Live broadcast recommendation method and device

Abstract

This application discloses a live broadcast recommendation method and device that eliminates the problem of interest matching feature failure caused by the large traffic of large V anchors selling a variety of products, improves the confidence of the important feature of product in matching anchors and users, and directly improves the recommendation efficiency. The live broadcast recommendation method of this application will enhance the user's experience and play a good supporting role for small and medium-sized anchors selling medium and long-tail vertical products, and will also significantly increase user click-through rate and order rate.

Description

Technical Field

[0001] This application relates to but is not limited to Internet technology, and in particular to a live broadcast recommendation method and device based on a live broadcast e-commerce recommendation scenario. Background Art

[0002] Livestreaming e-commerce is a new sales channel that emerged from the integration of "livestreaming + e-commerce." Merchants or influencers use livestreaming platforms to promote products to consumers. Compared to traditional e-commerce models that display products through images and text, livestreaming offers a more intuitive approach.

[0003] The interacting entities in livestream recommendation systems include users, products, and livestream rooms. In livestream e-commerce recommendation scenarios, recommending products of interest to users is key to improving user experience. However, how to provide targeted recommendations for livestream rooms to increase conversion rates among viewers has become a pressing issue. Summary of the Invention

[0004] The present application provides a live broadcast recommendation method and device, which can improve the accuracy of product-based matching and increase the conversion rate of viewers in the live broadcast room to achieve good business results.

[0005] An embodiment of the present invention provides a live broadcast recommendation method, comprising:

[0006] Determine the anchor's product category distribution based on the anchor's historical live broadcast products;

[0007] Determine the product-based matching degree between the anchor and the user based on the anchor's product categories and product category distribution;

[0008] Recommend anchors to users based on the determined matching degree.

[0009] In an exemplary embodiment, determining the product category distribution of the anchor includes:

[0010] Determining the commodity classification distribution based on the commodity classification features mined from the anchor;

[0011] The commodity classification distribution is greater than a preset threshold, indicating that the classified commodities corresponding to the commodity classification distribution are the main selling commodities of the anchor, and the anchor is a vertical category anchor; the commodity classification distribution is not greater than a preset threshold, indicating that the classified commodities corresponding to the commodity classification distribution are the non-main selling commodities of the anchor, and the anchor is a comprehensive category anchor.

[0012] In an exemplary embodiment, recommending a host to the user based on the determined matching degree includes:

[0013] The matching degrees are sorted, and the anchors corresponding to the product categories with high matching degrees are recommended to the user.

[0014] In an exemplary embodiment, recommending the anchor corresponding to the product category with a high matching degree to the user includes:

[0015] The anchor corresponding to the product category with the highest matching degree is recommended to the user.

[0016] In an exemplary embodiment, recommending the anchor corresponding to the product category with a high matching degree to the user includes:

[0017] The anchor with high matching degree and good preset additional factors is recommended to the user.

[0018] An embodiment of the present application further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute any of the above-mentioned live broadcast recommendation methods.

[0019] An embodiment of the present application further provides a device for implementing live broadcast recommendation, comprising a memory and a processor, wherein the memory stores the following instructions executable by the processor: for executing the steps of any of the above-described live broadcast recommendation methods.

[0020] The embodiment of the present application further provides a live broadcast recommendation device, comprising: a first processing module, a second processing module, and a matching recommendation module; wherein,

[0021] The first processing module is configured to determine the host's product category distribution based on the host's historical live broadcast products;

[0022] The second processing module is configured to determine the product-based matching degree between the anchor and the user based on the anchor's product category and product category distribution;

[0023] The matching recommendation module is configured to recommend anchors to users based on the determined matching degree.

[0024] In an exemplary embodiment, the first processing module is specifically configured to:

[0025] Determining the commodity classification distribution based on the commodity classification features mined from the anchor;

[0026] The commodity classification distribution is greater than a preset threshold, indicating that the classified commodities corresponding to the commodity classification distribution are the main selling commodities of the anchor, and the anchor is a vertical category anchor; the commodity classification distribution is not greater than a preset threshold, indicating that the classified commodities corresponding to the commodity classification distribution are the non-main selling commodities of the anchor, and the anchor is a comprehensive category anchor.

[0027] In an exemplary embodiment, the matching recommendation module is specifically configured to: sort the matching degrees and recommend the anchor corresponding to the product category with a high matching degree to the user.

