BACKGROUND

[0001 ] Machine learning may be used for purposes of predicting a future outcome. Machine learning regression addresses the problem of predicting a continuous value for a given variable conditioned on one or multiple other variable values.

BRIEF DESCRIPTION OF THE DRAWINGS

[0002] Fig 1 is a schematic diagram of a system to provide one or multiple recommendations to change an outcome predicted by a regression model according to an example implementation.

[0003] Figs. 2 and 3 are flow diagrams depicting techniques to provide one or multiple recommendations to change an outcome predicted by a regression model according to example implementations.

[0004] Fig. 4 is a table illustrating example recommendations provided by the recommendation engine of Fig. 1 according to an example implementation.

[0005] Fig. 5 is a schematic diagram of a physical machine according to an example implementation.

DETAILED DESCRIPTION

[0006] In general, regression analysis (linear regression analysis, for example) may be applied to a set of attributes (herein called a "record") to predict a continuous value based on the attributes. In machine learning, a regression model may be trained based on labeled training data (i.e., data producing observed, or labeled, continuous values) for purposes of predicting continuous values. As a more specific example, a regression model to predict house values may be trained with labeled house sales data. Using such a regression model, the regression model may be applied to a set of attributes (a square footage, a location, the number of floors, the year of construction, and so forth) to predict a price (a continuous value) for a house described by these attributes.

[0007] A challenge with using such a regression model is that the mere prediction of a value does not provide insight of what attributes affect to be changed, or the extent of how certain attributes affect the value. For example, in a house search, a particular user may use a regression model for purposes of predicting a house price for a given set of attributes. The user may be, for example, a potential home buyer that has a targeted house budgeted price (i.e., a target price accompanied by a range in which the actual price may vary). The user may be willing to change some of the attributes (location, number of bedrooms, and so forth), but the user may be unaware of which attributes to change to derive the desired change, as it may not be readily apparent which parameters can be changed to identify houses in the targeted price range.

[0008] Systems and techniques are discussed herein in which recommendations are provided for purposes of changing a continuous value that is predicted by a regression model. In particular, in accordance with example implementations, a system may receive a record that represents various input values that are provided to a regression model, such as parameters associated with a house search, for example. A regression model may be applied to this record for purposes of predicting a continuous value, such as a house price. The continuous value predicted by the regression model may not, however, be close enough to a target outcome for a given user. In accordance with example systems and techniques that are discussed herein, the original record is modified based on the training records (i.e., labeled data) that are used to train the regression model, and the regression model is applied to these modified training records to predict corresponding outcomes. The outcomes are compared to a target outcome, and based on these comparisons, one or multiple modified training records are identified and presented to the user. Accordingly, the user is provided recommendations for changing the outcome of the original record to provide a result that is close enough (within a predetermined range, as further described herein) to the target outcome.

[0009] In accordance with example implementations, a "target outcome" refers to a given target value and an associated error range. For example, for the case of predicting the price or value of a house, the target outcome may, in general, refer to a fixed price, in association with a range (a range of ± 5 percent, for example) about the price point.

[0010] As a more specific example, Fig. 1 depicts a system 100 in accordance with example implementations. In general, the system 100 includes a prediction engine 104, which receives a record 150 describing a set of attributes. A regression model 160, when applied to the attributes described by the record 150, may predict an outcome for the record 150. However, the predicted outcome may not be suitable, when compared to a target outcome 168. As such, the system 100 provides one or multiple recommendations 170 for purposes of guiding modification of the record 150 to produce a result within the target outcome 168.

