CN117556147B - Electronic commerce data classification recommendation system and method - Google Patents

Abstract

The present invention provides an e-commerce data classification recommendation system and method. Through the loss-aware feature attention mechanism network LAFAMN, the supervision network therein is used to weight different input feature groups, and the weights are self-regulated according to the prediction results of the network as a whole and each sub-network, so as to effectively solve the feature imbalance problem in the e-commerce data classification recommendation process; in addition, based on the double suppression loss function regulated by the classification confidence, the loss value of the large-class easy-to-judge samples is greatly suppressed, while ensuring that the loss value of the small-class difficult-to-judge samples is not reduced at all, so as to solve the category imbalance problem in two dimensions of the difficulty of judgment and the imbalance of the number of samples. The joint use of LAFAMN and the double suppression loss function can form a complete recommendation system, solve the class imbalance classification problem in three dimensions, and provide a complete solution process for the commodity recommendation problem.

Description

E-commerce data classification recommendation system and method

Technical Field

The present invention relates to the field of intelligent recommendation under the field of artificial intelligence technology, and more specifically, to an e-commerce data classification recommendation system and method.

Background Art

With the rapid development of the Internet, the coverage of the e-commerce industry is getting wider and wider. Almost all e-commerce platforms are using product recommendation systems to help customers quickly find what they are interested in. For e-commerce platforms, an effective recommendation algorithm can increase the profits of e-commerce platforms and reduce labor costs. However, due to the large number of e-commerce products, the small number of display locations, and the small number of users browsing products, there are class imbalance classification problems in different dimensions in the recommendation system, including: imbalance in the number of samples between categories, imbalance in the difficulty of sample discrimination, and imbalance in the number of sample features.

In the research field, the category with fewer samples is generally called the small category, and vice versa, the category with more samples is called the large category. In this regard, the traditional classification method aims to minimize the error rate to establish a classifier. Due to the small number of samples in the small category and the more complex feature distribution, the network tends to classify all samples as large categories. This classification method greatly reduces the generalization performance of the small category.

Therefore, analyzing the reasons why class imbalance in the e-commerce recommendation field causes reduced algorithm performance and exploring solutions to the class imbalance problem in recommendation algorithms has become one of the current research directions in the field of artificial intelligence recommendation.

Summary of the invention

In view of the multi-dimensional class imbalance problem of commodity data in the current recommendation system in the field of e-commerce, the purpose of the present invention is to provide an e-commerce data classification recommendation system and method, which uses the attention mechanism to integrate and improve the traditional multi-expert learning network to construct a LAFAM network, and based on the double suppression loss function (Suppression Loss) regulated by the classification confidence, suppresses the contribution of the loss value of large-category samples and easy-to-judge samples, while ensuring that the loss value of small-category difficult-to-judge samples is not reduced at all to solve the class imbalance problem in the existing recommendation algorithm.

The present invention provides an e-commerce data classification recommendation system, comprising:

A feature preprocessing unit, used to sort the first training data by importance and classify the features by a preset sorting method and a preset classification method, so as to obtain feature preprocessing data of the first training data;

A model training unit is used to perform learning and training processing on the training data through a preset LAFAM network to obtain each type of output and label, calculate the loss value of the output and label, and then return the weighted sum of the loss value to obtain a LAFAMN model after the loss value reaches the preset requirement; wherein, the training data includes the first training data and the feature preprocessing data, and the LAFAM network includes a supervision network and a preset number of sub-networks; the sub-network is used to train and learn the feature preprocessing data, and the supervision network is used to train and learn the first training data; and, the supervision network learns the proportion weight of the sub-network, and the sub-network completes the self-learning of the sub-network weight and the feature weight by dynamically adjusting the total loss; and, in the sub-network, a double suppression loss function is used to process highly unbalanced data sets;

The classification recommendation unit is used to perform classification recommendation processing on the e-commerce data through the LAFAMN model.

Among them, an optional solution is that the feature preprocessing unit includes:

An importance ranking unit, configured to perform a primary ranking process on the first training data by a preset ranking method, and weight the primary ranking process result to calculate a feature importance score, so as to obtain a final ranking result of the first training data;

The feature classification unit is used to classify the first training data according to a preset classification method.

Among them, the optional scheme is that the preset sorting method includes PS-smart, XGBoost, and GBDT; the first training data is divided into dense features, sparse features, time series features, and basic features through the preset classification method; wherein the basic features are other features that are not in the three categories of dense features, sparse features, and time series features after classification; and the subnetwork performs learning and training processing on different categories of input features respectively.

Among them, an optional solution is that the sub-network performs learning and training processing on different categories of input features respectively, including:

Input features of different categories after feature preprocessing of the first training data are respectively put into different sub-networks for learning, and all sub-networks output a score between 0 and 1, which represents the learning result of the current sample in the current sub-network; 0 means the sample is least likely to be recommended and 1 means the sample is most likely to be recommended;

Each sub-network will calculate a loss value through the sample's label value and prediction value. , in, is the total number of subnetworks; the loss value The larger the value of , the lower the corresponding sub-network's prediction confidence for the sample, and the lower its share in the final calculation of the output value.

Among them, an optional solution is that the output of the supervision network is the proportion of all sub-networks in the total output; and,

The label used in the supervised network training is the weight vector of the loss value l _i obtained after performing SoftMax calculation based on the loss value ,in, Internal element value , ,

Will As the input before mapping, SoftMax calculation is performed. sub-networks, and use SoftMax to estimate the confidence of its prediction success for each sub-network , as shown in formula (1): (1)

in, and Negatively correlated.

