CN110634027A - First-order user refined loss prediction method based on transfer learning - Google Patents

Abstract

The invention provides a first-order user refined loss prediction method based on transfer learning, which comprises the following steps: (1) data acquisition (2) data organization (3) transfer learning. According to the method, similar user groups based on target domain expansion in a source domain are focused through a Lookalike algorithm, the similar expansion user groups of the target domain and the target domain are used as training samples, a C5 decision tree is adopted, model building training is conducted by introducing a model-based transfer learning idea, the obtained model prediction accuracy is compared with the prediction accuracy obtained by calculating without adopting any transfer learning model building, and the prediction effect is improved to a certain extent.

Description

First-order user refined loss prediction method based on transfer learning

Technical Field

The invention relates to a user refined loss prediction method, in particular to a user refined loss prediction method based on transfer learning.

Background

In increasingly competitive situations, it is far more valuable to retain an old customer than attract a new customer. Effectively predicting and saving lost users and having very important significance for the survival and development of enterprises. In the case that the problem of loss prediction of the multi-order user is solved, it is very important to effectively predict the loss of the first-order user. In view of the fact that the first-purchase user has a short board with too few records of the same user behavior, the traditional prediction method has the defect of low accuracy.

Disclosure of Invention

The invention aims to provide a first-order user refined loss prediction method based on transfer learning so as to improve the prediction effect.

The invention adopts the following technical scheme:

a first-order user refined loss prediction method based on transfer learning is characterized by comprising the following steps:

(1) data acquisition

Obtaining an order record, taking a user who has ordered in a preset time period as an observation object, wherein a new user who has just come in a latest life cycle needs to be filtered out from the observation object, then taking all historical order behaviors of the observation object as initial training data,

among the observed subjects, attrition users are defined as: no subscription for >90 days, return 1: and (4) loss, otherwise, 0: the mixture is left for a long time,

(2) data organization

Calculating a data index user ID forming the following format by taking the daily consumption of the user as the minimum consumption granularity

Consumption No on the nth day.	Correlation index	Simple derivation index 1	Timing derivative index 2
				n＝1	√	√	-
n＞1	√	√	√

(3) Transfer learning

Defining all order record sets with the order times of single users being more than 1 as a source domain Dim_S(ii) a Data set with order number 1 as target domain Dim_TGradually completing the migration of the source domain knowledge to the target domain through the following steps:

3-1: dim is the dimension set_S∩Dim_TWherein: dim_SFor the computational dimension of the source domain data, Dim_TComputing dimensions for the target domain data;

3-2: if the treatment of removing the negative migration is needed, directly switching to the step 3-3;

sample set Smp ═ Data_S∪Data_TWherein: data_sFor source domain Data, Data_TIs target domain data;

3-3: the target sample set is processed to remove the negative migration, Smp _ Tr ═ Data_T∪Data_{T_Similar}，

Wherein: data_{T_Similar}For the Data of the source domain and the target domain_TA similar data set of (a);

Smp＝＝Smp_Tr

3-4: dividing a sample set Smp into training sets Train_SmpAccounting for 80 percent; and Test set Test_SmpAccounting for 20 percent;

3-5: by PCA, Dim is mapped to a low-dimensional subspace, i.e.: dim ═ > DimP, where: DimP is a factor that Dim maps in a low dimensional subspace;

3-6: the model DimP > Dim', where: dim' is an important factor relative to target Y;

3-7: model training: in the training set Train_SmpIn the method, Dim 'is used as an input, Train _ Y is used as an output, and C5 decision tree training modeling is used to obtain Train _ Y ═ F (Dim'); then at Test_SmpEvaluating the obtained model;

3-8: model obtained by training using Step7, for Data_TF (Dim') > DataT _ Y, DataT _ Y referring to Data_TAnd mapping the obtained output result according to the function F (Dim').

