WO2022007520A1 - 推荐项目确定方法、装置、设备及存储介质 - Google Patents

推荐项目确定方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2022007520A1
WO2022007520A1 PCT/CN2021/095567 CN2021095567W WO2022007520A1 WO 2022007520 A1 WO2022007520 A1 WO 2022007520A1 CN 2021095567 W CN2021095567 W CN 2021095567W WO 2022007520 A1 WO2022007520 A1 WO 2022007520A1
Authority
WO
WIPO (PCT)
Prior art keywords
item
sequence
items
recall
target
Prior art date
Application number
PCT/CN2021/095567
Other languages
English (en)
French (fr)
Inventor
胡忠平
Original Assignee
百果园技术(新加坡)有限公司
胡忠平
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from CN202010658808.4A external-priority patent/CN111859126B/zh
Application filed by 百果园技术(新加坡)有限公司, 胡忠平 filed Critical 百果园技术(新加坡)有限公司
Publication of WO2022007520A1 publication Critical patent/WO2022007520A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations

Definitions

  • the present application relates to the field of computer technology, for example, to a recommended item determination method, apparatus, device, and storage medium.
  • the present application provides a recommended item determination method, apparatus, device and storage medium, which can optimize the recommended item determination scheme.
  • a recommended item determination method includes:
  • the recall model corresponding to the current item set For each item set in the item library, according to the target user feature data corresponding to the target user and the recall model corresponding to the current item set, determine the recall item sequence corresponding to the current item set, wherein the item library contains at least three A collection of items corresponding to different heat levels;
  • a target recommended item sequence corresponding to the target user is determined according to the second candidate item sequence.
  • the recall item sequence determination module is set to determine the recall item sequence corresponding to the current item set according to the target user feature data corresponding to the target user and the recall model corresponding to the current item set for each item set in the item library, wherein all The above-mentioned project library contains at least three project sets corresponding to different heat levels;
  • a scoring and sorting module configured to perform scoring and sorting on the items in the recalled item sequence, and determine a first candidate item sequence
  • the score promotion module is configured to perform score promotion processing on the target items corresponding to the first preset heat level in the first candidate item sequence, to obtain a second candidate item sequence, wherein the first preset heat level includes the lowest heat level;
  • the target recommended item determination module is configured to determine a target recommended item sequence corresponding to the target user according to the second candidate item sequence.
  • a computer device including a memory, a processor, and a computer program stored on the memory and running on the processor, when the processor executes the computer program, the recommended item determination as provided by the embodiments of the present application is implemented. method.
  • a computer-readable storage medium which stores a computer program, and when the computer program is executed by a processor, implements the method for determining a recommended item provided by the embodiments of the present application.
  • FIG. 1 is a schematic flowchart of a method for determining a recommended item according to an embodiment of the present application
  • FIG. 2 is a schematic flowchart of another method for determining a recommended item provided by an embodiment of the present application
  • FIG. 3 is a schematic flowchart of a method for determining a recommended video provided by an embodiment of the present application
  • FIG. 4 is a schematic diagram of a recall model configuration provided by an embodiment of the present application.
  • FIG. 5 is a structural block diagram of an apparatus for determining a recommended item provided by an embodiment of the present application.
  • FIG. 6 is a structural block diagram of a computer device according to an embodiment of the present application.
  • FIG. 1 is a schematic flowchart of a method for determining a recommended item provided by an embodiment of the present application.
  • the method may be executed by an apparatus for determining a recommended item, where the apparatus may be implemented by software and/or hardware, and may generally be integrated in computer equipment such as a server.
  • the method includes the following steps.
  • Step 101 For each item set in the item library, according to the target user feature data corresponding to the target user and the recall model corresponding to the current item set, determine the recall item sequence corresponding to the current item set, wherein the item library contains At least three sets of items corresponding to different heat levels.
  • the items in the embodiments of the present application may include items published on the Internet (which may be embodied in various platforms) (such as commodities in online malls or items on second-hand trading websites, etc.) or resources (such as short videos). Short videos in the platform or news in the recommendation engine, etc.), the type of project is not limited.
  • the user can browse the items recommended by the platform through the client terminal corresponding to the platform, and the back-end server corresponding to the platform can make targeted item recommendation according to the characteristics of the user.