Among them, an optional solution is to use the loss value Calculate another set of forward parameters , so that the output of each sub-network is optimal, where The value of the element in the vector , , as shown in formula (2):

(2)

in, and There is a positive correlation.

The present invention also provides an e-commerce data classification recommendation method, which performs data class imbalance classification based on the e-commerce data classification recommendation system as described above, comprising:

Sorting the first training data by importance and classifying its features by a preset sorting method and a preset classification method to obtain feature preprocessing data of the first training data;

The training data is trained and learned through a preset LAFAM network to obtain each type of output and label. After the loss value of the output and label is calculated, the weighted sum of the loss value is returned for processing, so that the loss value reaches the preset requirement to obtain the LAFAMN model; wherein, the training data includes the first training data and the feature preprocessing data, and the LAFAM network includes a supervisory network and a preset number of sub-networks; the sub-network is used to train and learn the feature preprocessing data, and the supervisory network is used to train and learn the first training data; and, the supervisory network learns the proportion weight of the sub-network, and the sub-network completes the self-learning of the sub-network weight and the feature weight by dynamically adjusting the total loss; and, in the sub-network, a double suppression loss function is used to process highly unbalanced data sets;

The LAFAMN model is used to perform classification and recommendation processing on e-commerce data.

The present invention further provides an electronic device, comprising:

at least one processor; and,

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

The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can perform the steps in the e-commerce data classification recommendation method as described above.

From the above technical solutions, it can be seen that the e-commerce data classification recommendation system and method provided by the present invention firstly uses the loss-aware feature attention mechanism LAFAMN to weight different input feature groups using a supervised network, and self-regulates the weights according to the prediction results of the network as a whole and each sub-network, effectively solving the feature imbalance problem; in addition, the double suppression loss function based on the classification confidence regulation greatly suppresses the loss value of large-category easy-to-judge samples, while ensuring that the loss value of small-category difficult-to-judge samples is not reduced at all, so as to solve the category imbalance problem in two dimensions: difficulty of judgment and imbalance of sample quantity. The joint use of LAFAMN and the double suppression loss function can form a complete recommendation system, solve the class imbalance classification problem in three dimensions, and provide a complete solution process for the product recommendation problem.

BRIEF DESCRIPTION OF THE DRAWINGS

By referring to the following description in conjunction with the accompanying drawings, and with a more comprehensive understanding of the present invention, other objects and results of the present invention will become more apparent and easy to understand. In the accompanying drawings:

FIG1 is a schematic diagram of the logical structure of an e-commerce data classification recommendation system according to an embodiment of the present invention;

FIG2 is a schematic diagram of a framework of an e-commerce data classification recommendation system according to an embodiment of the present invention;

FIG3 is a schematic diagram of a theoretical structure of a LAFAM network according to an embodiment of the present invention;

FIG4 is a flow chart of an e-commerce data classification recommendation method according to an embodiment of the present invention;

FIG5 is a schematic diagram of an application structure of a LAFAM network in an example according to an embodiment of the present invention;

FIG6 is a schematic diagram of a grid search of a LAFAM network and a dual suppression loss function in an example according to an embodiment of the present invention;

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

DETAILED DESCRIPTION

In the following description, for the purpose of illustration, in order to provide a comprehensive understanding of one or more embodiments, many specific details are set forth. However, it is apparent that these embodiments may also be implemented without these specific details. In other examples, for ease of describing one or more embodiments, known structures and devices are shown in the form of block diagrams.

In view of the data class imbalance problem existing in the existing e-commerce product recommendation schemes, the present invention provides an e-commerce data classification recommendation system and method, which integrates the attention mechanism and the network architecture of hybrid multiple experts, constructs a double suppression loss function by using function suppression and power suppression, explores the influence of different networks and different loss functions on the performance of the recommendation system, and realizes the construction of a recommendation system model for data class imbalance. It is applied to the product recommendation of e-commerce to solve the problems of imbalance in the number of categories, imbalance in the difficulty of judgment, and imbalance in features in three dimensions.

In order to better illustrate the technical solution of the present invention, some technical terms involved in the present invention are briefly explained below.

GBDT (Gradient Boosting Decision Tree) is an additive model based on the idea of Boosting ensemble learning. It trains a series of weak learners (usually decision trees) step by step and iteratively. Each iteration attempts to correct the error of the previous iteration, and finally combines these weak learners into a strong learner. The GBDT model is highly interpretable and has good application effects. It is widely used in data mining, computational advertising, recommendation systems and other fields.

PS-Smart regression, parameter server PS (Parameter Server) is committed to solving large-scale offline and online training tasks, SMART (Scalable Multiple Additive Regression Tree) is an iterative algorithm implemented by GBDT based on PS. PS-smart can support training tasks with tens of billions of samples and hundreds of thousands of features, can run on thousands of nodes, and has failover capabilities and good stability; at the same time, PS-Smart supports multiple data formats, training targets and evaluation targets, as well as output feature importance, and includes optimizations such as histogram approximation to accelerate training.

XgBoost (eXtreme Gradient Boosting) is a type of synthetic algorithm that combines basis functions and weights to form a good data fitting effect. Unlike GBDT, XgBoost adds a regularization term to the loss function and uses the second-order Taylor expansion of the loss function as the fitting of the loss function. It is very powerful in parallel computing efficiency, missing value processing, and prediction performance. It also performs well in preventing overfitting and improving generalization ability, and can quickly and accurately solve many data science problems.

