CN117556147B

CN117556147B - Electronic commerce data classification recommendation system and method

Info

Publication number: CN117556147B
Application number: CN202410039143.7A
Authority: CN
Inventors: 武岳巍; 付睿翎; 邢彤彤; 余振宇; 吴肇良; 冯小丽; 殷复莲
Original assignee: Communication University of China
Current assignee: Communication University of China
Priority date: 2024-01-11
Filing date: 2024-01-11
Publication date: 2024-04-12
Anticipated expiration: 2044-01-11
Also published as: CN117556147A

Abstract

The invention provides an electronic commerce data classification recommending system and method, wherein a supervision network is used for weighting different input feature groups through a loss perception feature attention mechanism network LAFAMN, and the weight is automatically regulated and controlled according to the whole network and the prediction result of each sub-network, so that the problem of unbalanced features in the electronic commerce data classification recommending process is effectively solved; in addition, based on the double-suppression loss function regulated and controlled by the classification confidence, the loss value of the large-class easily-judged sample is greatly suppressed, and meanwhile, the loss value of the small-class difficultly-judged sample is ensured not to be reduced at all, so that the problem of unbalanced classification of two dimensions of difficulty in judgment and unbalanced sample quantity is solved. The combined use of the LAFAMN and the double-compression loss function can form a complete recommendation system, solve the problem of unbalanced classification of the class in three dimensions, and provide a complete solution flow for the problem of commodity recommendation.

Description

Electronic commerce data classification recommendation system and method

Technical Field

The invention relates to the field of intelligent recommendation in the technical field of artificial intelligence, in particular to an electronic commerce data classification recommendation system and method.

Background

With the rapid development of the internet, the coverage of the e-commerce industry is wider and wider, and almost all e-commerce platforms use product recommendation systems to help customers to quickly find what they are interested in. For the e-commerce platform, an effective recommendation algorithm can improve profits of the e-commerce platform and reduce labor cost, but because the e-commerce platform has huge commodity quantity, fewer display positions and fewer browsed commodity quantity, the e-commerce platform has the unbalanced classification problem of classes with different dimensions in a recommendation system, wherein the method comprises the following steps: sample quantity unbalance, sample discrimination difficulty unbalance and sample characteristic quantity unbalance among categories.

Aiming at the classification problem of data class imbalance, the class with small sample number is generally called as a minor class in the research field, whereas the class with large sample number is called as a major class. In this regard, the conventional classification method aims at minimizing the error rate to establish a classifier, and based on the characteristics of sparse sample number of subclasses and relatively complex feature distribution, the network tends to distinguish all the samples into the large class, and the classification method greatly reduces the generalization performance of the subclasses.

Therefore, the method analyzes the reason for the reduced algorithm performance caused by the class imbalance in the e-commerce recommendation field, explores the solution of the class imbalance problem in the recommendation algorithm, and becomes one of the research directions in the current artificial intelligence recommendation field.

Disclosure of Invention

In view of the above-mentioned multi-dimensional unbalanced problem of commodity data in the recommendation system in the current e-commerce field, the present invention aims to provide an e-commerce data classification recommendation system and method, which uses an attention mechanism to integrate and improve a traditional multi-expert learning network to construct a LAFAM network, and based on a double compression Loss function (support Loss) regulated by classification confidence, suppresses the contribution degree of a large class sample and a sample Loss value easy to judge, and simultaneously ensures that the Loss value of a small class sample difficult to judge is not reduced at all so as to solve the unbalanced problem of the class in the existing recommendation algorithm.

The invention provides an electronic commerce data classification recommendation system, which comprises:

the feature preprocessing unit is used for carrying out importance sorting and feature classification on the first training data through a preset sorting method and a preset classifying method so as to obtain feature preprocessing data of the first training data;

the model training unit is used for carrying out learning training treatment on training data through a preset LAFAM network so as to obtain each type of output and label, carrying out loss value calculation on the output and label, carrying out weighted summation on the loss values, and carrying out feedback treatment so that the loss values reach the preset requirements to obtain a LAFAMN model; wherein the training data comprises the first training data and the feature preprocessing data, and the LAFAM network comprises a supervision network and a preset number of sub-networks; the sub-network is used for training and learning the feature preprocessing data, and the supervision network is used for training and learning the first training data; the supervision network learns the duty ratio weight of the sub-network, and the sub-network completes self-learning of the sub-network weight and the characteristic weight by dynamically adjusting the total loss; and processing a highly unbalanced data set in the subnetwork using a double-hold loss function;

And the classification recommending unit is used for performing classification recommending processing on the e-commerce data through the LAFAMN model.

Wherein, optionally, the feature preprocessing unit includes:

the importance ranking unit is used for performing primary ranking processing on the first training data through a preset ranking method, and performing weighted calculation on the primary ranking processing result to obtain a final ranking result of the first training data;

and the feature classification unit is used for classifying the first training data according to a preset classification method.

Wherein, the optional scheme is that the preset sorting method comprises PS-smart, XGBoost, GBDT; dividing the first training data into dense features, sparse features, time sequence features and basic features by the preset classification method; wherein the basic features are other features which are not in the three types of dense features, sparse features and time sequence features after classification; and the subnetworks respectively perform learning training processing on the input features of different categories.

The optional scheme is that the sub-network respectively carries out learning training processing on input features of different categories, and the method comprises the following steps:

Respectively inputting different types of input features of the first training data after the feature preprocessing data into different sub-networks for learning, wherein all the sub-networks output a score between 0 and 1, and the score represents a learning result of a current sample in the current sub-network; wherein 0 is the least likely to be recommended and 1 is the most likely to be recommended;

each subnetwork will calculate a loss value from the tag value and the predicted value of the sample,，wherein (1)>Is the total number of sub-networks; said loss value->The larger the value of (c), the lower the predictive confidence of the corresponding sub-network for the sample, and the lower the duty cycle in the final calculation of the output value.

