RU2480828C1 - Method of predicting target value of events based on unlimited number of characteristics - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: method involves the following steps: using a computer to obtain data on events with indication for each event the target value and an arbitrary set of its numerical characteristics; performing segmentation; creating a predictive model and predicting the target value of the next similar event. The event characteristics are also numbered; their values are sorted; a sequence of cycles of segmenting events is performed; a segment is found for each event; the value of the economic effect of dividing into segments is calculated; the segments are divided based on the best characteristic having the maximum economic effect and its value. To predict the target value of the next similar event, values and characteristics of the similar event are successively compared with values of the best characteristics of segments. The segment with which the event is associated is determined from the value of the best characteristic. The average target value of such a segment serves as the prediction of the next similar event.

EFFECT: high accuracy and reliability of predicting target values of events, shorter and simpler procedure with large volumes of data.

5 cl, 5 dwg

Description

Technical field

The invention relates to computer technology and can be used to automatically conduct targeted campaigns with respect to customers, a forecast of a decrease and increase in the number of consumers of products, a forecast of the results of events for various objects and areas of economic activity.

State of the art

Known prognostic methods based on neural networks, Kohonen maps, decision trees, for which it is common to obtain at the input of the process of modeling a “feature vector” of fixed dimension after the “data preparation” stage (Data Preparation stage, see CRoss Industry Standard Process for Data Mining - CRISP -DM).

However, the real data accumulated by the companies contains a history of events of the objects under investigation with a variable and often a large number of events for each object, and only with large losses of information can be converted into a vector. In addition, each event may contain a number of characteristics that are potentially useful for forecasting.

For example, information on purchasing a product using a discount card simultaneously contains the purchase amount, the exact date, name and category of the product, the amount of the discount, the name and address of the store, the name of the department — up to 100 or more fields. Moreover, completely different characteristics may be significant for predicting the behavior of buyers in relation to specific stocks, for example:

- Amount of purchases with the word 'CHAMPAGNE' in 'GOODS NAME' in the MORNING

- The amount of purchases with the word 'MARTINI' FRIDAY,

- Number of purchases with the word 'PAMPERS' DAY,

- The number of words 'TOP' in the questionnaire,

- The number of visits to the webpage 'SCREWS AND SCREWS', etc.

Thus, both the values of text fields and the words in them, including in combination with numerical fields, can be used to form individual potentially useful characteristics. With an assortment of retail chains of tens and hundreds of thousands of items, the number of potentially useful characteristics of buyers often exceeds a million.

The amount of data of all possible characteristics of the events of millions of customers exceeds the size of the available RAM of modern computers, which forces developers of information systems to "reduce the dimension" of the analysis to several tens of thousands of variables, that is, by an expert method to discard potentially valuable information.

The algorithm of associative rules is also known, which is used to search for frequently encountered combinations of goods in shopping baskets (Association Rule and Quantitative Association Rule Mining among Infrequent Items MDM'07, Ling Zhou, Stephen Yau, 2007, San Jose, California, USA). But this algorithm does not allow using along with the main identifier (for example, the name of the product) other additional information (date and time of purchase, store address, group, subgroup, category and subcategory of the product).

A known method (US, 6839682) predicts the financial behavior of consumers (for example, buyers), which includes obtaining a number of input transactions for many consumers relative to many merchants; definition of at least one trading segment in which each merchant is combined with at least one of the specific trading segments; and for at least one consumer, applying computer input consumer transactions to at least one trading predictive segment model, for each trading definition of a forecasting segment model for a trading segment, the forecast functions between input transactions in the past time interval and financial behavior in the subsequent time interval, to produce for each consumer the predicted behavior in each of at least a subset of the trading segments.

This known method implements an automated forecast of a target indicator (CPU) of events, including computer receiving event data indicating for each event a target indicator and a set of characteristics of this event, segmentation, building a forecast model and forecasting a target indicator for subsequent similar events.

In this method, the event is a purchase, the target is the amount of sales in the trading segment, the characteristic is the vector of the sales amounts of the buyer in other trading segments, and all transactional data about the buyer is preliminarily converted into a revenue vector for the “trading segments”, after which a computer analysis based on this simplified data only.

