A product recommendation method, device and storage medium based on price-aware popularity debiasing

Abstract

The present invention discloses a product recommendation method, device and storage medium based on price-aware popularity debiasing, and belongs to the technical field of popularity debiasing product recommendation. The present invention proposes a new causal graph to consider the causal effects of price, popularity and user interest on user interaction behavior. Among them, for the price factor, the degree of matching of the user's consumption ability for the product is considered; for the user interest factor, the degree of matching between the user's characteristics and the product characteristics is considered. Secondly, based on the proposed causal graph, the decoupled learning model constructed during the recommendation process is implemented, and intervention operations are performed on the decoupled learning model during the inference stage to obtain more reasonable unbiased recommendation results. The present invention improves the diversity and novelty of the recommendation results, enhances the fairness of the recommendation system, improves the quality and accuracy of the recommendation results, and the technology is simple to implement, easy to implement in the project, does not introduce too much computing overhead, and is easy to implement and deploy in the project.

Description

Technical Field

[0001] The present invention relates to the technical field of popularity-debiased commodity recommendation, and in particular to a price-aware popularity-debiased commodity recommendation method, device and storage medium. Background Art

[0002] Recommendation systems aim to capture user preferences and provide users with appropriate items based on the analysis of user historical interaction data. However, current recommendation systems generally suffer from popularity bias. Popularity bias refers to the phenomenon that popular items are recommended more frequently than they should be, while less popular items that users may be interested in are under-recommended. The phenomenon of popularity bias is caused by the long-tail distribution of user interactions, that is, most user interactions occur with highly popular items. The popularity debiasing algorithm of existing recommendation systems mainly readjusts the frequency of recommendations based on the absolute size of popularity. However, this method has some problems: because items of different prices naturally have different popularity, for example, expensive items such as mobile phones and computers tend to have low popularity, while cheap items such as paper towels and beverages have high popularity. It is inappropriate to readjust the frequency of recommendations directly based on the absolute size of the popularity of the items. Existing methods will uniformly increase the recommendation frequency of expensive items such as mobile phones and computers, and uniformly reduce the recommendation frequency of cheap items such as paper towels and pens, but this is unfair to relatively unpopular brands among cheap items. Therefore, it is necessary to consider the price factor of the item when removing popularity bias. Summary of the invention

[0003] The present invention discloses a product recommendation method, device and storage medium based on price-aware popularity debiasing, so as to realize a product recommendation scheme combining popularity negative sampling with debiasing considering price factors, and improve the recommendation performance of the recommendation system.

[0004] In one aspect, the present invention discloses a method for recommending products based on price-aware popularity debiasing, the method comprising the following steps:

[0005] Step 1: construct a decoupled learning model for obtaining recommendation scores of products to be recommended;

[0006] The decoupled learning model includes a price module, a popularity module and a user interest module;

[0007] Among them, the price module is a dual-input branch, one branch is used to obtain the price index of the to-be-recommended product. The input of this branch is the price of the to-be-recommended product and the highest price of the historical products that the user has interacted with. The price index of the user for each to-be-recommended product is determined based on the price of the to-be-recommended product and the highest price. The other branch is used to obtain the user's consumption willingness. The input of this branch includes the user embedding expression and the product embedding expression. The user's consumption willingness is obtained based on the feedforward neural network and the activation function. The dimensions of the user embedding expression and the product embedding expression are consistent, and the embedding expression is generated by using random initialization. Then, the final embedding expression is learned through the training process of the model. The user's price index for each to-be-recommended product and the user's consumption willingness are fused through the calculation module to obtain the user's matching score sp for the price of each to-be-recommended product i. i , which is the effect score output by the price module;

[0008] The popularity module includes normalization and smoothing operations in sequence. Its input is the number of user interactions for each recommended product, that is, the number of users who have used the product as a historical product that the user has interacted with. Based on the smoothing operation, the popularity c of each recommended product i is output. i , which is the effect score output by the popularity module;

[0009] The input of the user interest module is the user embedding expression and the embedded expression of the intrinsic features of the recommended product except price and popularity. Other intrinsic features of the product include brand, category, quality and other features. These features are uniformly summarized and represented by an embedded expression and trained by the model. Based on the activation mapping of the inner product of the user embedding expression and the embedded expression of the intrinsic features of the recommended product, the matching score sn of the user interest on the intrinsic features of each recommended product i is obtained. i , which is the effect score output by the user interest module;

[0010] According to the preset intervention weight, the prevalence c i Perform debiasing to obtain the debiased popularity c′ i ; If the intervention weight is defined as δ, then c′ i =δ×c i .