  • the recommended timing can be designed according to the actual situation of the platform, such as when the user opens the platform, when the user refreshes the page, or when the user switches pages, etc., or it can be recommended regularly (for example, once every 20 seconds).
  • a target user can be understood as a user who currently needs the recommended item.
  • the target user characteristic data can be understood as relevant data such as attributes or behaviors that can reflect the target user's own characteristics.
  • User characteristic data may include, for example, inherent attribute data such as user name, user gender, user location or user occupation, and other related data such as user level, user social relationship, and user historical operation behavior.
  • it may further include determining a target user, and acquiring target user characteristic data corresponding to the target user.
  • the project library can be understood as a library that can be used to form recommended projects in the platform, and the existence form or storage location of the project library is not limited.
  • the popularity level can be understood as a dynamic attribute corresponding to a project. With the passage of time, the popularity level corresponding to a project can change. For a newly released project, the popularity level is the lowest. After a period of time, it is used for evaluation. Some factors of the heat rating may change, and the corresponding heat rating for the item will change.
  • the evaluation criteria of the popularity level can be set according to actual needs. For example, the popularity level can be determined based on at least one of the project release time, the number and/or frequency of project presentations, and the amount of user behavior data corresponding to the project.
  • the popularity level can be divided into a rising stage, a diffusion stage, a succession stage and a popular stage in order from low to high, wherein the succession stage and the popular stage can also be combined into one stage.
  • the items in the project library can belong to different project sets. For a project at the current moment, the project set to which it belongs is certain, but for a subsequent moment, the project may be due to the corresponding
  • the popularity level of the item changes and it is transferred to another item collection.
  • the concept of collection is used to describe the items. However, in the actual storage space, multiple items are not necessarily stored in a collection. Generally, they are input to different items due to different current corresponding heat levels.
  • the recall model and then have the opportunity to enter the corresponding recall item sequence.
  • recommender systems generally only use one recall model, which results in a relatively single data granularity and algorithm, and cannot provide more accurate and personalized recommendation services. Submerged in a large number of projects, cannot be reasonably diffused.
  • multiple heat levels are set, and corresponding recall models are set for different heat levels, that is, for items with different heat levels, the recall models used may be different. Therefore, It can design the recall model more targetedly, improve the pertinence and accuracy of item recall, and make the items entering the recall item sequence more reasonable.
  • Recall models corresponding to multiple item sets respectively are not limited in the embodiment of the present application, and may be set according to actual needs.
  • the target user feature data and the item information corresponding to the current item set may be input into the corresponding recall model, and the corresponding recall item sequence may be determined according to the output result of the recall model.
  • the item information may include related data such as attributes or characteristics corresponding to multiple items respectively.
  • the number of items in the recall item sequence can be set according to actual needs, and the number of items corresponding to different recall item sequences can be the same or different.
  • the number of items included in each recall item sequence is a corresponding set quantity, and there are at least two unequal set quantities, that is, there may be at least two recall item sequences with unequal numbers of items, Achieve more targeted recalls.
  • the number of items included in the recalled item sequence is the same, it can also be intercepted separately according to a preset configuration ratio in subsequent operations, and the execution position of the interception step is not limited.
  • Step 102 Score and sort the items in the recalled item sequence, and determine a first candidate item sequence.
  • all items of different popularity levels in the multiple recall item sequences can be scored uniformly, and sorted according to the scoring results from high to low, and the interception and ranking are in the top according to a certain number of requirements (such as topA, where, top refers to the top ranking, A refers to the ranking sequence number, and topA refers to the items with the ranking sequence numbers 1 to A), and the first candidate item sequence is obtained.
  • topA a certain number of requirements
  • Step 103 Perform score promotion processing on the target item corresponding to the first preset popularity level in the first candidate item sequence, to obtain a second candidate item sequence, wherein the first preset popularity level includes the lowest popularity level.
  • the score will generally be lower, and the ranking position will be relatively low.
  • the scores corresponding to the items of the first preset popularity level including the popularity level are improved, which can improve the probability or frequency of showing the items of the first preset popularity level including the lowest popularity level to users, so that new items can be effectively diffusion.