Thus, in this known method, information about the time of purchases, addresses of stores, retail chains within the “segment”, personal data of buyers and any other available data is not taken into account and is not used for forecasting.

A limitation of the known method is the insufficiently high accuracy and reliability of the forecast of target indicators (CPU) of events, as well as a long time and complex procedure when processing large amounts of data about events.

Disclosure of invention

The problem solved by the invention is the improvement of technical and operational characteristics and the expansion of the forecasting functionality while simplifying its procedure.

The technical result is to increase the accuracy and reliability of the forecast of event target indicators, as well as to reduce time and simplify the procedure for large amounts of observation data on various events by optimizing the distribution of event data by computer memory types and processing algorithms, which makes it possible to fully analyze all the necessary characteristics of events.

Modern computer processors can perform billions of operations per second, but their performance is limited by the speed of memory devices and data storage devices.

For this reason, the invention is based on a complete analysis of all possible characteristics of events, which is made possible by optimizing the distribution of event data by computer memory types and processing algorithms.

The construction of the model is carried out by sequentially constructing the levels of the decision tree by dividing the subset of events related to the tree node (segment) at each level into segments of the next level.

This approach allows for industrial applications to use only the basic initial levels of the decision tree from the constructed model, which give the maximum economic effect, taking into account the costs of their storage and use for hundreds and thousands of promotions, including at points of sale in devices with a limited amount of memory.

To solve the problem with the achievement of the specified technical result in the method, the computer receives event data indicating for each event a target indicator and an arbitrary set of its numerical characteristics, events are numbered by a computer, event characteristics are end-to-end numbering, characteristics values for events are sorted in the order of the characteristic number events - the value of the characteristics of the event, remember them in this order in the computer data storage device with the numbers of events, p after which, in the segmentation cycle, the set of events is sequentially divided into subsets — segments of the next level corresponding to nodes of the decision tree, during which they sequentially read characteristics, their values and event numbers from the data storage device, and for each segment aggregated data is calculated based on target indicators and the amount from the array of events from the computer’s RAM, and if the next characteristic of the event differs from the previous one or its value from the previous calculation t the economic effect of dividing into segments between the previous and next values of the previous characteristic, remember the best characteristic of the event and the value having the maximum economic effect of dividing into segments, after reading the last value of the characteristic from the data storage device, the segments are divided in which the economic effect of dividing into segments of the next level positive, on segments of the next level by the best characteristic and its value, if new segments are detected, then repeat the cycle of segments If there are no new segments, then to predict the target of a subsequent similar event, the characteristics of a similar event are successively compared with the values of the best characteristics, according to which the segment was divided into segments of the next level, and the segment to which the next one belongs is determined by the value of the best characteristic a similar event, the average value of the target indicator of such a segment serves as a forecast for the subsequent similar event.

An additional variant of the method is possible, according to which, when calculating the economic effect of dividing into segments, the economic effect is reduced by the value of the confidence interval of the economic effect, for which additional aggregated data is calculated, in particular, the sum of the squares of the target events. In this case, the maximum pessimistic economic effect is used as a criterion of the best division into segments. The economic effect may be, for example, savings in connection with the refusal to conduct a promotion in a non-target (unprofitable by the average target) segment. The pessimistic economic effect is adjusted downward by the value of the statistical error of the economic effect by calculating the confidence interval for the user-specified required confidence probability.

Variants of the method are also possible when the division of segments is performed by no more than a predetermined number of levels and / or at least a predetermined number of events in each segment of the next level. The limit on the number of events in a segment can be set based on a compromise between the required reliability of the segments and the available amount of data, which may not be enough to satisfy these requirements.

A variant of the method is also possible when the segment is divided into no more than two segments of the next level.

These advantages, as well as features of the present invention are illustrated by the best option for its implementation with reference to the accompanying figures.