[0011] Fusion matching score sp i , popularity c′ i and matching score sn i Get the recommendation score of each recommended product i;

[0012] Step 2: train the model parameters of the decoupled learning model based on the collected training data set, obtain the recommendation scores of each recommended product of the target user based on the trained decoupled learning model output, form a product recommendation list for the target user based on the recommendation score, and push it to the target user. For example, based on a given number of recommendations K, the previous K recommended products with the highest recommendation scores constitute the target user's product recommendation list.

[0013] Furthermore, the user's consumption intention is: Among them, e u represents the user embedding expression of user u, e i represents the product embedding expression of the recommended product i for user u, where e u With e i The dimensions of are all 1×d, the weight matrix W of the feedforward neural network is a d×d learnable parameter matrix, b is the bias term of the feedforward neural network, and the calculation formula of the activation function ELU′(·) is:

[0014]

[0015] Among them, e is the natural base, and x is the input parameter of ELU′(·).

[0016] Furthermore, the user's matching score sp for the price of each recommended product i i Set to:

[0017]

[0018] Among them, ω1 and ω2 are two preset weight coefficients, which can be iteratively updated based on the loss function of the decoupled learning model, e is the natural base, β represents the user's consumption willingness, α i Represents the price index of the recommended product i.

[0019] Furthermore, the normalization operation of the popularity module is specifically as follows: Among them, f i is the number of occurrences of the recommended product i, is the set of occurrence counts of all items to be recommended, f i ′ represents the normalized number of occurrences of the recommended product i.

[0020] Furthermore, the smoothing operation of the popularity module is as follows: i =f i ' γ , where f i ′ represents the normalized number of occurrences of the recommended product i, and γ represents the preset smoothing operation index, which can usually be set to a number greater than 0.

[0021] Furthermore, in step 2, the decoupled learning model uses the joint loss of BPR loss and popularity-based BPR loss during training: L = λ1L bpr +λ2L pop ;

[0022] Among them, L represents the joint loss, L bpr Indicates the loss using BPR, L pop represents the BPR loss based on popularity, λ1 and λ2 are two preset loss weighting coefficients;

[0023] Among them, L bpr and L pop Set to:

[0024]

[0025] Among them, N S Indicates the number of negative samples corresponding to each positive sample in the training data set, s i and j They represent the recommendation scores of positive samples and negative samples, respectively. Positive samples refer to products that the current user has interacted with (i.e., historical products that the current user has interacted with), and negative samples are randomly selected from all products that the current user has not interacted with. σ() represents the activation function.

[0026] According to sampling probability From the first screened negative sample set Extract product h from the dataset, and all extracted products h constitute a negative sample set Then based on the recommendation score s of the positive sample i , recommendation score s of product h h Calculate L pop ;

[0027] Among them, the first screening negative sample set The screening operation is: based on the recommendation score (also called prediction score) of each product in the training data set, for each positive sample i, if the recommendation score s of its corresponding negative sample j is j The recommendation score s is smaller than the positive sample i , and the popularity of the positive sample is c i Less than the popularity c of the negative sample j , then the current negative sample j is stored in the first screened negative sample set Among them, The initial value of is the empty set.

[0028] At the same time, the present invention also discloses a computer device, which includes a processor and a memory, in which at least one computer program is stored. The at least one computer program is loaded and executed by one or more processors to implement a product recommendation method based on price-aware popularity debiasing disclosed in the present invention.

[0029] On the other hand, the present invention also discloses a computer-readable storage medium, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor of a computer device to implement a product recommendation method based on price-aware popularity debiasing disclosed in the present invention.

[0030] The technical solution provided by the present invention brings at least the following beneficial effects:

[0031] (1) Improved diversity and novelty of recommendation results. By removing the popularity bias in the recommendation scenario, the present invention allows more niche or novel content to be recommended, increases the diversity of results, and enables users to access a wider range of products, taps into more potential points of interest for users, and improves user experience.

[0032] (2) The fairness of the recommendation system is enhanced. For product providers, the debiased recommendation opportunities of the present invention are more fair, which is conducive to maintaining a good ecological development.