  • the second candidate item sequence can be obtained by re-sorting according to the score.
  • Step 104 Determine a target recommended item sequence corresponding to the target user according to the second candidate item sequence.
  • an item sequence with a higher score (such as topB, where B is less than A) can be cut out from the second candidate item sequence according to the number of items corresponding to the target recommended item sequence, and the target user corresponding to the item sequence can be obtained.
  • Target recommendation item sequence can be cut out from the second candidate item sequence according to the number of items corresponding to the target recommended item sequence, and the target user corresponding to the item sequence can be obtained.
  • the method for determining recommended items determines, for each item set in the item database, the corresponding recall item sequence according to the target user feature data corresponding to the target user and the recall model corresponding to the current item set, wherein,
  • the item library contains at least three item sets corresponding to different heat levels, the items in the recall item sequence are scored and sorted, and the first candidate item sequence is determined, and the first candidate item sequence corresponding to the first preset heat level is sorted.
  • the target item is subjected to a score promotion process to obtain a second candidate item sequence, wherein the first preset popularity level includes the lowest popularity level, and a target recommended item sequence corresponding to the target user is determined according to the second candidate item sequence.
  • the recall item sequence is determined according to the corresponding recall model for at least three item sets corresponding to different heat levels.
  • the recall model in the heat stage is used to determine the recall item sequence, which makes the recall item sequence more targeted and accurate. After the items in the recall item sequence are scored and sorted, the target items that contain the lowest heat level items are scored. It can effectively increase the display frequency of low-fat projects, and even if cold projects are distributed more reasonably.
  • there is at least one item set corresponding to at least two recall models that is, there is at least one target item set, and the number of recall models corresponding to the target item set is at least two;
  • Each recall model corresponds to a sequence of recall items.
  • the item set corresponding to at least two recall models may include the item set corresponding to the lowest popularity level, and the item with the lowest popularity level can be understood as being in the initial stage of the recommendation system, and generally there is less user interaction data. If a single conventional recall model is used, it generally cannot spread well in the crowd. Using multiple channels can better display cold start items.
  • the recall model corresponding to the item set corresponding to the lowest popularity level includes at least one of the following: a cold-start recall model based on social relations, a cold-start recall model based on geographic location, a cold-start recall model based on user interests, and a cold-start recall model based on user interests.
  • a cold-start recall model based on social relations includes at least one of the following: a cold-start recall model based on social relations, a cold-start recall model based on geographic location, a cold-start recall model based on user interests, and a cold-start recall model based on user interests.
  • Personalized cold start recall model Personalized cold start recall model.
  • the structures and related parameters of multiple recall models are not limited, and can be set according to actual needs. A schematic illustration of the above-mentioned recall models is given below.
  • the basic strategy can be to publish new projects by project publishers who have preset social relations with target users (such as friend relations, follow or follow relations, or are in the same communication group, etc.). (corresponding to the lowest popularity level) is recommended to target users.
  • a first preset strategy can be used for sorting.
  • the first preset strategy can be based on the intimacy with the target user and whether the target user is We are paying attention to the corresponding booth and other factors to set. Exemplarily, taking a short video as an example, it can be understood that the amount is started through social relations or paying attention to booths.
  • the platform can select the target user according to the social relations of the target user according to a certain ratio. New videos published by the user's friends are recommended to the user.
  • the platform can also recommend items to the target user according to the intimacy between the publisher and the target user corresponding to the video of the following booth.
  • the basic strategy can be to recommend new projects (corresponding to the lowest popularity level) published by project publishers whose geographic location is within a preset distance from the target user's current geographic location to the target user .
  • new projects corresponding to the lowest popularity level
  • the target user whose geographic location is within a preset distance from the target user's current geographic location to the target user .
  • a short video it can be understood that by obtaining the geographic location of the target user, new videos published by users near the target user are recommended to the target user.
  • the basic strategy can be to determine the interest degree of the target user in different project publishers based on the historical behavior of the target user, and determine the project publisher whose interest degree is higher than the preset interest threshold as the interested project release , recommends new items (corresponding to the lowest popularity level) published by the item publisher whose similarity with the interested item publisher meets the preset similarity requirement to the target user.