Dense features refer to features in which most elements are non-zero. This type of feature is usually used for data types such as images and audio. Dense features are slower to calculate, but they can provide more information and therefore perform better in some tasks.

Sparse features refer to features in which only a few elements are non-zero. This type of feature is usually used for data types such as text and recommendation systems. Sparse features are faster to calculate.

Time series features are features extracted or counted based on time series.

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

The 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 embodiments is merely illustrative and is not intended to limit the present invention and its application or use. Technologies and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such technologies and devices should be considered part of the specification.

In order to illustrate the e-commerce data classification recommendation system and method provided by the present invention, Figures 1, 2, 3, 4, 5, and 6 exemplarily illustrate the logical structure of the e-commerce data classification recommendation system, the system architecture, the theoretical structure of the LAFAM network, the flow chart of the e-commerce data classification recommendation method, the application structure of the LAFAM network in the example, and the grid search of the LAFAM network and the dual suppression loss function in the example; Figure 7 exemplarily illustrates the e-commerce data classification recommendation method of the embodiment of the present invention.

The e-commerce data classification recommendation system provided by the present invention is based on the LossAware Feature Attention Mechanism Network (LAFAMN) and is implemented in combination with a classification recommendation system framework of a double suppression loss function to jointly solve the problems of imbalanced number of categories, imbalanced difficulty of discrimination, and imbalanced features. For the convenience of expression, in the following embodiments of the present invention, the LAFAM network is also referred to as a network.

1 and 2 respectively show the logical structure and system architecture of the e-commerce data classification recommendation system according to an embodiment of the present invention.

As shown in both FIG. 1 and FIG. 2 , the e-commerce data classification recommendation system 100 provided by the present invention mainly includes three parts: a feature preprocessing unit 110, a model training unit 120 and a classification unit 130. The corresponding system architecture is a universal procedural system, which is divided into three stages: feature preprocessing, LAFAMN model training and classification.

The feature preprocessing unit 110 is used to sort the first training data by importance and classify the features by a preset sorting method and a preset classification method to obtain feature preprocessing data of the first training data;

The model training unit 120 is used to perform learning and training processing on the training data through a preset LAFAM network to obtain each type of output and label, calculate the loss value of the output and label, and then perform weighted summation of the loss value and then return the loss value to obtain a LAFAMN model after the loss value reaches a preset requirement;

The classification unit 130 is used to perform classification and recommendation processing on the e-commerce data through the LAFAMN model.

The above three parts will be described in detail below in conjunction with specific embodiments.

Before feature preprocessing, it is necessary to first collect training data (sample collection). In the present invention, a part of the collected training data is used as the first training data for learning and training. After feature preprocessing, the first training data is input into the model training unit sub-network for learning and training. For the supervision network of the model training unit, all the first training data are learned and trained.

In addition, since the solution of the present invention is mainly aimed at the multi-dimensional class imbalance problem existing in the commodity data in the current recommendation system in the e-commerce field, in the embodiment shown in FIG2 , the collected training data includes item features, coupons, commissions, history, prices and other data related to the commodity data in the e-commerce field.

The collected training data can be centrally stored in a sample database, and after the sample collection is completed, part of the training data can be preprocessed by the feature preprocessing unit 110. The present invention uses a variety of preset algorithms to complete the sorting and classification of part of the training data to replace the feature selection and fusion in the traditional classification method.

In a specific 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 used to perform primary ranking processing on the first training data by a preset ranking method, and weight the primary ranking processing result to calculate the feature importance score to obtain the final ranking result of the first training data; the feature classification unit 112 is used to classify the first training data according to the preset classification method.

Specifically, as an example, in the sorting part, PS-smart, XGBoost, and GBDT are used to sort the feature importance. The feature importance scores are weighted based on the sorting results given by the three methods, and the final sorting is given. In addition, in order to improve the sorting efficiency, a certain degree of truncation can be performed on the sorting given in this process, and features with a contribution of almost 0 can be discarded. The specific truncation position should be adjusted according to the specific problem. In the feature classification part, in order to better cooperate with the operation of the LAFAM network, in a specific embodiment of the present invention, the features are divided into basic features, dense features, sparse features, and time series features, and different sub-networks are used to process different features. The basic features are other features that are not in these three categories of features (dense features, sparse features, and time series features) after classification.

The model training unit 120 performs learning and training processing on the training data through a preset LAFAM network. The LAFAM network includes a supervision network and a preset number of sub-networks, wherein the sub-network is used to train and learn the features preprocessed by the feature preprocessing unit 110, and the supervision network is used to train and learn all training data that have not been preprocessed, wherein the supervision network learns the weight of the sub-network.

During the training process of the LAFAMN model, all the data obtained above are first input for learning. The sorted and classified feature data are input into each corresponding sub-network, and all the features that have not been pre-processed are input into the supervision network, which will learn the weight of the sub-network. In this way, the training method of the LAFAMN model solves the problem of feature imbalance and allocates suitable networks for different types of features; and in the present invention, the self-learning of sub-network weights and feature weights is completed by dynamically adjusting the total loss, so that the network can make each feature make its greatest contribution to the prediction result without discarding any feature.

In addition, in order to solve the class imbalance problem, the present invention uses a dual suppression loss function in an appropriate sub-network to process highly unbalanced data sets. By combining the LAFAM network and the dual suppression loss function, the three problems of feature imbalance, quantity imbalance, and discrimination difficulty imbalance are perfectly solved.

