Resource recommendation method, deep learning model training method, device and equipment

Abstract

The invention provides a resource recommendation method, and relates to the technical field of artificial intelligence, in particular to the technical field of deep learning and intelligent recommendation. According to the specific implementation scheme, the method comprises the following steps: determining recent preference characteristics of a target user according to interaction behavior data of the target user in a historical time window of a live broadcast room; determining cross features according to a matching relationship between the recent preference features and respective resource features of the plurality of live broadcasting rooms; determining at least one target live broadcast room from the plurality of live broadcast rooms according to the recent preference characteristics, the respective resource characteristics of the plurality of live broadcast rooms and the cross characteristics; and outputting the information of the target live broadcast room. The invention further provides a deep learning model training method and device, electronic equipment and a storage medium.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to the field of deep learning and intelligent recommendation technology. More specifically, the present disclosure provides a resource recommendation method, a training method for a deep learning model, an apparatus, an electronic device, a storage medium, and a computer program product. Background Art

[0002] In the current e-commerce recommendation system, the user-side features of the recommendation model are more focused on the basic features of users, and the resource-side features are more based on the basic information of resources and products. These features lack timeliness and have limitations, resulting in a single recommendation model and a lack of personalization. Summary of the invention

[0003] The present disclosure provides a resource recommendation method, a deep learning model training method, an apparatus, an electronic device, a storage medium, and a computer program product.

[0004] According to a first aspect, a resource recommendation method is provided, the method comprising: determining a target user's recent preference characteristics based on the target user's interactive behavior data within a historical time window of a live broadcast room; determining cross-features based on a matching relationship between the recent preference characteristics and respective resource characteristics of a plurality of live broadcast rooms; determining at least one target live broadcast room from a plurality of live broadcast rooms based on the recent preference characteristics, respective resource characteristics of the plurality of live broadcast rooms, and cross-features; and outputting information of the target live broadcast room.

[0005] According to the second aspect, a training method for a deep learning model is provided, the method comprising: determining recent preference features of sample users based on the interactive behavior data of sample users within a historical time window of a live broadcast room; determining cross-features based on a matching relationship between recent preference features and respective resource features of a plurality of live broadcast rooms; using a deep learning model to determine recommendation indicators for each of the plurality of live broadcast rooms based on recent preference features, respective resource features of the plurality of live broadcast rooms, and cross-features; determining the loss of the deep learning model based on the recommendation indicators; and adjusting parameters of the deep learning model based on the loss to obtain a trained deep learning model.

[0006] According to a third aspect, a resource recommendation device is provided, which includes: a first preference feature determination module, used to determine the recent preference features of a target user based on the interactive behavior data of the target user within a historical time window of the live broadcast room; a first cross-feature determination module, used to determine the cross-feature based on the matching relationship between the recent preference features and the respective resource features of multiple live broadcast rooms; a target live broadcast room determination module, used to determine at least one target live broadcast room from multiple live broadcast rooms based on the recent preference features, the respective resource features of the multiple live broadcast rooms, and the cross-features; and an output module, used to output the information of the target live broadcast room.

[0007] According to a fourth aspect, a training device for a deep learning model is provided, the device comprising: a second preference feature determination module, for determining recent preference features of sample users based on interactive behavior data of sample users within a historical time window of a live broadcast room; a second cross-feature determination module, for determining cross-features based on a matching relationship between recent preference features and respective resource features of a plurality of live broadcast rooms; a first recommendation index determination module, for determining recommendation indexes of respective live broadcast rooms based on recent preference features, respective resource features of the plurality of live broadcast rooms, and cross-features using a deep learning model; a loss determination module, for determining the loss of the deep learning model based on the recommendation index; and an adjustment module, for adjusting parameters of the deep learning model based on the loss to obtain a trained deep learning model.

[0008] According to a fifth aspect, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method provided according to the present disclosure.

[0009] According to a sixth aspect, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute the method provided according to the present disclosure.

[0010] According to a seventh aspect, a computer program product is provided, comprising a computer program, wherein the computer program is stored on at least one of a readable storage medium and an electronic device, and wherein the computer program implements the method provided according to the present disclosure when executed by a processor.

[0011] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0013] Figure 1 is a schematic diagram of an exemplary system architecture to which a resource recommendation method and a training method for a deep learning model can be applied according to an embodiment of the present disclosure;

[0014] Figure 2 is a flow chart of a resource recommendation method according to an embodiment of the present disclosure;

[0015] Figure 3 is a flow chart of a resource recommendation method according to another embodiment of the present disclosure;

[0016] Figure 4 is a flowchart of a method for training a deep learning model according to an embodiment of the present disclosure;

[0017] Figure 5 is a schematic diagram of a training method for a deep learning model according to an embodiment of the present disclosure;

[0018] Figure 6 is a block diagram of a resource recommendation device according to an embodiment of the present disclosure;

[0019] Figure 7 is a block diagram of a training apparatus for a deep learning model according to an embodiment of the present disclosure; and

[0020] Figure 8 It is a block diagram of an electronic device according to at least one of a resource recommendation method and a deep learning model training method according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0021] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0022] In the current e-commerce recommendation system, user-side features are more focused on basic user information such as user basic information, live broadcast user portraits, and market user portraits. Resource-side information is more based on basic information about resources and products. In addition, the model will also introduce some business-side information, such as resources distributed by non-natural traffic and artificially promoted resources.

[0023] Although the current resource recommendation model has introduced a large number of user-side features and resource-side features, it still has some obvious drawbacks in the e-commerce live broadcast scenario. These drawbacks are mainly reflected in the following aspects:

[0024] Lack of timeliness: Live streaming is a very time-sensitive business, so it requires very fast feedback of timeliness information. In the current e-commerce recommendation system, user-side features are often based on user basic portraits, which can only reflect users' long-term preferences, but not their recent preferences. Therefore, the products recommended by the current model to users often lack effectiveness and cannot meet users' recent needs.

[0025] Feature limitations: Currently, most features focus on describing user-side behaviors and preferences for live content. Although these features can reflect users’ shopping tendencies to a certain extent, they often ignore more complex and detailed user behavior patterns. For example, the user’s dwell time, click frequency, and browsing order when browsing products can provide more information about user interests. In addition, live content characterizes user preferences for live content, such as lifestyle live broadcasts and sports live broadcasts. The description of live content features often focuses on achieving the goal of attracting users to watch the live broadcast room, rather than on achieving the goal of getting users to buy products in the live broadcast room.

[0026] The resource-side features also lack very detailed category descriptions. In e-commerce live streaming, there are many types of products, and different product categories may have completely different user groups and purchasing behaviors. Since the resource-side features cannot accurately describe product categories and only contain broad content information such as life and sports, the model cannot accurately match user needs when making recommendations.