Further, the invention discloses a migration learning-based first-order user refined loss prediction method, which is characterized by comprising the following steps:

the specific steps of step 3-3 are as follows:

data of target domain Data set (i.e. record ordered for the first time)_TAs seed Data, expansion is carried out on the basis of seed Data, i.e. the record Data most similar to the seed Data is screened out from the source domain_{S_Similar}Obtaining Smp _ Tr ═ Data_T∪Data_{S_Similar}The Smp _ Tr is the data set after the negative migration processing,

the migration sample consists of two parts:

Smp_Tr＝Data_T∪Data_{S_Similar}

wherein: data_TIs a target domain data set;

Data_{S_Similar}for data sets in a source domain similar to a target data source

3-3-1: for each target domain Data_TThe seed data of (2) are respectively calculated according to the following formulas (1) and (2), and the similarity between the seed data and each member in the source domain data is calculated for each target domain member;

pearson similarity

Sim＝1-R (1)

Distance

The similarity distance is calculated by calculating the obtained PCA mapping factor in the data set of the combination of the source domain and the target domain by using the formula,

wherein:

the PCA factor in the source domain is: fs ═ Fs₁，Fs₂，......Fs_nn is the number of factors of the PCA mapping factor, FS: a set of PCA factors in the source domain comprising n factors Fs1, Fs 2. FT ═ Ft₁，Ft₂，......Ft_nn is the number of factors for the PCA mapping factor FT: a set of PCA factors in the target domain, comprising n factors Ft1, Ft 2.. Ftn,

3-3-2: r, D, respectively taking Sim 0.2and D1 as values, and obtaining a target domain data record ID set, wherein Sim is a similarity measure calculated by formula (1);

3-3-3: removing the duplication of the target data record ID set obtained in the Step2, and adding the record ID in the source domain;

3-3-4: the record ID obtained at Step3 is deduplicated again, and the record set corresponding to the record ID set obtained later is Data_T∪Data_{T_Similar}I.e. to form a processed extended data set Smp _ Tr.

Advantageous effects of the invention

According to the method for predicting the fine loss of the first-time ordered user based on the transfer learning, similar user groups based on the target domain expansion in a source domain are focused through a Lookalike algorithm, the similar expansion user groups of the target domain and the target domain are used as training samples, a C5 decision tree is adopted, model-based transfer learning idea modeling training is introduced, the prediction accuracy of an obtained model is compared with the prediction accuracy obtained by adopting no transfer learning modeling calculation, and the prediction effect is improved to a certain extent.

Drawings

FIG. 1 is a schematic diagram of the partitioning of a data source.

Detailed Description

The technical scheme of the invention is described in detail by combining the specific examples as follows:

1. model training method based on model transfer learning

(1) Data acquisition

For the convenience of understanding of the description of the method, all historical airline ticket ordering records provided by a international airline ticket search platform G are taken as an example and are described herein

Referring to fig. 1, the users who have over-ordered behaviors during 2015.12.01-2018.3.11 through the platform, historical order data during 2015.12.01-2018.10.11 are used as observation objects, new users who have just come in the latest life cycle need to be filtered out of the observation objects, then all historical order behaviors of the observation objects are used as initial training data,

in this embodiment, the time from 2018.03.11 to 2018.10.11 in the reservation map is for the user to go through a complete lifecycle N. In other embodiments, the complete lifecycle should be determined separately from the industry-traffic lifecycles of the different products. For airline ticket subscriptions, the user's lifecycle N is 180 days.

On this basis, the attrition user defines: no subscription for >90 days, (1: attrition), otherwise (0: retention).

And during this time period the new user's behavior is not used as a training sample.

(2) Data organization

And calculating data indexes forming the following format by taking the daily consumption of the user as the minimum consumption granularity:

table 1: data index

Consumption No on the nth day.	Correlation index	Simple derivation index 1	Timing derivative index 2
				n＝1	√	√	—
n>1	√	√	√

The specific indexes are summarized and detailed in table 2:

table 2: specific data index

(3) Migration learning

Defining all order record sets with the order times of single users being more than 1 as a source domain Dim_S(ii) a Data set with order number 1 as target domain Dim_TGradually completing the migration of the source domain knowledge to the target domain through the following steps:

step 1: dim is the dimension set_S∩Dim_TWherein: dim_SFor the computational dimension of the source domain data, Dim_TComputing dimensions for the target domain data;

step 2: if the processing of removing the negative migration is needed, directly switching to Step 3;

sample set Smp ═ Data_S∪Data_TWherein: data_sFor source domain Data, Data_TIs target domain data;

step 3: the target sample set is processed to remove the negative migration, Smp _ Tr ═ Data_T∪Data_{T_Similar}，