Wherein, the output of the supervision network is the duty ratio of all sub-networks in the total output; and, in addition, the processing unit,

the label used in the supervision network training is obtained by performing SoftMax calculation according to the loss valuel _i Specific gravity vector of (2)Wherein->The value of the internal element->，，

Will beSoftMax computation as input before mapping for the firstSub-networks, confidence level of prediction success of each sub-network is estimated by using SoftMax>As shown in formula (1): （1）

Wherein,and->And has negative correlation.

Wherein, alternatively, the loss value is usedRecalculating a set of forward parametersSo that the output of each sub-network is optimized, wherein->Element values in vectors，As shown in formula (2):

（2）

wherein,and->And shows positive correlation.

The invention also provides an electronic commerce data classification recommending method, which is used for carrying out data class unbalanced classification based on the electronic commerce data classification recommending system and comprises the following steps:

performing importance sorting and feature classification on the first training data through a preset sorting method and a preset classifying method to obtain feature preprocessing data of the first training data;

training and learning the training data through a preset LAFAM network to obtain each type of output and label, calculating the loss value of the output and label, and carrying out feedback processing after weighting and summing the loss value to enable the loss value to reach the preset requirement to obtain a LAFAMN model; wherein the training data comprises the first training data and the feature preprocessing data, and the LAFAM network comprises a supervision network and a preset number of sub-networks; the sub-network is used for training and learning the feature preprocessing data, and the supervision network is used for training and learning the first training data; the supervision network learns the duty ratio weight of the sub-network, and the sub-network completes self-learning of the sub-network weight and the characteristic weight by dynamically adjusting the total loss; and processing a highly unbalanced data set in the subnetwork using a double-hold loss function;

And carrying out classified recommendation processing on the e-commerce data through the LAFAMN model.

The invention also provides an electronic device comprising:

at least one processor; the method comprises the steps of,

a memory communicatively coupled to the at least one processor; wherein,

the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the steps in the e-commerce data classification recommendation method as described above.

According to the technical scheme, the e-commerce data classification recommendation system and method provided by the invention firstly weight different input feature groups through a loss perception feature attention mechanism LAFAMN by using a supervision network, and automatically regulate and control the weight according to the whole network and the prediction result of each sub-network, so that the problem of unbalanced features is effectively solved; in addition, based on the double-suppression loss function regulated and controlled by the classification confidence, the loss value of the large-class easily-judged sample is greatly suppressed, and meanwhile, the loss value of the small-class difficultly-judged sample is ensured not to be reduced at all, so that the problem of unbalanced classification of two dimensions of difficulty in judgment and unbalanced sample quantity is solved. The combined use of the LAFAMN and the double-compression loss function can form a complete recommendation system, solve the problem of unbalanced classification of the class in three dimensions, and provide a complete solution flow for the problem of commodity recommendation.

Drawings

Other objects and attainments together with a more complete understanding of the invention will become apparent and appreciated by referring to the following description taken in conjunction with the accompanying drawings. In the drawings:

FIG. 1 is a schematic diagram of a logical structure of an e-commerce data classification recommendation system according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a framework of an e-commerce data classification recommendation system according to an embodiment of the invention;

fig. 3 is a schematic diagram of a theoretical structure of a LAFAM network according to an embodiment of the present invention;

FIG. 4 is a flowchart of an e-commerce data classification recommendation method according to an embodiment of the present invention;

fig. 5 is a schematic diagram of an application structure of an LAFAM network in an example according to an embodiment of the present invention;

fig. 6 is a grid search schematic diagram of an LAFAM network and a double hold-down loss function in an example according to an embodiment of the invention;

fig. 7 is a schematic diagram of an electronic device according to an embodiment of the invention.

Detailed Description

In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more embodiments. It may be evident, however, that such embodiment(s) may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form in order to facilitate describing one or more embodiments.

Aiming at the problem of unbalanced data class in the existing e-commerce product recommendation scheme, the invention provides an e-commerce data classification recommendation system and method, which are used for solving the problems of unbalanced class quantity, unbalanced discrimination difficulty and unbalanced characteristic three dimensions by integrating an attention mechanism and a mixed multi-expert network architecture, constructing a double-compression loss function in a function compression and exponentiation manner, exploring the influence of different networks and different loss functions on the performance of a recommendation system, realizing the construction of a recommendation system model oriented to unbalanced data class, and applying the recommendation system model to the commodity recommendation of e-commerce.

In order to better explain the technical scheme of the invention, the following will briefly explain some technical terms related to the invention.

GBDT (Gradient Boosting Decision Tree, gradient lifting decision tree) is an additive model based on Boosting ensemble learning ideas, by training a series of weak learners (usually decision trees) step by step and iteratively, each iteration tries to correct the errors of the previous iteration, and finally combines the weak learners into a strong learner. The GBDT model has the advantages of strong interpretation, good application effect and wide application in the fields of data mining, advertisement calculation, recommendation systems and the like.

PS-Smart regression, parameter server PS (Parameter Server) is dedicated to solving large-scale offline and online training tasks, SMART (Scalable Multiple Additive Regression Tree) is the iterative algorithm of GBDT based PS implementation. The PS-smart can support training tasks with billions of samples and hundreds of thousands of characteristics, can run on thousands of nodes, has a fault transfer function, and has good stability; meanwhile, PS-Smart supports various data formats, training targets and evaluation targets, and outputs feature importance, and comprises optimization of acceleration training such as histogram approximation.

