Recommendation method, recommendation system and related device based on benefit balance

Abstract

The invention relates to the technical field of recommendation, provides a recommendation method and system based on benefit balance and a related device, and can realize benefit balance between merchants and users. The method comprises the following steps: constructing a user node, a merchant node, a commodity node and edge relationships according to user information, merchant information and commodity information, wherein the edge relationships are an edge relationship between the user node and the commodity node and an edge relationship between the merchant node and the commodity node; using the user node and the merchant node to guide and aggregate the commodity node so as to obtain a target weight corresponding to the commodity node commonly guided by the user node and the merchant node; determining a feature combination of the commodity node according to the target weight; determining a score of a commodity corresponding to the commodity node according to the feature combination of the commodity node and the user node; and generating a recommendation list according to the scores of the commodities.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of recommendation, in particular to a recommendation method based on benefit balance, a recommendation system and related devices. BACKGROUND

[0002] When the existing recommendation system recommends goods to users, it mainly evaluates the matching score of user demand and goods characteristics, and recommends goods with high matching score to users, focusing on enhancing user satisfaction, but often ignoring the interests of merchants in the system, which is not conducive to merchants and even affects user satisfaction. The existing recommendation system is difficult to meet the interests of both merchants and users, and cannot comprehensively measure the interests of both, lacking bilateral fairness. SUMMARY

[0003] The present application aims to provide a recommendation method based on benefit balance, a recommendation system and related devices, which takes into account the needs of both merchants and users, guides the aggregation process of goods characteristics, and realizes the balance of interests between merchants and users.

[0004] The first aspect of the embodiment of the present application provides a recommendation method based on benefit balance, comprising:

[0005] According to user information, merchant information and goods information, user nodes, merchant nodes, goods nodes and edge relationships are constructed, and the edge relationships are the edge relationships between the user nodes and the goods nodes and the edge relationships between the merchant nodes and the goods nodes;

[0006] The user nodes and the merchant nodes are used to guide and aggregate the goods nodes to obtain target weights corresponding to the goods nodes guided by the user nodes and the merchant nodes;

[0007] According to the target weights, the feature combination of the goods nodes is determined;

[0008] According to the feature combination of the goods nodes and the user nodes, the score of the goods corresponding to the goods nodes is determined;

[0009] According to the score of the goods, a recommendation list is generated.

[0010] The second aspect of the embodiment of the present application provides a recommendation system based on benefit balance, comprising:

[0011] A construction unit is configured to construct user nodes, merchant nodes, goods nodes and edge relationships according to user information, merchant information and goods information, and the edge relationships are the edge relationships between the user nodes and the goods nodes and the edge relationships between the merchant nodes and the goods nodes;

[0012] The guiding aggregation unit is configured to guide aggregation of the commodity node by using the user node and the merchant node, so as to obtain a target weight corresponding to the commodity node guided by the user node and the merchant node;

[0013] The first determination unit is configured to determine a feature combination of the commodity node according to the target weight;

[0014] The second determination unit is configured to determine a score of a commodity corresponding to the commodity node according to the feature combination of the commodity node and the user node;

[0015] The generating unit is configured to generate a recommendation list according to the score of the commodity.

[0016] The third aspect of the embodiment of the present application provides a computer device, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method in the first aspect.

[0017] The fourth aspect of the embodiment of the present application provides a computer program product, including instructions, and the instructions are executed by the processor to realize the steps of the method in the first aspect.

[0018] The fifth aspect of the embodiment of the present application provides a computer storage medium, including instructions, and when the instructions are executed on the computer, the computer executes the steps of the method in the first aspect.

[0019] Compared with the related art, the embodiment provided by the present application guides aggregation of the commodity node by using the user node and the merchant node, gives the target weight to the commodity node, gives greater weight to the commodity meeting the interests of the merchant and the user, makes the feature combination of the commodity node reflect the interests of the merchant and the user at the same time, and obtains the recommendation based on the feature combination of the commodity node, so that the interests of the merchant and the user can be effectively balanced. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on the drawings shown.

[0021] Figure 1 The flowchart of the recommendation method based on interest balance provided by the embodiment of the present application is shown.

[0022] Figure 2A schematic diagram of a user node, a merchant node, a commodity node and an edge relationship provided for an embodiment of the present application is shown in FIG. 1.

[0023] Figure 3 A schematic diagram of a user benefit evaluation index provided for an embodiment of the present application is shown in FIG. 2.

[0024] Figure 4 A schematic diagram of a benefit balance construction index provided for an embodiment of the present application is shown in FIG. 3.