[001 1 ] More specifically, in accordance with example implementations, the prediction engine 104 accesses (reads data from a memory, for example) a set 164 of one or multiple training records 166. In this regard, a given training record 166 corresponds to labeled training data and describes a set of attributes having a known outcome (i.e., an outcome that has been observed). The regression model 160 is trained against the training records 166 of the set 164. The prediction engine 104 modifies the training records 166 of the set 164 to produce a candidate set 108 of modified records 1 10. In this manner, in accordance with example implementations, a modified record 1 10 is derived from a combination of the record 150 and one of the training records 166 of the set 164. The record 150 may have one or multiple attributes that are "modifiable" and one or more attributes that are not modifiable. More specifically, a modifiable attribute refers to an attribute that may be changed for purposes of steering the outcome toward the target outcome 168. For the case of house values or prices, a record 150 may represent such attributes as house location and number of rooms, but, as an example, the number of floors may not be modifiable/adjustable. For example, for the particular application, a house search may be performed by an elderly person who does not desire a house with more than one floor. As another example, the record 150 may represent attributes used by a user for purposes of searching for a certain salary (relating to employment). As such, the record may have such modifiable attributes as a level of education or a job location. However, attributes, such as gender, ethnicity, and so forth may not be modifiable.

[0012] In accordance with example implementations, the prediction engine 104 constructs a given modified record 1 10 by combining the non-modifiable attributes from the record 150 with the modifiable attributes from a given training record 166.

[0013] Not all of the modified records 1 10 of the candidate set 108 may, in accordance with some implementations, be suitable to produce an outcome that satisfies the target outcome 168. As such, the system 100 may include a filtering engine 125, which filters the modified records 1 10 of the candidate set 108 to produce a filtered subset 120 of the modified records 1 10. The outcomes 1 12 of the modified records 1 10 of the filtered subset 120 are consistent with the target outcome 168. However, some of the modified records 1 10 of the filtered subset 120 produce outcomes that are better or closer to the target outcome than others.

[0014] In accordance with example implementations, a recommendation engine 130 of the system 100 ranks the modified records 1 10 of the filtered subset 120 to produce one or multiple recommendations 170. In accordance with some

implementations, the recommendations 170 are a ranked list of the modified records 1 10, detailing the changes for producing the closest match to the target value.

[0015] Among its other features, in accordance with some implementations, the prediction engine 104, the filtering engine 125 and the recommendation engine 130 may include hardware processors 106, 126 and 132, respectively. Moreover, as further described herein, in accordance with example implementations, the system 100 may receive various other user inputs, such as the number of top results to provide as the recommendations 170, a designation of which attributes are modifiable, and so forth.

[0016] Thus, referring to Fig. 2, in accordance with example implementations, a technique 200 includes applying (block 204) a regression model to a given record to predict an outcome and based at least in part on the application of the regression model, a determination is made (decision block 208) whether to recommend changes to the record based on the target outcome and the predicted outcome. If so, pursuant to the technique 200, the given record is modified (block 212) based on one or multiple training records to provide one or multiple modified records. Pursuant to the technique 200, the regression model may then be applied to the modified record(s) to predict one or multiple outcomes for the modified record(s) pursuant to block 216. One or multiple recommendations may then be provided (block 220) based at least in part on a target outcome and the outcome(s) for the modified record (s).

[0017] As a more specific example, in accordance with some example

implementations, the system 100 may perform a technique 300 that is illustrated in Fig. 3. Referring to Fig. 3 in conjunction with Fig. 1 , the system 100 applies (block 310) regression prediction based at least in part on the regression model 160, the target outcome 168 and the record 150. As shown in Fig. 3, the target outcome 168 may include a target value 302, as well as an error range 304, defining a range about the target value 302.

[0018] Based on the result of the regression prediction, the prediction engine 104 determines (decision block 314) whether the absolute value of the difference between the predicted value and the target value is less than or equal to the error range. If so, then the current record 150 and its associated values are sufficient, and the technique 300 terminates. Otherwise, prediction engine 104 accesses the next record from the training set 164, pursuant to block 320, and modifies (block 326), the records, as determined by data 322 indicating the attributes that may be modified.

[0019] Next, the prediction engine 104 applies (block 330) regression-based prediction based on the regression model 160 and determines (decision block 334) whether the absolute value of the difference between the predicted value and the target value is within the error range. If not, the prediction engine 104 discards the modified record 1 10, and control proceeds to block 320 to begin another iteration to produce another modified record 1 10. If, however, the prediction engine 104 determines (decision block 334) that the absolute value of the difference between the predicted and target values are within the error range, then the prediction engine 104 calculates a record distance, pursuant to block 338. In this manner, a "record distance" refers to a distance between attributes of the original record 150 and the modified record 1 10. Next, the prediction engine 104 determines (decision block 340) whether all of the training records have been processed. If not, control returns to block 320.

