CN117635182A - Target recommendation method and device - Google Patents

Abstract

The disclosure relates to the technical field of computers and provides a target recommendation method and device. The method comprises the following steps: acquiring user project data, wherein the user project data comprises user characteristics, historical behavior characteristics and interaction characteristics relative to a current project of a current user; inputting user item data into a preset target recommendation model to obtain recommendation data of a current user relative to the current item, wherein the target recommendation model is obtained by carrying out model fusion on at least two pre-established pre-selected models through a ranking weighting strategy, and the at least two pre-selected models are established according to target values of target recommendation. According to the technical scheme, the recommendation effect of the target recommendation system can be improved.

Description

Target recommendation method and device

Technical Field

The disclosure relates to the field of computer technology, and in particular relates to a target recommendation method and device.

Background

In the related art, a marketing business system of an enterprise may accumulate a large amount of online behavior data of users. With the expansion of the user group, the method is limited by the number and the energy of workers, and when the behavior data on the user line is mined, the mined value behaviors are limited only by means of business experience and basic data analysis.

Meanwhile, with the development of the service, the APP (application program) can be updated in an iterative manner, APP pages and behavior buried points can be changed, and historical user data, project data and online data can be different. The target recommendation system related to the marketing business of the enterprise can use the target recommendation model to find potential users with strong purchase intention from the user group, and because the target recommendation model needs to leave off-line evaluation time before being on line, the difference between on-line data and data used by model training after the target recommendation model is on line can be larger, so that the recommendation effect of the target recommendation model and the target recommendation system is rapidly reduced.

How to improve the recommendation effect of the target recommendation system is a technical problem to be solved currently.

Disclosure of Invention

In view of the above, the embodiments of the present disclosure provide a target recommendation method, apparatus, electronic device, and readable storage medium, so as to solve the technical problem in the prior art that the recommendation effect of the target recommendation system is reduced.

In a first aspect of an embodiment of the present disclosure, there is provided a target recommendation method, including: acquiring user project data, wherein the user project data comprises user characteristics, historical behavior characteristics and interaction characteristics relative to a current project of a current user; and inputting the user item data into a preset target recommendation model to obtain recommendation data of the current user relative to the current item, wherein the target recommendation model is obtained by carrying out model fusion on at least two pre-established pre-selected models through a ranking weighting strategy, and the at least two pre-selected models are established according to target values of target recommendation.

In a second aspect of the embodiments of the present disclosure, there is provided a target recommendation apparatus, including: the system comprises an acquisition module, a storage module and a display module, wherein the acquisition module is used for acquiring user project data, and the user project data comprises user characteristics, historical behavior characteristics and interaction characteristics relative to a current project of a current user; and the recommendation module is used for inputting the user item data into a preset target recommendation model to obtain the recommendation data of the current user relative to the current item, wherein the target recommendation model is obtained by carrying out model fusion on at least two pre-established pre-selected models through a ranking weighting strategy, and the at least two pre-selected models are established according to target values of target recommendation.

In a third aspect of the disclosed embodiments, an electronic device is provided, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the steps of the above method when executing the computer program.

In a fourth aspect of the disclosed embodiments, there is provided a readable storage medium storing a computer program which, when executed by a processor, implements the steps of the above method.

Compared with the prior art, the embodiment of the disclosure has the beneficial effects that: according to the technical scheme, the pre-selection model is built in advance according to the target value of the target recommendation model, and model fusion is conducted on the pre-selection model through a ranking weighting strategy to obtain the target recommendation model, so that generalization capability and stability of the model are improved, and recommendation effect of a recommendation system is improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings that are required for the embodiments or the description of the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present disclosure, and other drawings may be obtained according to these drawings without inventive effort for a person of ordinary skill in the art.

Fig. 1 is a schematic flow chart of a target recommendation method according to an embodiment of the disclosure;

FIG. 2 is a flowchart of a method for fusing target recommendation models according to an embodiment of the present disclosure;

FIG. 3 is a schematic diagram of a positive and negative sample sampling provided by an embodiment of the present disclosure;

FIG. 4 is a flowchart of another objective recommendation method according to an embodiment of the present disclosure;

Fig. 5 is a schematic structural diagram of a target recommendation device according to an embodiment of the present disclosure;

fig. 6 is a schematic structural diagram of an electronic device according to an embodiment of the present disclosure.

Detailed Description

In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular system configurations, techniques, etc. in order to provide a thorough understanding of the disclosed embodiments. However, it will be apparent to one skilled in the art that the present disclosure may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present disclosure with unnecessary detail.

