure they yield the aggregate unpredictability of the multi-target model. A representation of the stochastic multifaceted nature of the linear models and the unpredictability of the polynomial structure for the multi-target case is given in Table 3. For this model we utilize the sigmeareal dataset . This dataset has 2 targets and 8 properties, which incorporate X.Y. point, visualAngle, and minDistance.

As an option in contrast to the MDL conspire depicted above, we propose to use as a heuristic the error of a polynomial equation as evaluated on inconspicuous information. The strategy is fundamentally the same as the method for cross-approval. We split the train set into 10 sections. We assemble a model on 9 sections and we utilize the tenth part for approval. Let the squared relative error on the 9 sections for a given model be reTrain2 and the squared relative error on the tenth part be reTest2. For each model, we keep the tuple (reTrain2, reTest2). The heuristic estimation of the model is max (reTrain2, reTest2) (the littler the better). We alter the shaft seek technique in ClPER so that a (tyke) model produced from this model (utilizing the refinement administrator) can enter the pillar just if its heuristic esteem is littler than min (reTrain2, reTest2).

The components needed to calculate L according to Equation 20, i.e., and as well as those needed to calculate W according to Equation 16, i.e., and arc also given. The number of independent, Table 3: The stochastic complexity of the linear model for the first target - 2W1, for the second target - 2W2, the complexity of the polynomial structure L and the total complexity of the multitarget polynomial model MDL for several models learned on the sigmareal dataset. While we could consider just the error on the approval set, the ClPER look technique creates numerous new models from a solitary model. The likelihood that some of them will have a littler error on the approval set than the parent model, just by happenstance, is high. To evade over-fitting we likewise consider the error on the train set. The error on the train set ought to be littler than the error on the approval set. In the event that this isn't the situation, we won't create refined models from this model: The refinement administrator won't think about this polynomial later on ages. We can in any case over-fit, however the likelihood of this would be littler.

After we manufacture 10 models, leaving every one of the ten sections out, we have to make the last model. We normal every one of the models to make one last model. The last model can be seen as troupe of polynomial models. The way toward developing the gathering is represented in Figure 5.

Prema Kumari1* Dr. Aswini Kumar2 2

tuple where t is the quantity of targets. Since we utilize relative error, the more awful estimation of the qualities in the heuristic is at 1 and greater. The best estimation of the heuristic is at 0. We will allude to this heuristic as CV heuristic. The best dimension diagram of the CV ClPER algorithm. First the Data is part into a few folds. For every F of the folds, first the pillar Q is introduced either with the most straightforward polynomial equation P = C. or then again with a client indicated polynomial. In each pursuit cycle, an arrangement of new, more perplexing polynomials is created from the polynomials in the shaft by utilizing a refinement administrator. The coefficients previously the terms in a polynomial are fitted on the train information F.train of the overlay F. For every one of the created polynomials T, we compute the estimation of the heuristic (T.ErrorOnTrain.T.ErrorOnValidation). Toward the finish of the emphasis, the equations with the littlest heuristic qualities are held in the pillar. The scan assessment for this overlap F stops when the refinement administrator cannot create any new equations. It can likewise stop if the substance of the shaft is unaltered in the last cycle. Such a circumstance happens when each polynomial created in the last emphasis has a more regrettable CV heuristic gauge than the polynomials as of now in the shaft. The best equation e from the bar Q is attached to the arrangement of best equations E. One such equation is created for every one of the folds. The normal of this equations is the last polynomial equation R.

The first ClPER can just deal with numeric characteristics. To take into consideration discrete (ostensible) properties, we utilize a method that first proselytes the discrete ascribes to parallel traits. For this, we utilize the overparametrized method. Tabic 4: A best dimension layout of the CVCiper algorithm. Q is the arrangement of best b equations (the shaft) and Qr is the set. of refined equations. S is the set. of ten overlap that the Data is part into. Each overlap F is a tuple (Train, Validation) where F.Train is utilized for both preparing and approval, while equation. For each of these tuples, a polynomial equation is produced. E is the arrangement of these equations. The subsequent polynomial equation R is the normal of all equations in E.

Let the values of a discrete attribute X be where k stands for the number of different values the attribute X has. We replace the discrete attribute X with k binary attributes in the transformed dataset. Each binary attribute X, is defined as meaning: The original ClPER handles binary attributes as numeric. Note however, that each binary attribute X has the property that Xd =X for an arbitrary degree d. Thus, in the new ClPER we modified the search procedure to consider only the first degree of the binary variables, since all the higher degrees are equivalent to the first degree. The first ClPER handles twofold properties as numeric. Note notwithstanding, that every parallel quality X has the property that Xd =X for a subjective degree d. In this manner, in the new ClPER we

Prema Kumari1* Dr. Aswini Kumar2 2

degrees are proportional to the first degree.

This investigation presents Ciper, a method for proficient induction of polynomial equations that can be utilized as prescient regression models. Ciper utilizes heuristic shaft seek through the space of competitor polynomial equations. The pursuit depends on a refinement administrator with low stretching component that makes it significantly more appropriate for polynomial regression contrasted with much complex refinement administrators utilized in stepwise regression methods and MARS. Assessment of Ciper on various standard prescient regression errands demonstrates that it is better than linear regression and stepwise regression methods and in addition regression trees. Ciper gives off an impression of being aggressive to model trees as well. The intricacy of the induced polynomials, as far as number of parameters, is much lower than the multifaceted nature of piecewise models.

1. Brazdil, P., Soares, C., da Costa, J. (2003). Ranking learning algorithms: Using IBL and meta-learning on accuracy and time results. Machine Learning 50(3), pp. 251–277 2. E. Frank and I. H. Witten (1999). Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. Morgan Kaufmann, San Mateo, CA, 1999. 3. Ghahramani, Z. (2004). Unsupervised learning. In: Advanced Lectures on Machine Learning. Pp. 72–112 (Springer, New York, NY, 2004). 4. Kumar, K.; Alsaleh, M. (1996). Application of Hankel matrices in polynomial regression. Appl. Math. Comput. 1996, 77, pp. 205–211. 5. L. Todorovski, S. D_zeroski, and P. Ljubi_c. Discovery of polynomial equations for regression. In Proceedings of the Sixth International Multi-Conference Information Society (Volume A), pages 151{154, Ljubljana, Slovenia, 2003. Jozef Stefan Institute. 6. L. Todorovski, S. Dˇzeroski, and P. Ljubiˇc. (2003). Discovery of polynomial equations for regression. In Proceedings of the Sixth Conference on Information Society (Intelligent and Computer Systems Volume), pages 151–154. Joˇzef Stefan Institute, Ljubljana, Slovenia, 2003. Methods, Abstract and Applied Analysis, ID 648047, 2014 8. S. Dˇzeroski, L. Todorovski, and P. Ljubiˇc. (2003). Inductive databases of polynomial equations. In Proceedings of the Second International Workshop on Knowledge Discovery in Inductive Databases (at ECML/PKDD-2003), pages 28–43. Rudjer Boˇskovi´c Institute, Zagreb, Croatia, 2003. 9. Todorovski L., Ljubic P., Dzeroski S. (2004). ―Inducing polynomial equations for regression‖ // In: Lecture notes in computer science, Lecture notes in artificial intelligence, 3201, Springer, Berlin, pp. 441-452. 10. Y. Wang and I. H. Witten (1997). Induction of model trees for predicting continuous classes. In The Proceedings of the Poster Papers of the Eighth European Conference on Machine Learning, pages 128{137, University of Economics, Faculty of Informatics and Statistics, Prague, 1997.

Research Scholar, OPJS University, Churu, Rajasthan