[0028] The live broadcast recommendation method and device provided in the embodiments of this application eliminates the problem of interest matching feature failure caused by the large traffic of large V anchors selling a variety of products, improves the confidence of the important feature of products in matching anchors and users, and directly improves the recommendation efficiency. The live broadcast recommendation method of this application will enhance the user's experience and play a good supporting role for small and medium-sized anchors selling medium and long-tail vertical products, and will also significantly improve user click-through rate and order rate.

[0029] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings are used to provide a further understanding of the technical solution of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application and do not constitute a limitation on the technical solution of the present application.

[0031] Figure 1 This is a flowchart of the live broadcast recommendation method in an embodiment of the present application;

[0032] Figure 2 This is a schematic diagram showing historical information on the host side and the user side in an embodiment of the present application;

[0033] Figure 3 Schematic diagram of the composition structure of the live broadcast recommendation device in an embodiment of the present application. DETAILED DESCRIPTION

[0034] To make the purpose, technical solutions and advantages of this application more clear, the embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be noted that, unless there is a conflict, the embodiments and features in the embodiments of this application can be combined with each other in any way.

[0035] In a typical configuration of the present application, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0036] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0037] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include non-transitory media such as modulated data signals and carrier waves.

[0038] The steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions. Also, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be performed in an order different from that shown here.

[0039] The product categories that have received the most exposure, clicks, and / or orders from the anchor over the past period are considered to be a very important feature. This feature is believed to describe the anchor from the perspective of the product and is the main characterization and expression of the anchor's product attributes. This feature is used to query the user's product browsing / search / add-to-cart history, hoping to use this feature to extract the degree of match with the user's product interests. However, for anchors who specialize in selling a certain category, using this feature to extract the degree of match with the user's product interests is not a problem. For example, anchors who specialize in selling jewelry, beef, a certain type of seafood, etc., the categories of products listed by these anchors are relatively simple and change very rarely. Therefore, using this feature to activate users' product interests is reasonable and accurate, and the resulting product-based interest matches are very helpful for overall anchor recommendations.

[0040] However, for influencers with a wide range of product categories, also known as top comprehensive influencers, the situation is different. For example, based on the statistics of the categories of product promotion / clicks / orders placed by this influencer over the past 14 days, it is found that the top categories at the top of the rankings vary greatly and the distribution of each category is not high. The statistics of the product promotion categories over the past 14 days include: beauty and skin care (13%), snacks (12%), women's clothing (8%), and fast food (6%). If the top product categories discovered by this influencer are directly used to recommend to users, the following problems will arise:

[0041] First, the distribution of product categories among the top 1, top 2, and even top 3 items varies very little. The top product categories generated by statistics in different time windows are different, and the categories themselves vary greatly. This leads to large differences in the effectiveness of the TOP1 product category feature in activating the same user's product history, causing the effect of this feature to be unstable. Second, these big V anchors are basically all-inclusive hodgepodge anchors (i.e., comprehensive anchors). Users who watch and buy mostly pursue the mentality of low prices across the entire network. They buy whatever type of product the anchor promotes, and the correlation strength with users' own product interests is far less than that of vertical anchors (i.e., anchors who focus on selling a certain type of product). In the recommendation scenario, this big V anchor has high exposure, and the amount of offline training samples generated is large, while the exposure of vertical mid- and long-tail anchors is relatively low. Ultimately, these features that are very effective for vertical mid- and long-tail anchors have little effect on the samples corresponding to the top anchors, causing the features learned by the model to become invalid and lose their effectiveness for vertical anchors. In summary, if only the mined products of livestreamers are used to represent livestreamers, neither the top nor the bottom products corresponding to influential livestreamers will accurately reflect the category to which they belong, resulting in inaccurate or ineffective features, which in turn affects the accuracy of product interests across all livestreamers. In other words, when using a user's product interests to query livestreamers, it's impossible to characterize whether the product is the livestreamer's main selling item. This can lead to inaccurate product-based interest matching for influential livestreamers. Furthermore, the large sample size of influential livestreamers generally weakens the accuracy of product-based matching between livestreamers and users.

[0042] To mitigate the impact of ineffective product category feature mining by live streamers, this application provides a live stream recommendation method that incorporates the distribution of categories corresponding to the live streamer's product hits, clicks, and orders. This method improves the accuracy of product-based matching between live streamers and users by identifying whether the live streamer is a vertical or comprehensive streamer.

[0043] Figure 1 This is a flow chart of the live broadcast recommendation method in the embodiment of the present application. Figure 1 Shown, including:

[0044] Step 100: Determine the host's product category distribution based on the host's historical live broadcast products.