[0027] Simplified recommendation content: Due to feature limitations, current e-commerce recommendation systems can only provide simple matching recommendations based on user history and product attributes, resulting in simplification of recommendation content and lack of diversity and innovation. This recommendation method may fail to attract user interest and even lead to user loss.

[0028] Lack of personalization: Although the model has introduced a series of user-side features, due to the lack of more detailed user behavior analysis and product category characterization, current recommendation systems often find it difficult to provide truly personalized recommendations. In e-commerce live broadcasts, users tend to have a higher interest in products in a specific category. If the recommendation system cannot accurately grasp the user's interests, the recommended products may not meet the user's needs.

[0029] In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0030] In the technical solution of the present disclosure, the user's authorization or consent is obtained before obtaining or collecting the user's personal information.

[0031] Figure 1 is a schematic diagram of an exemplary system architecture to which a resource recommendation method and a deep learning model training method can be applied according to an embodiment of the present disclosure. It should be noted that: Figure 1 What is shown is merely an example of a system architecture to which the embodiments of the present disclosure can be applied, in order to help those skilled in the art understand the technical content of the present disclosure, but it does not mean that the embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios.

[0032] like Figure 1 As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0033] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Terminal devices 101, 102, 103 can be various electronic devices, including but not limited to smart phones, tablet computers, laptops, etc.

[0034] At least one of the resource recommendation method and the deep learning model training method provided in the embodiments of the present disclosure can generally be executed by the server 105. Accordingly, the resource recommendation device and the deep learning model training device provided in the embodiments of the present disclosure can generally be set in the server 105. The resource recommendation method and the deep learning model training method provided in the embodiments of the present disclosure can also be executed by a server or a server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105. Accordingly, the resource recommendation device and the deep learning model training device provided in the embodiments of the present disclosure can also be set in a server or a server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105.

[0035] Figure 2 is a flowchart of a resource recommendation method according to an embodiment of the present disclosure.

[0036] like Figure 2 As shown, the resource recommendation method 200 includes operations S210 to S240.

[0037] In operation S210, recent preference characteristics of the target user are determined based on the target user's interactive behavior data within a historical time window of the live broadcast room.

[0038] For example, you can collect the target user's search history, click history, live broadcast viewing history, etc. within the historical time window to determine the integrity and accuracy of the data. The historical time window can be set as needed, such as 14 days. For the collected target user's interactive behavior data, you can perform data cleaning to remove duplicate, invalid or abnormal data from the collected data, such as removing deviating values ​​and data with incorrect formats, to ensure data quality.

[0039] Next, the interactive behavior characteristics of the target users can be extracted from the cleaned data, such as product card clicks, attention, interaction, shopping bag clicks, and live broadcast viewing time. For the continuous feature of duration, bucket processing can be performed, such as setting corresponding markers for each interval of a certain duration range, and processing the duration as a discrete feature. For example, 0~12s is marked as 1, 12~24s is marked as 2, and so on. The discretization of duration can capture the behavioral characteristics of target users in different duration ranges.

[0040] Next, the resource information associated with the above-mentioned behavior of the target user can be determined. For example, the list of anchor IDs and the list of product categories for which the target user has behaviors such as following / clicking / collecting / watching / searching / interacting / converting. The above-mentioned interactive behavior characteristics, as well as the list of anchor IDs and the list of product categories associated with the interactive behavior characteristics can be determined as the recent preference characteristics of the target user. For example, the recent preference characteristics of the target user can include the list of anchor IDs that the target user follows, the list of categories of products that the target user clicks, the list of anchor IDs that the target user has converted, the list of product categories that the target user has converted, and so on.

[0041] The above-mentioned recent preference features can be obtained by collecting and extracting the live broadcast room interaction behavior data of the target user at each historical time window, and can reflect the recent preferences of the target user.

[0042] In operation S220, a cross feature is determined based on a matching relationship between the recent preference feature and the resource features of each of the plurality of live broadcast rooms.

[0043] For each live broadcast room, it can be determined whether the anchor ID list and product category list in the recent preference characteristics of the target user match the anchor ID and product category of the live broadcast room. For example, if the anchor ID list in the recent preference characteristics of the target user contains the anchor ID of the live broadcast room, or the product category list in the recent preference characteristics of the target user contains the product category of the live broadcast room, it can be determined that the recent preference characteristics of the target user match the resource characteristics of the live broadcast room. If the anchor ID list and product category list in the recent preference characteristics of the target user do not contain the anchor ID and product category of the live broadcast room, it can be determined that the recent preference characteristics of the target user do not match the resource characteristics of the live broadcast room.

[0044] According to the matching between the recent preference characteristics of the target user and the resource characteristics of the live broadcast room, a corresponding flag can be set to indicate the cross-feature. For example, when the recent preference characteristics of the target user match the resource characteristics of the live broadcast room, flag 1 can be set to indicate the cross-feature. Otherwise, flag 0 can be set to indicate the cross-feature.

[0045] For each live broadcast room, the cross-features between the target user and the live broadcast room can be determined based on the matching between the recent preference features of the target user and the resource features of the live broadcast room. Multiple cross-features can be determined for multiple live broadcast rooms.

[0046] It should be noted that the resource features of the live broadcast room can be new resource-side features that introduce anchor ID and product category. It can be understood that the original basic resource-side features of the live broadcast room generally only focus on the basic attribute information of the resource and lack detailed category description.

[0047] The embodiments of the present disclosure divide the product categories on the resource side into fine categories, for example, fine categories into primary categories and secondary categories, to obtain fine category characteristics, which can more accurately describe the characteristics of the resource side.

[0048] The embodiments of the present disclosure determine cross-features based on the matching relationship between recent preference features and resource features, which can capture the interaction between recent preference features and resource-side features and enhance the model's memory of recent preference features of target users.

[0049] In operation S230, at least one target live broadcast room is determined from the plurality of live broadcast rooms according to the recent preference feature, the resource features of the plurality of live broadcast rooms, and the cross feature.

[0050] For example, the recent preference characteristics of the target user, the resource characteristics of multiple live broadcast rooms, and the cross-features between the target user and each live broadcast room can be input into the resource recommendation model to obtain indicators such as the target user's click-through rate, conversion rate, and total transaction amount for each live broadcast room. These indicators represent the probability that the target user is interested in the live broadcast. Based on these indicators, the comprehensive evaluation value of the live broadcast room can be determined, and the live broadcast room can be recommended to the target user based on the comprehensive evaluation value. For example, at least one live broadcast room with the highest comprehensive evaluation value can be recommended to the target user.