XgBoost (eXtreme Gradient Boosting, extreme gradient lifting) is a synthetic algorithm with good data fitting effect formed by combining a basis function and a weight, and is different from GBDT, the XgBoost adds a regularization term to a loss function, and the second-order Taylor expansion of the loss function is used as the fitting of the loss function. The method is very powerful in parallel computing efficiency, missing value processing and prediction performance, simultaneously is excellent in preventing overfitting and improving generalization capability, and can quickly and accurately solve a plurality of data science problems.

Dense features, meaning that most elements in a feature vector are non-zero, are commonly used for data types such as image, audio, etc. The computation speed of dense features is slower but can provide more information and thus perform better in certain tasks.

Sparse features, which means that only a few elements of a feature vector are non-zero, are commonly used for data types such as text, recommendation systems, and the like. The computation speed of the sparse feature is faster.

Time series characteristics, characteristics extracted or counted based on time series.

F-measure, a statistic, also called F-Score, is a Precision and Recall weighted harmonic mean, is a commonly used evaluation criterion in the IR (information retrieval) field, and is commonly used to evaluate the quality of classification models.

Specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

It should be noted that the following description of the exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses. Techniques and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but should be considered part of the specification where appropriate.

In order to illustrate the e-commerce data classification recommendation system and method provided by the invention, fig. 1, fig. 2, fig. 3, fig. 4, fig. 5 and fig. 6 show exemplary labeling of a logic structure, a system architecture, a theoretical structure of a LAFAM network, a flow chart of the e-commerce data classification recommendation method, an application structure of the LAFAM network in an example, and grid searching of the LAFAM network and a double compression loss function in the example of the e-commerce data classification recommendation system according to the embodiment of the invention; fig. 7 illustrates an exemplary method for recommending e-commerce data classification according to an embodiment of the present invention.

The e-commerce data classification recommendation system provided by the invention takes a Loss perception feature attention mechanism network (Loss Aware Feature Attention Mechanism Network, LAFAMN) as a main body and is realized by combining a classification recommendation system framework of a double compression Loss function (support Loss), so that the problems of unbalanced category number, unbalanced discrimination difficulty and unbalanced feature are solved together. For convenience of description, in the following embodiments of the present invention, the LAFAM network is also simply referred to as a network.

Fig. 1 and 2 illustrate an e-commerce data classification recommendation system logic structure and a system architecture, respectively, according to an embodiment of the present invention.

As shown in fig. 1 and fig. 2 together, the e-commerce data classification recommendation system 100 provided by the invention mainly comprises three parts, namely a feature preprocessing unit 110, a model training unit 120 and a classification unit 130, and the corresponding system architecture is a general flow system and is divided into three stages of feature preprocessing, LAFAMN model training and classification.

The feature preprocessing unit 110 is configured to perform importance ranking and feature classification on the first training data by using a preset ranking method and a preset classification method, so as to obtain feature preprocessing data of the first training data;

The model training unit 120 is configured to perform learning training processing on training data through a preset LAFAM network to obtain each type of output and label, calculate a loss value of the output and label, and perform feedback processing after weighting and summing the loss value to make the loss value reach a preset requirement to obtain a LAFAM model;

and the classification unit 130 is used for performing classification recommendation processing on the e-commerce data through the LAFAMN model.

The three parts will be described in detail with reference to specific embodiments.

Before feature preprocessing, training data acquisition (sample acquisition) is first required. In the invention, part of the acquired training data is used as first training data for learning training, the first training data is input into a sub-network of a model training unit for learning training after feature preprocessing, and for a supervision network of the model training unit, all the first training data is subjected to learning training.

In addition, since the solution of the present invention mainly aims at the problem of multi-dimensional imbalance of commodity data in the recommendation system in the current e-commerce field, in the embodiment shown in fig. 2, the collected training data includes data such as commodity characteristics, coupons, commissions, histories, prices, and the like related to the commodity data in the e-commerce field.

The collected training data may be stored in a sample database in a centralized manner, and after the sample is collected, a part of the training data may be preprocessed by the feature preprocessing unit 110. The invention finishes the sorting and classifying of part of training data through a plurality of preset algorithms so as to replace the feature selection and fusion in the traditional classifying method.

In one embodiment of the present invention, the feature preprocessing unit 110 includes an importance ranking unit 111 and a feature classification unit 112. The importance ranking unit 111 is configured to perform a primary ranking process on the first training data by using a preset ranking method, and perform weighted calculation on the primary ranking result to obtain a final ranking result of the first training data; the feature classification unit 112 is configured to perform classification processing on the first training data according to a preset classification method.

Specifically, as an example, the feature importance is ranked in the ranking section using three ways of PS-smart, XGBoost, GBDT, the feature importance score is weighted against the ranking results given by the three ways, and the final ranking is given. In addition, in order to improve the sorting efficiency, a certain degree of truncation can be performed on the sorting given in the process, the discarding operation is performed on the feature with the contribution degree of almost 0, and the specific truncation position should be adjusted according to the specific problem. In the feature classification section, to better coordinate with the operation of the LAFAM network, in one embodiment of the present invention, features are classified into basic features, dense features, sparse features, and timing features, and different sub-networks are used to process different features. The basic features are other features that are not within the three classes of features (dense features, sparse features, and timing features) after classification.

The model training unit 120 performs learning training processing on the training data through a preset LAFAM network. The LAFAM network comprises a supervision network and a preset number of sub-networks, wherein the sub-networks are used for training and learning the features obtained by preprocessing by the feature preprocessing unit 110, the supervision network is used for training and learning all training data which are not subjected to feature preprocessing, and the supervision network learns the duty ratio weights of the sub-networks.