[0033] (3) Improved quality and accuracy of recommendation results. Based on removing popularity bias, the present invention can make the recommendation results pay more attention to the quality of the product itself and the matching degree of the user, thereby reducing the excessive recommendation of low-quality but popular content and improving the recommendation quality.

[0034] (4) The technology is simple to implement and easy to implement in engineering. Compared with other debiasing methods, the present invention has good compatibility with the existing recommendation framework, does not introduce excessive computational overhead, and is easy to implement and deploy in engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0036] Figure 1 A schematic diagram of the structure of a decoupled learning model used in an embodiment of the present invention;

[0037] Figure 2 A cause-and-effect diagram of factors that may affect the interaction between a user and a product, as used in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the drawings in the implementation of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present invention.

[0039] Since commodities of different prices naturally have different popularity, the practice of directly readjusting the recommendation frequency according to the absolute value of popularity by the popularity debiasing algorithm of the existing recommendation system will lead to unreasonable situations. Therefore, the embodiment of the present invention provides a commodity recommendation method based on price-aware popularity debiasing to achieve the removal of price-aware popularity deviation.

[0040] In the embodiment of the present invention, a new causal graph is first proposed to consider the causal effects of price, popularity and user interest on user interaction behavior. Among them, for the price factor, the embodiment of the present invention considers the matching degree of the user's consumption ability for the product; for the user interest factor, the embodiment of the present invention considers the matching degree of user characteristics and product characteristics. Secondly, based on the proposed causal graph, the embodiment of the present invention proposes a product recommendation method based on price-aware popularity debiasing, which is implemented based on the constructed decoupled learning model during the recommendation process. Performing intervention operations on the decoupled learning model in the inference stage can obtain more reasonable unbiased recommendation results. Finally, the embodiment of the present invention also proposes a new popularity-based negative sampling method for use in recommendation processing, and proposes a popularity-based BPR (Bayesian Personalized Ranking) loss function to select some negative samples with large amounts of information to decouple user behavior.

[0041] As a possible implementation, see Figure 1 The two main processes of the commodity recommendation method based on price-aware popularity debiasing provided by the embodiment of the present invention are debiasing based on a decoupled learning model and negative sampling based on popularity. The specific implementation includes:

[0042] (1) Debiasing based on decoupled learning model.

[0043] In the embodiment of the present invention, the factors that may affect the interaction between users and products in the recommendation scenario are described using a causal graph, such as Figure 2As shown in the figure, node "I" represents the product, "U" represents the user, and node "P" represents the product price. Node "C" represents the popularity of the product, node "N" represents other intrinsic features of the product besides price and popularity, nodes "M1" and "M2" represent matching scores, and node "Y" represents the probability of interaction. Figure 2 The meaning of the cause-effect diagram shown is described as follows:

[0044] 1) I→P, I→C, and I→N represent the price, popularity, and other intrinsic characteristics of a product, respectively.

[0045] 2)(P,U)→M2→Y represents the impact of price factors on user interaction behavior. The matching score M2 is determined by the matching degree between the product price and the user's own consumption willingness. On the one hand, lower-priced products may be more popular with users, representing a higher matching score. On the other hand, the user's consumption willingness also determines their tolerance for price, so the stronger the consumption willingness, the higher the matching score.

[0046] 3)(N,U)→M1→Y describes the impact of user interests on user interaction behavior. The matching score M1 is used to measure the degree of match between user interests and the intrinsic characteristics of the product (excluding price and popularity). A higher matching score means that the product is more attractive to the user.

[0047] 4) I→C→Y represents the impact of popularity on user interaction behavior. Due to popularity bias, popular items are more likely to be recommended and interacted with. Therefore, higher popularity indicates a higher probability of interaction.

[0048] Based on this causal graph, the embodiment of the present invention proposes a decoupled learning model to decouple the effects of commodity price, popularity, and user interest on user interaction behavior. Figure 1 As shown in Figure 1, the decoupled learning model can be divided into three modules, namely price module, popularity module, and user interest module. The specific settings of each module are:

[0049] (a) Price module

[0050] Considering that the relative price of a product to a user provides a more accurate description of the impact of price factors than the absolute price of the product, the embodiment of the present invention designs a price index, expressed as α i , to represent the relative price of goods. Price index α i The calculation of is as follows:

[0051]

[0052] Among them, price i Indicates the price of product i, price uIndicates the highest price among the historical commodities that user u has interacted with.