With the development of internet technology, online business activities are becoming more and more popular, and various business applications accumulate a large amount of user behavior data. In this context, the business recommendation system, i.e., the target recommendation system, becomes a key technology for analyzing user data and mining potential purchase will, thereby improving conversion rate and profitability. The business opportunity recommendation system can effectively find potential users with strong purchase will in a huge user group, and can improve screening and identification capacity of business opportunities and user conversion efficiency so as to overcome the limitation caused by only relying on business experience and basic data analysis.

There are some commercial intelligent solutions and recommendation systems on the market that make target recommendations, which are mainly implemented based on traditional machine learning methods and simple data processing techniques, such as linear regression, logistic regression, decision trees, etc. However, under the conditions of APP iterative update, page and behavior buried point change and the like, prediction accuracy and stability of the business opportunity recommendation system cannot be guaranteed, and differences among historical user data, project data and online data have a large influence on model effects.

Specifically, a conventional machine learning-based business recommendation system provides personalized recommendations for potential customers by feature extraction and model training of user behavior data. According to the technical scheme, data preprocessing is performed through technologies such as cleaning, integration and standardization of original user data, and in the feature extraction process, useful features are extracted from the original data according to service requirements and data characteristics. And then, training a machine learning model to conduct business prediction by using the extracted characteristics and the historical data, and generating a personalized recommendation list for the user by applying the trained model.

In the above data processing and feature extraction processes, the prior art generally adopts a fixed feature processing method, and lacks feature diversity for different models, so that the models may not be able to sufficiently capture useful information in the data. In terms of model training, the prior art generally only uses a single machine learning algorithm for training, and lack of diversity may lead to insufficient generalization capability and stability of the model. In addition, the technical scheme in the prior art cannot adapt to challenges of iterative updating and data change of an application program, and the prediction performance and the recommendation effect of the model are easily affected.

In order to solve the above problems, the embodiments of the present disclosure provide a target recommendation scheme, so as to ensure that in the practical application of a model after being online, the model can adapt to the difference between online data and training data, keep higher prediction performance and recommendation effect, and avoid rapid decline of the model effect.

The object recommendation method and apparatus according to the embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.

Fig. 1 is a flowchart of a target recommendation method according to an embodiment of the present disclosure. The method provided by the embodiments of the present disclosure may be performed by any electronic device, such as a terminal or server, having computer processing capabilities. As shown in fig. 1, the target recommendation method includes:

step S101, user item data is acquired, where the user item data includes user characteristics of the current user, historical behavior characteristics, and interaction characteristics with respect to the current item.

Step S102, inputting user item data into a preset target recommendation model to obtain recommendation data of a current user relative to the current item, wherein the target recommendation model is obtained by carrying out model fusion on at least two pre-established pre-selected models through a ranking weighting strategy, and the at least two pre-selected models are established according to target values of target recommendation.

As shown in fig. 1, in the application reasoning process of the target recommendation model provided in the embodiment of the present disclosure, model training needs to be performed to obtain a pre-selected model and a pre-selected model set before the target recommendation model is applied, and the pre-selected models in the pre-selected model set are fused to obtain the target recommendation model.

In the disclosed embodiment, the original model of the pre-selected model is selected based on a target value, which may be, but is not limited to, the number of target customers visited in the future 14 days or the number of target customers visited in the future 28 days.

From the perspective of model fusion, modeling using different target values can improve overall predictive performance by fusing the predicted results of multiple models. The client behavior and the house buying will under different time scales may have different, and the output of the two models is fused by using a weighted average equal fusion strategy, so that the prediction results of the two models can be combined, and the more accurate prediction results are obtained.

From the perspective of the business value that may be generated, modeling using different target values may provide different prediction results for different business needs, thereby improving the practicality of the prediction results. For example, if the goal is to convert the customer to actual visit as soon as possible, then the future 14 days of visit can be used as the target value for modeling, thereby obtaining more accurate prediction results and improving the conversion rate. If the long-term value of the customer is more focused, then whether future 28 days of visit is used as the target value modeling can be used, so that the long-term conversion potential of the customer is predicted better, and the life cycle value of the customer is improved.

In embodiments of the present disclosure, at least two pre-selected models may be trained according to the following training method: and iterating the original model by using the training data until the evaluation function value of the original model is not increased or reaches the set iteration times, so as to obtain the target recommended model.