  • the target user comes to the platform to browse again, according to the historical preference of the target user Paike, search the nearest unknown Paike, and recommend the new video released by the Paike to target users, so as to achieve the purpose of video cold start.
  • the basic strategy can be to use a preset model to calculate the feature vector of the new item based on the attributes of the new item (corresponding to the lowest popularity level) and the behavior data of other users.
  • Set the model to calculate the feature vector of the target user, and then calculate the interest score based on the feature vector of the target user and the feature vector of the new item, and recommend new items to the target user according to the score.
  • the basic attributes of the new video and a small amount of behavior are used as feature input, and the trained Deep Neural Networks (DNN) model calculates the feature vector of the new video.
  • DNN Deep Neural Networks
  • the target user accesses the platform , using the same DNN model to calculate the target user feature vector, then calculate the distance between the target user feature vector and the new video feature vector, score interest, and recommend topC new videos to the target user for cold start according to the interest score.
  • the recall model corresponding to the item set corresponding to the next lowest popularity level includes at least one of the following: a recall model based on a user's collaborative filtering algorithm and a recall model based on a follow-up action.
  • a recall model based on a user's collaborative filtering algorithm For the items with the second lowest popularity level, although the accumulated user behaviors are not many, compared with the items with the lowest popularity level, the user behaviors are relatively more, and other recall models that refer to user behaviors can be used for recall.
  • the structures and related parameters of multiple recall models are not limited, and can be set according to actual needs. The following two kinds of recall models are schematically explained.
  • the user behavior corresponding to the item with the next lowest popularity level is more than the user behavior corresponding to the item with the lowest popularity level.
  • the basic idea of the recall model (User Collaboration Filter, UserCF) based on the user's collaborative filtering algorithm is that many users can work together to continuously interact with the platform, so that their recommendation list can continuously filter out the items that they are not interested in. So as to meet their needs more and more.
  • User-based may refer to calculating which users have similar interests by analyzing user behaviors on items (such as browsing, favorites, likes, and forwarding, etc.), and then recommending items concerned by users with similar interests to each other.
  • the basic strategy can be to obtain the feature vector of the target user, calculate a plurality of seed users whose similarity with the target user meets the preset similarity requirement according to the feature vector, use the multiple seed users as seeds, and obtain the seed users generated for the item set.
  • Behavior data score and sort multiple items in the item set according to the behavior data, and obtain the recall item sequence.
  • the feature vector of the target user is obtained through the DNN model, the topD users most similar to the target user are calculated according to the feature vector, and these similar users are used as seeds to search for the likes, concerns, and shares of the seed users.
  • Unpopular videos (items with the next lowest popularity level) for behaviors such as , download, and completion are weighted and scored according to different behaviors, and finally returned to the topE video to recommend to users.
  • the basic strategy can be to obtain the target user's watch list, and recommend the items in the watch list that users have positive behaviors to the target user.
  • the user's watch list is obtained, and the videos in the watch list where the user has positive behaviors are weighted and scored, and then recommended to the user, wherein the positive behaviors may include, for example, like, follow, share, download and the completion of the broadcast and other behaviors.
  • mainstream recall models may be employed.
  • youtube DNN models and recall models for item-based collaborative filtering algorithms ItemCF can be included, among others.
  • scoring and sorting the items in the recall item sequence and determining the first candidate item sequence includes: using a preset sorting and scoring model to score the items in the recall item sequence, and obtaining Sorting points corresponding to items in the recall item sequence; sorting items in the recall item sequence from high to low according to the sorting points to obtain a first sequence; A number of items are set, and a first candidate item sequence is determined according to the acquired items.
  • the number and type of the preset sorting and scoring models are not limited. For example, they may include a first sorting and scoring model for rough sorting and a second sorting and scoring model for fine sorting.
  • the sorting and scoring model may be a deep interest model ( Deep Interest Network, DIN).
  • the acquiring a preset number of items in the first order, and determining the first candidate item sequence according to the acquired items includes: acquiring the preset number of items in the top order, and according to the acquired items Items determine a second sequence; after adding items in the first sequence and not in the second preset heat level to the second sequence, a first candidate item sequence is obtained, wherein , the second preset heat level includes the lowest heat level.