[0053] That is, in this embodiment, based on Figure 1 The calculation module (Div) shown in the figure calculates the α of each product based on the input product price and the highest product price that the user has interacted with. i .

[0054] Then, the consumption willingness modeling of the user is discussed. A higher consumption willingness indicates a greater possibility of interactive behavior. Therefore, the embodiment of the present invention models the consumption willingness of the user as follows:

[0055]

[0056] in, and is the embedded expression of users and products, is a learnable parameter matrix, is a learnable parameter (i.e., bias term), d represents the dimension of the embedded expression, and ELU′(·) is a variant of the ELU activation function. The calculation formula of ELU′(·) is as follows:

[0057]

[0058] Among them, e is the natural base, and x is the input of ELU′(·).

[0059] Finally, calculate the matching score between the user and the product price:

[0060]

[0061] Among them, ω1 and ω2 are hyperparameters.

[0062] (b) Popularity Module

[0063] The popularity module is used to model the impact of popularity on user interaction behavior. First, the number of occurrences of all products in the training set is counted, expressed as Among them, n represents the number of goods, f i (i=1,…,n) represents the number of occurrences of product i, that is, the total number of times the product has been interacted with by users in the dataset;

[0064] Second, normalize the number of occurrences of the item to a range between 0 and 1, as follows:

[0065]

[0066] Among them, Max(·) represents the maximum element in the set.

[0067] Finally, since the frequency of occurrence of different items may vary greatly, which may not be suitable for model learning, a smoothing operation is performed on the occurrence count as follows:

[0068] c i =f i ' γ

[0069] Among them, c i It is used to indicate the influence of the popularity of commodity i, and γ indicates the preset smoothing operation index.

[0070] (c) User Interest Module

[0071] The user interest module is designed to calculate the matching score between user interests and intrinsic features of commodities (except price and popularity). u ) and product embedding (e i ) to calculate the matching score. Formally, the matching score is calculated as follows:

[0072]

[0073] Among them, x i and x u denote the user embedding of user u and the embedding of the intrinsic features of item i except price and popularity, respectively.

[0074] Effect score calculation and intervention analysis based on three decoupled modules:

[0075] The final score of the product is calculated according to the formula:

[0076] S(y i )=s i =sp i ·c i ·sn i

[0077] Among them, S(y i ) represents the interaction probability between the user and the i-th item, sp i represents the price effect (the matching score between user and product price), c i represents the popularity effect, sn i Represents the user interest effect.

[0078] Intervention analysis explores the potential results of manipulating a variable. In the proposed decoupling model, the embodiment of the present invention intervenes in the popularity module to reduce the popularity bias. By reducing the influence of popularity, a debiased recommendation result considering price factors can be obtained. The score after intervention can be calculated as follows:

[0079] S(yi |do(C))=S m (y i |C′)=sp i ·c′ i ·sn i

[0080] Among them, c′ i represents the popularity effect after debiasing. If the intervention weight is defined as δ, then c′ i =δ×c i , to scale the popularity effect.

[0081] In the field of recommendation systems, the BPR loss function is a widely used loss function in many recommendation works. BPR loss takes into account the relative order of items. BPR loss aims to maximize the difference between positive samples and negative samples. This enables the model to better learn the characteristics of positive samples. Formally, BPR loss is expressed as:

[0082]

[0083] Among them, N S Represents the number of negative samples corresponding to each positive sample, s i and j Represents the scores of positive samples and negative samples, where negative samples are randomly selected from all products that the user has not interacted with.

[0084] However, the interaction between users and commodities is affected by multiple factors at the same time, including price, popularity and user interest, and the influence of these factors is mixed with each other. Therefore, the label of the interaction behavior only represents the overall influence of these three factors. The proposed model cannot simply separate and learn the specific influence of these three factors by using only BPR loss. In order to enable the decoupled learning model to better learn the specific influence of factors, an embodiment of the present invention proposes a popularity-based negative sampling and a popularity-based BPR loss function.

[0085] First, in order to capture the specific impact of factors, the popularity-based negative sampling proposed in the embodiment of the present invention aims to sample important samples that truly reflect the user's interest in the product and the user's willingness to consume the product. In each training cycle, the prediction scores of all products can be obtained.