[0020] Otherwise, the reduced set of modified records has been produced, and the recommendation engine 130 ranks the remaining modified records for purposes of providing the recommendations 170. More specifically, in accordance with example implementations, the recommendation engine 130 receives user data 354, which designates a number "K" of recommendations. The recommendation engine 130 then determines (block 350) the "top K" records whose record distances to the record 150 are the closest. As illustrated in Fig. 3, in accordance with example implementations, the recommendations 170 include the selected modified records 1 10 of the set 120, the associated distance(s) between the predicted and target outcomes 360, the distance(s) between the original and modified records 362 and the predicted outcome(s) 364.

[0021 ] Fig. 4 is an illustration of recommendations 400 produced, where the record represents to a set of attributes related to a house search, and the outcome is a desired price or value. As shown in Fig. 4, the recommendations 400 may be formatted in a table having such columns as a ranked number column 410, a price column 412, a location column 414, a rooms column 416, a floors column 420, a size column 422, an outcome distance column 424 and a record distance column 426. As illustrated in Fig. 4, in accordance with some implementations, each row of the table corresponds to one of the modified records, and the records are ranked based on the record distances. In other words, the top row assigned number one corresponds to the modified record having the smallest record distance (i.e., the modified record that is the closest to the original record). The record distance refers to the degree of change between the original record 150 and the corresponding record shown in the table. It is noted that the highest ranked modified record may be associated with an outcome distance that is larger than a lower ranked modified record.

[0022] Referring to Fig. 5, in accordance with some implementations, the system 100 includes a physical machine 500. The physical machine 500 is an actual machine that is made up of actual hardware 504 and machine executable instructions 520, or "software."

[0023] The hardware 504 may include, for example, one or multiple central processing units (CPUs) 510, a memory 514, a network interface 512 and a display 516. Depending on the particular implementation, the processors 106, 126 and 132 (see Fig. 1 ) may be formed from one or multiple CPUs 510, one or multiple processing cores of the CPU(s) 510, and so forth. In general, the memory 514 may be a non-transitory storage medium that may store data, program instructions, data structures, and so forth, depending on the particular implementation. The memory 514 may be formed from semiconductor storage devices, phase change memory devices, magnetic storage devices, optical storage devices, memristors, and so forth, as well as one or more of these device types.

[0024] In accordance with example implementations, the memory 514 may store program instructions that when executed by the CPU(s) 510 cause the CPU(s) 510 to form one or more components 530 of the system 100.

[0025] Therefore, in accordance with example implementations, one or multiple components of the system 100 (the prediction engine 104, the filtering engine 125, the recommendation engine 130, and so forth, for example) may be software components, i.e., components formed by at least one processor executing machine executable instructions, or software. In further example implementations, one or multiple of the components 530 may be formed from dedicated hardware (one or more integrated circuits that contain logic configured to perform any of the functions of the components 530, as described herein). Thus, the components 530 may take on many different forms and may be based on software and/or hardware, depending on the particular implementation.

[0026] The memory 514 may store other data, in accordance with example implementations, such as data 540 representing the record 150; regression model data 542 representing parameters of the regression model 160; data 544

representing the training record set 164; data 545 representing the modified records 1 10; data 546 representing the outcome distances 360; data 548 representing the record distances 362; and so forth. [0027] The machine executable instructions 520 may contain other instructions that are executed by one or multiple CPU(s) 510 for purposes of forming one or more additional software components, such as an operating system 526, one or multiple applications 528, and so forth.

[0028] While the present invention has been described with respect to a limited number of embodiments, those skilled in the art, having the benefit of this disclosure, will appreciate numerous modifications and variations therefrom. It is intended that the appended claims cover all such modifications and variations as fall within the true spirit and scope of this present invention.