Specifically, PR AUC (Precision-Recall Area Under Curve, area under Precision-Recall curve) can be used as an evaluation function, which evaluates the performance of a model by plotting Precision-Recall curves under different thresholds and calculating the area under the curves, and which focuses more on the predictive effect of positive samples when evaluating a data set with unbalanced distribution of positive and negative samples, i.e., an unbalanced data set, relative to ROC AUC (Receiver Operating Characteristic Area Under Curve, area under subject working characteristics). In practical applications, there is a greater need to care about the predictive effect of the model on the sample, i.e., the high potential customer. PR AUC is more concerned about accuracy and recall, and the two indexes are directly related to the prediction effect of the positive sample, so that the PR AUC can more accurately evaluate the performance of the model on the whole number of users, can reflect the overall ordering effect of the model, namely whether the model can accurately predict the preference of the users, and has important significance for examining the ordering capability and recommendation effect of the model. In contrast, ROC AUC considers both true and false positive rates, where false positive rates are greatly affected by the number of negative samples, resulting in a lower degree of concern for the predictive effect of ROC AUC on positive samples.

In embodiments of the present disclosure, the original model of the pre-selected model may be any of the following: random forest, lightGBM (Light Gradient Boosting Machine, lightweight gradient hoist), XGBoost (eXtreme Gradient Boosting, extreme gradient lift tree), catBoost (Categorical Boosting, category feature lift), neural network TabNet for table data. Wherein RandomForest, lightGBM, XGBoost is a tree model and TabNet is a deep learning model. The different original models have respective advantages in terms of processing different types of features and data structures, so that the model obtained by fusing the different pre-selected models can fully utilize the advantages of each model. For example, random forest is suitable for processing nonlinear relations, XGBoost and LightGBM train faster on a large-scale data set, catoost has stronger processing capacity for class features, and TabNet is suitable for high-dimensional sparse data. Through model fusion, the advantages can be fully exerted, and the overall recommendation effect is improved.

Further, random forest may be replaced with other tree models, such as, but not limited to Extremely Randomized Trees (extremely random tree), adaBoost (Adaptive Boosting, adaptive lifting), gradientBoosting, and the like. TabNet can be exchanged for other deep learning models such as NODE (Neural Oblivious Decision Ensembles, neural forgetting decision set), CNN (Convolutional Neural Networks, convolutional neural network), RNN (Recurrent Neural Network ), etc. to capture more complex feature interactions.

In the disclosed embodiment, the original training data is first acquired before training the pre-selected model. The features contained in the raw training data may be: visit feature, sign-up feature, project feature, subscriber line feature, property membership tag feature, user liveness feature, project liveness feature, time interval feature, and policy information. The user line features and the property member tag features are user features, the user activity features, the project activity features and the time interval features are historical behavior features, and the visit features, the subscription features, the project features and the policy information are interaction features.

From the raw training data, a training data set may be obtained. In this case, the multi-day data set may be selected to form a training data set including three days of data from the beginning of the month, the middle of the month, and the end of the month, respectively. Because the data distribution at different time points may have certain difference, the overall distribution deviation can be reduced by splicing the multi-day data sets, so that the generalization capability and stability of the model are improved. In addition, the training set and the testing set time slices of the training data set obtained through splicing are not overlapped, and two different continuous months are selected, so that the generalization capability of the model can be effectively improved, and overfitting is avoided. Because the time slices of the training set and the test set do not overlap, the model does not rely too much on specific data features in the training set, but rather learns a more generalized law. Shown in table 1 is a distribution of different points in time of the training data set.

TABLE 1 distribution of different time points of training data sets

In embodiments of the present disclosure, before the training data is used to iterate the original model, i.e., during the feature processing after the dataset is sampled, similar feature processing pairs may be constructed such that each pre-selected model is model trained according to a random feature set, where the random feature set is obtained by feature processing training data from one similar feature processing scheme randomly selected from the similar feature processing pairs. Specifically, each model may randomly select a similar feature processing scheme to process. By using different feature processing schemes, different feature sets can be generated for each model, so that the diversity of the models can be increased, model combinations with higher diversity can better make up for the defects of each other, the effect of model integration is improved, and better prediction performance is obtained.

For example, different models can be built using both the future 14 days and the future 28 days of visit. In addition, when the feature processing is carried out, the potential advantages of each method under a specific scene can be discovered by randomly selecting the feature processing method, so that the effect of each method is exerted to the maximum extent, and the overall prediction effect is improved. For feature filtering, for example, variance filtering may be used: the features with smaller variances are removed by calculating the variance of each feature. The smaller the variance, the smaller the range of variation of the value of the feature in the sample, and the smaller the contribution to the predictive power of the model. In general, features with variances less than a certain threshold will be considered redundant and can be culled directly. Correlation coefficient filtering may also be used: highly correlated features are removed by calculating correlation coefficients between the features. Highly correlated features contribute similarly to the predictive power of the model, so only one need be retained. The results of the two filtering features are different, and it is difficult to determine which final performance is better, and one of the two filtering features can be randomly selected for use. Wherein the similar feature processing pair is shown in table 2 below:

Table 2 similar characteristic treatment pairs

Specifically, as shown in table 2, when outlier processing is performed, one may be randomly selected at K-means (K-means) bucket and outlier cutoff. Wherein, through K-means barrel algorithm, the outlier and normal value can be divided into different clusters. Outliers are typically outliers that have low similarity to other data points, so they can be separated from other data points by setting the appropriate K value. After the data points are grouped into different clusters, the data points in each cluster may be processed, for example, by calculating the mean or median of the data points within the cluster to replace outliers. The method can maintain the overall distribution characteristics of the data and reduce the influence of abnormal values on the performance of the model. In addition, outlier truncation is another method of handling outliers. The core idea of this approach is to identify outliers by setting a threshold. For data points above or below the threshold, they can be set to the threshold, thereby eliminating the effect of outliers on the model.

Further, in terms of outlier processing, other outlier detection methods such as a distribution-based method, e.g., Z-score (standard score), IQR (interquartile range, quartile range), a clustering-based method DBSCAN (Density-Based Spatial Clustering of Applications with Noise, density-based clustering method with noise), or a Density-based method LOF (Local Outlier Factor ) may be tried.

As shown in table 2, in performing feature filtering, a filtering method may be randomly selected between variance filtering and correlation coefficient filtering. Variance filtering is a filtering method based on characteristic variances. The core idea is as follows: if the variance of a feature is smaller, the variance of the feature is smaller, the value change range of the feature in the data set is smaller, the contribution to the prediction capability of the model is smaller, and the feature with the variance smaller than a certain threshold can be eliminated. Correlation coefficient filtering is a filtering method based on correlation between features or between features and a target variable, which aims to eliminate highly correlated features, thereby reducing co-linearity between features.

As shown in table 2, one way in which to randomly select a feature based on model feature selection and challenge verification may be performed. Model-based feature selection the importance of a feature is determined by training a model. The importance of features may be obtained by feature importance scores provided by the model itself or by coefficients of the model, from which the most important feature subset is selected. The feature importance score provided by the features can be scores of feature importance provided by tree models such as decision trees, random forests and the like, and coefficients of the models can be coefficients of linear regression models, support vector machines and the like. Challenge verification is a method for evaluating and selecting features whose core idea is to predict whether data is from a training set or verification set by training a classification model. The basic assumption of this approach is that if a feature is distributed differently in the training set and the validation set, then this feature may negatively impact the generalization ability of the model.

In addition, other feature selection methods such as RFE (Recursive Feature Elimination ) may be used in the feature selection.

As shown in table 2, when feature encoding is performed, a mode may be randomly selected between single thermal encoding, which is a discrete feature processing method mainly applied to features having category attributes, and LabelEncoder. The principle of this approach is to convert the class features into a single thermal vector in binary form. For features with n different classes, one-hot encoding creates an n-dimensional one-hot vector, where only one element is a 1, representing the current class, and the remaining elements are 0. For example, if there are three possible values for a class feature: { A, B, C }, it can be converted into using one-hot encoding: a: [1, 0], B: [0,1,0], C: [0,0,1]. LabelEncoder is another discrete feature processing method, and is also applicable to features with category attributes. Unlike the unicode, the labelen codec represents class information by assigning an integer value to each class. This means that the original class feature will be replaced by a numerical value. For example, for the class feature { A, B, C } described above, it can be converted using LabelEncoder into: a:0, b:1, C:2.

In the embodiment of the disclosure, the training data may be subjected to enhancement processing. For example, whether the features of the online user of the training data are added may be based on the determination result of determining whether the online features are all 0 or null values among the features of the user.

Specifically, when the on-line features are both 0 or null, it may be set whether the feature value of the feature of the on-line user is no. If the on-line feature is not 0 or null, it is set whether the feature value of the on-line user feature is yes.

In the method for enhancing the training data according to the embodiment of the present disclosure, null filling may be performed on discrete null values and continuous null values of the features of the training data, respectively.

Specifically, discrete value nulls may be filled with "others" and continuous value nulls may be filled with "0".

In the method for enhancing the training data according to the embodiment of the present disclosure, the feature of the user behavior duration ratio of the training data may be further increased according to the ratio of the user behavior duration of the user behavior in different time periods to the total time period duration obtained by statistics.

Taking real estate projects as an example, the occupation ratio of the time periods of visit, subscription, activity, judgment and project sales in different time periods to the total time period can be counted. After the duty ratio of the events in different time periods is obtained, the number of the features can be more stable by using the duty ratio, the time period is not influenced, and meanwhile, the comparison between the different time periods can be carried out.