In the classification recommendation unit, the e-commerce data is classified and recommended using the LAFAMN model.

In addition, in a specific embodiment of the present invention, after the LAFAMN model is obtained through model training, the parameters of the LAFAM network are also searched by grid search. , and And the parameters of the double suppression loss function , , , And optimize, combine the value range of each parameter, generate a parameter grid and select the best combination of the accuracy of network classification recommendation and F-measure parameters to obtain the classification threshold, and optimize the parameters of the trained LAFAMN model according to the classification threshold. Specifically, as an embodiment, in the classification recommendation process, the present invention refers to the products whose daily sales volume ranks at the top of the industry and total sales volume as "hot products", that is, the products have a relatively hot sales market. At the same time, according to the LAFAMN model score, the possibility of the products being hot is ranked. In the classification process, the present invention also pays attention to the top 50/100/200 winning rates (HR@50, HR@100 and HR@200) parameters used in the project to maximize them.

FIG3 is a schematic diagram of the theoretical structure of a LAFAM network according to an embodiment of the present invention.

In order to solve the problem of feature imbalance, the present invention proposes a loss-aware feature attention mechanism network LAFAMN. The theoretical structure of LAFAMN is shown in FIG3 .

In the LAFAMN shown in FIG3 , first, multiple subnetworks are set up to learn features. The number and type of subnetworks can be adjusted and changed according to different problems. The first three subnetworks in FIG3 are shown as multiple different subnetworks.

The input features of different categories after feature preprocessing of the first training data are respectively put into different sub-networks for learning. All sub-networks output a score between 0 and 1, which represents the learning result of the current sample in the current network, 0 is the least likely to be recommended, and 1 is the most likely to be recommended. Each sub-network will calculate a loss value based on the label value and prediction value of the sample. , , where is the difference between the output prediction value of the ith subnetwork and the current sample label value, and N is the total number of subnetworks (excluding the supervisory network). This loss value is not only fed back as a loss function, but more importantly, it is used as a criterion for determining the output confidence. The larger the value, the lower the prediction confidence of the subnet for the sample, and the network should strengthen its learning of the current subnet, so it should be given a greater weight in the return loss. The larger the value, the higher the error probability of its prediction. Therefore, when the output value is finally calculated, its proportion should be lower. Secondly, according to the above characteristics, the present invention sets a supervision network, as shown in the bottom sub-network in Figure 3. The output of this network is the proportion of all sub-network outputs in the total output. The label used in its training is the one obtained after the SoftMax calculation based on the loss value. The weight vector , where N is the total number of sub-networks (excluding the supervision network), and Internal element value , .

The SoftMax function is generally used to solve multi-classification problems and map the outputs of multiple neurons to a range. The present invention uses its characteristic of exponentially mapping the predicted values of multiple categories to As the input before mapping, SoftMax calculation is performed. sub-networks, and use SoftMax to estimate the confidence of its prediction success for each sub-network , as shown in formula (1):

(1)

right In terms of loss The larger the value, the lower the confidence level of its prediction success, that is, the greater the gap between the predicted value and the label value. The smaller it is, the lower the contribution of the sub-network to the overall output of the network. and Negatively correlated. However, if the confidence level is used alone to determine its proportion in the output, a phenomenon will occur, that is, as long as one sub-network has a relatively good prediction result, the entire network will almost completely rely on the result of this network for prediction, which will cause poor robustness and stability of the network. Therefore, the present invention hopes that each sub-network in the LAFAM network will do its best to deliver, that is, in an overall learning, the output of each sub-network will be optimized. Therefore, through Calculate another set of forward parameters , where N is the total number of sub-networks (excluding the supervision network), and The value of the element in the vector , , as shown in formula (2):

(2)

right In terms of loss The larger the value, the lower the confidence level of its prediction success, that is, the greater the gap between the predicted value and the label value. The larger it is, the higher the contribution of the sub-network to the overall loss value of the network is expected to be. and There is a positive correlation.

As shown in Figure 3, in the LAFAM network, there are three types of outputs and labels, namely:

Overall network output and the label value of the original sample ;

Output of the subnetwork and the label value of the original sample ;

Output of the supervised network And the label value obtained by subnet loss SoftMax .

Each of the three types of outputs needs to calculate the loss value separately. In this embodiment, the loss values of the three types of outputs are recorded as , and , the loss is calculated by the output value and label value of each class. After obtaining the three loss values, they are weighted and summed and then transmitted back so that each type of network can obtain the corresponding loss value and perform optimization. Therefore, the total loss function is defined in this embodiment (return all), in order to calculate and prove the feedback coefficient, the square loss function is used as an example. The specific process is shown in formula (3):

(3)

in, , and is the parameter of the LAFAM network, which can be adjusted and set using grid search or other parameter adjustment methods. y _j is the output of the jth sub-network; is the temperature coefficient used in the SoftMax operation, You can adjust the gap between the values after SoftMax calculation. Generally, . The smaller it is, the steeper the SoftMax curve is, and the gap between the values output after the operation will be larger, which will amplify the similarity and make it easier to distinguish difficult samples; on the contrary, The larger the value, the smoother the SoftMax curve, and the smaller the difference between the values output after the operation. To control the loss of the sub-network Weighted balance.

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

(4)

in, is the parameter in the output vector given by the supervisory network. The network diagram shown in Figure 3 is the entire process required for the LAFAM network to be trained and predicted. During training, the LAFAM network will calculate three outputs separately, calculate the loss for each output and send it back. However, in the final prediction process, the parameters output by the supervisory network are directly used for the weighted calculation of the total output to obtain the LAFAM network output y . All dotted parts in Figure 3 are no longer run.