[0025] Figure 5 A schematic diagram of a user benefit guidance target and a merchant benefit guidance target guiding an aggregation process provided for an embodiment of the present application is shown in FIG. 4.

[0026] Figure 6 A virtual structure schematic diagram of a benefit balance-based recommendation system provided for an embodiment of the present application is shown in FIG. 5.

[0027] Figure 7 A hardware structure schematic diagram of a server provided for an embodiment of the present application is shown in FIG. 6. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0029] In current recommendation systems, traditional recommendation systems mainly focus on user satisfaction and commodity sales volume, but less consider the balance of benefits. This approach has many problems, such as causing conflicts between user satisfaction and merchant benefits, over-optimizing user satisfaction may cause low utilization of merchant commodities; only focusing on short-term user satisfaction will damage the long-term benefits of the system, reduce the participation of merchants and the overall activity of the platform; and ignoring the benefits of merchants will make the recommendation results biased towards certain specific commodities, limiting the selection range of users. Therefore, the interests of users, merchants and platforms are imbalanced.

[0030] To solve these problems, the embodiments of the present application propose a benefit balance-based recommendation method, which considers both merchant benefits and user benefits in the consideration range, guides the aggregation process of commodity features, and then generates more accurate recommendation results and conducts comprehensive quality evaluation. In this way, the embodiments of the present application overcome the tendency of one-sided pursuit of maximizing one-sided benefits in current recommendation systems, achieve dynamic balance and common growth of benefits between merchants and users, and thus improve the benefits of the platform.

[0031] The benefit balance-based recommendation method will be described below from the perspective of a benefit balance-based recommendation system. Please refer to Figure 1, which is a flowchart of a recommendation method based on interest balance provided by an embodiment of the present invention, including:

[0032] 101. Construct user nodes, merchant nodes, product nodes and edge relationships based on user information, merchant information and product information. The edge relationships are the edge relationships between user nodes and product nodes and the edge relationships between merchant nodes and product nodes.

[0033] In this embodiment, the recommendation system can construct user nodes, merchant nodes, product nodes and edge relationships based on user information, merchant information and product information. User nodes represent user personalized preferences and historical behaviors, merchant nodes represent merchant operation goals, and product nodes represent specific recommended products provided by merchants. The edge relationship between user nodes and product nodes represents the interactive relationship between users and products, and the edge relationship between merchant nodes and product nodes represents the interactive relationship between merchants and products. Figure 2 shown.

[0034] It's understandable that historical behavior includes browsing, clicking, purchasing, and rating, and a user's personalized preferences can be derived from historical behavior. For example, if a user has recently frequently searched for "outdoor equipment," clicked on trekking poles multiple times, and spent a significant amount of time browsing blue trekking poles, this indicates that the user has recently preferred trekking poles within the outdoor gear category, and their preferred color is blue.

[0035] Merchants can be item providers, service providers, resource providers, text, music, video, image creators, etc. Correspondingly, operational goals include retention rate, exposure, utilization rate, etc. Items to be recommended can be items, videos, articles, etc., and the items to be recommended include type, brand, rating information, etc.

[0036] 102. User nodes and merchant nodes are used to guide and aggregate product nodes to obtain target weights corresponding to product nodes jointly guided by user nodes and merchant nodes.

[0037] In this embodiment, the recommendation system uses user nodes and merchant nodes to guide the aggregation of commodity nodes to obtain target weights corresponding to commodity nodes jointly guided by user nodes and merchant nodes. By taking user nodes and merchant nodes as guide nodes, the features of commodity nodes are aggregated, and user interests and merchant interests are comprehensively considered to assign weights to commodity nodes. Specifically, the user interests corresponding to the user nodes and the merchant interests corresponding to the merchant nodes can be taken as the guiding target to guide the aggregation of the features of the commodity nodes and assign weights to the commodity nodes. The target weights corresponding to the commodity nodes guided by user interests and merchant interests can reflect user interests and merchant interests at the same time, and realize the balance of multi-party interests. For example, the user nodes give personalized preferences of the user themselves, provide guidance information for the aggregation process of the commodity nodes, so that the commodity nodes with high matching degree of user interests obtain higher weights in the aggregation process, thereby strengthening the feature expression of related commodities and making the recommendation results more in line with user needs; the merchant nodes provide guidance information for the aggregation process of the commodity nodes based on the utilization rate of the commodities to be recommended, and give greater weights to commodities with low sales or interaction to improve the probability of recommendation of these low utilization rate commodities, balance the interests among merchants, and avoid excessive concentration of popular commodities in recommendation.