  • the order of adding items to be added can be determined with reference to the order in the first sequence, that is, for the first item and the second item added to the second sequence, if the first item in the first sequence is relatively If the second item is higher, then in the first candidate item sequence, the first item is still higher than the second item.
  • performing a score boosting process on a target item corresponding to a first preset popularity level in the first candidate item sequence to obtain a second candidate item sequence includes: targeting the first candidate item sequence For each target item corresponding to the first preset popularity level in the , the ranking score corresponding to the current target item is improved based on the number of presentations of the current target item, and a new ranking score corresponding to the current target item is obtained; Sort the items in the first candidate item sequence to obtain a second candidate item sequence.
  • the ranking score corresponding to the current target item is improved based on the number of presentations of the current target item to obtain a new ranking score corresponding to the current target item, including: using an upper confidence bound (Upper Confidence Bound).
  • UCB Upper Confidence Bound
  • UCB UCB algorithm performs promotion processing on the ranking score corresponding to the current target item based on the number of presentations of the current target item, and obtains a new ranking score corresponding to the current target item, wherein the first item corresponding to the UCB algorithm includes the According to the sorting score, the second item corresponding to the UCB algorithm includes the reciprocal of the number of presentations.
  • the advantage of this setting is that the UCB algorithm includes a sorting part and a weight escalation part, but the sorting part algorithm is relatively simple and has insufficient precision.
  • the weight promotion part improves the ranking score of the target item, realizes the improvement of the UCB algorithm, and can obtain a more ideal final ranking result.
  • the re-sorting the items in the first candidate item sequence according to the new sorting score to obtain the second candidate item sequence includes: re-sorting the first candidate items according to the new sorting score
  • the items in the sequence are sorted to obtain a third sequence; the first display position adjustment and/or the interval adjustment of adjacent display positions are performed on the target items of the third preset popularity level included in the third sequence to obtain a second candidate item sequence, wherein the third preset heat level includes the lowest heat level.
  • the placement position can also be understood as the ranking position in the second candidate item sequence.
  • the advantage of this setting is that, after obtaining the third sequence reordered according to the new sorting score processed by the score promotion, the placement control can be performed for items with a low popularity level to prevent cold items from being placed too early or cold for the first time. Item placement is too dense, causing user discomfort and reducing negative impact on users.
  • the scoring and sorting the items in the recalled item sequence and determining the first candidate item sequence includes: performing the same item fusion processing and/or the violation item on the items in the recalled item sequence
  • the filtering process is performed to obtain a sequence of items to be sorted; the items in the sequence of items to be sorted are scored and sorted, and a first candidate item sequence is determined.
  • the same item fusion processing can be understood as deduplication processing.
  • the same item may exist in more than two recall item sequences, and deduplication processing can be performed.
  • projects can be reviewed and processed before sorting, and non-compliant projects can be filtered out to ensure a healthy environment for the platform.
  • it can also be marked, and when calculating the sorting score, it can be considered to appropriately increase the sorting score for the items with the mark.
  • the method further includes: recommending items in the target recommended item sequence to the target user .
  • the recommendation may be performed according to a preset recommendation strategy, and the preset recommendation strategy may be set according to actual needs, for example, it may include a recommendation time point and a recommended position in the interface, and the like.
  • FIG. 2 is a schematic flowchart of another recommended item determination method provided by an embodiment of the present application. As shown in FIG. 2 , the method may include the following steps.
  • Step 201 Obtain target user feature data corresponding to the target user.
  • Step 202 For each item set in the item library, determine a recall item sequence corresponding to the current item set according to the target user feature data and the recall model corresponding to the current item set.
  • the item library contains at least three item sets corresponding to different heat levels.
  • the popularity level is determined based on the item release time, the number and/or frequency of item presentations, and the amount of user behavior data corresponding to the item.
  • Step 203 Perform the same item fusion processing and/or the illegal item filtering processing on the items in the recalled item sequence to obtain the item sequence to be sorted.
  • Step 204 Use a preset sorting and scoring model to score the items in the item sequence to be sorted, and obtain the sorting scores corresponding to the items in the item sequence to be sorted.
  • the DIN model is used to calculate the ranking score corresponding to each item in the sequence of items to be sorted.