[0086] Next, in order to select products that truly reflect user interests and consumption intentions, use the following inequality to make a selection:

[0087] s j i

[0088] c j >c​i

[0089] Among them, i refers to the positive sample, j refers to the negative sample, s is the predicted score of the product, and c is the popularity effect value of the product. If the score of the negative sample product j is lower than the score of the positive sample, but the popularity effect value is higher than the positive sample, it means that the values ​​of the other two effects of the positive sample are lower than the other two effects of the positive sample. According to these two inequalities, the negative sample set is screened out.

[0090] Therefore, the embodiment of the present invention proposes a BPR loss function based on popularity as follows:

[0091]

[0092]

[0093] Among them, product h is selected from the negative sample set The sampling probability is Pr h , represents the negative sample set selected based on the above two inequalities, and o represents the negative sample set The element, s o represents the score of negative sample o (i.e., the recommendation score based on the model), s i represents the recommendation score of the recommended product i, s h represents the recommendation score of product h, L pop represents the BPR loss based on popularity, σ() represents the activation function, and η represents the preset weighting coefficient.

[0094] In the popularity-based negative sampling process proposed in the embodiment of the present invention, firstly, considering that the higher scores of negative samples represent samples that are more difficult to process and more informative, a larger sampling probability should be assigned to these samples when sampling negative samples from the negative sample set. This is because the higher prediction scores of negative samples indicate that the model has failed to learn the sample well. Giving these products a higher sampling probability will provide better guidance for model training. Secondly, compared with the selected negative samples, users interact with positive products because their interests and consumption intentions are consistent with the positive samples. Therefore, by expanding the difference between positive samples and selected negative samples through the proposed popularity-based BPR loss, the interest module and the price module can learn the actual causal effects.

[0095] In summary, the final loss function can be expressed as:

[0096] L=λ1L bpr +λ2L pop

[0097] Among them, λ1 and λ2 are preset loss weighting coefficients.

[0098] In order to further demonstrate the recommendation performance of the commodity recommendation method based on price-aware popularity debiasing provided by an embodiment of the present invention, further verification is performed through experimental simulation.

[0099] The data set, comparison index, and comparison method of this experimental simulation are set as follows:

[0100] Dataset: This simulation experiment uses two data sets, namely Amazon-Toys and Games and Amazon-Tools and Home Improvement. The data of the three data sets are shown in the following table:

[0101] Table 1: Statistics of the dataset

[0102]

[0103]

[0104] Comparison indicators: Three evaluation indicators widely used in the recommendation system popularity deviation removal task are used: MRR@K, NDCG@K, and HIT@K.

[0105] Comparison methods: IPS (a widely used baseline method), DICE (a decoupled learning method with outstanding results), PDA (one of the best causal debiasing models currently), FAiR (the best model for two-end debiasing currently), and InvPref (one of the best causal debiasing models currently).

[0106] Experimental results: As can be seen from Table 2, the price-aware popularity-debiased product recommendation method (PAPBM algorithm for short) proposed in this embodiment has achieved the best results on both datasets. Specifically, on the toy dataset, PAPBM outperforms the best baseline model (i.e., PDA) on MRR@1 with a relative improvement of 18.3%, 15.7% on MRR@10, 15.2% on MRR@20, 14% on NDCG@10, 11% on NDCG@20, 9.6% on HIT@10, and 7.8% on HIT@20. On the tool dataset, PAPBM outperforms the best baseline model (i.e., InvPref) with a relative improvement of 33.3% on MRR@1, 23.3% on MRR@10, 22.3% on MRR@20, 19.1% on NDCG@10, 16% on NDCG@20, 10% on HIT@10, and 8.6% on HIT@20. The improvement of the proposed method on all datasets may be attributed to the consideration of price factors and the proposed popularity-based negative sampling method, which helps the decoupled learning model to better understand and learn how interactions occur.

[0107] Table 2: Experimental results (Recall@20, NDCG@20)

[0108]

[0109]

[0110] *Improv. indicates the relative improvement of the method proposed in this embodiment over other methods.

[0111] In an exemplary embodiment, an embodiment of the present invention further provides a computer device, the computer device comprising a processor and a memory, wherein at least one computer program is stored in the memory. The at least one computer program is loaded and executed by one or more processors to implement any of the above-mentioned commodity recommendation methods based on price-aware popularity debiasing.

[0112] In an exemplary embodiment, an embodiment of the present invention further provides a computer-readable storage medium, in which at least one computer program is stored, and the at least one computer program is loaded and executed by a processor of a computer device to implement any of the above-mentioned product recommendation methods based on price-aware popularity debiasing.