[0051] In operation S240, information of the target live broadcast room is output.

[0052] For example, the target live broadcast room may be at least one live broadcast room with the highest comprehensive evaluation value, and information such as the identification or entrance of these live broadcast rooms may be sent to the terminal of the target user so that the terminal of the target user can display these live broadcast rooms to the target user.

[0053] The embodiments of the present disclosure determine the recent preference characteristics of the target user based on the interactive behavior data of the target user within the historical time window of the live broadcast room, determine the cross-features based on the matching relationship between the recent preference characteristics and the resource-side characteristics, and recommend the live broadcast room to the target user based on the recent preference characteristics, resource characteristics and cross-features of the target user, thereby improving the timeliness and accuracy of resource recommendations.

[0054] The embodiments of the present disclosure refer to recent preference features and recommend live broadcast rooms to target users based on the recent preference features, which can better meet the recent needs of target users and improve the timeliness of recommended resources. The embodiments of the present disclosure introduce cross features, which characterize the matching relationship between the recent preference features of target users and resource-side features, which can enhance the model's memory of the recent preference features of target users and improve the accuracy of resource recommendations.

[0055] Figure 3 is a flowchart of a resource recommendation method according to another embodiment of the present disclosure.

[0056] like Figure 3 As shown, this embodiment includes operations S311 to S314, operations S321 to S322, and operations S331 to S332. Among them, operations S311 to S314 are steps for constructing user-side features, operations S321 to S322 are steps for constructing resource-side features, and operations S331 to S332 are steps for using the recommendation model to recommend resources based on the constructed user-side features and resource-side features.

[0057] In operation S311, the interactive behavior data of the target user within the historical time window of the live broadcast room is collected.

[0058] For example, you can collect the target user's search history, click history, live broadcast viewing history, etc. within a historical time window (such as 14 days). Clean the collected data to remove duplicate, invalid or abnormal data.

[0059] In operation S312, the interactive behavior features and the host identification list and product category list associated with the interactive behavior features are extracted from the interactive behavior data to obtain recent preference features.

[0060] According to an embodiment of the present disclosure, interactive behavior features and at least one of an anchor identifier and an object category associated with the interactive behavior features are extracted from interactive behavior data; and recent preference features are determined based on the interactive behavior features and at least one of the anchor identifier and the object category.

[0061] For example, the interactive behavior features of the target user can be extracted from the cleaned data, such as product card clicks, attention, interaction, shopping bag clicks, and live broadcast room viewing time. Then determine the anchor ID list and product category list associated with the above-mentioned behaviors of the target user. The above-mentioned interactive behavior features and the anchor ID list and product category list associated with the interactive behavior are used as recent preference features. For example, recent preference features can include the anchor ID list that the target user follows, the category list of goods clicked by the target user, the anchor ID list that the target user has conversion behavior, the category list of goods that the target user has conversion behavior, and so on.

[0062] In operation S313 , a cross feature and a recommendation weight are determined according to a matching relationship between the recent preference feature and the resource-side feature.

[0063] According to an embodiment of the present disclosure, for each live broadcast room, the cross-features between the target user and the live broadcast room are determined based on the matching relationship between at least one of the anchor identification and the object category in the recent preference feature and at least one of the anchor identification and the object category in the live broadcast room. In response to the target anchor identification and at least one of the target object category in the recent preference feature matching at least one of the anchor identification and the object category in the live broadcast room, the number of conversions of the conversion behavior is determined as the recommendation weight.

[0064] In order to capture the relationship between the recent preference features of the target user and the resource-side features, the embodiment of the present disclosure designs a feature cross-layer to determine the matching relationship between the recent preference features and the resource-side features.

[0065] For example, the feature cross-layer is used to determine whether recent preference features (such as anchor ID list / product category list) match the anchor ID and product category on the resource side, that is, whether the anchor ID list / product category list that the target user recently prefers contains the anchor ID / product category of the live broadcast room on the resource side.

[0066] In addition, the characterization of the features of the embodiments of the present disclosure focuses more on achieving the goal of getting the target user to purchase the product in the live broadcast room. Therefore, the conversion behavior of the target user in the historical time window is more valuable for reference. Therefore, when it is found that the recent preference features of the target user match the anchor ID or product category of the live broadcast room on the resource side, the number of conversions of the target user on the anchor or category in the historical time window can be counted as the recommendation weight. The recommendation weight can also be introduced into the recommendation model as a feature of a dimension.

[0067] It should be noted that the resource-side features that match the target user's recent preference features can be new resource-side features that introduce the live broadcast room host ID and live broadcast room product category features. In other words, the categories of the live broadcast room products are accurately described on the resource side, making the resource-side features more detailed.

[0068] In operation S314, the recent preference feature is added to the historical user-side feature to obtain a new user-side feature.

[0069] According to an embodiment of the present disclosure, recent preference features are added to historical user-side features to obtain new user-side features, where the historical user-side features are obtained based on basic user-side features of the target user and recent preference features of at least one historical time window.

[0070] Adding recent preference features to historical user-side features may include adding interactive behavior features and a host ID list and a product category list associated with the interactive behavior features to the historical user-side features to obtain new user-side features.

[0071] Historical user-side features can be obtained by regularly adding recent preference features to the basic user-side features of the target user. Basic user-side features can be basic user information, live broadcast user portraits, market user portraits, etc.

[0072] For example, the recent preference features in this historical time window can be added to the basic user-side features to obtain new user-side features. For the next time window, the new user-side features obtained by this update become historical user-side features. The historical user-side features represent the long-term preferences of the target user. The new user-side features include historical user-side features and recent preference features, and can take into account both the long-term and recent preferences of the target user.

[0073] The embodiments of the present disclosure dynamically update the interest profile of the target user, including the preferred anchor identification and category, based on the target user's recent behavior data, and can set a suitable update frequency to ensure the timeliness of the target user's interest profile, thereby improving the timeliness of resource recommendations.

[0074] In operation S321, the anchor ID and product category characteristics of the live broadcast room are determined.

[0075] In operation S322, the anchor ID and product category features are added to the basic resource side features to obtain new resource side features.

[0076] Operation S321 to operation S322 are the construction of the live broadcast room characteristics on the resource side. The basic resource side characteristics may include basic information on the live broadcast content and product attributes.

[0077] The embodiments of the present disclosure can accurately describe the product categories on the resource side by introducing anchor ID and product category features on the resource side. Compared with the simple matching recommendation based on user historical behavior and product attributes in the related art, which leads to the single recommended content, the embodiments of the present disclosure accurately match the product categories that the target user is interested in, so that the recommended content is more in line with the needs of the target user, and realizes truly personalized recommendations.