In the method for enhancing the training data according to the embodiment of the present disclosure, the aggregate feature of the training data may be further increased according to the variance, the maximum value, the average value obtained by counting the continuous features according to the item category and the number of times obtained by counting the number of occurrences of the discrete features.

In the method for enhancing the training data in the embodiment of the disclosure, normalization processing can be performed on days of opening the project so far, and the normalization processing can be used as new and old descriptions of the project. The normalization operation can unify the data in different numerical ranges to the same scale, and the influence of different characteristics on different models in the data range is avoided.

Training the original model by using training data to obtain a preselected model, obtaining a preselected model set, and then carrying out model fusion according to the preselected model set to obtain a target recommendation model. As shown in fig. 2, the fusion method of the target recommendation model includes:

s201, sorting at least two pre-selected models from large to small according to the evaluation function values of the at least two pre-selected models to obtain a pre-selected model set.

S202, traversing a pre-selected model set, selecting a pre-selected model with the minimum sequence number in the pre-selected model set as a basic model in a first cycle, taking a fusion model of the previous cycle in other cycles as a basic model, and executing the following cycle contents until a fusion model with the highest evaluation function value is obtained, wherein the fusion model is taken as a target recommendation model:

And acquiring a first evaluation function value of a first fusion model after weighted fusion of the first pre-selected model and the basic model according to a pre-constructed weight list, acquiring the first fusion model when the difference value of the first evaluation function value and the evaluation function value of the basic model is greater than or equal to a set threshold value, and deleting the first pre-selected model from a pre-selected model set, wherein the first pre-selected model is a pre-selected model in the pre-selected model set.

Specifically, the pre-selected model with the smallest sequence number in the pre-selected model set is the pre-selected model with the largest evaluation function value. And fusing the preselected model with the largest evaluation function value with other preselected models, deleting the corresponding preselected model in the preselected model set when the evaluation function value of the obtained fused model is larger than that of the basic model, taking the obtained fused model as the basic model of the next cycle, and traversing the preselected model set according to the cycle until the fused model with the highest evaluation function value is obtained.

Further, in the cyclic process, when the first evaluation function value of the first fusion model after the weighted fusion of the first pre-selected model and the basic model is obtained according to the pre-constructed weight list, the first weight in the weight list can be obtained, the weight of the basic model is taken as the first weight, the weight of the first pre-selected model is taken as the second weight, the basic model and the first pre-selected model are weighted and summed, the first fusion model is obtained, and the first evaluation function value of the first fusion model is obtained. Wherein the second weight is the difference between 1 and the first weight.

Specifically, the weight list may be a list in which weights are sequentially increased. In the first cycle, the first weight in the weight list may be selected as the weight of the base model, and in each subsequent cycle, the weight in the corresponding order of the cycle order in the weight list may be selected as the weight of the base model.

The model fusion process in the technical scheme of the embodiment of the disclosure is an automatic fusion process based on a climbing algorithm. The model is automatically fused based on the hill climbing algorithm, so that the fusion process is more efficient and adaptive.

Specifically, pre-selected models previously trained through different data sets, feature extraction, algorithms are structured into a model list m_list, with the models ordered in the list from large to small by the value of PR AUC. Selecting the model with the biggest PR AUC as a basic model M _base The corresponding PR AUC takes the value of Score _base . Selecting models M in sequence in a model list _i The following loop contents are performed:

constructing a weight list w= [0.01,0.02, … …,0.99]Starting from 0.01, ending at 0.99, and selecting weights w sequentially with a step size of 0.01 _n Setting a basic model M _base The weight is w _n ，M _i The weight is 1-w _n Calculate the basic model M _base And model M _i PR AUC after weighted fusion, noted Score _{i_n} Calculate the difference epsilon=score _{i_n} -Score _base If ε is not less than 0.0001, record w _n 、Score _{i_n} New fusion model M _{base_i_n} . The difference epsilon is the set threshold, the value is not limited to 0.0001, and i and n are serial numbers.

According to the circulating content, traversing all models in the model list, and selecting Score _{i_n} Highest M _{base_i_n} Model as new M _base And reject M in model list _i . And if the new fusion model is not constructed in the current cycle, ending the cycle.