[0038] The merchant interest, the merchant interest evaluation index, the user interest, and the user interest evaluation index will be described in detail below.

[0039] The merchant interest evaluation index reflects the performance of the merchant in the recommendation system and its contribution to the platform value in the recommendation system. The interests of the merchant directly affect the ecological health and sustainable development of the platform. Reasonably measuring the interests of the merchant not only helps to optimize the recommendation strategy, but also promotes the win-win of the merchant and the platform and improves the overall market competitiveness. The measurement of the interests of the merchant involves multiple factor indexes, including user retention rate, conversion rate, commodity utilization rate, exposure, exposure fairness, and resource fairness, etc.

[0040] Retention rate is the probability that a user remains active after a certain period of time. High retention rate represents that the platform can continuously attract users, which means that the platform has innovative improvements in user experience, content value, etc., and is conducive to long-term development. It can be calculated as follows:

[0041]

[0042] Among them, Remaining Active Users refers to the number of users who are still active after a period of time from the initial number of users; Initial Users refers to the initial number of users.

[0043] Conversion rate measures the probability of a user performing a key operation (such as clicking, purchasing, registering, etc.) in the recommendation system. It can be calculated as follows:

[0044]

[0045] where, U convert is the number of users who complete the target behavior, U exposed is the number of users reached by the recommended content.

[0046] Product utilization rate, measures the proportion of products provided by the merchant that are actually consumed (purchased, played, downloaded, etc.). It can be calculated as follows:

[0047]

[0048] where, PUR q is the product utilization rate of product q; Y q denotes the product set owned by product q; |Y q | is the number of products owned by product q; B i is the purchase frequency of product i, indicating the sales situation of the product; S i is the inventory of product i; ∈ is a very small number to avoid division by zero error; α is used for exponential weight adjustment, when α > 1, the weight of products with higher purchase rate increases, emphasizing hot-selling products, when α < 1, all products are balanced, avoiding extreme products dominating the calculation.

[0049] Exposure, indicates the degree of publicity of the merchant's products in the recommendation system, which can be evaluated by considering the position of each product in the recommendation list of different users. The following formula is used to calculate:

[0050]

[0051] where, e s is the exposure of merchant s; |γ s | is the product set owned by merchant s; γ s is the number of products owned by merchant s; U is the set of all users; r u,i is the position of product i in the recommendation list of user u, starting from 1, if product i is not in the recommendation list of user u, r u,i is 0; is the indicator function, if product i is in the recommendation list of user u, it is 1, otherwise it is 0.

[0052] Exposure fairness, used to measure the balance of different content in the recommendation system. This index can avoid the overexposure of some popular content and bury long-tail content, and improve content diversity. The Max-Min Fairness of the minimum number of times a product is recommended in all time steps is used to measure, and the calculation formula is as follows:

[0053]

[0054] Where T represents the time step or the length of the interaction sequence; N is the total number of user-item interaction sequences in the test data; Refers to user u at time t t The set of the top K recommended products; S refers to the set of all merchants; is an indicator function, which is 1 if product i belongs to merchant s, otherwise it is 0; s is the total number of products of merchant s.

[0055] Resource fairness refers to the degree of balance in the platform's resource allocation. This allows more content to have equal opportunities, promotes diversity in the content ecosystem, and brings more recommendations. It can be calculated as follows:

[0056]

[0057] Among them, x i Resources obtained for the i-th merchant or content merchant; It is the average number of resources owned by merchants or content merchants on the platform.

[0058] User benefit evaluation indicators, in the recommendation system, reflect the user's experience, satisfaction and trust in the system. Figure 3 As shown in Figure 1, measuring user interest involves multiple factors, including relevance, differentiation, diversity, novelty, unpopularity, high quality, and randomness. These factors influence user experience, satisfaction, and trust in the system across multiple dimensions, including precise matching, content richness, exploration, personalized recommendations, quality assurance, and serendipity. An excellent recommendation system must strike a balance between these factors, ensuring that recommendations align with user interests and improve interaction efficiency while avoiding overly restrictive recommendations and providing diverse and high-quality options. This strengthens user engagement, enhances the overall user experience, and promotes long-term user retention.

[0059] The relevance factor (rel) is the basis for measuring user benefits. Based on different application scenarios, we flexibly select similarity measurement methods between target users and recommended products to ensure that recommended content is closely related to user needs, effectively avoid invalid recommendations, and lay the foundation for a satisfactory user experience. The following formula is proposed to measure the relevance factor:

[0060]

[0061] CTR@K represents the average click-through rate (CTR) over a long period of time and is an indicator of user preference. is user u t The preference score for item i; K is the list length.