[0078] In operation S331, new user-side features, new resource-side features, cross-features, and recommendation weights are integrated and input into a recommendation model to obtain at least one recommendation index.

[0079] The new user-side features, cross features, and recommendation weights constructed in operations S311 to S314 and the new resource-side features constructed in operations S321 to S322 are integrated to obtain integrated features. The integrated features are input into the recommendation model, and the recommendation model can output at least one recommendation indicator, such as click-through rate, conversion rate, and total transaction amount.

[0080] In operation S332, the live broadcast room is recommended to the target user according to the recommendation index.

[0081] According to an embodiment of the present disclosure, an evaluation value of a live broadcast room is determined based on at least one of a click-through rate, a conversion rate, and a total transaction amount, and at least one target live broadcast room is determined from the multiple live broadcast rooms based on their respective evaluation values.

[0082] Click-Through-Rate (CTR), Conversion Rate (CVR) and Gross Merchandise Volume (GMV) all represent the probability that the target user is interested in the live broadcast room. The comprehensive evaluation value of a live broadcast room can be determined based on the click-through rate, conversion rate and total transaction amount. For example, the comprehensive evaluation value can be obtained by multiplying the click-through rate, conversion rate and total transaction amount. Each live broadcast room can obtain its own comprehensive evaluation value, and at least one live broadcast room with the highest comprehensive evaluation value can be selected from multiple live broadcast rooms to recommend to the target user.

[0083] Compared with the related art, the user-side features only focus on the basic user information, and the resource-side features only focus on the live broadcast content and basic product information. The embodiments of the present disclosure add the target user's recent preference features for anchor ID and product categories on the user side, such as the anchor ID list / product category list associated with the target user's attention, click, conversion and other behaviors. On the resource side, the primary and secondary category features of the product are also accurately described. A feature cross-layer is also designed to capture whether the recent preference features match the resource-side features, that is, whether the target user's preferred anchor ID list / product category list contains the anchor ID / product category on the resource side. The introduction of the above series of features can more accurately match the needs of the target user and improve the accuracy of resource recommendations.

[0084] According to an embodiment of the present disclosure, the present disclosure also provides a training method for a deep learning model.

[0085] Figure 4 It is a flowchart of a method for training a deep learning model according to an embodiment of the present disclosure.

[0086] like Figure 4 As shown, the training method 400 of the deep learning model includes operations S410 to S450. The deep learning model of the embodiment of the present disclosure may be a resource recommendation model.

[0087] In operation S410, recent preference characteristics of the sample user are determined based on the interactive behavior data of the sample user within a historical time window of the live broadcast room.

[0088] For example, you can collect the sample user's search records, click records, live broadcast viewing records, etc. within a historical time window (such as 14 days). Clean the collected data to remove duplicate, invalid or abnormal data.

[0089] Next, the interactive behavior features of the sample users can be extracted from the cleaned data, such as product card clicks, attention, interaction, shopping bag clicks, and live broadcast room viewing time. Then determine the anchor ID list and product category list associated with the above-mentioned behaviors of the sample users. The interactive behavior features of the above-mentioned sample users and the anchor ID list and product category list associated with the interactive behavior features can be used as the recent preference features of the sample users. For example, the recent preference features of the sample users include the anchor ID list followed by the sample users, the category list of the goods clicked by the sample users, the anchor ID list for the sample users to have conversion behaviors, the category list of the goods for which the sample users have conversion behaviors, and so on.

[0090] In operation S420, a cross feature is determined based on a matching relationship between the recent preference feature and the resource features of each of the plurality of live broadcast rooms.

[0091] In order to capture the relationship between the recent preference features of sample users and the resource-side features, the embodiments of the present disclosure design a feature cross-layer to determine the matching relationship between the recent preference features and the resource-side features.

[0092] For example, the feature cross-layer is used to determine whether recent preference features (such as anchor ID list / product category list) match the anchor ID and product category on the resource side, that is, whether the anchor ID list / product category list that the sample user recently prefers contains the anchor ID / product category of the live broadcast room on the resource side.

[0093] It should be noted that the resource-side features matched with the recent preference features of the sample users can be new resource-side features that introduce the live broadcast room host ID and live broadcast room product category features. In other words, the categories of the live broadcast room products are accurately described on the resource side, making the resource-side features more detailed.

[0094] The embodiments of the present disclosure determine cross-features based on the matching relationship between recent preference features and resource features, which can capture the interaction between user-side features and resource-side features and enhance the model's memory of the preference features of sample users.

[0095] In operation S430, a recommendation index for each of the plurality of live broadcast rooms is determined based on recent preference features, resource features of each of the plurality of live broadcast rooms, and cross-features using a deep learning model.

[0096] For example, the recent preference characteristics of the sample user, the resource characteristics of multiple live broadcast rooms, and the cross-features between the sample user and each live broadcast room can be input into the resource recommendation model to obtain indicators such as the click-through rate, conversion rate, and total transaction amount of the sample user for each live broadcast room. These indicators represent the probability that the sample user is interested in the live broadcast.

[0097] In operation S440, the loss of the deep learning model is determined according to the recommended indicator.

[0098] For example, the actual click-through rate, conversion rate, and total transaction amount of the sample user in the live broadcast room can be determined based on the historical interactive behavior data of the sample user in the live broadcast room. Then, the click-through rate, conversion rate, and total transaction amount output by the deep learning model are compared with the actual click-through rate, conversion rate, and total transaction amount to calculate the loss, and obtain the click-through rate loss, conversion rate loss, and total transaction amount loss. Based on the click-through rate loss, conversion rate loss, and total transaction amount loss, the overall loss of the deep learning model is determined.

[0099] In operation S450, the parameters of the deep learning model are adjusted according to the loss to obtain a trained deep learning model.

[0100] According to the above overall loss, the parameters of the deep learning model can be adjusted to complete a round of training. Next, for the next or next batch of sample users, repeat the above operation until the preset stop condition is reached to obtain a trained deep learning model. The above stop condition can be, for example, that the accuracy of the model reaches a threshold, or the number of training times reaches a preset number of times, etc.

[0101] The above-mentioned trained deep learning model introduces the recent preference features of sample users for anchor IDs and product categories on the user side, and introduces anchor IDs and product categories on the resource side. In addition, it also introduces cross-features that characterize the matching relationship between recent preference features and resource-side features. The above-mentioned series of introduced features can improve the accuracy of recommendation model predictions, thereby improving the accuracy of resource recommendations.

[0102] Compared with the related art, in which user-side features only focus on basic user information and resource-side features only focus on live content and basic product information, the embodiments of the present disclosure focus on deepening user behavior analysis to more accurately depict user interests and preferences, thereby optimizing the recommendation effect of the recommendation system.