List of drawings

Figure 1 depicts a diagram of the formation and processing of data to build a model and forecast the target indicator of the claimed method;

Figure 2 - cycle by level of segmentation in figure 1;

Figure 3 - screen for the formation of events (observations) based on the purchase history of the buyer for the claimed method;

Figure 4 - screen of the result of building a model in the form of a decision tree for conducting a promotion;

Figure 5 - the same as figure 4, for personal recommendations of the film "Pretty Woman" on the data of NetFlix.

The best embodiment of the invention

The claimed method can be implemented using the following algorithm (figure 1 and figure 2) of the computing device during which a binary decision tree is built, that is, each segment is divided into no more than two segments of the next level. In Figs. 1 and 2, the data blocks and the transmission path are highlighted by dashes, solid lines show the data processing blocks, control units, and the path. The blocks highlighted in bold lines in the diagram (figure 2) characterize the main cycle, providing maximum performance.

Based on the event log 1 received by the computer (Fig. 1), block 2 generates event data consisting of a numerical target indicator (CPU) for each event and an arbitrary set of numerical characteristics with textual names. Events are formed in accordance with the business logic necessary for a particular task, for example, the economics of promotions, postal or CMC mailings, distribution of coupons or personal offers. Hereinafter, the concept of "number" refers to real numbers that, as a special case, could be obtained by assigning comparable non-numerical characteristics to successive numerical values, for example, dates, education levels or knowledge of a foreign language from the customer’s questionnaire and so on.

In the process of constructing a model when it enters the processing of each event, it is numbered. Each newly discovered characteristic (end-to-end numbering for all events) is also sequentially numbered by block 3, as a result of which the event may look, for example, like this:

Event No. 928712 Target = -1023.3 Feature №1000200 = 56.4 Feature No. 56223222 = 20.1 ... Feature No. 23001234 = -23112.8

After that, the value of each characteristic for each event is sorted and stored in the data storage device 4 in the order "Characteristic number - characteristic value - event number", for example:

Feature Number Value Event number one 112334.1 899772 one 111222.2 100223 one 110445.0 223344 one 99223.9 887778 one 12899.0 22333 2 56.3 822109 2 -0.1 112223 2 -12.9 122 ... 56223222 23.6 766553 56223222 22.6 122233 56223222 22.5 666554 56223222 20.1 928712 56223223 190.6 956712

The value of the characteristic from the event example for block 3 is highlighted in bold (sorting by value is carried out in descending order).

After that, in the random access memory (RAM) of the computer create block 5 of the array of events, indexed by the event number indicating the target and the segment number.

Next, in the segmentation cycle of block 6, the total values of the aggregated data of the segments of block 7 are calculated (Fig. 2). Aggregated data can be the total number of events, the sum of the targets, the sum of the squares of the targets, and others selected by the developer.

An example of an array of segments and aggregated data about them (block 7):

Field / Segment No. one 2 ... Segment ID R0110 R00 Number of events, total 65422 875092 Current events 56 78921 CPU total 34982.67 187430.0 Current CPU Amount -12932.1 45638.0 The sum of the squares of the CPU total (double precision) 425166667 566033498.0 The sum of the squares of the CPU is current (double precision) 5368123.01 45984023.8 The best pessim. Result 2941.45 98.88 Best Feature, No. 5 87469 The best value of the characteristic -56.12 Absent Updated Yes No

To do this, in block 8 (figure 2) in a cycle to the end of the data from the drive of block 4 is buffered reading of the next record of values of the characteristics. Each record of the characteristic value is checked for the coincidence of the characteristic number and the characteristic value from the previous record in decision block 9. If they do not differ, then the event target obtained from the array of block 5 is immediately used in block 10 to accumulate aggregated data for the segment of block 7 to which the event belongs. For example, a unit is added to the current value of the number of events, an event target value is added to the sum of the segment’s target indicators (CPU), the square of the current CPU is added to the sum of the squares of the CPU, etc. In this case, the segment is marked as updated by setting the flag Updated = "Yes" in block 7.