[0062] The difference factor (diff) focuses on the degree of deviation of the recommended content from the user's historical behavior. By deeply mining the characteristics of the user's historical behavior data, it breaks the user's regular expectations and to some extent, it helps to broaden the user's interest range.

[0063] The diversity factor (div) emphasizes the heterogeneity between recommended products. The recommendation system enriches the recommended content from multiple dimensions to meet the user's needs in different scenarios and improve the user's overall satisfaction with the recommendation system.

[0064] The novelty factor (nov) mainly focuses on newly launched products by the system. In specific implementation, user portraits can be used to accurately match user interest points and push the latest online content to users, allowing them to access cutting-edge information first, satisfy their curiosity for new things, and enhance user experience.

[0065] The unpopularity factor (unpop) focuses on mining niche products in the long-tail distribution. By analyzing user interaction behavior, the popularity index of the project is calculated to filter out products that are less popular but may have unique value, and recommend them to users to satisfy their desire to explore niche and unknown areas.

[0066] The high-quality factor (high-level) is characterized by high ratings and good reputation of products. It can guide the recommendation system to preferentially recommend high-quality content that is highly consistent with user interests, ensuring that users have an excellent experience when using the recommendation service and fundamentally improving user recognition of recommended content.

[0067] The randomness factor (random) injects uncertainty into the recommendation system. In specific implementation, all candidate items can be given equal candidate probability to break the predictability of the recommended results, making each recommendation potentially bring unexpected new discoveries to users, increasing the interest and freshness of the recommendation process, and meeting user expectations for the unknown.

[0068] It can be understood that the values of the merchant interest evaluation index and the user interest evaluation index can be obtained from user historical behavior data and platform global interaction data. The recommendation method based on interest balance can be widely applied to various recommendation scenarios in different fields. Those skilled in the art can select different dimensions of merchant interest and different dimensions of user interest as guidance targets to guide the aggregation of commodity nodes according to actual application scenarios. Similarly, different factor indicators can be selected as the merchant interest evaluation index and the user interest evaluation index, such as those shown in Table 1. Figure 4

[0069] ​For example, on an e-commerce platform, with the help of benefit evaluation indicators, the platform can accurately measure the resource coverage, exposure rate, and purchase or rental rate of the merchant's goods. The recommendation model based on graph neural network can integrate the needs of merchants and users to provide accurate product recommendation services for users, thereby optimizing the shopping experience and promoting the growth of platform transactions.

[0070] For example, in terms of content recommendation systems, the content exposure of content creators (similar to merchants) can be measured, such as being consumed by users (similar to purchase or rental), and other factors, and the cost-effective factors such as user preference for content quality and type can be combined to push content that matches the user's interests to the user, improve the user's stickiness to the platform, promote the creativity of content creators, and achieve a win-win situation for the platform, users, and creators. In addition, in scenarios involving the needs of both merchants and users, such as shared resource platforms and online education course recommendations, the interests of both parties can be balanced to optimize resource allocation and recommendation results, and improve the overall operational efficiency and user satisfaction of the system.

[0071] It can also be understood that for different guidance targets, a trade-off indicator can be used to measure the performance of different guidance targets in the recommendation system. The trade-off indicator can be an accuracy indicator (recall, nDCG), a fairness indicator (IED@k), a benefit balance indicator, and the like.

[0072] For example, a fusion indicator with multi-factor indicator fusion weighting is used as a trade-off indicator for measuring benefit trade-off performance. Specifically, the following 13 surprise factors are used for fusion weighting, 1-6 are merchant benefit evaluation indicators, and 7-13 are user benefit evaluation indicators. The trade-off indicator for benefit trade-off performance determined by the user benefit evaluation indicator and the merchant benefit evaluation indicator is:

[0073] Index number i 1 2 3 4 5 6 7 8 9 10 11 12 13 Index m Retention rate Conversion rate Utilization rate Exposure rate Exposure fairness Resource fairness Relevance Difference Diversity Novelty Unpopularity High quality Randomness

[0074]

[0075] where m i is the i-th factor indicator, λ i is the weight corresponding to the i-th factor indicator.

[0076] It should be noted that the combination method of indicator weighting fusion has lower computational complexity compared to the mathematical optimization combination method, and not only effectively handles the fairness problem between multiple interest groups, but also maintains the balance of recommendation quality and accuracy, without causing a decrease in system utility. Therefore, the method of weighted fusion used in this embodiment has strong flexibility and operability, and can balance the surprise utility indicators on the user side and the platform side, thereby achieving benefit balance.