[0103] Combine the following Figure 5 The training method of the deep learning model disclosed in the present invention is described.

[0104] Figure 5 It is a schematic diagram of a training method for a deep learning model according to an embodiment of the present disclosure.

[0105] like Figure 5 As shown, input features 510 are input into a deep learning model 520 to obtain multiple recommendation indicators, such as click-through rate, conversion rate, total transaction amount, shallow conversion rate, etc. According to the difference between each indicator output by the deep learning model 520 and the actual indicator, the overall loss 530 is determined, and the parameters of the deep learning model are adjusted according to the overall loss 530.

[0106] Input features 510 include constructed new user-side features, which include recent preference features and historical user-side features. Recent preference features may include recent (such as within the most recent historical time window) interactive behavior features with the user and a list of anchor IDs and a list of product categories associated with the interactive behavior features. Historical user-side features are obtained based on the basic user-side features of sample users and recent preference features of multiple historical time windows. In other words, the user's recent preference features are calculated regularly and added to the historical user-side features to update the user-side features and ensure the timeliness of the user-side features.

[0107] In addition, the embodiments of the present disclosure not only stop at the simple record of user behavior history, but also further explore the potential value of these data through cross-feature layers. By taking the anchor list and product category list preferred by the user in a specific historical time window, and exploring the matching relationship between them and the anchor identification or primary and secondary categories on the resource side. Once a match is found, it is also possible to count the number of conversions that occurred to this user on this specific anchor or category in the historical time window. This data will undoubtedly provide a valuable reference for the recommendation model. Therefore, the number of conversions is also introduced into the recommendation model as a recommendation weight to further improve the model performance.

[0108] The embodiments of the present disclosure introduce a series of new features into the recommendation model, which are based on the user's behavior data in key scenarios such as search, click, and live broadcast in the past historical time window. By capturing the list of anchor IDs in which the user has converted in these scenarios, as well as the list of related primary and secondary categories, the user's conversion interest history can be more comprehensively portrayed.

[0109] This measure is particularly critical when users are cold-starting, because for users who are new to or rarely use live streaming, product category portraits associated with user search / follow / collection behaviors can significantly improve the accuracy of recommendations. The embodiments of the present disclosure connect these features to the recommendation model, which can improve the recommendation effect of the model for cold-start users.

[0110] The deep learning model 520 is a multi-task learning model, including a click-through rate task sub-model 521 , a conversion rate task sub-model 522 , a transaction amount task sub-model 523 , a shallow conversion rate task sub-model 524 , and a shared network 525 .

[0111] Each task sub-model includes its own expert network and tower network, and the tower network may include multiple MLPs (Multilayer Perceptron). Input features 510 are input to the expert network of each sub-model, and the expert network of each sub-model performs expert processing and outputs expert features. The expert network of each sub-model refers to a network used to specifically process features for each task. These expert networks may have the same structure, but may have different parameter values ​​because they learn different tasks.

[0112] The expert features of each expert network are input into the corresponding tower network. In addition, the input features 510 are also input into the shared network 525 to obtain shared features. The shared features are also input into the tower network of each task sub-model.

[0113] Each sub-model's respective tower network processes the input expert features and shared features, and finally outputs its own task indicators. For example, the click-through rate task sub-model 521 outputs the click-through rate, the conversion rate task sub-model 522 outputs the conversion rate, the transaction amount task sub-model 523 outputs the transaction amount, and the shallow conversion rate task sub-model 524 outputs shallow conversion rate indicators such as whether the product card is clicked and the viewing time.

[0114] Based on the interactive behavior data of sample users within the historical time window, we can determine the users’ actual click-through rate, conversion rate, total transaction amount, whether they clicked on the product card, viewing time and other shallow conversion rate indicators for the live broadcast room.

[0115] According to the difference between each recommendation index output by the deep learning model 520 and the actual recommendation index of the sample user, the loss of each recommendation index can be determined, such as click-through rate loss, conversion rate loss, transaction amount loss, and shallow conversion rate loss. According to the loss of each index, the overall loss 530 of the deep learning model 520 can be determined. For example, the loss of each index can be weighted and summed to obtain the overall loss 530. According to the overall loss 530, the parameters of each network in the deep learning model 520 can be adjusted.

[0116] For example, the parameters of the expert network and tower network of each task sub-model in the deep learning model 520 may be adjusted, and the parameters of the shared network may also be adjusted, so that each task sub-model may have the same network structure but different parameter values, so that each task sub-model may output indicators of different tasks.

[0117] The embodiments of the present disclosure can understand the personalized needs of users in more detail by introducing new user-side features, new resource-side features, cross-features, and recommendation weights. By performing model training and parameter adjustment based on the introduced features, the model can more accurately predict the user's future indicators for the live broadcast room, thereby achieving a comprehensive improvement in the performance of the recommendation system.

[0118] During the model training process, the best model configuration is found by adjusting hyperparameters and optimization algorithms. At the same time, feature importance evaluation methods can be used to determine the contribution of each feature to model performance and eliminate features that have a negative impact on model gain.

[0119] According to an embodiment of the present disclosure, based on the interactive behavior data of the verified user within the historical time window of the live broadcast room, the recommendation index output by the trained deep learning model for the verified user is determined; the gain in the accuracy of the recommendation index is determined; and in response to the gain being negative, at least one of the anchor identification and object category in the recent preference feature is removed from the new user-side feature to obtain a new historical user-side feature.

[0120] For example, the recent preference features of the verified user can be determined based on the interactive behavior data of the verified user in the historical time window of the live broadcast room; the cross-features can be determined based on the matching relationship between the recent preference features of the verified user and the resource features of each of the multiple live broadcast rooms. The recommendation indicators of each of the multiple live broadcast rooms can be determined based on the recent preference features, the resource features of each of the multiple live broadcast rooms, and the cross-features using an untrained deep learning model and a trained deep learning model. Based on the interactive behavior data of the verified user in the historical time window of the live broadcast room, the actual recommendation indicators of the verified user for the multiple live broadcast rooms can also be determined.

[0121] Next, the prediction accuracy of the untrained model is determined based on the recommended indicators output by the untrained model and the actual recommended indicators. The prediction accuracy of the trained model is determined based on the recommended indicators output by the trained model and the actual indicators.

[0122] Then compare the prediction accuracy of the untrained model with the prediction accuracy of the trained model to determine whether the accuracy of the trained model has been improved. If so, it can be determined that the newly introduced recent preference features have a positive gain on the prediction effect of the model, and the newly introduced user preference features can be retained in the new user-side features. Otherwise, it can be determined that the newly introduced recent preference features have a negative gain on the prediction effect of the model, and at least one of the anchor ID, product category, and interactive behavior features in the recent preference features can be removed from the new user-side features.