Blocks 8, 9 and 10 in the process of building the model are used most often than others when performing operations, since the data is sorted by numbers and values of characteristics, and it is highly likely that neighboring records in the drive of block 4 have the same values of characteristics (for example, a thousand buyers bought one each unit of goods and the corresponding characteristic for all such events will be the same and equal to one). For these operations, only a comparison of the data in the RAM of the computer and arithmetic operations by its processor to update the aggregated data is required, which allows to increase the productivity and speed of building the model, despite the fact that the volume of these characteristics for real volumes of events, as a rule, does not fit in the operational memory.

The pessimistic economic effect is used as a criterion for the best division into segments.

The economic effect may be, for example, the savings resulting from the refusal to conduct a promotion in a non-target (unprofitable by the average target) segment, in which case the economic effect will be equal to the sum of the target indicators in the unprofitable segment with a minus sign. Another example of the economic effect is the amount of excess CPU in the daughter segment compared to the average for the parent segment.

Other indicators of the optimal division of the segment, which are not strictly economic, can be used, but in any case, for industrial applicability, they should somehow lead to the calculation of the economic effect, therefore we consider them as a special case of economic criteria.

To obtain a pessimistic economic effect, the economic effect is adjusted downward by the value of the statistical error of the economic effect by calculating the confidence interval for a user-specified confidence probability.

Thus, [Segmentation Result]

= [Pessimistic economic effect] =

[Economic effect] - [Confidence interval of Economic effect]

[Confidence interval of Economic effect] can be calculated based on the standard deviation of the target indicator and the number of events in the segment, taking into account Student's coefficients or other statistical methods using the aggregated data necessary for them and accumulated in block 7, depending on the task assigned to the algorithm developer.

Segmentation is performed if the pessimistic economic effect of it is greater than zero.

If the characteristic number or value is different (as well as after reading the last record from the drive 4), then before updating the segment data, in the cycle of block 11 for all updated segments in block 12 it is checked whether the result of the possible division of this segment into segments of the next level is improved by the value between the differing values of one characteristic, the comparison of which was carried out in block 9, or if a record with a different characteristic was read in block 9, then by the value “Present-Absent” the previous character sticks.

For example, if the previous and next record in drive 4 belong to the same characteristic:

Feature Number Value Observation number 94512 8,456 305612 94512 7,232 2309145

then the division should be carried out according to the value of characteristic No. 94512 between 8.456 and 7.232.

Since there can be no characteristic between the previous and next value of the characteristic with the sorting used, a more precise boundary between the segments of the next level cannot be determined and can be set in this interval. For example, the exact boundary can be defined as the average, rounded to the minimum number of significant digits that keep it within the interval, in this case 8.0.

For the option of calculating the economic result with the rejection of segments with a negative average value of the target indicator, the economic effect of dividing the segment into two segments of the next level by the value of characteristic 94512 equal to 8.0 will be 12932.1 with the following aggregated values in the array of segments of block 7:

Number of events, total 65422 Current events 56 CPU total 34982.67 Current CPU Amount -12932.1

If the result of a possible division of a segment (for simplicity in this example is calculated without a confidence interval) is improved, then the following values can be set in the data of this segment of block 7:

Field Value The best pessim. result 12932.1 Best Feature, No. 94512 The best value of the characteristic 8.0 Updated Yes

If the previous and next record in drive 4 belong to different characteristics:

Feature Number Value Observation number 94512 8,456 305612 94513 96205467,223 2309145

then the division should be carried out according to the value “Present” - “Missing” of characteristic No. 94512, and if the result of a possible division of the segment is also improved to 12932.1, then the following values should be set in the data of this segment of block 7:

Field Value The best pessim. result 12932,1 Best Feature, No. 94512 The best value of the characteristic Is present Updated Yes

In the case of improving the economic result of a possible division of the segment in block 13, the found new values of the result, the best characteristic and its best value are set for the segment.

The calculations of block 12 require more computer processor operations than the cycle of blocks 8, 9, and 10, but are performed less frequently, which eliminates the need to go to the data storage device during the calculations and provides a computational speed comparable with the speed of sequential buffered data reading from the hard disk, there is close to the maximum possible.