[0077] Product utilization directly reflects actual product consumption, avoiding attribution uncertainty in user behavior metrics. This reflects users' true choices and reduces interference from recommendation strategies. Merchants can also improve utilization by optimizing inventory and product quality, making this approach highly actionable. Users' core need for recommendations is to obtain content that matches their interests, and relevance factors can be used to measure user interest.

[0078] Therefore, the following uses the relevance factor and utilization rate as examples of user benefit evaluation indicators and merchant benefit evaluation indicators, and the trade-off indicators for measuring the benefit trade-off performance determined by the relevance factor and utilization rate are:

[0079] r@K=PUR p +λCTR@K

[0080] Where λ is the weight of the correlation factor.

[0081] In some embodiments, such as Figure 5 As shown, the user node and the merchant node are used to guide and aggregate the product node to obtain the target weight corresponding to the product node jointly guided by the user node and the merchant node. Specifically, the following may be included:

[0082] Determine user interest guidance targets and merchant interest guidance targets based on user nodes and merchant nodes;

[0083] The user interest guidance goal is used to guide the aggregation of product nodes to obtain the first weight corresponding to the product node;

[0084] The merchant interest guidance goal is used to guide and aggregate the product nodes to obtain the second weight corresponding to the product nodes;

[0085] A target weight corresponding to the commodity node is determined according to the first weight and the second weight.

[0086] It is understandable that based on the user node and the merchant node, corresponding user interest guidance goals and merchant interest guidance goals can be determined for different user characteristics and merchant characteristics.

[0087] In further embodiments, the merchant is a commodity provider, a service provider or a resource provider, the relevance factor is taken as an example as the user interest evaluation index, and the utilization rate is taken as an example as the merchant interest evaluation index. In the recommendation system, the commodity node is an instance node of the feature aggregation tree structure, and the user node and the merchant node jointly guide the aggregation of the commodity tree. From the user side, the user node assigns weights to the commodity node according to the user's own user demand. In the operation logic of the recommendation system, the commodities corresponding to these high-weight nodes will be preferentially selected and recommended, thereby effectively improving the adaptability and satisfaction of the commodities obtained by the user. From the merchant side, the merchant node takes the commodity utilization rate as the guidance basis, and gives greater weight to the commodity instances with lower commodity utilization rate. This strategy is based on the consideration of system resource management. By recommending these low-utilization-rate commodities, the load situation of the overall commodity can be balanced, avoiding the situation that some commodities are in an idle state for a long time, thereby improving the utilization efficiency of the entire commodity pool.

[0088] It can be understood that the user preference is the embodiment of the user demand, and the user preference is quantified by the relevance factor.

[0089] Taking the user demand or interest preference as the user interest guidance target, and using the user interest guidance target to guide and aggregate the commodity node to obtain the first weight corresponding to the commodity node can specifically include:

[0090] The first weight is calculated by the following formula:

[0091]

[0092] Wherein, u is a user node, t u is a feature vector of a user preference or a feature vector of a user demand, c is a commodity node, f c is a feature vector of the commodity node c, a p,c |u is the weight of the commodity node c under the guidance of the user node u relative to its parent node p, τ1 is a smoothing parameter for regulating the smoothing degree of the weight distribution, N p represents a set of neighbor nodes of the node p, i.e., a set of child nodes of p, c` represents all child nodes of the parent node p, f c` is a feature vector of the commodity node c`, and <,> represents an inner product operation.

[0093] It can be understood that the commodity node c is the node corresponding to the commodity c, and the commodity c is any one of the candidate commodity set C. The first weight a p,cThe weight value is higher, indicating that the user demand is met. In the aggregation process, the node with high weight is given higher priority, so that the recommendation system can preferentially screen and recommend the corresponding resource, thereby effectively improving the adaptation of the resource to the user demand and the user satisfaction.

[0094] The utilization rate is used as the merchant interest guidance target. The merchant interest guidance target is used to guide the aggregation of the commodity node to obtain the second weight corresponding to the commodity node. Specifically, the method can include the following steps:

[0095] The second weight is calculated by the following formula:

[0096]

[0097] Wherein, s is a merchant node, r c is the utilization rate of the commodity corresponding to the commodity node c, a p,c is the weight of the commodity node c relative to its parent node p under the guidance of the merchant node s, and t2 is a smoothing parameter. r c` is the utilization rate of the commodity corresponding to the commodity node c.

[0098] It can be understood that the second weight a p,c uses 1-r c to construct such a reverse association, giving greater weight to commodity instances with lower utilization rates. Therefore, from the perspective of merchant resource management, this weight reflects the importance of the commodity node c in the aggregation process. Through this weight distribution method, the load of the resource can be effectively balanced, avoiding excessive idle of part of the resource, and thus improving the utilization efficiency of the entire commodity pool.