[0123] The new user-side features after retaining or removing the newly introduced user preference features can be used to train the model again until it is determined that the retained user-side features have all produced positive gains for the model, and the trained model can be deployed in the production environment. The new user-side features after retaining or removing the newly introduced user preference features become the new historical user-side features for updating and application of the model in the next time window.

[0124] After the model training is completed, the model can be deployed to the production environment and the performance of the model can be monitored in real time. Based on the real-time data feedback, the model can be updated regularly to maintain its timeliness and accuracy.

[0125] According to an embodiment of the present disclosure, the present disclosure also provides a resource recommendation device and a deep learning model training device.

[0126] Figure 6 is a block diagram of a resource recommendation device according to an embodiment of the present disclosure.

[0127] like Figure 6As shown, the resource recommendation device 600 includes a first preference feature determination module 610, a first cross-feature determination module 620, a target live broadcast room determination module 630 and an output module 640.

[0128] The first preference feature determination module 610 is used to determine the recent preference features of the target user based on the target user's interactive behavior data within a historical time window of the live broadcast room.

[0129] The first cross-feature determination module 620 is used to determine the cross-feature according to the matching relationship between the recent preference feature and the resource features of each of the multiple live broadcast rooms.

[0130] The target live broadcast room determination module 630 is used to determine at least one target live broadcast room from multiple live broadcast rooms based on recent preference characteristics, resource characteristics of each of the multiple live broadcast rooms, and cross-characteristics.

[0131] The output module 640 is used to output the information of the target live broadcast room.

[0132] The first preference feature determination module 610 includes a first feature extraction unit and a first feature determination unit.

[0133] The first feature extraction unit is used to extract the interactive behavior feature and at least one of the anchor identification and the object category associated with the interactive behavior feature from the interactive behavior data.

[0134] The first feature determination unit is used to determine a recent preference feature according to the interactive behavior feature and at least one of the anchor identifier and the object category.

[0135] The first cross-feature determination module 620 is used to determine the cross-features between the target user and the live broadcast room for each live broadcast room based on the matching relationship between at least one of the anchor identification and object categories in the recent preference features and at least one of the anchor identification and object categories in the live broadcast room.

[0136] According to an embodiment of the present disclosure, the interactive behavior includes a conversion behavior. The target live broadcast room determination module 630 includes a target determination unit, a recommendation weight determination unit and a first target live broadcast room determination unit.

[0137] The target determination unit is used to determine the target anchor identifier and target object category associated with the conversion behavior.

[0138] The recommendation weight determination unit is used to determine the number of conversion behaviors for each live broadcast room, in response to matching at least one of the target anchor identification and target object category in the recent preference characteristics with at least one of the anchor identification and object category of the live broadcast room, as the recommendation weight.

[0139] The first target live broadcast room determination unit is used to determine at least one target live broadcast room from multiple live broadcast rooms according to recent preference characteristics, resource characteristics of each of the multiple live broadcast rooms, cross-features, and recommendation weights.

[0140] The target live broadcast room determination module 630 includes a first user-side feature updating unit and a second target live broadcast room determination unit.

[0141] The first user-side feature updating unit is used to add recent preference features to historical user-side features to obtain new user-side features, wherein the historical user-side features are obtained based on basic user-side features of the target user and recent preference features of at least one historical time window.

[0142] The second target live broadcast room determination unit is used to determine at least one target live broadcast room from the multiple live broadcast rooms according to the new user side characteristics, the new resource side characteristics of the multiple live broadcast rooms respectively, and the cross characteristics.

[0143] Among them, the new resource-side feature is obtained by adding at least one of the anchor identification and object category of the live broadcast room to the basic resource-side feature of the live broadcast room.

[0144] The second target live broadcast room determination unit includes an evaluation value determination subunit and a target live broadcast room determination subunit.

[0145] The evaluation value determination subunit is used to determine the evaluation value of each live broadcast room according to new user-side features, new resource-side features of the live broadcast room, and cross-features between the target user and the live broadcast room.

[0146] The target live broadcast room determination subunit is used to determine at least one target live broadcast room from multiple live broadcast rooms according to respective evaluation values ​​of the multiple live broadcast rooms.

[0147] An evaluation value determination subunit is used to determine at least one of the target user's click-through rate, conversion rate and total transaction amount for the live broadcast room based on new user-side characteristics, new resource-side characteristics of the live broadcast room, and cross-characteristics between the target user and the live broadcast room; and to determine the evaluation value of the live broadcast room based on at least one of the click-through rate, conversion rate and total transaction amount.

[0148] Figure 7 is a block diagram of a training device for a deep learning model according to an embodiment of the present disclosure.

[0149] like Figure 7 As shown, the training device 700 of the deep learning model includes a second preference feature determination module 710, a second cross feature determination module 720, a first recommendation index determination module 730, a loss determination module 740 and an adjustment module 750.

[0150] The second preference feature determination module 710 is used to determine the recent preference features of the sample user based on the interactive behavior data of the sample user within the historical time window of the live broadcast room.

[0151] The second cross-feature determination module 720 is used to determine the cross-feature according to the matching relationship between the recent preference feature and the resource features of each of the multiple live broadcast rooms.

[0152] The first recommendation index determination module 730 is used to determine the recommendation index of each of the multiple live broadcast rooms based on recent preference features, resource features of each of the multiple live broadcast rooms, and cross-features using a deep learning model.

[0153] The loss determination module 740 is used to determine the loss of the deep learning model according to the recommended indicator.

[0154] The adjustment module 750 is used to adjust the parameters of the deep learning model according to the loss to obtain a trained deep learning model.

[0155] The second preference feature determination module 710 includes a second feature extraction unit and a second feature determination unit.

[0156] The second feature extraction unit is used to extract the interactive behavior feature and at least one of the anchor identification and the object category associated with the interactive behavior feature from the interactive behavior data.

[0157] The second feature determination unit is used to determine the recent preference feature according to the interactive behavior feature and at least one of the anchor identifier and the object category.

[0158] The first recommendation index determination module 730 includes a second user-side feature updating unit and a recommendation index determination unit.

[0159] The second user-side feature updating unit adds the recent preference feature to the historical user-side feature to obtain a new user-side feature, wherein the historical user-side feature is obtained based on the basic user-side feature of the sample user and the recent preference feature of at least one historical time window.

[0160] The recommendation index determination unit is used to use a deep learning model to determine the recommendation index of each of the multiple live broadcast rooms according to new user-side features, new resource-side features of each of the multiple live broadcast rooms, and cross-features.