Wherein: data_{T_Similar}For the Data of the source domain and the target domain_TA similar data set of (a);

the specific calculation steps are detailed as follows: the method part of the Lookalike relieving negative migration is described, and the forming process of the Smp _ Tr of the data processing part is described;

Smp＝Smp_Tr；

step 4: dividing a sample set Smp into training sets Train_Smp(80%), and Test set Test_Smp(20％)；

Step 5: by PCA, Dim is mapped to a low-dimensional subspace, i.e.: dim ═ > DimP, where: DimP is a factor that Dim maps in a low dimensional subspace;

step 6: the model DimP > Dim', where: dim' is an important factor relative to target Y;

step 7: model training in training set Train_SmpIn the method, Dim 'is used as an input, Train _ Y is used as an output, and C5 decision tree training modeling is used to obtain Train _ Y ═ F (Dim'); then at Test_SmpEvaluating the obtained model;

step 8: model obtained by training using Step7, for Data_TF (Dim') > DataT _ Y; wherein DataT _ Y refers to Data_TAnd mapping the obtained output result according to the function F (Dim').

2. Lookalike-based negative migration mitigation method

Data of target domain Data set (i.e. record ordered for the first time)_TAs seed Data, expansion is carried out on the basis of seed Data, i.e. the record Data most similar to the seed Data is screened out from the source domain_{S_Similar}Obtaining Smp _ Tr ═ Data_T∪Data_{S_Similar}. Smp _ Tr is the data set after the negative migration processing.

The migration sample consists of two parts:

Smp_Tr＝Data_T∪Data_{S_Similar}

wherein: data_TIs a target domain data set;

Data_{S_Similar}for data sets in a source domain similar to a target data source

Step 1: for each target domain Data_TThe seed data of (2) are calculated respectively according to the following formulas (1) and (2), and the seed data and the source domain data are calculated for each target domain memberSimilarity between members of (a);

pearson similarity

Sim＝1-R (1)

Distance

The similarity distance is calculated using the above formula by calculating the resulting PCA mapping factor in the dataset of the source domain and target domain combinations.

Wherein: ft: is the set of PCA factors in the target domain, Fs: is a set of PCA factors in the source domain, the PCA factors in the source domain being: FS ═ Fs₁，Fs₂，......Fs_nn is the number of factors of the PCA mapping factor, FS: a set of PCA factors in the source domain comprising n factors Fs1, Fs 2. The PCA factor in the target domain is: FT ═ Ft₁，Ft₂，......Ft_nn is the number of factors of the PCA mapping factor, FT: the set of PCA factors in the target domain consists of n factors Ft1, Ft 2.

Step 2: r, D, respectively taking Sim 0.2and D1, and obtaining a target domain data record ID set;

step 3: removing the duplication of the target data record ID set obtained in the Step2, and adding the record ID in the source domain;

step 4: the record ID obtained at Step3 is deduplicated again, and the record set corresponding to the record ID set obtained later is Data_T∪Data_{T_Similar}I.e. to form a processed extended data set Smp _ Tr.

3. And (3) according to the new first-purchase user, predicting whether the user will lose by using the trained model and predicting the new first-purchase user by using the model constructed in the parts 1 and 2 to obtain a loss prediction conclusion.

Table 3: accuracy of each prediction method

The present invention is not limited to the above description or what is shown in the drawings, and modifications and changes may be made within the scope of the present invention without departing from the gist of the present invention.