[0045] There are three interactive entities in the live broadcast recommendation system: users, products, and live broadcast rooms (i.e., anchors). In the embodiment of the present application, considering that the big V anchors have large traffic, the types of products they sell are diverse and vary greatly, and the mined product classification features are inaccurate, then, after using this feature to interact with the user's product history sequence, there will be problems such as inaccurate characterization and expression, resulting in the failure or error of the product expression that should have played an important role. In the embodiment of the present application, based on the product classification features mined from the anchor, the distribution of the product classification on all the products of the anchor is further determined to describe whether the product classification is the main product of the anchor. By introducing the product classification distribution, it is possible to accurately perceive which products in the anchor's historical product sequence are the main products and which products are non-main products with a relatively small overall proportion, thereby increasing the accuracy of subsequent product-based matching.

[0046] In an exemplary embodiment, determining the product category distribution of the anchor in step 100 may include:

[0047] Determine the product category distribution based on the product category features mined from the anchor;

[0048] The determined product category distribution is greater than the preset threshold, indicating that the category products corresponding to the product category distribution are the main selling products of the anchor. For this type of products, the anchor can be a vertical category anchor; the determined product category distribution is not greater than the preset threshold, indicating that the category products corresponding to the product category distribution are the non-main selling products of the anchor. For this type of products, the anchor can be a comprehensive category anchor.

[0049] In one embodiment, determining the product category distribution based on the product category features mined from the anchor may include:

[0050] Based on the anchor's live broadcast history data over the past period of time (such as 90 days), the number of each category of goods sold during the live broadcast is counted. Then, the number of goods in each category is divided by the total number of goods to obtain the proportion or distribution of each category of goods.

[0051] In this embodiment, the distribution of a host's product categories is truncated at different thresholds. Only product categories exceeding the threshold are meaningful in representing the host's products. This allows us to clearly identify whether a host is a vertical or comprehensive category anchor for a particular product category. This discretized category distribution can then be used to activate product interest based on the user's top interest categories.

[0052] In an illustrative example, Figure 2As shown, the anchor's historical live broadcast product information may include: the static feature on the right side of the dotted line, which is the static expression of the anchor side, and is composed of various discrete ID-type features to characterize the anchor's broadcast time, broadcast frequency, age, category and other portrait information; and the dynamic feature list on the right side of the dotted line, which is the dynamic type information on the anchor side, composed of various on-air product sequences since the anchor started broadcasting. The products in the sequence may include auxiliary information (sideinformation) such as classification, price, number of views, and purchase volume.

[0053] Step 101: Determine the product-based matching degree between the host and the user based on the host's product categories and product category distribution.

[0054] In an exemplary embodiment, a deep interest network (DIN) or a transformer can be used to determine the degree of matching between the host and the user based on the product, so that the interest intensity information of the host's product category in the user's product history information can be obtained. In the embodiment of the present application, in terms of the user-host matching strength based on the product, the distribution of the corresponding product category in the host's historical sales is considered, which enhances the effectiveness of the expression matching results, removes non-interest matching caused by the host being a big V host, and improves the accuracy of the features. DIN or transformer is a sequence modeling method that can easily process the similarity between the host's on-air products and the user's historical behavior, thereby obtaining a matching degree based on the user and the host based on the product, for example, the matching degree obtained is 70%. Through the host's product category, the distribution or proportion of the category in all the host's product categories can be queried. Assuming it is 10%, then 10% multiplied by 70% can be used to correct the matching degree between the user and the host based on the product.

[0055] In an illustrative example, Figure 2 As shown, the user's product history information may include: the dynamic feature list on the left side of the dotted line, which is the user's dynamic information, which may include the user's product browsing and / or shopping cart and / or search history information over the past period of time. The user's product history information contains side information such as the number of times the user browsed, searched, and added to cart for the product, which describes the user's interests and hobbies in products over the past period of time; and the static feature on the left side of the dotted line, which is the static expression of the user side, which uses various discrete ID-type features to describe the user's age, gender, regional purchasing power and other portrait information. Figure 2The information on the right represents the static attributes of the host's products, such as price and category. These attributes are used to describe the host's live and featured products.