  • Step 205 Sort items in the item sequence to be sorted from high to low according to the sorting score to obtain a first sequence.
  • Step 206 Acquire the first preset number of items in the first sequence, and determine the second sequence according to the acquired items.
  • Step 207 After adding the items of the second preset popularity level that are in the first sequence and not in the second sequence to the second sequence, a first candidate item sequence is obtained.
  • the second preset heat level includes the lowest heat level.
  • Step 208 For each target item corresponding to the first preset popularity level in the first candidate item sequence, use the UCB algorithm to improve the ranking score corresponding to the current target item based on the number of presentations of the current target item to obtain the current target item. The corresponding new sorting score.
  • the first preset heat level includes the lowest heat level.
  • the new ranking is divided into score
  • the original ranking is divided into rank_score (that is, the ranking score in step 204)
  • the score can be calculated based on the following deformed UCB algorithm formula:
  • T represents the number of presentations of the item (that is, the current presentation number of the current target item)
  • boost_score is the first preset value (it can be an empirical value or a fixed value called through experiments, and the general value range is 1 to 10)
  • beta is the second preset value (it can be an empirical value or a fixed value called through experiments, and the general value range is 0 to 1).
  • Step 209 Re-sort the items in the first candidate item sequence according to the sorting score to obtain a third sequence.
  • Step 210 Adjust the first display position and/or the interval between adjacent display positions for the target item of the third preset popularity level included in the third sequence to obtain a second candidate item sequence.
  • the third preset heat level includes the lowest heat level.
  • the first preset heat level, the second preset heat level, and the third preset heat level may be the same or different.
  • Step 211 Acquire the second preset number of items ranked first in the second candidate item sequence, and obtain a target recommended item sequence corresponding to the target user.
  • Step 212 Recommend the items in the target recommended item sequence to the target user according to the preset recommendation strategy.
  • the recall item sequence is determined according to the corresponding recall model for at least three item sets corresponding to different heat levels, wherein one or some item sets may correspond to different recall models, which can be used It is more in line with the diversified recall models of different heat stages to determine multiple recall item sequences, which makes the recall item sequence more targeted, accurate and richer.
  • the UCB algorithm is improved, and the preset sorting and scoring model is used to calculate the sequence of items. Sort points, and based on the reciprocal of the number of impressions, the items with lower popularity levels will be improved, and the placement of low-level items will be effectively controlled to effectively improve the display frequency of low-level items. Therefore, it can more reasonably solve the E&E problem in the recommender system.
  • FIG. 3 is a schematic flowchart of a method for determining a recommended video provided by an embodiment of the present application, taking a short video recommendation scenario as an example for description, which may be based on video release time, video presentation times and/or frequency, and user behavior data corresponding to the video
  • the popularity level corresponding to the short video is set as the starting stage, the diffusion stage, the succession stage and the popular stage.
  • the method may include the following steps.
  • Step 301 Obtain target user feature data corresponding to the target user.
  • Step 302 For each video set in the video library, determine a recall video sequence corresponding to the current video set according to the target user feature data and the recall model corresponding to the current video set.
  • the video library contains four sets of videos corresponding to different popularity levels.
  • Each recall model can be regarded as a recall channel.
  • the corresponding recall models include a social relationship-based cold-start recall model, a geographic-location-based cold-start recall model, a user-interest-based cold-start recall model, and a personalization-based cold-start recall model.
  • These recall channels in the start-up stage each have their own unique advantages. They complement each other and work together when the resources are cold-started.
  • video resources generally cannot pass through a single, Conventional channels spread well among the crowd. Displaying cold start resources to users through multiple channels and forms is conducive to mining more high-quality content and promoting the sound development of the system ecology.
  • the multi-channel cold start method improves the diversity of resources.
  • the corresponding recall model includes a user-based collaborative filtering algorithm recall model and a follow-up action-based recall model.
  • the new video has a certain starting stage and has some behaviors, it can be diffused through the above channel.
  • the mainstream recall channel can be used for database recall.
  • FIG. 4 is a schematic diagram of a recall model configuration provided by an embodiment of the present application.
  • the recall models that can be used in the starting stage include the cold-start recall model based on social relationship and follow (follow) booth, the cold-start recall model based on geographic location (nearby), and the cold-start recall model based on user interests.