In the model fusion process in the embodiment of the disclosure, the model is weighted and averaged according to the ranking instead of the score, so that the scale influence of the model evaluation function value can be reduced. Specifically, the model evaluation function values are not directly used for weighted average during model fusion, but are ranked from large to small according to each model evaluation function value, and the ranks are normalized and then weighted average is obtained. Different models may differ in the scale of the evaluation function values, and averaging the evaluation function values directly may result in some models having too much effect on the results. By averaging the normalized ranks, the difference in the scale of the evaluation function values can be eliminated, so that the contribution of each model to the fusion result is more balanced. In addition, the weighted average is performed according to the ranking, so that the relative order of clients can be focused more, and the accuracy of the recommendation system is improved. In particular, in a recommendation system, the relative order of customers, i.e., which customers are more likely to make purchases, is often of greater concern. The fusion model obtained by carrying out ranking weighted average on the evaluation function values is more concerned with ranking than the absolute size of the evaluation function values, so that the requirement of a target recommendation system is met.

In embodiments of the present disclosure, resampling may be performed when multiple models are built. As shown in fig. 3, for positive samples, different SEED (SEEDs) are selected, 80% of positive samples are randomly extracted, negative samples are divided into 20 barrels, and each time different barrels are randomly selected, so that the diversity of training data can be increased, the adaptability of the model to different conditions can be improved, and the model difference can be increased.

In addition, data sampling methods such as self-service sampling (Bootstrap), hierarchical sampling (Stratified Sampling) and the like can be adopted, so that the diversity of training data and the generalization capability of a model can be improved.

As shown in fig. 4, a target recommendation method in an embodiment of the present disclosure includes the following steps:

in step S411, the original feature is acquired.

Step S412, select the date of the data.

In step S413, the training set verifies set division.

In step S414, the positive sample is sampled.

Step S415, negative sample sampling.

Step S416, stitching the data sets.

Step S421, null filling.

Step S422, add whether the online user feature.

Step S423, normalizing the opening days.

In step S424, the project time-period activity statistics is performed.

Step S425, outlier processing.

Step S426, feature filtering.

In step S427, a feature is selected.

Step S428, feature encoding.

In step S431, a model is selected. Specifically, a pre-selected model can be selected from random forests, lightweight gradient lifts, extreme gradient lift trees, class-type feature lifts, tabNet.

Step S432, training a model.

Step S433, the verification set evaluates whether to promote, if yes, step S432 is executed, and if not, step S434 is executed.

Step S434, the training of the acquisition model is stopped.

Step S435, if the maximum number of search times is reached, step S436 is executed, and if not, step S432 is executed.

In step S436, a model is acquired.

Step S437, if a set number of models are obtained, step S438 is executed, and if not, step S431 is executed.

In step S438, the model list is ordered.

And S439, fusing the models.

In step S441, model prediction is performed.

In step S442, the prediction result is pushed, and specifically, the prediction result is the recommendation data output by the target recommendation model.

The technical scheme of the embodiment of the disclosure provides an advanced and stable business opportunity recommendation system, which is beneficial to enterprises to find potential clients with purchase will from huge user groups, and improves screening and identification capabilities of business opportunities and user conversion efficiency. The business opportunity recommendation system comprises rich feature construction, diversified data set construction and sampling strategies, similar feature processing pairs, fusion of various preselected models, model fusion strategies based on greedy algorithm and the like, so that the prediction performance, screening recognition capability and user conversion efficiency of the business opportunity recommendation system are effectively improved in practical application. Specifically, the system can build and sample strategies through diversified data sets, and process similar characteristics and the like, so that the adaptability of the model to different conditions is improved, and the diversity of the model is increased. And combining a plurality of original models such as random forests, XG Boost, lightGBM, catBoost, tabNet and the like, and adopting a model fusion technology to improve the generalization capability and stability of the model. The recommendation system can cope with challenges of iterative updating of application programs and data change, and adapts to new data distribution when models are fused by adopting strategies such as ranking weighted average and the like, so that the effect of the recommendation system in practical application is ensured to be continuously excellent. The recommendation system automatically completes the model fusion process by using greedy algorithms such as a mountain climbing algorithm, and the like, so that the complexity of model fusion is reduced, and the recommendation system is easier to deploy and maintain.

According to the target recommendation method, the pre-selection model is built in advance according to the target value of the target recommendation model, and model fusion is conducted on the pre-selection model through the ranking weighting strategy to obtain the target recommendation model, so that generalization capability and stability of the model are improved, and recommendation effect of a recommendation system is improved.

The following are device embodiments of the present disclosure that may be used to perform method embodiments of the present disclosure. The target recommending apparatus described below and the target recommending method described above may be referred to correspondingly to each other. For details not disclosed in the embodiments of the apparatus of the present disclosure, please refer to the embodiments of the method of the present disclosure.