In the training process of the LAFAMN model, input learning is firstly carried out on all the obtained data. The feature data subjected to sorting and classification is input into each corresponding sub-network, and all features which are not subjected to feature preprocessing are input into a supervision network, and the supervision network learns the duty ratio weights of the sub-networks. In this way, the training mode of the LAFAMN model solves the problem of unbalanced characteristics, and suitable networks are allocated for different types of characteristics; in the invention, the self-learning of the subnet weight and the feature weight is also completed by dynamically adjusting the total loss, so that the network can make the largest contribution of each feature to the prediction result under the condition of not discarding any feature.

Furthermore, to address the class imbalance problem, the present invention uses a double-hold loss function in the appropriate subnetwork to handle highly unbalanced data sets. By matching the LAFAM network and the double-compression loss function, three problems of unbalanced characteristics, unbalanced quantity and unbalanced difficulty in distinguishing are perfectly solved.

And in the classified recommending unit, the e-commerce data is subjected to classified recommending processing through the LAFAMN model.

In addition, in one embodiment of the invention, parameters of the LAFAM network are also obtained by a grid search mode after the LAFAMN model is obtained by model training、And->And parameters of the double compression loss function +.>、、、And optimizing, combining the value ranges of the parameters, generating a parameter grid, selecting the combination which enables the accuracy rate of network classification recommendation and F-measure parameters to reach the optimum to obtain a classification threshold, and carrying out parameter optimization processing on the trained LAFAMN model according to the classification threshold. Specifically, as an embodiment, in the process of classified recommendation, the invention refers to a commodity with daily sales volume arranged at the top of industry and total sales volume as an "explosion commodity", that is, a commodity with a relatively fire-explosion sales market. Meanwhile, the commodity explosion probability ranking is given according to the LAFAMN model scoring, and the invention also attaches importance to the parameters of the first 50/100/200 of the bid rates (HR@50, HR@100 and HR@200) used in the engineering in the classifying process so as to maximize the parameters.

Fig. 3 is a schematic diagram of a theoretical structure of a LAFAM network according to an embodiment of the present invention.

In order to solve the problem of unbalanced characteristics, the invention provides a loss perception characteristic attention mechanism network LAFAMN, and the theoretical structure of the LAFAMN is shown in figure 3.

In the LAFAMN shown in fig. 3, first, a plurality of sub-networks are set to learn characteristics, and the number and types of the sub-networks can be adjusted and changed according to different problems, and the first three sub-networks are shown as a plurality of different sub-networks in fig. 3.

Input features of different categories of the first training data after feature preprocessing data are respectively put into different sub-networks to be learned, all the sub-networks output a score between 0 and 1, the score represents a learning result of a current sample in the current network, and 0 is the least possible to be recommended, and 1 is the most possible to be recommended. Each subnet calculates a loss value from the label value and the predicted value of the sample，Wherein, namely the firstiA gap between the output predicted value and the current sample tag value of the sub-network, whereinNIs the total number of subnetworks (excluding the supervisory network). This loss value is passed back not only as a loss function, but more importantly as a decision criterion for the output confidence. / >The larger the value, the lower the prediction confidence of the subnet for the sample, the more the learning strength of the network for the current subnet should be enhanced, thus giving it more weight in the loss of backhaul. At the same time (I)>A larger value indicates a higher probability of error in its prediction, and therefore its duty cycle should be lower when the output value is finally calculated. Next, according to the above features, the present invention sets a supervisory network, as shown in the lowermost sub-network in fig. 3, whose output is all sub-networksThe duty ratio of the complex output in the total output is calculated by SoftMax according to the loss value, and the label used in training is about +.>Specific gravity vector->WhereinNFor the total number of subnetworks (excluding the supervision network), additionally +.>The value of the internal element->，。

SoftMax functions are commonly used to address the multi-classification problem, mapping the output of multiple neurons into a range. The invention uses the characteristic that the predictive value of a plurality of categories can be mapped exponentiallySoftMax calculations were performed as input before mapping. Then for->Sub-networks, confidence level of prediction success of each sub-network is estimated by using SoftMax>As shown in formula (1):

（1）

for a pair ofLoss- >The larger the size of the container,the lower the confidence of the prediction success, i.e. the larger the difference between the predicted value and the tag value, the +.>The smaller the contribution of the sub-network to the overall output of the network is, the lower the +.>And->And has negative correlation. However, if the confidence is simply used to determine the duty ratio of the output, a phenomenon is caused that if only one sub-network has a better prediction result, the whole network is predicted almost completely depending on the result of the one network, which causes poor robustness and stability of the network. Thus, the present invention contemplates best effort delivery of each sub-network in the lam network, i.e., optimizing the output of each sub-network during a single overall study. Thus pass->Recalculating a set of forward parametersWhereinNFor the total number of subnetworks (excluding the supervision network), additionally +.>Element values in vectors，As shown in formula (2):

（2）

for a pair ofLoss->The larger the confidence of the prediction success is lower, namely the larger the difference between the predicted value and the tag value is, and +.>The larger the sub-network is, the higher the contribution degree of the sub-network to the overall loss value of the network is expected to be, the +.>And->And shows positive correlation.