Claims (2)

1. A first-order user refined loss prediction method based on transfer learning is characterized by comprising the following steps:

(1) data acquisition

Obtaining an order record, taking a user who has ordered in a preset time period as an observation object, wherein a new user who has just come in a latest life cycle needs to be filtered out from the observation object, then taking all historical order behaviors of the observation object as initial training data,

among the observed subjects, attrition users are defined as: no subscription for >90 days, return 1: run-off, otherwise return 0: the mixture is left for a long time,

(2) data organization

Calculating a data index user ID forming the following format by taking the daily consumption of the user as the minimum consumption granularity

Consumption No on the nth day. Correlation index Simple derivation index 1 Timing derivative index 2 n＝1 √ √ — n>1 √ √ √

(3) Transfer learning

Number of single user subscriptions>1 as the source domain Dim_S(ii) a Data set with order number 1 as target domain Dim_TGradually completing the migration of the source domain knowledge to the target domain through the following steps:

3-1: dim is the dimension set_S∩Dim_TWherein: dim_SFor the computational dimension of the source domain data, Dim_TComputing dimensions for the target domain data;

3-2: if the treatment of removing the negative migration is needed, directly switching to the step 3-3;

sample set Smp ═ Data_S∪Data_TWherein: data_SFor source domain Data, Data_TIs target domain data;

3-3: the target sample set is processed to remove the negative migration, Smp _ Tr ═ Data_T∪Data_{T_Similar}，

Wherein: data_{T_Similar}For the Data of the source domain and the target domain_TThe similar data set of (a) is,

Smp＝Smp_Tr；

3-4: dividing a sample set Smp into training sets Train_SmpAccount for 80%, and Test set Test_Smp20 percent of the total weight of the composition;

3-5: by PCA, Dim is mapped to a low-dimensional subspace, i.e.: dim ═ DimP, where: DimP is a factor that Dim maps in a low dimensional subspace;

3-6: the model DimP ═ > Dim', where: dim' is an important factor relative to target Y;

3-7: model training: in the training set Train_SmpIn the method, Dim' is used as input and Train _ Y is used as output, so thatTraining and modeling by using a C5 decision tree to obtain Train _ Y ═ F (Dim'); then at Test_SmpEvaluating the obtained model;

3-8: model obtained by training using Step7, for Data_TIs predicted, F (Dim') ═ is predicted>DataT _ Y, DataT _ Y refers to Data_TAnd mapping the obtained output result according to the function F (Dim').

2. The migration learning based first-order user refinement churn prediction method of claim 1, wherein:

the specific steps of step 3-3 are as follows:

data of target domain Data set (i.e. record ordered for the first time)_TAs seed Data, expansion is carried out on the basis of seed Data, i.e. the record Data most similar to the seed Data is screened out from the source domain_{S_Similar}Obtaining Smp _ Tr ═ Data_T∪Data_{S_Similar}The Smp _ Tr is the data set after the negative migration processing,

the migration sample consists of two parts:

Smp_Tr＝Data_T∪Data_{S_Similar}

wherein: data_TIs a target domain data set;

Data_{S_Similar}for data sets in a source domain similar to a target data source

3-3-1: for each target domain Data_TThe seed data of (2) are respectively calculated according to the following formulas (1) and (2), and the similarity between the seed data and each member in the source domain data is calculated for each target domain member;

pearson similarity

Sim＝1–R (1)

Distance

The similarity distance is calculated by calculating the obtained PCA mapping factor in the data set of the combination of the source domain and the target domain by using the formula,

wherein:

the PCA factor in the source domain is: FS ═ Fs₁，Fs₂，……，Fs_nn is the number of factors of the PCA mapping factor, FS: the set of PCA factors in the source domain, comprising n factors Fs1, Fs2, … … Fsn,

the PCA factor in the target domain is: FT ═ Ft₁，Ft₂，......Ft_nn is the number of factors for the PCA mapping factor FT: the set of PCA factors in the target domain, comprising n factors Ft1, Ft2, … … Ftn,

3-3-2: r, D, respectively taking Sim (0.2 and D (1) as values, and obtaining a target domain data record ID set, wherein Sim is a similarity measure calculated by formula (1);

3-3-3: removing the duplication of the target data record ID set obtained in the Step2, and adding the record ID in the source domain;

3-3-4: the record ID obtained at Step3 is deduplicated again, and the record set corresponding to the record ID set obtained later is Data_T∪Data_{T_Similar}I.e. to form a processed extended data set Smp _ Tr.

Images