After each cycle, according to the segmentation level of block 6 (Fig. 1), in decision block 14 it is checked whether new segments are detected at this level, if they are found, then new segments in block 15 are added to block 16 of the predictive model in the form of a binary decision tree.

After that, the division of segments in which segments of the next level are identified, that is, the assignment of events of these segments in block 5 to new segments is carried out, for which data from drive 4 are read according to the characteristics used for division (included in block model 16), compared with the value according to which the division is performed, and depending on the result of the comparison, the event belongs to the corresponding segment. At the same time, the average values of the target indicator in the segments are calculated, which are indicated for each segment in the model of block 16, after which it is saved for future use.

If no new segments are identified, model building ends.

At the time of forecasting the target indicator (possibly in real time), event data is generated by block 2 according to the same data and rules as for model building, which guarantees the applicability of the constructed model of block 16 for forecasting. To predict the target indicator of such an event in block 17, the corresponding values of its characteristics are sequentially compared with the values of these characteristics in the nodes of the tree of block 16 until the event is assigned to a segment that does not have segments. The average value of the target indicators of events in this segment used to build the model is a forecast of the target indicator of such an event.

Figure 3 shows an example of a potential promotion event “BATON SUNNY 400 GR” (field 30) to a potential customer similar to existing customer No. 5 (field 31), and the procedure for calculating the target indicator (field 32) and characteristics (table 33 )

In this case, the business model of the action consists in the free provision to customers of the promoted (target) product (“sunny loaf”) for testing. The target in field 32 is to estimate the profit from the promotion in relation to a buyer who is similar to an existing one but who has not yet purchased the product, calculated in fields 34 and 35. All the event characteristics that appear on the screen of table 33 are generated by automatically calculating the amount of purchases for different time periods periods with the name of the product contained in the first record 36 in the purchase journal of table 37: “Cottage cheese cake“ CHOCOLATE 340 GR ”, or the words in the name of this product. For other products, the corresponding characteristics are also formed in the purchase table, they are not included on the screen.

To build a model in accordance with the claimed method, from each event, the target indicator in the field 38 and the characteristics of the event are indicated in the form of a table of 33 real values with text names. The user must enter only the parameters of the business model in the sections "Revenues" 39 and "Costs" 40, and also in one of the variants of the method a confidence probability can be indicated.

As a result, the computer built a predictive model (figure 4) in the form of a decision tree. The field 50 “Decision tree - information” displays the presence of 200 events in a segment of 687 in the higher node of the tree (parent segment), while the average economic effect in field 51 in the form of expected profit per contact in the segment is 0.51 rubles, and in the higher the node of the tree is negative and equal to 1.39 rubles. The division into segments was carried out according to the characteristic “Sum of purchases with the word 'Ф / П' in the name”, by the value “greater than 68.71”. The indication 'F / P' in the name of the goods is short for “Fin. Package ”, that is, plastic packaging for low-cost dairy products, thus,“ SUN BATON ”is bought by the same customers who buy low-cost dairy products.

Table 52 “Best Dividing Parameters” shows the aggregated data of potential segmentation according to the characteristics “Sum of purchases of products with a word in the name GR 'on Mondays”. The abbreviation GR 'is used to indicate the weight in grams of packaged goods. At the same time, the number of observations in possible segments of the next level is 137 and 63, respectively, and the average profit in them is +1.39 and -1.41. However, taking into account the standard deviation and Student's coefficient, the pessimistic economic effect of the possible segmentation (RAP) 53 turned out to be negative, as a result of which further division into segments was not performed.

For customers who have not previously purchased the target product, but whose characteristic “Sum of purchases with the word 'Ф / П' in the name” is present and more than 68.71, the average value in this segment, that is, 0.51 rubles per contact, is the forecast of the target indicator - economic effect.

Segmentation can also be controlled by parameter 54 “maximum number of segmentation levels”.