As shown in fig. 3, in the lam network, there are three types of outputs and labels in total, namely:

Network overall outputAnd tag value of original sample->；

Output of a subnetworkAnd tag value of original sample->；

Output of supervisory networkAnd tag value from subnet loss SoftMax。

The three types of output are respectively calculated as loss values, and the loss values of the three types of output are respectively recorded as、And->And performing loss calculation by using the output values and the tag values of various types. After obtaining the three loss values, the three loss values are weighted and summed and then returned so that each type of network can take the corresponding loss value and optimize the loss value. Therefore, the total loss function is defined in this embodiment +.>(return all) for calculation and demonstration of the return derivative, an example is made using a square loss function, the specific procedure is shown in equation (3):

（3）

wherein,、and->For parameters of the lam network, grid search or other parameter tuning methods can be used to adjust and set the parameters.y _j Is the firstjAn output of the sub-network;For the temperature coefficient used in SoftMax operation, +.>The differences between the values after SoftMax calculation can be adjusted, typically,/->。The smaller the SoftMax curve is, the steeper the difference between the values output after operation is, and the similarity is amplified so as to be convenient for distinguishing samples which are difficult to separate; on the contrary, the- >The larger the SoftMax curve is, the smoother the difference between the values output after the operation is. In this embodiment by setting the coefficient +.>To control sub-network losses->Weighted equalization.

The final output of the network is shown in equation (4):

（4）

wherein,is a parameter in the output vector given by the supervisory network. The network diagram shown in fig. 3 is the whole process required by the lam network in training and prediction, and during training, the lam network calculates three outputs respectively, and calculates the loss for each output and returns the loss. However, when the prediction process is finally performed, the parameters of the supervision network output are directly used for the weighted calculation of the total output to obtain the LAFAM network outputyAll dash-dot line portions in fig. 3 are no longer running.

The invention also provides a double compression Loss function (support) for optimizing and solving the unbalance problem of two dimensions, namely the unbalance of the number of samples among the categories and the unbalance of the difficulty in discriminating the samples, existing in the existing classification scheme.

Often, the imbalance of data in classification problems often occurs in a large class of data far more than a small class. At this point, if the global loss is directly calculated in a general manner (e.g., square loss, cross entropy loss, etc.), the sample training loss of a large class dominates. In the intelligent electronic commodity recommendation scene, the explosion commodities are classified into subclasses, and finding out the explosion commodities is an important target for research. Aiming at the problem of unbalanced sample number, a lower loss contribution degree is distributed for samples approaching to a large class in a function mode, and meanwhile, the loss of samples approaching to a small class is guaranteed to be free from attenuation as much as possible, so that better optimization learning can be conveniently carried out on the samples.

Such ideas can also be used to identify the problem of difficulty unbalance. The larger the difference between the sample prediction output value and the label value, the more distant the prediction value is from the label value, the larger the error rate is, and the lower the prediction confidence is, the fewer the number of the samples is, but the importance and difficulty of the prediction are. It is therefore desirable to boost the contribution of samples with large gaps between predicted output values and tag values to total loss while suppressing the contribution of samples with high confidence to total loss.

In summary, the double-compression loss function provided by the embodiment of the invention is composed of three parts, wherein the first part is used for compressing the contribution degree of the loss of a large class of samples by means of the function; the second part is to suppress the contribution of the easily separable samples through the same function and different parameters; the last part is to suppress the contribution degree of the easily separable samples in a mode of expanding the output gap through a high power function, and the form is shown as a formula (5):

（5）

wherein,outputting a final prediction for the network;Taking a value of 0 or 1 as a label value of an original sample;is the distance between the sample predicted output value and the tag value, in this embodiment referred to as the prediction bias +.>；、、、And->Parameters of the double-suppression loss function can be set by combining actual data set problems through parameter adjustment modes such as a grid search method and the like.

In a specific application process, the LAFAMN model provided by the invention can calculate the probability that the commodity becomes an explosion commodity. To evaluate the effectiveness of the LAFAMN model, four evaluation indexes, namely root mean square error RMSE (Root Mean Squared Error), average absolute error MAE (Mean Absolute Error), weighted average absolute error percentage WMAPE (Weighted Mean Absolute Percentage Error) and the Rate of winning hr@50/100/200 (Hire rate@50/100/200) of the prior 50/100/200, are used for performing performance evaluation on the LAFAMN model provided by the invention in regression problem. The first three evaluation indexes are used for considering the model from the regression algorithm side, and the last evaluation index is used for considering the model from the perspective of the accuracy of the commodity prediction sequencing. In the classification problem, three evaluation indexes, namely Precision, accuracy Accuracy and F value F-measure, are used in the invention.

The RMSE index can better reflect the influence of the extreme value on the error, and the calculation mode is shown as a formula (6):

（6）

where n is the number of samples,is the predicted model output, +.>Is the true tag value.

The MAE index can very intuitively show the regression error condition, and the calculation mode is shown as a formula (7):

（7）

The WMAPE index is less influenced by extreme values and individuals, the overall prediction condition of the network can be displayed more uniformly, and the calculation mode is shown as a formula (8):

（8）

the HR index can reflect the accuracy of model recommendation, that is, whether the requirement item of the user is included in the recommendation item of the model, and the calculation mode is shown in formula (9):

（9）

wherein,is all test set, +.>Is the sum of the number of items belonging to the test set in each user top-k list.

The Precision index can effectively reflect the Precision of the subclass, and the calculation mode is shown as a formula (10):

（10）

wherein,representing the number of samples predicted to be positive and actually also positive, ++>Representing the number of samples predicted as negative but actually positive, < >>Representing the number of samples predicted as positive but actually negative, < >>The number of samples predicted as negative examples and actually as negative examples is represented.

The Accuracy index reflects the global Accuracy, can ensure that the network does not make the prediction of the deflection, and the calculation mode is shown as a formula (11):

（11）

the F-measure index is used for measuring the overall performance of an algorithm in the data class unbalance classification problem, and the calculation modes are shown in a formula (12) and a formula (13):

（12）

（13）

Wherein,for a factor that adjusts the specific gravity of Precision to Recall, typically 1 is taken.

For the three indexes of RMSE, MAE and WMAPE, the smaller the numerical value is, the better the recommendation system effect is. For the four indexes of HR@50/100/200, precision, accuracy and F-measure, the larger the numerical value is, the better the classification recommendation system effect is indicated.