Fig. 5 is a schematic diagram of a target recommendation device provided in an embodiment of the disclosure. As shown in fig. 5, the target recommendation device includes:

the obtaining module 501 is configured to obtain user item data, where the user item data includes user characteristics of a current user, historical behavior characteristics, and interaction characteristics relative to the current item.

The recommendation module 502 is configured to input user item data into a preset target recommendation model to obtain recommendation data of a current user relative to the current item, where the target recommendation model is obtained by performing model fusion on at least two pre-established pre-selected models through a ranking weighting strategy, and the at least two pre-selected models are established according to a target value of target recommendation.

In an embodiment of the present disclosure, the target recommendation device may further include a fusion module, configured to fuse the pre-selected model to obtain a target recommendation model. Specifically, the fusion module may be used to:

sequencing at least two pre-selected models from large to small according to the evaluation function values of the at least two pre-selected models to obtain a pre-selected model set, traversing the pre-selected model set, selecting a pre-selected model with the smallest sequence number in the pre-selected model set as a basic model in a first cycle, selecting a fusion model of the previous cycle as the basic model in other cycles, and executing the following cycle contents until a fusion model with the highest evaluation function value is obtained as a target recommendation model: and acquiring a first evaluation function value of a first fusion model after weighted fusion of the first pre-selected model and the basic model according to a pre-constructed weight list, acquiring the first fusion model when the difference value of the first evaluation function value and the evaluation function value of the basic model is greater than or equal to a set threshold value, and deleting the first pre-selected model from a pre-selected model set, wherein the first pre-selected model is a pre-selected model in the pre-selected model set.

The fusion module may also be used to: the method comprises the steps of obtaining a first weight in a weight list, taking the first weight as the weight of a basic model, taking a second weight as the weight of a first pre-selected model, weighting and summing the basic model and the first pre-selected model to obtain a first fusion model, and obtaining a first evaluation function value of the first fusion model, wherein the second weight is a difference value between 1 and the first weight.

In the disclosed embodiments, the target value may include the number of target customers visited in the future 14 days or the number of target customers visited in the future 28 days.

In an embodiment of the present disclosure, the target recommendation device may further include a training module for training at least two pre-selected models according to the following training method: and iterating the original model by using the training data until the evaluation function value of the original model is not increased or reaches the set iteration times, so as to obtain the target recommended model.

In an embodiment of the present disclosure, the original model of the at least two pre-selected models comprises at least any one of: random forests, lightweight gradient lifts, extreme gradient lift trees, class-type feature lifts, neural networks for tabular data.

In an embodiment of the present disclosure, the target recommendation device may further include a construction module for: the pair of similar feature processes is structured such that each pre-selected model is model trained from a random feature set obtained by characterizing training data from a similar feature processing scheme randomly selected from the pair of similar feature processes.

According to the target recommendation device disclosed by the embodiment of the disclosure, the pre-selection model is pre-established according to the target value of the target recommendation model, and the pre-selection model is subjected to model fusion through the ranking weighting strategy to obtain the target recommendation model, so that the generalization capability and stability of the model are improved, and the recommendation effect of a recommendation system is improved.

Fig. 6 is a schematic diagram of an electronic device 6 provided by an embodiment of the present disclosure. As shown in fig. 6, the electronic device 6 of this embodiment includes: a processor 601, a memory 602 and a computer program 603 stored in the memory 602 and executable on the processor 601. The steps of the various method embodiments described above are implemented by the processor 601 when executing the computer program 603. Alternatively, the processor 601 may implement the functions of the modules in the above-described device embodiments when executing the computer program 603.

The electronic device 6 may be a desktop computer, a notebook computer, a palm computer, a cloud server, or the like. The electronic device 6 may include, but is not limited to, a processor 601 and a memory 602. It will be appreciated by those skilled in the art that fig. 6 is merely an example of the electronic device 6 and is not limiting of the electronic device 6 and may include more or fewer components than shown, or different components.

The processor 601 may be a central processing unit (Central Processing Unit, CPU) or other general purpose processor, digital signal processor (Digital Signal Processor, DSP), application specific integrated circuit (Application Specific Integrated Circuit, ASIC), field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, or the like.

The memory 602 may be an internal storage unit of the electronic device 6, for example, a hard disk or a memory of the electronic device 6. The memory 602 may also be an external storage device of the electronic device 6, for example, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash Card (Flash Card) or the like, which are provided on the electronic device 6. The memory 602 may also include both internal and external storage units of the electronic device 6. The memory 602 is used to store computer programs and other programs and data required by the electronic device.

It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-described division of the functional units and modules is illustrated, and in practical application, the above-described functional distribution may be performed by different functional units and modules according to needs, i.e. the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-described functions. The functional units and modules in the embodiment may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit, where the integrated units may be implemented in a form of hardware or a form of a software functional unit.