[0056] For example, suppose that the commodity category statistics of Big V anchor A include: beauty and skin care (51%), snacks (12%), women's clothing (8%), and fast food (6%); another anchor B is an anchor who specializes in fast food sales. In this embodiment, assuming that a user's interest product is fast food, then, if the method of determining the interest intensity information of the anchor's commodity category in the user's commodity history information in the relevant technology is used, it will be considered that Big V anchor A has a high degree of matching with the user based on fast food. However, if the live broadcast recommendation method provided in the embodiment of the present application is used, it will be considered that the commodity category distribution of fast food in Big V anchor A is relatively low. Assuming that the preset threshold is 50%, then, in this embodiment, for commodity categories with a distribution greater than 50%, it is considered that the anchor's main selling commodity. For such commodities with a distribution greater than 50%, the anchor is a vertical category anchor. In this way, in the embodiment of the present application, it will be considered that anchor B (with a fast food distribution of 100%) has a high degree of matching with the user based on fast food. Obviously, recommending a host who focuses on fast food to users who are interested in fast food is accurate and reasonable, and will also improve user satisfaction.

[0057] In an exemplary embodiment, the specific implementation of step 101 can be based on the widely used mined user head interests as a query, and a matching method for extracting the host's real-time interest in the products being broadcast is used through DIN or Transformer, which introduces the information of the host's product distribution. For example, the distribution of the host's historical product categories mined in advance is used to obtain the distribution information of the host's real-time products. In this way, the distribution ratio is discretized and bucketed. When calculating the interest matching degree between the user's head interests and the host's real-time product listings, the accuracy of the product matching is supplemented by additionally introducing the corresponding distribution of each product in the host's product listing history. In one embodiment, obtaining the distribution information of the host's real-time products may include: based on the host's live broadcast history data over a period of time (such as 90 days), counting the number of each category of products sold in the live broadcast, and then dividing the number of products in each category by the total number of products to obtain the proportion or distribution of each product category, and recording the distribution of each category. In this way, for the products being broadcast by the host in real time, the corresponding distribution can be queried by its category.

[0058] The diverse product categories of celebrity anchors naturally match users' product interests well. This match isn't due to the actual products themselves, but rather to the wide variety of products the celebrity anchors offer. However, in this embodiment of the application, while using DIN to express the intensity of interest in a anchor's product category within a user's product history, the product category distribution is also utilized, increasing the accuracy of the interest intensity calculated by DIN.

[0059] It reduces this type of false matching, reduces the mismatching of big V live broadcasts, and increases the correct matching of products for real vertical anchors, thereby improving user experience and supporting mid- and long-tail anchors.

[0060] Step 102: Recommending hosts to the user based on the determined matching degree.

[0061] In an exemplary embodiment, the matching degrees are sorted, and the hosts corresponding to the product categories with high matching degrees are recommended to the user.

[0062] In one embodiment, the anchor corresponding to the product category with the highest matching degree may be recommended to the user.

[0063] In one embodiment, preset additional factors may also be considered to recommend to users hosts whose matching degree is not the highest but whose comprehensive preset additional factors are better. That is to say, hosts with high matching degree and good preset additional factors are recommended to users. For example, for a product such as eggs with matching degree 1> matching degree 2> matching degree 3, the preset additional factor is eggs laid by free-range chickens. Assuming that the eggs corresponding to matching degree 1 are not laid by free-range chickens, and the eggs corresponding to matching degrees 2 and matching degrees 3 are laid by free-range chickens, then, if the comprehensive additional factor is eggs laid by free-range chickens, the hosts corresponding to the eggs with matching degree 2 may be recommended to users.

[0064] The live streaming recommendation method provided in the embodiments of this application eliminates the problem of interest matching failure caused by the large number of traffic and the diverse products sold by big V anchors. It increases the confidence of the important feature of product in matching anchors and users, and directly improves the efficiency of recommendation. The live streaming recommendation method of this application will enhance the user experience and play a good supporting role for small and medium-sized anchors selling medium and long-tail vertical products. The user click-through rate and order rate will also be significantly improved.

[0065] The present application also provides a computer-readable storage medium storing computer-executable instructions for executing Figure 1 Any one of the live broadcast recommendation methods.

[0066] The present application further provides a device for implementing live broadcast recommendation, comprising a memory and a processor, wherein the memory stores the following instructions that can be executed by the processor: Figure 1The steps of any one of the live broadcast recommendation methods.

[0067] Figure 3 This is a schematic diagram of the structure of the live broadcast recommendation device in the embodiment of the present application. Figure 3 As shown, it includes: a first processing module, a second processing module, and a matching recommendation module; wherein,

[0068] The first processing module is configured to determine the host's product category distribution based on the host's historical live broadcast products;

[0069] The second processing module is configured to determine the product-based matching degree between the anchor and the user based on the anchor's product category and product category distribution;

[0070] The matching recommendation module is configured to recommend anchors to users based on the determined matching degree.