[0161] Among them, the new resource-side feature is obtained by adding at least one of the anchor identification and object category of the live broadcast room to the basic resource-side feature of the live broadcast room.

[0162] The deep learning model training device 700 also includes a second recommendation index determination module, a gain determination module and a elimination module.

[0163] The second recommendation indicator determination module is used to determine the recommendation indicator output by the trained deep learning model for the verified user based on the interactive behavior data of the verified user within the historical time window of the live broadcast room.

[0164] The gain determination module is used to determine the gain of the accuracy of the recommendation indicator.

[0165] The elimination module is used to remove the interactive behavior feature in the recent preference feature and at least one of the anchor identification and the object category from the new user-side feature in response to the gain being negative, so as to obtain a new historical user-side feature.

[0166] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0167] Figure 8 A schematic block diagram of an example electronic device 800 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0168] like Figure 8 As shown, the device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 to a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the device 800 can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0169] A number of components in the device 800 are connected to the I / O interface 805, including: an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0170] The computing unit 801 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 801 performs the various methods and processes described above, such as at least one of the resource recommendation method and the training method of the deep learning model. For example, in some embodiments, at least one of the resource recommendation method and the training method of the deep learning model may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of at least one of the resource recommendation method and the training method of the deep learning model described above may be performed. Alternatively, in other embodiments, the computing unit 801 may be configured to execute at least one of the resource recommendation method and the deep learning model training method in any other appropriate manner (for example, by means of firmware).

[0171] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0172] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0173] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0174] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0175] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0176] A computer system may include clients and servers. Clients and servers are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship to each other.

[0177] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0178] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A resource recommendation method, comprising: Determine the recent preference characteristics of the target user based on the target user's interactive behavior data within the historical time window of the live broadcast room; Determine a cross feature according to a matching relationship between the recent preference feature and respective resource features of the plurality of live broadcast rooms; Determine at least one target live broadcast room from the multiple live broadcast rooms according to the recent preference feature, the resource features of the multiple live broadcast rooms, and the cross feature; and Output the information of the target live broadcast room.

2. The method according to claim 1, wherein: Determining the recent preference characteristics of the target user according to the target user's interactive behavior data within the historical time window of the live broadcast room includes: Extracting an interactive behavior feature and at least one of a host identifier and an object category associated with the interactive behavior feature from the interactive behavior data; and The recent preference feature is determined according to the interactive behavior feature and at least one of the anchor identifier and the object category.

3. The method according to claim 2, wherein: Determining the cross-feature according to the matching relationship between the recent preference feature and the resource features of each of the plurality of live broadcast rooms includes: For each live broadcast room, the cross-features between the target user and the live broadcast room are determined based on the matching relationship between at least one of the anchor identification and the object category in the recent preference features and at least one of the anchor identification and the object category in the live broadcast room.

4. The method according to claim 3, wherein: The interactive behavior includes a conversion behavior; and determining at least one target live broadcast room from the multiple live broadcast rooms according to the recent preference feature, the resource features of the multiple live broadcast rooms, and the cross feature includes: Determine a target anchor identifier and a target object category associated with the conversion behavior; For each live broadcast room, in response to at least one of the target anchor identifier and the target object category in the recent preference feature being matched with at least one of the anchor identifier and the object category of the live broadcast room, the number of conversion behaviors is determined as a recommendation weight; and At least one target live broadcast room is determined from the multiple live broadcast rooms according to the recent preference characteristics, the resource characteristics of each of the multiple live broadcast rooms, the cross-features, and the recommendation weights.

5. According to the method of any one of claims 2 to 4, determining at least one target live broadcast room from the multiple live broadcast rooms according to the recent preference feature, the resource features of the multiple live broadcast rooms, and the cross feature comprises: Adding the recent preference feature to the historical user-side feature to obtain a new user-side feature, wherein the historical user-side feature is obtained based on the basic user-side feature of the target user and the recent preference feature of at least one historical time window; Determine at least one target live broadcast room from the multiple live broadcast rooms according to the new user-side feature, the new resource-side features of each of the multiple live broadcast rooms, and the cross feature; Among them, the new resource-side feature is obtained by adding at least one of the anchor identification and object category of the live broadcast room to the basic resource-side feature of the live broadcast room.

6. The method according to claim 5, wherein: The determining at least one target live broadcast room from the multiple live broadcast rooms according to the new user-side feature, the new resource-side features of the multiple live broadcast rooms, and the cross-feature comprises: For each live broadcast room, determining an evaluation value of the live broadcast room according to the new user-side features, the new resource-side features of the live broadcast room, and the cross-features between the target user and the live broadcast room; and At least one target live broadcast room is determined from the multiple live broadcast rooms according to the respective evaluation values ​​of the multiple live broadcast rooms.

7. The method according to claim 6, wherein: For each live broadcast room, determining the evaluation value of the live broadcast room according to the new user-side feature, the new resource-side feature of the live broadcast room, and the cross-feature between the target user and the live broadcast room includes: for each live broadcast room, Determine at least one of the click rate, conversion rate and total transaction amount of the target user for the live broadcast room according to the new user-side feature, the new resource-side feature of the live broadcast room, and the cross-feature between the target user and the live broadcast room; and The evaluation value of the live broadcast room is determined according to at least one of the click-through rate, conversion rate and total transaction amount.

8. A method for training a deep learning model, comprising: Determine the recent preference characteristics of the sample users based on the interactive behavior data of the sample users in the historical time window of the live broadcast room; Determine a cross feature according to a matching relationship between the recent preference feature and respective resource features of the plurality of live broadcast rooms; Determine the recommendation index of each of the multiple live broadcast rooms based on the recent preference feature, the resource feature of each of the multiple live broadcast rooms, and the cross feature using a deep learning model; Determining the loss of the deep learning model according to the recommended indicator; as well as According to the loss, the parameters of the deep learning model are adjusted to obtain a trained deep learning model.

9. The method according to claim 8, wherein: Determining the recent preference characteristics of the sample user according to the interactive behavior data of the sample user in the historical time window of the live broadcast room includes: Extracting an interactive behavior feature and at least one of a host identifier and an object category associated with the interactive behavior feature from the interactive behavior data; and The recent preference feature is determined according to the interactive behavior feature and at least one of the anchor identifier and the object category.