The integrated modules may be stored in a readable storage medium if implemented in the form of software functional units and sold or used as a stand-alone product. Based on such understanding, the present disclosure may implement all or part of the flow of the method of the above-described embodiments, or may be implemented by a computer program to instruct related hardware, and the computer program may be stored in a readable storage medium, where the computer program may implement the steps of the method embodiments described above when executed by a processor. The computer program may comprise computer program code, which may be in source code form, object code form, executable file or in some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), an electrical carrier signal, a telecommunications signal, a software distribution medium, and so forth. It should be noted that the content of the computer readable medium can be appropriately increased or decreased according to the requirements of the jurisdiction's jurisdiction and the patent practice, for example, in some jurisdictions, the computer readable medium does not include electrical carrier signals and telecommunication signals according to the jurisdiction and the patent practice.

The above embodiments are merely for illustrating the technical solution of the present disclosure, and are not limiting thereof; although the present disclosure has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the disclosure, and are intended to be included in the scope of the present disclosure.

Claims (10)

1. A target recommendation method, the method comprising:

acquiring user project data, wherein the user project data comprises user characteristics, historical behavior characteristics and interaction characteristics relative to a current project of a current user;

and inputting the user item data into a preset target recommendation model to obtain recommendation data of the current user relative to the current item, wherein the target recommendation model is obtained by carrying out model fusion on at least two pre-established pre-selected models through a ranking weighting strategy, and the at least two pre-selected models are established according to target values of target recommendation.

2. The method of claim 1, wherein the fusing method of the target recommendation model comprises:

sequencing the at least two pre-selected models from large to small according to the evaluation function values of the at least two pre-selected models to obtain a pre-selected model set;

traversing the pre-selected model set, selecting a pre-selected model with the minimum sequence number in the pre-selected model set as a basic model in a first cycle, taking a fusion model of the previous cycle in other cycles as a basic model, and executing the following cycle contents until a fusion model with the highest evaluation function value is obtained, wherein the fusion model is used as the target recommendation model:

acquiring a first evaluation function value of a first fusion model obtained by weighting and fusing a first pre-selected model and the basic model according to a pre-constructed weight list, acquiring the first fusion model when the difference value of the first evaluation function value and the evaluation function value of the basic model is larger than or equal to a set threshold value, and deleting the first pre-selected model from the pre-selected model set, wherein the first pre-selected model is a pre-selected model in the pre-selected model set.

3. The method of claim 2, wherein obtaining a first evaluation function value of a first fusion model of a first pre-selected model and a weighted fusion of the base model from a pre-constructed weight list comprises:

Acquiring a first weight in the weight list;

taking the first weight as the weight of the basic model, and taking the second weight as the weight of the first pre-selected model to carry out weighted summation on the basic model and the first pre-selected model to obtain the first fusion model, wherein the second weight is a difference value between 1 and the first weight;

and acquiring a first evaluation function value of the first fusion model.

4. The method of claim 1, wherein the target value comprises a number of target clients visited in a future 14 days or a number of target clients visited in a future 28 days.

5. The method of claim 2, wherein prior to ranking the at least two pre-selected models from large to small according to their evaluation function values, the method further comprises: training the at least two pre-selected models according to the following training method: and iterating the original model by using training data until the evaluation function value of the original model is not increased or reaches the set iteration times, so as to obtain the target recommended model.

6. The method of claim 1, wherein the original model of the at least two pre-selected models comprises at least any one of:

Random forests, lightweight gradient lifts, extreme gradient lift trees, class-type feature lifts, neural networks for tabular data.

7. The method of claim 5, wherein prior to iterating the original model using the training data, the method further comprises: the pair of similar feature processes is structured such that each pre-selected model is model trained from a random feature set obtained by feature processing training data from a similar feature processing scheme randomly selected from the pair of similar feature processes.

8. An object recommendation device, characterized in that the device comprises:

the system comprises an acquisition module, a storage module and a display module, wherein the acquisition module is used for acquiring user project data, and the user project data comprises user characteristics, historical behavior characteristics and interaction characteristics relative to a current project of a current user;

and the recommendation module is used for inputting the user item data into a preset target recommendation model to obtain the recommendation data of the current user relative to the current item, wherein the target recommendation model is obtained by carrying out model fusion on at least two pre-established pre-selected models through a ranking weighting strategy, and the at least two pre-selected models are established according to target values of target recommendation.

9. An electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any of claims 1 to 7 when the computer program is executed.

10. A readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method according to any one of claims 1 to 7.