[0099] According to the first weight and the second weight, the target weight corresponding to the commodity node is determined. Specifically, the method can include the following steps:

[0100] The first weight and the second weight are weighted and fused to calculate the target weight.

[0101] The calculation formula is specifically:

[0102] a p,s,c = b a p,c |u+ (1-b) a p,c |s

[0103] Wherein, b is a weight coefficient, and the value range is 0≤b≤1. Its role is to adjust the relative importance of user demand and merchant resource management in the aggregation process. a p,s,c is the aggregation weight of the commodity node c relative to its parent node p and the merchant node s.

[0104] 103、According to the target weight, the feature combination of the commodity node is determined.

[0105] In this embodiment, the recommendation system uses user nodes and merchant nodes to guide the aggregation of commodity nodes to obtain target weights of commodity nodes guided by user nodes and merchant nodes, and then determines the feature combination of the commodity nodes according to the target weights. The determination of the feature combination of the commodity nodes according to the target weights specifically includes:

[0106] Based on the target weights and the commodity nodes, the final feature combination of the commodity nodes is determined through convolution operation and layer combination.

[0107] It should be noted that the convolution operation aggregates the features of the commodity nodes in different layers to realize the aggregation and update of the features, and at the same time, the interests of users and merchants are taken into account. The layer combination operation combines the feature combinations of the commodity node c in different layers (from 0 to k layers) to comprehensively consider the information of each layer.

[0108] The convolution operation formula is specifically as follows:

[0109]

[0110] wherein, is the feature vector of the node p in the j+1 layer under the guidance of the user u and the merchant s, j is any layer in the convolution layer, and 0≤j≤k-1.

[0111] It can be understood that, is the feature vector of the j+1 layer of the commodity node c` aggregated by all child nodes of the commodity node c`, which is essentially the feature aggregation of the commodity node c in the j+1 layer.

[0112] The layer combination formula is specifically as follows:

[0113]

[0114] wherein, is the final feature combination of the commodity node, Comb is a combination operator, is the original embedding, is the feature vector of the commodity corresponding to the commodity node c in the 1st layer, is the feature vector of the commodity corresponding to the commodity node c in the kth layer.

[0115] It can be understood that for each commodity c in the candidate commodity set C, a final feature combination is obtained.

[0116] 104. Determine the score of the commodity corresponding to the commodity node according to the feature combination of the commodity node and the user node.

[0117] 105. Generate a recommendation list according to the score of the commodity.

[0118] In this embodiment, the recommendation system can determine the score of the commodity corresponding to the commodity node by calculating the similarity between the feature combination of the commodity node and the user node, and select the commodity with high similarity as the commodity recommended to the user.

[0119] It can be understood that the feature combination of the commodity node obtained by the recommendation system through the above steps reflects the interests of the user and the interests of the merchant, and by calculating the similarity between the feature combination of the commodity node and the user node, the commodity with high similarity is recommended to the user, so that a recommended list with balanced interests and high recommendation accuracy can be obtained.

[0120] The recommended list can be generated by the following formula:

[0121]

[0122] Wherein, sim is a similarity calculation function, TopK is a function of selecting K highest similarity commodities, f u is the user node feature.

[0123] It should be noted that the value of K can be set by the person skilled in the art according to actual needs, which is not limited here.

[0124] Compared with the related art, the embodiment of the present application can accurately measure the interests of the merchant and the user by guiding the aggregation of the commodity node by the user node and the merchant node, improve the resource utilization rate on the merchant side, improve the satisfaction on the user side, and guide the aggregation of commodity features according to the needs of both parties, based on the graph convolution network, so that the recommendation result is more accurate. At the same time, by dynamically adjusting the parameters to balance the interests of both parties, the recommendation system can be optimized to realize the win-win of merchants and users, provide strong support for the long-term stable development of related platforms, and effectively balance the interests of merchants, users and platforms.

[0125] The above describes the embodiments of the present application from the perspective of the recommendation method based on interest balance, and the following describes the embodiments of the present application from the perspective of the recommendation system based on interest balance.

[0126] Please refer to Figure 6 , the virtual structure schematic diagram of the recommendation system based on interest balance provided by the embodiment of the present application, the recommendation system based on interest balance 200 includes:

[0127] The construction unit 201 is used for constructing the user node, the merchant node, the commodity node and the edge relationship according to the user information, the merchant information and the commodity information, and the edge relationship is the edge relationship between the user node and the commodity node, the edge relationship between the merchant node and the commodity node.