For the imbalance problems in two dimensions, namely, imbalance in the number of samples between categories and imbalance in the difficulty of sample discrimination, existing in existing classification schemes, the present invention also proposes a double suppression loss function (Suppression Loss) to optimize and solve them.

Normally, data imbalance in classification problems is often manifested in that the number of large-category data is far greater than that of small-category data. At this time, if the global loss is directly calculated in a general way (such as square loss, cross entropy loss, etc.), the training loss of large-category samples will dominate. In the scenario of intelligent recommendation of electronic products, the best-selling products are small-category products, and finding the best-selling products is an important research goal. In order to solve the problem of imbalanced sample quantity, a function is used to assign a lower loss contribution to samples that are closer to large categories, while trying to ensure that the loss of samples close to small categories is not attenuated, which can facilitate better optimization and learning.

This idea can also be used to solve the problem of imbalanced discrimination. The larger the gap between the sample prediction output value and the label value, the farther the prediction value is from the label value, the greater the error rate, and the lower the prediction confidence. Such samples are often small in number, but they are the focus and difficulty of prediction. Therefore, it is necessary to increase the contribution of samples with a large gap between the prediction output value and the label value to the total loss, while suppressing the contribution of samples with high confidence to the total loss.

In summary, the dual suppression loss function proposed in the embodiment of the present invention consists of three parts. The first part is to suppress the loss contribution of large-class samples by means of functions; the second part is to suppress the contribution of easy-to-separate samples by using the same function and different parameters; the last part is to suppress the contribution of easy-to-separate samples by expanding the output gap through a high-power function, and its form is shown in formula (5):

(5)

in, is the final prediction output of the network; is the label value of the original sample, which takes the value 0 or 1; is the distance between the sample prediction output value and the label value, which is called prediction deviation in this embodiment ; , , , and is the parameter of the double suppression loss function, which can be set by adjusting parameters such as grid search method in combination with the actual data set problem.

In the specific application process, the LAFAMN model proposed in the present invention can calculate the probability of a product becoming a hot-selling product. In order to evaluate the effectiveness of the LAFAMN model, in the regression problem, four evaluation indicators are used to evaluate the performance of the LAFAMN model provided by the present invention. The four evaluation indicators are Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Weighted Mean Absolute Percentage Error (WMAPE), and Hire Rate @50/100/200. Among them, the first three evaluation indicators consider the model from the regression algorithm side, and the last evaluation indicator considers the model from the perspective of the accuracy of product prediction and ranking. In the classification problem, the present invention uses three evaluation indicators, namely Precision, Accuracy, and F-measure.

The RMSE indicator can better reflect the impact of extreme values on errors. Its calculation method is shown in formula (6):

(6)

Where n is the number of samples, is the predicted model output, is the actual label value.

The MAE indicator can very intuitively show the regression error. Its calculation method is shown in formula (7):

(7)

The WMAPE indicator is less affected by extreme values and individuals, and can show the overall prediction of the network more evenly. Its calculation method is shown in formula (8):

(8)

The HR indicator can reflect the accuracy of the model recommendation, that is, whether the user's demand items are included in the model's recommended items. Its calculation method is shown in formula (9):

(9)

in, is the set of all tests, It is the sum of the number of items in each user's top-k list that belong to the test set.

The precision index can effectively reflect the precision rate of small categories, and its calculation method is shown in formula (10):

(10)

in, It indicates the number of samples that are predicted to be positive and are actually positive. Indicates the number of samples predicted to be negative but actually positive. Indicates the number of samples predicted to be positive but actually negative. Indicates the number of samples that are predicted to be negative and are actually negative.

The Accuracy indicator reflects the global accuracy and can ensure that the network does not make biased predictions. Its calculation method is shown in formula (11):

(11)

The F-measure indicator is used to measure the overall performance of the algorithm in the data class imbalance classification problem. Its calculation method is shown in formula (12) and formula (13):

(12)

(13)

in, It is a coefficient that adjusts the ratio of Precision to Recall, and is usually set to 1.

For the three indicators of RMSE, MAE and WMAPE, the smaller the value, the better the recommendation system. For the four indicators of HR@50/100/200, Precision, Accuracy and F-measure, the larger the value, the better the classification recommendation system.

As shown in FIG. 4 , the present invention further provides an e-commerce data classification recommendation method, which performs classification recommendation of e-commerce data based on the e-commerce data classification recommendation system 100 as described above, including:

S1: sorting the importance and classifying the features of the first training data by a preset sorting method and a preset classification method to obtain feature preprocessing data of the first training data;

S2: The training data is trained and processed through the preset LAFAM network to obtain each type of output and label, and after the loss value of the output and label is calculated, the weighted sum of the loss value is returned for processing, so that the loss value reaches the preset requirement to obtain the LAFAMN model; wherein, the training data includes the first training data and the feature preprocessing data, and the LAFAM network includes a supervision network and a preset number of sub-networks; the sub-network is used to train and learn the feature preprocessing data, and the supervision network is used to train and learn the first training data; and, the supervision network learns the proportion weight of the sub-network, and the sub-network completes the self-learning of the sub-network weight and the feature weight by dynamically adjusting the total loss; and, in the sub-network, a double suppression loss function is used to process highly unbalanced data sets;

S3: Classify and recommend e-commerce data using the LAFAMN model.