[0071] In an exemplary embodiment, the first processing module may be specifically configured to:

[0072] Determine the product category distribution based on the product category features mined from the anchor;

[0073] The determined product category distribution is greater than the preset threshold, indicating that the category products corresponding to the product category distribution are the main selling products of the anchor. For this type of products, the anchor can be a vertical category anchor; the determined product category distribution is not greater than the preset threshold, indicating that the category products corresponding to the product category distribution are the non-main selling products of the anchor. For this type of products, the anchor can be a comprehensive category anchor.

[0074] In an exemplary embodiment, the second processing module can be specifically configured to use DIN or Transformer to determine the product-based match between the anchor and the user.

[0075] In one exemplary embodiment, the matching recommendation module can be configured to sort the matching degrees and recommend livestreamers corresponding to product categories with high matching degrees to the user. In one embodiment, livestreamers corresponding to product categories with the highest matching degrees are recommended to the user. In another embodiment, preset additional factors can also be considered to recommend livestreamers whose matching degrees are not the highest but who have a better overall performance based on the preset additional factors.

[0076] The live streaming recommendation device provided in the embodiments of this application eliminates the problem of interest matching failure caused by the large number of traffic and diverse products sold by big V anchors. It increases the confidence of the important feature of product in matching anchors and users, and directly improves the efficiency of recommendation. The live streaming recommendation device of this application will enhance the user experience and play a good supporting role for small and medium-sized anchors selling medium and long-tail vertical products. The user click-through rate and order rate will also be significantly improved.

[0077] Although the embodiments disclosed in this application are as described above, the contents described are merely embodiments adopted to facilitate understanding of this application and are not intended to limit this application. Any person skilled in the art to which this application belongs may make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in this application. However, the scope of patent protection of this application shall still be based on the scope defined by the attached claims.

Claims

1. A live broadcast recommendation method, comprising: Determining the product classification distribution of the anchor based on the anchor's historical live broadcast products includes: determining the product classification distribution based on the product classification features mined for the anchor; if the product classification distribution is greater than a preset threshold, it indicates that the classified products corresponding to the product classification distribution are the anchor's main selling products, and the anchor is a vertical category anchor; if the product classification distribution is not greater than the preset threshold, it indicates that the classified products corresponding to the product classification distribution are the anchor's non-main selling products, and the anchor is a comprehensive category anchor; Determine the product-based matching degree between the anchor and the user based on the anchor's product categories and product category distribution; Recommend anchors to users based on the determined matching degree.

2. The live broadcast recommendation method according to claim 1, wherein: The recommending a host to the user according to the determined matching degree includes: The matching degrees are sorted, and the anchors corresponding to the product categories with high matching degrees are recommended to the user.

3. The live broadcast recommendation method according to claim 2, wherein: The recommending the anchor corresponding to the product category with a high matching degree to the user includes: The anchor corresponding to the product category with the highest matching degree is recommended to the user.

4. The live broadcast recommendation method according to claim 2, wherein: The recommending the anchor corresponding to the product category with a high matching degree to the user includes: The anchor with high matching degree and good preset additional factors is recommended to the user.

5. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the live broadcast recommendation method according to any one of claims 1 to 4.

6. A device for implementing live broadcast recommendation, comprising a memory and a processor, wherein: The memory stores the following instructions that can be executed by the processor: used to execute the steps of the live broadcast recommendation method described in any one of claims 1 to 4.

7. A live broadcast recommendation device, comprising: A first processing module, a second processing module, and a matching recommendation module; wherein, The first processing module is configured to determine the host's product classification distribution based on the host's historical live broadcast products, including: determining the product classification distribution based on the product classification features mined for the host; if the product classification distribution is greater than a preset threshold, it indicates that the classified products corresponding to the product classification distribution are the host's main selling products, and the host is a vertical category host; if the product classification distribution is not greater than the preset threshold, it indicates that the classified products corresponding to the product classification distribution are the host's non-main selling products, and the host is a comprehensive category host; The second processing module is configured to determine the product-based matching degree between the anchor and the user based on the anchor's product category and product category distribution; The matching recommendation module is configured to recommend anchors to users based on the determined matching degree.

8. The live broadcast recommendation device according to claim 7, wherein: The matching recommendation module is specifically configured to sort the matching degrees and recommend the anchor corresponding to the product category with a high matching degree to the user.

Images