A similar example of the constructed model is a decision tree for personal recommendations of the film “PRETTY WOMAN” on NetFlix data (Fig. 5). The difference between this example is that, as a base for numerical characteristics, instead of the purchase amount, the film is rated by the viewer on a 5-point scale. In this case, the minimum rating at which a recommendation should be made is taken to be 4 and specified in field 60. The best criterion (characteristic and value) in the decision tree node 61 for the recommendations of the PRETTY WOMAN movie at the first segmentation level is the rating of the film DIRTY DANCING »More than 3.5. At the same time, the average forecast rating in field 62 in this segment exceeds the required one, equal to 4, by 0.37, while in the higher node of the tree (parent segment) it is lower than the required one by 0.09 points.

The pessimistic effect of dividing this segment into the next level segments in field 63 is also positive, and the same characteristic of the film “DIRTY DANCING”, but already by the value of 4.5, became the best criterion for dividing.

In both examples, no manual data input or preparation was performed, segmentation was carried out according to all automatically generated characteristics, the total number of which reached hundreds of thousands or more, which is clear from the values in fields 54 and 64 “Decision tree - number of unique parameters + values”.

Thus, in the claimed method for constructing the next level of the forecast model in the form of a binary decision tree, the pessimistic economic effect of dividing each segment into segments of the next level is taken into account, taking into account the confidence interval.

The values of the characteristics of events are stored in sorted form in the data storage device, which allows you to free up computer RAM for targets and segments, as well as through sequential reading of data from the storage device, as quickly as possible to fully exhaust all values of the set of characteristics for all segments of each level of the binary decision tree.

The constructed model can be used for instant forecasting of the target indicator of events, including those that did not participate in its construction. The average value of the target indicator in the segment to which the event will be assigned is used as a forecast by sequentially comparing the characteristics of the event with the values in the nodes of the decision tree.

The statistical methods for calculating the confidence interval, in particular, based on Student's coefficients, as shown by studies on a large amount of data, can guarantee the identification of dependencies in the data for millions of indicators, automating the process of “reducing dimension” by analyzing all calculated characteristics and all their values, at the same time practically excluding the human factor in predicting events and their assessment.

Industrial applicability

The most successful invention is applicable to the formation of targeted offers to target segments of customers on industrial data volumes in various industries.

Claims (5)

1. A method for predicting a target indicator of events, which consists in the fact that the computer receives event data indicating for each event a target indicator and an arbitrary set of its numerical characteristics, segmentation, create a predictive model and predict the target indicator of a subsequent similar event, characterized in that events are numbered using a computer, numbering of event characteristics is carried out end-to-end, the values of characteristics for events are sorted in the order the event characteristic number - the characteristics of the event, they are stored in this order in the computer data storage device with the indication of the event numbers, after which, in the segmentation cycle, the set of events is sequentially divided into subsets — segments of the next level corresponding to nodes of the decision tree, during which the characteristics and their values are sequentially read and event numbers from the data store, and for each segment, aggregated data is calculated based on the targets and the number from the array of events from the operational computer memory, and if the next characteristic of the event differs from the previous one or its value from the previous one, the economic effect of dividing into segments between the previous and next values of the previous characteristic is calculated, the best characteristic of the event and the value having the maximum economic effect of dividing into segments are remembered, after reading the last value of the characteristic from the data storage device, segments are divided in which the economic effect of dividing into segments of the next level is positive, by segment The next level of the best characteristic and its value, if new segments are detected, then repeat the segmentation cycle, if new segments are not found, then to predict the target of a subsequent similar event, the values of the characteristics of a similar event are successively compared with the values of the best characteristics, according to which the division was made segments into segments of the next level, and by the value of the best characteristic, determine the segment to which the subsequent similar event belongs, The value of the target indicator of such a segment serves as a forecast for the subsequent similar event. 2. The method according to claim 1, characterized in that when calculating the economic effect of dividing into segments, the economic effect is reduced by the value of the confidence interval of the economic effect. 3. The method according to claim 1, characterized in that the division of the segments perform no more than a given number of levels. 4. The method according to claim 1, characterized in that the division of the segments is performed provided that each segment of the next level contains at least a predetermined number of events. 5. The method according to claim 1, characterized in that the sequential division of the set of events into subsets - segments of the next level is performed by dividing a segment of the previous level into two segments of the next level.

Images