The above-mentioned e-commerce data classification recommendation method is an implementation method corresponding to the above-mentioned e-commerce data classification recommendation system. Its specific execution steps can refer to the specific implementation example of the above-mentioned e-commerce data classification recommendation system, and will not be described in detail here.

It can be seen from the above embodiments that the loss-aware feature attention mechanism network (LAFAMN) proposed in the present invention uses the attention mechanism to integrate and improve the traditional multi-expert learning network for the multi-dimensional class imbalance problem existing in the recommendation system in the current e-commerce field, namely feature imbalance, imbalance in discrimination difficulty and imbalance in sample quantity. Complementarily, a double suppression loss function (SuppressionLoss) based on classification confidence regulation is also proposed to suppress the contribution of the loss value of large-class samples and easy-to-judge samples, while ensuring that the loss value of small-class difficult-to-judge samples is not reduced at all. The combined application of the LAFAM network and the double suppression loss function can form a complete recommendation system to solve the above three-dimensional class imbalance classification problem. Experiments have shown that the e-commerce data classification recommendation system and method proposed in the present invention have significant improvements in regression problem indicators and classification problem indicators compared with existing recommendation algorithms.

The e-commerce data classification recommendation system provided by the present invention will be described in more detail below with reference to a specific application example.

In terms of the construction of the sample database, in this embodiment, an e-commerce data set from an e-commerce company is used for experimental verification, which is recorded as the Tmall data set in the subsequent description. The total storage size of the data set is 10.7GB, and it contains a total of 51,134,193 rows of data. There are 453 original data features. Since many features are missing or incorrectly filled, in the experiment, this embodiment selects 286 features that have been trusted in the system algorithm (after more than 1 year of verification, the system fills in the data with a correct rate greater than 90%) for experimental use.

Since the final output value given by the LAFAMN model is a continuous value within the range of values, this embodiment first considers this problem as a regression problem for experimental verification. In this process, the RMSE, MAE, WMAPE, and HR@50/100/200 parameters are used to verify the effect and verify the score ranking given by the network. However, from the perspective of practical application, whether a product is a hot-selling product or not is an absolute judgment. Therefore, in order to make the prediction problem practical, the present invention also converts the sorting regression problem into a classification problem for learning and verification. After normalizing the sales volume in the data set, 0.15 is selected as the threshold for binary classification. After classification, the number of categories 0 (major category, non-hot-selling product) and 1 (minor category, hot-selling product) is 70,184,766 and 20,483 respectively, and the imbalance is 3426:1.

In this embodiment, the 286 features in the Tmall data set are divided into basic features, price features, discount features and time series features. Price features refer to 41 features such as the absolute value of the price of the current product, the average price under the leaf category, the average price of competing products, the lowest historical price and its corresponding sales volume, the highest historical price and its corresponding sales volume. Discount features refer to 28 features such as the discount value given to the current product, the historical average discount value, the average discount value of competing products, the highest historical discount value and its corresponding sales volume, the lowest historical discount value and its corresponding sales volume. Time series features refer to 10 features such as the daily price, daily sales volume, and daily product rating of the product within 30 days. The 207 features other than the above three features belong to the basic features. The features of the above four categories will cover the entire 286 features, each feature belongs to and only belongs to one category, and the features belonging to each category do not overlap or repeat.

Therefore, the LAFAM network structure used in the example of the present invention is shown in Figure 5, and its network parameters correspond to the parameter names in Figure 3 one by one, showing the state of the time, that is, setting four sub-networks and a supervision network for learning. The outermost solid line frame in Figure 5 is the training model, and the dotted line frame is the prediction model. Features are input into the four sub-networks by category, and all features are learned using the deep cross network DCN commonly used in recommendation algorithm engineering. It consists of two parts, Deep and Cross. The Deep part is a fully connected network for extracting the depth characteristics of features, and the Cross part performs feature cross calculations for extracting the breadth characteristics of features. It can be well compatible with vertical category recommendations and extensive hobby explorations, and its operation efficiency is high, so it is used as a learning network for all features. Price features and discount features use a simple fully connected neural network, which can extract high-dimensional features of samples for recommendation prediction. The time series features are learned using a gated recurrent neural network GRU, which has a simple structure and can overcome the long dependency problem of the RNN network.

In this embodiment, the square loss function is used as the basic return loss function of the network to verify the performance of the LAFAM network in the regression problem. The performance is compared with other traditional networks (such as GBDT, DCN, and MoE). The experimental results are shown in Table 1:

Table 1 Comparison of performance data of different networks in regression problems

First, we conduct a macro analysis of Table 1. We can see that the LAFAM network has obvious advantages in the HR@50/100/200 parameters. At the same time, in terms of RMSE, MAE, and WMAPE, the LAFAM network has no obvious disadvantages compared to other networks, and even occupies the best position in MAE and WMAPE parameters. The RMSE parameter is reduced because in highly unbalanced data sets, large categories of samples contribute a large number of error values. Therefore, from the results, when considering regression problems, the LAFAM network can well optimize the problem of reduced prediction accuracy caused by feature imbalance and data category imbalance, and has a certain prediction success rate advantage compared to other traditional networks.