As shown in fig. 4, the present invention further provides a method for classifying and recommending e-commerce data, which performs classification and recommendation of e-commerce data based on the e-commerce data classifying and recommending system 100 as described above, and includes:

s1: performing importance sorting and feature classification on the first training data through a preset sorting method and a preset classifying method to obtain feature preprocessing data of the first training data;

s2: training and learning the training data through a preset LAFAM network to obtain each type of output and label, calculating the loss value of the output and label, and carrying out feedback processing after weighting and summing the loss value to enable the loss value to reach the preset requirement to obtain a LAFAMN model; wherein the training data comprises the first training data and the feature preprocessing data, and the LAFAM network comprises a supervision network and a preset number of sub-networks; the sub-network is used for training and learning the feature preprocessing data, and the supervision network is used for training and learning the first training data; the supervision network learns the duty ratio weight of the sub-network, and the sub-network completes self-learning of the sub-network weight and the characteristic weight by dynamically adjusting the total loss; and processing a highly unbalanced data set in the subnetwork using a double-hold loss function;

S3: and carrying out classified recommendation processing on the e-commerce data through the LAFAMN model.

The above-mentioned electronic commerce data classification recommendation method is an implementation method corresponding to the above-mentioned electronic commerce data classification recommendation system, and specific implementation steps thereof may refer to specific embodiments of the above-mentioned electronic commerce data classification recommendation system, which will not be described in detail herein.

The embodiment can be seen that the loss-aware feature attention mechanism network (LAFAMN) provided by the invention integrates and improves the traditional multi-expert learning network by using an attention mechanism aiming at the multi-dimensional unbalanced problem of commodity data in the recommendation system in the current e-commerce field, namely, the feature unbalance, the difficulty in distinguishing unbalance and the sample number unbalance. Complementary to the above, a classification confidence regulation-based double compression Loss function (suppresion Loss) is also provided, so that the contribution degree of the Loss value of the large class sample and the easily-judged sample is inhibited, and meanwhile, the Loss value of the small class sample difficult to judge is ensured not to be reduced at all. The combined application of the LAFAM network and the double-suppression loss function can form a complete recommendation system, and the problem of unbalanced class classification in the three dimensions is solved. Experiments prove that the e-commerce data classification recommendation system and method provided by the invention are obviously improved in regression problem indexes and classification problem indexes compared with the existing recommendation algorithm.

The electronic commerce data classification recommendation system provided by the invention will be described in more detail below with a specific application embodiment.

In terms of construction of the sample database, in the present embodiment, experimental verification is performed using an e-commerce dataset from a certain e-commerce, which is denoted as a Tmall dataset in the following description. The data is stored in 10.7GB, the total data comprises 51,134,193 lines, the original data features are 453, and because of the phenomena of missing, misplugging and the like of many features, 286 features which are confidence in a system algorithm (the system fills out data with the accuracy more than 90% after more than 1 year of time verification) are selected for experimental use in the embodiment in an experiment.

Because the final output value given by the LAFAMN model is a continuous value within the value range, in the embodiment, the problem is considered as a regression problem to carry out experimental verification, RMSE, MAE, WMAPE and HR@50/100/200 parameters are used for carrying out effect verification in the process, and the score ordering given by the network is verified. However, from the practical application point of view, whether a commodity is an explosion commodity or not is an explosion commodity is an absolute judgment, so that in order to materialize a prediction problem, the invention also converts a sorting regression problem into a classification problem to learn and verify, normalizes sales in a data set, selects 0.15 as a threshold value to perform two classifications, and the number of classified classes of 0 (large class, non-explosion commodity) and 1 (small class, explosion commodity) is 70,184,766 and 20,483 respectively, and the unbalance degree is 3426:1.

The 286 features in the Tmall dataset are divided into a base feature, a price feature, a discount feature, and a timing feature in this embodiment. The price characteristics refer to 41 characteristics such as the absolute value of the price of the current commodity, the average value of the price under the category of leaves, the average value of the price of the commodity of the bid product, the lowest value of the historical price and the corresponding sales quantity, the highest value of the historical price and the corresponding sales quantity. The discount features refer to 28 features of discount values given by current commodities, historical average discount values, average discount values of bid commodities, historical highest discount values and corresponding sales thereof, historical lowest discount values and corresponding sales thereof. The time sequence features refer to 10 features of daily price, daily sales, daily commodity score and the like of commodity in 30 days. 207 features other than the above three features are attributed to the basic features. The features of the above four categories will cover the entire 286 features, each belonging to and among only one category, the features to which each category belongs are not intersected and repeated.

Therefore, the LAFAM network structure applied in the embodiment of the present invention is shown in fig. 5, where the network parameters correspond to the parameter names in fig. 3 one by one, and the state is shown when four sub-networks and one supervision network are set for learning. In fig. 5, the training model is shown in the outermost solid line frame, and the prediction model is shown in the dash-dot line frame. The features are input into four sub-networks according to categories, the full features are learned by using a Deep crossover network DCN common in recommended algorithm engineering, the Deep part is a full-connection network and is used for extracting the depth characteristics of the features, the Cross part is used for carrying out feature crossover calculation and is used for extracting the breadth characteristics of the features, vertical category recommendation and wide preference heuristics can be well compatible, and the operation efficiency is high, so that the full-feature Deep crossover network is used as a full-feature learning network. Price and discount features, i.e., using a simple fully-connected neural network, can extract high-dimensional features of the sample to make recommendation predictions. The time sequence features are learned by using the gate-controlled cyclic neural network GRU, which has a simple structure and can overcome the long dependence problem of the RNN.