10. The method according to claim 9, wherein: The using of the deep learning model to determine the recommendation index of each of the multiple live broadcast rooms based on the recent preference feature, the resource features of each of the multiple live broadcast rooms, and the cross-feature includes: Adding the recent preference feature to the historical user-side feature to obtain a new user-side feature, wherein the historical user-side feature is obtained based on the basic user-side feature of the sample user and the recent preference feature of at least one historical time window; Determine a recommendation index for each of the multiple live broadcast rooms according to the new user-side feature, the new resource-side feature of each of the multiple live broadcast rooms, and the cross-feature; Among them, the new resource-side feature is obtained by adding at least one of the anchor identification and object category of the live broadcast room to the basic resource-side feature of the live broadcast room.

11. The method according to claim 10, further comprising: Determine the recommendation index output by the trained deep learning model for the verified user according to the interactive behavior data of the verified user in the historical time window of the live broadcast room; Determining the gain in accuracy of the recommendation indicator; and In response to the gain being negative, the interactive behavior feature in the recent preference feature and at least one of the anchor identifier and the object category are removed from the new user-side feature to obtain a new historical user-side feature.

12. A resource recommendation device, comprising: A first preference feature determination module is used to determine the recent preference features of the target user based on the interaction behavior data of the target user in the historical time window of the live broadcast room; A first cross-feature determination module is used to determine the cross-feature according to the matching relationship between the recent preference feature and the resource features of each of the plurality of live broadcast rooms; A target live broadcast room determination module, used to determine at least one target live broadcast room from the multiple live broadcast rooms according to the recent preference feature, the resource features of the multiple live broadcast rooms, and the cross feature; as well as The output module is used to output the information of the target live broadcast room.

13. The device according to claim 12, wherein: The first preference feature determination module includes: A first feature extraction unit is used to extract an interactive behavior feature and at least one of a host identifier and an object category associated with the interactive behavior feature from the interactive behavior data; and The first feature determination unit is used to determine the recent preference feature according to the interactive behavior feature and at least one of the anchor identifier and the object category.

14. The device according to claim 13, wherein: The first cross-feature determination module is used to determine, for each live broadcast room, the cross-features between the target user and the live broadcast room based on the matching relationship between at least one of the anchor identification and the object category in the recent preference features and at least one of the anchor identification and the object category in the live broadcast room.

15. The device according to claim 14, wherein: The interactive behavior includes a conversion behavior; the target live broadcast room determination module includes: A target determination unit, used to determine a target anchor identifier and a target object category associated with the conversion behavior; a recommendation weight determination unit, configured to determine, for each live broadcast room, the number of conversion behaviors as the recommendation weight in response to at least one of the target anchor identifier and the target object category in the recent preference feature being matched with at least one of the anchor identifier and the object category of the live broadcast room respectively; and The first target live broadcast room determination unit is used to determine at least one target live broadcast room from the multiple live broadcast rooms according to the recent preference characteristics, the resource characteristics of the multiple live broadcast rooms, the cross-features, and the recommendation weights.

16. The device according to any one of claims 13 to 15, wherein: The target live broadcast room determination module includes: A first user-side feature updating unit, configured to add the recent preference feature to the historical user-side feature to obtain a new user-side feature, wherein the historical user-side feature is obtained based on the basic user-side feature of the target user and the recent preference feature of at least one historical time window; A second target live broadcast room determination unit, configured to determine at least one target live broadcast room from the multiple live broadcast rooms according to the new user-side feature, the new resource-side feature of each of the multiple live broadcast rooms, and the cross feature; Among them, the new resource-side feature is obtained by adding at least one of the anchor identification and object category of the live broadcast room to the basic resource-side feature of the live broadcast room.

17. The device according to claim 16, wherein: The second target live broadcast room determination unit includes: An evaluation value determination subunit, configured to determine, for each live broadcast room, an evaluation value of the live broadcast room according to the new user-side features, the new resource-side features of the live broadcast room, and the cross-features between the target user and the live broadcast room; and The target live broadcast room determination subunit is used to determine at least one target live broadcast room from the multiple live broadcast rooms according to the respective evaluation values ​​of the multiple live broadcast rooms.

18. The device according to claim 17, wherein: The evaluation value determination subunit is used to determine at least one of the click-through rate, conversion rate and total transaction amount of the target user for the live broadcast room based on the new user-side characteristics, the new resource-side characteristics of the live broadcast room, and the cross-characteristics between the target user and the live broadcast room; and to determine the evaluation value of the live broadcast room based on at least one of the click-through rate, conversion rate and total transaction amount.

19. A training device for a deep learning model, comprising: A second preference feature determination module is used to determine the recent preference features of the sample user based on the interaction behavior data of the sample user in the historical time window of the live broadcast room; A second cross-feature determination module is used to determine the cross-feature according to the matching relationship between the recent preference feature and the resource features of each of the plurality of live broadcast rooms; A first recommendation index determination module, configured to use a deep learning model to determine a recommendation index for each of the plurality of live broadcast rooms based on the recent preference feature, the resource features of each of the plurality of live broadcast rooms, and the cross-feature; A loss determination module, used to determine the loss of the deep learning model according to the recommended indicator; as well as An adjustment module is used to adjust the parameters of the deep learning model according to the loss to obtain a trained deep learning model.

20. The device according to claim 19, wherein The second preference feature determination module includes: A second feature extraction unit is used to extract an interactive behavior feature and at least one of a host identifier and an object category associated with the interactive behavior feature from the interactive behavior data; and The second feature determination unit is used to determine the recent preference feature according to the interactive behavior feature and at least one of the anchor identifier and the object category.

21. The device according to claim 20, wherein: The first recommendation indicator determination module includes: A second user-side feature updating unit is configured to add the recent preference feature to the historical user-side feature to obtain a new user-side feature, wherein the historical user-side feature is obtained based on the basic user-side feature of the sample user and the recent preference feature of at least one historical time window; A recommendation index determination unit, configured to determine a recommendation index for each of the plurality of live broadcast rooms according to the new user-side feature, the new resource-side feature of each of the plurality of live broadcast rooms, and the cross-feature using a deep learning model; Among them, the new resource-side feature is obtained by adding at least one of the anchor identification and object category of the live broadcast room to the basic resource-side feature of the live broadcast room.

22. The apparatus according to claim 21, further comprising: A second recommendation index determination module is used to determine the recommendation index output by the trained deep learning model for the verified user according to the interactive behavior data of the verified user in the historical time window of the live broadcast room; A gain determination module, used to determine the gain of the accuracy of the recommendation indicator; as well as The elimination module is used for removing the interactive behavior feature in the recent preference feature and at least one of the anchor identification and the object category from the new user-side feature in response to the gain being negative, so as to obtain a new historical user-side feature.

23. An electronic device, comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 11.

24. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 11.

25. A computer program product, comprising a computer program, wherein the computer program is stored on at least one of a readable storage medium and an electronic device, and when the computer program is executed by a processor, implements the method according to any one of claims 1 to 11.

Images