Secondly, we conduct a detailed analysis of the three winning rate parameters. First, we compare different parameters of the same algorithm, that is, the horizontal comparison of the table. It can be seen that from HR@200→HR@100→HR@50, that is, from the widest to the most precise, the accuracy of GBDT is constantly decreasing, and the LAFAM network is constantly improving. In other words, the LAFAM network is more accurate in top prediction, that is, the identification of small class samples. Then we compare different algorithms with the same parameters, that is, the vertical comparison of the table. It can be seen that the LAFAM network is 8.4% higher than the second highest MoE network in HR@200, 10.4% higher than the MoE network in HR@100 parameters, and 34.8% higher than the MoE network in HR@50 parameters. The more precise the range, the better the performance of LAFAMN compared with other networks. It has obvious advantages in predicting small class samples and can complete regression sorting and recommendation well in the case of class imbalance.

The following is the classification verification of the LAFAM network and the double suppression loss function Suppression Loss.

First, the grid search method is used to determine the network parameters. The grid search optimization of some parameters is shown in Figure 6, where the black dots are the selected parameter values. The final network parameters are shown in Table 2. The table only shows the optimal parameters of the proposed LAFAM network and the dual suppression loss function Suppression Loss. The remaining networks and loss functions are adjusted in the same way to obtain the optimal solution.

Table 2 Summary of parameter settings

The square loss function, focal loss function, shrinkage loss function, and double suppression loss function were verified under different networks. In terms of comparison network selection, the MLP, GRU, DCN, MoE, and LAFAM networks were selected for verification. The first four are sub-networks used in the LAFAM network, so they are tested separately. While serving as baseline indicators, they are also equivalent to completing the ablation experiment, proving that the joint network has corresponding improvements and effect optimization compared to each sub-network. The experimental results are shown in Table 3. Each column in the table is a comparison of the verification results of four different loss functions under the same network, and the corresponding order in each column is a comparison of the verification results of different networks with the same loss function.

Table 3 Comparison of experimental results of different loss functions in different networks

From Table 3, we can see that in terms of network, LAFAM network has a more obvious improvement on highly unbalanced datasets than other networks. In the 1:20 unbalanced dataset, the three parameters of LAFAM network are in the optimal position, which is about 10% higher than other networks on average, and only about 7% lower than the optimal network on average in the 1:5 and 1:10 unbalanced datasets. Therefore, LAFAM network is stable and has obvious advantages when used in class imbalance and feature imbalance datasets. In terms of loss function, it can be clearly found that in each network, the double suppression loss function Suppression Loss has the best performance, because it can well solve the problems of unbalanced number of categories and unbalanced sample discrimination confidence. The experimental effect of the double suppression loss function is about 15% higher than other loss functions on average, which verifies the performance on a variety of networks and unbalanced datasets, so its adaptability is also relatively good. From the experimental data, the stability of the double suppression loss function is acceptable.

As shown in FIG. 7 , the present invention further provides an electronic device, the electronic device comprising:

at least one processor; and,

a memory communicatively connected to at least one processor; wherein,

The memory stores a computer program that can be executed by at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the steps in the aforementioned e-commerce data classification recommendation method.

Those skilled in the art will appreciate that the structure shown in FIG. 7 does not limit the electronic device 1 and may include fewer or more components than shown in the figure, or combine certain components, or arrange the components differently.

For example, although not shown, the electronic device 1 may also include a power source (such as a battery) for supplying power to various components. Preferably, the power source may be logically connected to the at least one processor 10 through a power management device, so that the power management device can realize functions such as charging management, discharging management, and power consumption management. The power source may also include any components such as one or more DC or AC power sources, recharging devices, power failure detection circuits, power converters or inverters, and power status indicators. The electronic device 1 may also include a variety of sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

Furthermore, the electronic device 1 may also include a network interface. Optionally, the network interface may include a wired interface and/or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), which is generally used to establish a communication connection between the electronic device 1 and other electronic devices.

Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), or a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, and an OLED (Organic Light-Emitting Diode) touch device. The display may also be appropriately referred to as a display screen or a display unit, which is used to display information processed in the electronic device 1 and to display a visual user interface.

It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.

The e-commerce data classification recommendation program 12 stored in the memory 11 of the electronic device 1 is a combination of multiple instructions. When running in the processor 10, it can achieve:

Sorting the first training data by importance and classifying its features by a preset sorting method and a preset classification method to obtain feature preprocessing data of the first training data;

The training data is trained and learned through a preset LAFAM network to obtain each type of output and label, and after the loss value of the output and label is calculated, the weighted sum of the loss value is returned for processing, so that the loss value reaches the preset requirement to obtain the LAFAMN model; wherein, the training data includes the first training data and the feature preprocessing data, and the LAFAM network includes a supervisory network and a preset number of sub-networks; the sub-network is used to train and learn the feature preprocessing data, and the supervisory network is used to train and learn the first training data; and, the supervisory network learns the proportion weight of the sub-network, and the sub-network completes the self-learning of the sub-network weight and the feature weight by dynamically adjusting the total loss; and, in the sub-network, a double suppression loss function is used to process highly unbalanced data sets;

The LAFAMN model is used to perform classification and recommendation processing on e-commerce data.

Specifically, the specific implementation method of the processor 10 for the above instructions can refer to the description of the relevant steps in the corresponding embodiment of Figure 4, which will not be repeated here.

Furthermore, if the module/unit integrated in the electronic device 1 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

As described above, the e-commerce data classification recommendation system and method proposed in the present invention are described by way of example with reference to the accompanying drawings. However, those skilled in the art should understand that various improvements can be made to the e-commerce data classification recommendation system and method proposed in the present invention without departing from the content of the present invention. Therefore, the protection scope of the present invention should be determined by the content of the attached claims.