In this embodiment, experiments were performed to verify the performance of the LAFAM network in the regression problem by using the square loss function as the basic backhaul loss function of the network, and compared with the performance of other conventional networks (e.g. GBDT, DCN, moE), the experimental results are shown in table 1:

table 1 comparison of performance data of different networks in regression problem

First, by performing a macroscopic analysis of table 1, it can be seen that the lam network has significant advantages in the hr@50/100/200 parameters, while at the same time, the lam network has no significant disadvantages over other networks in terms of RMSE, MAE, WMAPE, and even occupies optimal positions on the MAE, WMAPE parameters, the RMSE parameters are reduced because in highly unbalanced data sets, large classes of samples contribute a large number of error values. Therefore, from the results, when the LAFAM network is considered as a regression problem, the LAFAM network can well optimize the problem of prediction accuracy reduction caused by characteristic imbalance and data class imbalance, and has certain prediction winning rate advantages compared with other traditional networks.

And secondly, carrying out detailed analysis on three winning rate parameters. Firstly, comparing different parameters between the same algorithms, namely transversely comparing tables, it can be seen that from HR@200→HR@100→HR@50, namely from the widest to the most precise, the accuracy of GBDT is continuously reduced, and the LAFAM network is continuously improved, in other words, the LAFAM network is more precise in top prediction, namely the alignment of subclass samples. And then comparing the same parameters with different algorithms, namely longitudinally comparing the tables, the LAFAM network is 8.4% higher than the MoE network arranged at the second highest in HR@200, 10.4% higher than the MoE network in HR@100 parameter, 34.8% higher than the MoE network in HR@50 parameter, the more accurate the range, the better the LAFAMN performs compared with other networks, the obvious advantage of predicting subclass samples is realized, and regression sequencing and recommendation can be completed under the condition of unbalanced classification.

Classification verification of the lam network and the double compression Loss function support Loss is performed below.

First, network parameters are determined using a grid search method, and grid search optimization of some parameters is shown in FIG. 6, where black dot locations are selected parameter values. The network usage parameters finally obtained are shown in table 2, only the optimal parameters of the proposed LAFAM network and the double compression Loss function support are shown in the table, and the rest networks and the Loss function are all used for parameter tuning in the same mode and the optimal solution is obtained.

Table 2 summary of parameter settings

Under different networks, square Loss function Square, focal Loss function Focal Loss, shrinkage Loss function Sringage Loss and double compression Loss function support Loss are respectively verified, MLP, GRU, DCN, moE and LAFAM networks are respectively verified in comparison network selection, the former networks are sub-networks used in the LAFAM networks respectively, and therefore the LAFAM networks are independently tested, and when being used as a base line index, an ablation experiment is equivalently completed, so that the combined network is proved to have corresponding improvement and effect optimization compared with each sub-network. The experimental results are shown in table 3, each column in the table is the comparison of the verification results of four different loss functions under the same network, and the corresponding sequence in each column is the comparison of the verification results of different networks of the same loss function.

TABLE 3 comparison of experimental results of different loss functions in different networks

From the observation of table 3, in terms of network, the LAFAM network has a more obvious improvement on the data set with high unbalance degree compared with other networks, three parameters of the LAFAM network are all in optimal positions in the 1:20 unbalance data set, the average is about 10% higher than that of other networks, and the average is about 7% lower than that of the optimal network in the 1:5 and 1:10 unbalance data sets, so that the LAFAM network is used in the unbalance-like and characteristic unbalance data sets, and has stable effect and obvious advantage. In terms of the Loss function, it can be clearly found that in each network, the double compression Loss function support has the best performance, because it can well solve the problems of unbalanced category number and unbalanced sample discrimination confidence. The experimental effect of the double-compression loss function is improved by about 15% in average compared with other loss functions, and the performance of various networks and unbalance degree data sets is verified, so that the suitability of the double-compression loss function is excellent. From experimental data, the stability of the double compression loss function is acceptable.

As shown in fig. 7, the present invention also provides an electronic device, including:

At least one processor; the method comprises the steps of,

a memory communicatively coupled to the at least one processor; wherein,

the memory stores a computer program executable by at least one processor to enable the at least one processor to perform the steps of the aforementioned e-commerce data classification recommendation method.

It will be appreciated by those skilled in the art that the structure shown in fig. 7 is not limiting of the electronic device 1 and may include fewer or more components than shown, or may combine certain components, or may be arranged in different components.

For example, although not shown, the electronic device 1 may further include a power source (such as a battery) for supplying power to each component, and preferably, the power source may be logically connected to the at least one processor 10 through a power management device, so that functions of charge management, discharge management, power consumption management, and the like are implemented through the power management device. The power supply may also include one or more of any of a direct current or alternating current power supply, recharging device, power failure detection circuit, power converter or inverter, power status indicator, etc. The electronic device 1 may further include various sensors, bluetooth modules, wi-Fi modules, etc., which will not be described herein.

Further, the electronic device 1 may also comprise a network interface, optionally the network interface may comprise a wired interface and/or a wireless interface (e.g. WI-FI interface, bluetooth interface, etc.), typically used for establishing a communication connection between the electronic device 1 and other electronic devices.

The electronic device 1 may optionally further comprise a user interface, which may be a Display, an input unit, such as a Keyboard (Keyboard), or a standard wired interface, a wireless interface. Alternatively, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch, or the like. The display may also be referred to as a display screen or display unit, as appropriate, for displaying information processed in the electronic device 1 and for displaying a visual user interface.

It should be understood that the embodiments described are for illustrative purposes only and are not limited to this configuration in the scope of the patent application.

The electronic commerce data classification recommendation program 12 stored in the memory 11 of the electronic device 1 is a combination of a plurality of instructions, which when executed in the processor 10, may implement:

Specifically, the specific implementation method of the above instruction by the processor 10 may refer to the description of the relevant steps in the corresponding embodiment of fig. 4, which is not repeated herein.