[0113] In one possible implementation, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, an optical data storage device, or the like.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0115] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A product recommendation method based on price-aware popularity debiasing, characterized in that: The following steps are involved: Step 1: construct a decoupled learning model for obtaining recommendation scores of products to be recommended; The decoupled learning model includes a price module, a popularity module and a user interest module; Among them, the price module is a dual-input branch, one branch is used to obtain the price index of the product to be recommended. The input of this branch is the price of the product to be recommended and the highest price of the historical products that the user has interacted with. The user's price index for each product to be recommended is determined based on the price of the product to be recommended and the highest price; the other branch is used to obtain the user's consumption intention. The input of this branch includes the user embedding expression and the product embedding expression. The user's consumption intention is obtained based on the feedforward neural network and the activation function, wherein the dimensions of the user embedding expression and the product embedding expression are consistent; the user's price index for each product to be recommended and the user's consumption intention are integrated through the calculation module to obtain the user's matching score for the price of each product to be recommended i ; The popularity module includes normalization and smoothing operations in sequence. Its input is the number of occurrences of each recommended product. Based on the smoothing operation, it outputs the popularity of each recommended product i. , which is the effect score output by the popularity module; The input of the user interest module is the user embedding expression and the embedding expression of the intrinsic features of the recommended product. Based on the activation mapping of the inner product of the user embedding expression and the embedding expression of the intrinsic features of the recommended product, the matching score of the user interest to the intrinsic features of each recommended product i is obtained. ; Wherein, the embedded expression of the intrinsic features of the recommended product is the embedded expression of the intrinsic features except price and popularity; According to the preset intervention weights, the prevalence Perform debiasing to obtain the popularity after debiasing ; Fusion Matching Score , popularity and matching score Get the recommendation score of each recommended product i; Among them, the user's consumption willingness is: ,in, represents the user embedding expression of user u, represents the product embedding expression of the recommended product i for user u, where and The dimensions of , the weight matrix of the feedforward neural network for The learnable parameter matrix, is the bias term of the feedforward neural network, the activation function The calculation formula is: ； Among them, e is the natural base, for Input parameters of The matching score of the user's price for each recommended product i Set to: ； in, and are two preset weight coefficients, which can be iteratively updated based on the loss function of the decoupled learning model. is the natural base, Indicates the user's consumption willingness. Represents the price index of the recommended product i for: ,in, represents the price of commodity i, Indicates the highest price of historical products that user u has interacted with; Step 2: training the model parameters of the decoupled learning model based on the collected training data set, and obtaining the recommendation scores of the target user's recommended products based on the trained decoupled learning model output, and forming a product recommendation list for the target user based on the recommendation scores and pushing it to the target user; The decoupled learning model uses a joint loss of BPR loss and popularity-based BPR loss during training: ; in, represents the joint loss, Indicates the loss using BPR, represents the popularity-based BPR loss, , are two preset loss weighting coefficients; in, and Set to: ， ； in, Indicates the number of negative samples corresponding to each positive sample in the training data set, and They represent the recommendation scores of positive samples and negative samples respectively. Positive samples refer to products that the current user has interacted with, and negative samples are randomly selected from all products that the current user has not interacted with. represents the activation function, Indicates the preset weighting coefficient; According to sampling probability From the first screened negative sample set Extract products from , all the goods extracted Constructing negative sample set , and then based on the recommendation score of the positive sample ,commodity Recommended score Calculated ; Among them, the first screening negative sample set The filtering operation is: based on the recommendation score of each product in the training data set, for each positive sample , if the recommendation score of a corresponding negative sample j The recommendation score is smaller than the positive sample , and the popularity of the positive sample Less than the popularity of the negative sample , then the current negative sample j is stored in the first screened negative sample set Among them, The initial value of is the empty set.

2. The method according to claim 1, characterized in that Intrinsic characteristics of a product include brand, category and / or quality.

3. The method according to claim 1, characterized in that The normalization operation of the popularity module is as follows: ,in, is the number of occurrences of product i to be recommended, is the set of occurrence counts of all products to be recommended, Represents the normalized number of occurrences of the recommended product i.

4. The method according to claim 1, characterized in that The smoothing operation of the popularity module is as follows: ,in, represents the normalized number of occurrences of the recommended product i, Indicates the preset smoothing operation index.

5. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images