[0128] The guiding aggregation unit 202 is configured to perform guiding aggregation on the commodity node by using the user node and the merchant node, so as to obtain a target weight corresponding to the commodity node guided by the user node and the merchant node;

[0129] The first determination unit 203 is configured to determine a feature combination of the commodity node according to the target weight;

[0130] The second determination unit 204 is configured to determine a score of a commodity corresponding to the commodity node according to the feature combination of the commodity node and the user node;

[0131] The generation unit 205 is configured to generate a recommendation list according to the score of the commodity.

[0132] In a possible design, the guiding aggregation unit 202 is specifically configured to:

[0133] determine a user benefit guiding target and a merchant benefit guiding target according to the user node and the merchant node;

[0134] perform guiding aggregation on the commodity node by using the user benefit guiding target, so as to obtain a first weight corresponding to the commodity node;

[0135] perform guiding aggregation on the commodity node by using the merchant benefit guiding target, so as to obtain a second weight corresponding to the commodity node;

[0136] determine a target weight corresponding to the commodity node according to the first weight and the second weight.

[0137] In a possible design, the guiding aggregation unit 202 is specifically configured to:

[0138] calculate the first weight by using the following formula:

[0139]

[0140] wherein u represents the user node, t u represents a feature vector of the user, c represents the commodity node, f c represents a feature vector of the commodity node c, a p,c |u represents a weight of the commodity node c relative to a parent node p of the commodity node c under the guidance of the user node u, τ1 represents a smoothing parameter, and is used to control a smoothing degree of the weight distribution, N p represents a set of neighbor nodes of the node p, that is, a set of child nodes of the node p, c` represents all child nodes of the parent node p, f c` represents a feature vector of the commodity node c`, and <,> represents an inner product operation.

[0141] In a possible design, the guiding aggregation unit 202 is specifically configured to:

[0142] calculate the second weight by using the following formula:

[0143]

[0144] wherein s is a merchant node, r c is the utilization rate of the commodity corresponding to the commodity node c, a p,c |s is the weight of the commodity node c relative to its parent node p under the guidance of the merchant node s, and t2 is a smoothing parameter, r c` is the utilization rate of the commodity corresponding to the commodity node c.

[0145] In a possible design, the guidance aggregation unit 202 is specifically configured to:

[0146] weight the first weight and the second weight to calculate a target weight.

[0147] The calculation formula is specifically as follows:

[0148] a p,s,c = β · a p,c |u+ (1-β) · a p,c |s

[0149] wherein β is a weight coefficient, and the value range of β is 0≤β≤1, and the role of β is to adjust the relative importance of user demand and merchant resource management in the aggregation process, a p,s,c is the aggregation weight of the commodity node c relative to its parent node p and the merchant node s.

[0150] In a possible design, the first determination unit 203 is specifically configured to:

[0151] based on the target weight and the commodity node, determine a final feature combination of the commodity node through a convolution operation and layer combination.

[0152] The convolution operation formula is specifically as follows:

[0153]

[0154] wherein, is a feature vector of the node p at the j+1 layer under the guidance of the user u and the merchant s, j is any layer in the convolution layer, and 0≤j≤k-1.

[0155] The layer combination formula is specifically as follows:

[0156]

[0157] wherein, is the final feature combination of the commodity node, Comb is a combination operator, is the original embedding.

[0158] In a possible design, the generation unit 205 is specifically configured to:

[0159] The recommendation list is generated by the following formula:

[0160]

[0161] wherein sim is a similarity calculation function, TopK is a function of selecting K highest similarity commodities, f u is a user node feature.

[0162] The embodiment of the present application also provides a server, as shown in the figure, the server 300 of the embodiment comprises at least one processor 301, at least one network interface 304 or other user interface 303, a memory 305, and at least one communication bus 302. The server 300 can optionally contain a user interface 303, including a display, a keyboard or a clicking device. The memory 305 can contain a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory. The memory 305 stores execution instructions, when the server 300 runs, the processor 301 communicates with the memory 305, and the processor 301 calls the instructions stored in the memory 305 to execute the recommendation method described above. The operating system 306 contains various programs for implementing various basic services and processing tasks according to hardware. Figure 7 The server provided by the embodiment of the present application can execute the technical solutions of the embodiments of the recommendation method described above, and the implementation principles and technical effects are similar, which will not be described here.

[0163] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a computer to implement the method process related to the recommendation system in any method embodiment described above. Correspondingly, the computer can be the recommendation system described above.