Further, the modules/units integrated in the electronic device 1 may be stored in a computer readable storage medium if implemented in the form of software functional units and sold or used as separate products. The computer readable medium may include: any entity or device capable of carrying the 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).

The e-commerce data classification recommendation system and method according to the present invention are described above by way of example with reference to the accompanying drawings. However, it will be appreciated by those skilled in the art that various modifications may be made to the e-commerce data classification recommendation system and method set forth in the present invention described above without departing from the spirit of the present invention. Accordingly, the scope of the invention should be determined from the following claims.

Claims

1. An electronic commerce data classification recommendation system, comprising:

The model training unit is used for carrying out learning and training treatment on training data through a preset loss perception feature attention mechanism network (Loss Aware Feature Attention Mechanism Network, LAFAMN) so as to obtain each type of output and label, carrying out loss value calculation on the output and label, carrying out weighted summation on the loss values and carrying out back transmission treatment on the loss values, so that the loss values reach preset requirements, and then obtaining a LAFAMN model; wherein the training data comprises the first training data and the feature pre-processing data, and the LAFAMN comprises a supervisory network and a preset number of sub-networks; the sub-network is used for training and learning the feature preprocessing data, and the supervision network is used for training and learning the first training data; the supervision network learns the duty ratio weight of the sub-network, and the sub-network completes self-learning of the sub-network weight and the characteristic weight by dynamically adjusting the total loss; and processing a highly unbalanced data set in the subnetwork using a double-hold loss function;

the classified recommending unit is used for performing classified recommending processing on the e-commerce data through the LAFAMN model;

The output and label of the LAFAMN include:

network overall output, includingAnd tag value of original sample->；

Sub-network outputs includingAnd tag value of original sample->；

Supervising network output, includingAnd tag value from subnet loss SoftMax；

Wherein the losses of three types of outputThe loss values are respectively recorded as、And->Define the total loss function->The process of calculating the total loss function using the square loss function is shown in equation (3):

（3）

wherein,、and->As a parameter of the LAFAMN,y _j is the firstjAn output of the sub-network;For the temperature coefficient used in SoftMax operation, by setting the coefficient +.>To control the sub-network loss value->Weighted equalization;

the final output of LAFAMN is shown in equation (4):

（4）

wherein,is a parameter in the output vector given by the supervisory network;

the double compression loss function comprises three parts, wherein,

the first part is to suppress the loss contribution degree of a large class of samples by means of a function;

the second part is to suppress the contribution of the easily separable samples through the same function and different parameters;

the last part is to suppress the contribution degree of the easily separable samples in a mode of expanding the output gap through a high power function, and the form is shown as a formula (5):

（5）

Wherein,final prediction output for LAFAMN;Taking a value of 0 or 1 as a label value of an original sample;is the distance between the sample predicted output value and the tag value;、、、And->Is a parameter of the double compression loss function.

2. The e-commerce data classification recommendation system of claim 1, wherein the feature preprocessing unit comprises:

3. The e-commerce data classification recommendation system of claim 2,

the preset sorting method comprises PS-smart, XGBoost, GBDT;

dividing the first training data into dense features, sparse features, time sequence features and basic features by the preset classification method; wherein the basic features are other features which are not in the three types of dense features, sparse features and time sequence features after classification;

And the subnetworks respectively perform learning training processing on the input features of different categories.

4. The e-commerce data classification recommendation system of claim 3 wherein the sub-network performs a learning training process on different classes of input features, respectively, comprising:

each sub-network calculates a loss value from the label value and the predicted value of the sample，Wherein->Is the total number of sub-networks; said loss value->The larger the value of (c), the lower the predictive confidence of the corresponding sub-network for the sample, and the lower the duty cycle in the final calculation of the output value.

5. The e-commerce data classification recommendation system of claim 4,

the output of the supervision network is the duty ratio of all sub-networks in the total output; and, in addition, the processing unit,

the label used in the supervision network training is obtained by performing SoftMax calculation according to the loss value l _i Specific gravity vector of (2)Wherein->The value of the internal element->，，

Will beA SoftMax calculation is performed as input before mapping for +.>Sub-networks, confidence level of prediction success of each sub-network is estimated by using SoftMax>As shown in formula (1): （1）

wherein,and->And has negative correlation.

6. The e-commerce data classification recommendation system of claim 5, further comprising:

by the loss valueA set of forward parameters is calculated again>So that the output of each sub-network is optimized, wherein->Element value in vector +.>，As shown in formula (2):

（2）

wherein,and->And shows positive correlation.

7. An electronic commerce data classification recommendation method based on classification recommendation of electronic commerce data by the electronic commerce data classification recommendation system according to any one of claims 1 to 6, comprising:

training and learning the training data through a preset LAFAMN to obtain each type of output and label, calculating the loss value of the output and label, and carrying out feedback processing after weighting and summing the loss value to enable the loss value to reach the preset requirement to obtain a LAFAMN model; wherein the training data comprises the first training data and the feature pre-processing data, and the LAFAMN comprises a supervisory network and a preset number of sub-networks; the sub-network is used for training and learning the feature preprocessing data, and the supervision network is used for training and learning the first training data; the supervision network learns the duty ratio weight of the sub-network, and the sub-network completes self-learning of the sub-network weight and the characteristic weight by dynamically adjusting the total loss; and processing a highly unbalanced data set in the subnetwork using a double-hold loss function;

8. An electronic device, the electronic device comprising:

at least one processor; the method comprises the steps of,

a memory communicatively coupled to the at least one processor; wherein,

the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the steps in the e-commerce data classification recommendation method of claim 7.