Claims (8)

1. An e-commerce data classification recommendation system, characterized by comprising: A feature preprocessing unit, used to sort the first training data by importance and classify the features by a preset sorting method and a preset classification method, so as to obtain feature preprocessing data of the first training data; A model training unit is used to perform learning and training processing on the training data through a preset Loss Aware Feature Attention Mechanism Network (LAFAMN) to obtain each type of output and label, and after calculating the loss value of the output and label, the weighted sum of the loss value is returned for processing, so that the loss value reaches the preset requirement to obtain the LAFAMN model; wherein, the training data includes the first training data and the feature preprocessing data, and the LAFAMN includes a supervision network and a preset number of sub-networks; the sub-network is used to train and learn the feature pre-processing data, and the supervision network is used to train and learn the first training data; and, the supervision network learns the proportion weight of the sub-network, and the sub-network completes the self-learning of the sub-network weight and the feature weight by dynamically adjusting the total loss; and, in the sub-network, a double suppression loss function is used to process highly unbalanced data sets; A classification recommendation unit, used for performing classification recommendation processing on e-commerce data through the LAFAMN model; The outputs and labels of the LAFAMN include: The overall output of the network includes and the label value of the original sample ; Subnetwork output, including and the label value of the original sample ; The output of the supervised network includes And the label value obtained by subnet loss SoftMax ; Among them, the loss values of the three types of output are recorded as , and , define the total loss function , the process of calculating the total loss function using the square loss function is shown in formula (3): (3) in, , and is the parameter of LAFAMN, yj is the output of _the jth sub-network; The temperature coefficient used in the SoftMax operation is set by To control the loss value of the sub-network Weighted balance; The final output of LAFAMN is shown in formula (4): (4) in, are the parameters in the output vector given by the supervisory network; The dual suppression loss function includes three parts, among which, The first part is to suppress the loss contribution of large-class samples by means of functions; The second part is to suppress the contribution of easy-to-divide samples through the same function and different parameters; The last part is to suppress the contribution of easy-to-divided samples by expanding the output gap through a high-power function, which is shown in formula (5): (5) in, It is the final prediction output of LAFAMN; is the label value of the original sample, which takes the value 0 or 1; is the distance between the sample prediction output value and the label value; , , , and is the parameter of the double suppression loss function. 2. The e-commerce data classification recommendation system according to claim 1, wherein the feature preprocessing unit comprises: An importance ranking unit, configured to perform a primary ranking process on the first training data by a preset ranking method, and weight the primary ranking process result to calculate a feature importance score, so as to obtain a final ranking result of the first training data; The feature classification unit is used to classify the first training data according to a preset classification method. 3. The e-commerce data classification recommendation system according to claim 2, characterized in that: The preset sorting methods include PS-smart, XGBoost, and GBDT; The first training data is divided into dense features, sparse features, time series features and basic features by the preset classification method; wherein the basic features are other features that are not included in the three categories of dense features, sparse features and time series features after classification; The sub-networks perform learning and training processing on input features of different categories respectively. 4. The e-commerce data classification recommendation system according to claim 3, wherein the sub-networks perform learning and training processing on input features of different categories respectively, including: Input features of different categories after feature preprocessing of the first training data are respectively put into different sub-networks for learning, and all sub-networks output a score between 0 and 1, which represents the learning result of the current sample in the current sub-network; 0 means the sample is least likely to be recommended and 1 means the sample is most likely to be recommended; Each sub-network will calculate a loss value through the sample's label value and prediction value , ,in, is the total number of subnetworks; the loss value The larger the value of , the lower the corresponding sub-network's prediction confidence for the sample, and the lower its share in the final calculation of the output value. 5. The e-commerce data classification recommendation system according to claim 4, characterized in that: The output of the supervisory network is the proportion of all sub-networks in the total output; and The label used in the supervised network training is the weight vector of the loss value l _i obtained after performing SoftMax calculation based on the loss value ,in, Internal element value , , Will As the input before mapping, SoftMax calculation is performed. sub-networks, and use SoftMax to estimate the confidence of its prediction success for each sub-network , as shown in formula (1): (1) in, and Negatively correlated. 6. The e-commerce data classification recommendation system according to claim 5, further comprising: By the loss value Calculate another set of forward parameters , so that the output of each sub-network is optimal, where The value of the element in the vector , , as shown in formula (2): (2) in, and There is a positive correlation. 7. An e-commerce data classification recommendation method, based on the e-commerce data classification recommendation system according to any one of claims 1 to 6, comprising: Sorting the first training data by importance and classifying its features by a preset sorting method and a preset classification method to obtain feature preprocessing data of the first training data; The training data is trained and learned by the preset LAFAMN to obtain each type of output and label. After the loss value of the output and label is calculated, the weighted sum of the loss value is returned for processing, so that the loss value reaches the preset requirement to obtain the LAFAMN model; wherein, the training data includes the first training data and the feature preprocessing data, and the LAFAMN includes a supervisory network and a preset number of sub-networks; the sub-network is used to train and learn the feature pre-processing data, and the supervisory network is used to train and learn the first training data; and, the supervisory network learns the proportion weight of the sub-network, and the sub-network completes the self-learning of the sub-network weight and the feature weight by dynamically adjusting the total loss; and, in the sub-network, a double suppression loss function is used to process highly unbalanced data sets; The LAFAMN model is used to perform classification and recommendation processing on e-commerce data. 8. An electronic device, characterized in that the electronic device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the steps in the e-commerce data classification recommendation method as described in claim 7.