[0164] The embodiment of the present application also provides a computer program or a computer program product comprising a computer program, which is executed on a computer, and will make the computer implement the method process related to the recommendation system in any method embodiment described above. Correspondingly, the computer can be the recommendation system described above.

[0165] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware. The foregoing program can be stored in a computer readable storage medium. The program is executed to execute the steps of the above-mentioned method embodiments, and the foregoing storage medium includes ROM, RAM, magnetic disc or optical disc and various storage program codes.

[0166] ​

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A recommendation method based on interest balance, characterized in that: include: Constructing user nodes, merchant nodes, product nodes, and edge relationships based on user information, merchant information, and product information, wherein the edge relationships are the edge relationships between the user node and the product node, and the edge relationships between the merchant node and the product node; Using the user node and the merchant node to guide and aggregate the product node to obtain a target weight corresponding to the product node jointly guided by the user node and the merchant node; Determining a feature combination of the commodity node according to the target weight; Determining a score of the product corresponding to the product node according to a feature combination of the product node and the user node; Generate a recommendation list based on the scores of the products.

2. The method according to claim 1, characterized in that The step of guiding and aggregating the commodity nodes by using the user node and the merchant node to obtain a target weight corresponding to the commodity node jointly guided by the user node and the merchant node includes: Determine a user interest guidance target and a merchant interest guidance target according to the user node and the merchant node; performing guidance aggregation on the product nodes by using the user interest guidance goal to obtain the first weight corresponding to the product nodes; Using the merchant interest guidance target to guide and aggregate the product nodes to obtain a second weight corresponding to the product nodes; A target weight corresponding to the commodity node is determined according to the first weight and the second weight.

3. The method according to claim 2, characterized in that The user interest guidance target is a user demand or interest preference, and the adopting of the user interest guidance target to guide aggregation of the product nodes to obtain a first weight corresponding to the product node includes: The first weight is calculated by the following formula: Among them, u is the user node, t u is the feature vector of user preference or user demand, c is the product node, f c is the feature vector of commodity node c, α p,c |u is the weight of the commodity node c relative to its parent node p under the guidance of the user node u, τ1 is the smoothing parameter, N p represents the set of neighboring nodes of the parent node p, c` represents all child nodes of the parent node p, and f c` is the feature vector of the commodity node c`, and <,> represents the inner product operation.

4. The method according to claim 2, characterized in that The merchant interest guidance target is utilization rate, and the step of guiding and aggregating the commodity nodes using the merchant interest guidance target to obtain a second weight corresponding to the commodity nodes includes: The second weight is calculated by the following formula: Among them, s is the merchant node, r c is the utilization rate of the commodity corresponding to commodity node c, α p,c |s is the weight of the commodity node c relative to its parent node p under the guidance of the merchant node s, τ2 is the smoothing parameter, N p represents the set of neighboring nodes of the parent node p, c` represents all child nodes of the parent node p, r c` is the utilization rate of the commodity corresponding to commodity node c'.

5. The method according to claim 2, characterized in that Determining the target weight corresponding to the commodity node according to the first weight and the second weight includes: The target weight is calculated by weighting and fusing the first weight and the second weight.

6. The method according to claim 1, characterized in that Determining the feature combination of the commodity node according to the target weight includes: Based on the target weight and the product node, the final feature combination of the product node is determined by convolution operation and layer combination. The convolution operation formula is: in, It is the feature vector of node p in the j+1 layer under the guidance of user u and merchant s, j is any layer in the convolution layer, and 0≤j≤k-1, α p,s,c is the target weight of the product corresponding to the product node c, The layer union formula is: in, is the final feature combination of the product node, Comb is the combination operator, is the original embedding of the product corresponding to the product node c, is the feature vector of the product corresponding to the product node c in the first layer, is the feature vector of the product corresponding to product node c in the kth layer.

7. A recommendation system based on interest balance, characterized in that: include: A construction unit, configured to construct a user node, a merchant node, a commodity node, and an edge relationship based on the user information, the merchant information, and the commodity information, wherein the edge relationship is an edge relationship between the user node and the commodity node, and an edge relationship between the merchant node and the commodity node; a guidance aggregation unit, configured to use the user node and the merchant node to perform guidance aggregation on the commodity node, so as to obtain a target weight corresponding to the commodity node jointly guided by the user node and the merchant node; A first determining unit, configured to determine a feature combination of the commodity node according to the target weight; A second determining unit, configured to determine a score of the product corresponding to the product node according to a feature combination of the product node and the user node; A generating unit is configured to generate a recommendation list according to the scores of the products.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

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

10. A computer program product comprising instructions, characterized in that When the instruction is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images