  • recall models that can be used in the diffusion stage include User-Based Collaborative Filtering (UserCF) and follow-action-based recall models Model (follow action); recall models that can be used in the succession stage include DNN recall model and DNN-based cold start system recommendation recall model (DNN EE); recall models that can be used in the hot stage include recall models based on popular videos (hot, Such as the common popular leaderboard), the recall model of item-based collaborative filtering algorithm (ItemCF) and the youtube DNN model.
  • the above is just an example. In practical application, other recall models can be used to replace any one of the recall models. Compared with fewer or more recall models in Figure 4, the popularity of multiple recall models can also be used.
  • the level can also be set freely, which is not limited in the embodiments of the present application.
  • each recall model that is, each recall channel, respectively outputs a recall video sequence
  • the size of each recall video sequence (that is, the number of videos contained) may be equal.
  • truncation processing can be performed according to the proportion of each channel configuration to obtain a processed recall video sequence.
  • the advantage of this setting is that when the size of the recall video sequence corresponding to a recall model needs to be changed, the recall model does not need to be changed, and the corresponding ratio can be directly adjusted, so that the sequence size can be adjusted quickly and accurately to avoid errors. The same goes for projects other than video.
  • Step 303 Perform the same video fusion processing and the illegal video filtering processing on the videos in the recalled video sequence to obtain the to-be-sorted video sequence.
  • Step 304 Use a preset sorting and scoring model to score the videos in the video sequence to be sorted, and obtain a sorting score corresponding to the videos in the video sequence to be sorted.
  • Step 305 Sort the videos in the video sequence to be sorted from high to low according to the sorting score to obtain a first sequence.
  • Step 306 Acquire the first preset number of videos in the first sequence, and determine the second sequence according to the acquired videos.
  • the video of topN may be obtained to obtain the second sequence, and N may be 1000, for example.
  • Step 307 After adding the videos in the starting stage and the diffusion stage in the first sequence but not in the second sequence to the second sequence, a first candidate video sequence is obtained.
  • Step 308 For each target video corresponding to the starting stage and the diffusion stage in the first candidate video sequence, use the UCB algorithm to improve the ranking score corresponding to the current target video based on the number of presentations of the current target video to obtain the current target video. The corresponding new sorting score.
  • the new ranking is divided into score, and the ranking is divided into rank_score, and the score can be calculated based on the following deformed UCB algorithm formula:
  • VV represents the number of presentations of the video (that is, the current presentation number of the current target video)
  • boost_score is the first preset value (it can be an empirical value or a fixed value called through experiments, and the general value range is 1 to 10)
  • beta is the second preset value (it can be an empirical value or a fixed value called through experiments, and the general value range is 0 to 1).
  • Step 309 Re-sort the videos in the first candidate video sequence according to the ranking score to obtain a third sequence.
  • Step 310 Adjust the first display position and the interval between adjacent display positions for the target videos in the starting stage and the diffusion stage included in the third sequence to obtain a second candidate video sequence.
  • the adjustment methods for the target videos in the starting stage and the spreading stage may be different.
  • the target video in the starting stage may be displayed at a later position than the target video in the spreading stage. can be larger than the target video in the Diffusion phase, etc.
  • Step 311 Acquire the second preset number of items ranked first in the second candidate item sequence, and obtain a target recommended video sequence corresponding to the target user.
  • the video of topM can be obtained to obtain the target recommended video sequence, and M can be 20, for example.
  • Step 312 Recommend the video in the target recommended video sequence to the target user according to the preset recommendation strategy.
  • the recall video sequences are determined according to the corresponding recall models for video sets of different popularity levels, and multiple recall models are set for each popularity level, that is, multiple recall channels are set , these channels complement each other and are in charge of different functions.
  • the multi-stage and multi-channel architecture can meet the needs of different degrees of cold start of video, and jointly complete the distribution of cold video, which can quickly sense user feedback, filter out high-quality videos, and improve the efficiency of cold start.
  • the method of rights escalation based on the reciprocal of the number of presentations, improves the scores of videos with low popularity levels, and controls the distribution density of cold videos to reduce the negative impact on users. On the premise of meeting certain personalized needs, it can effectively improve the low level.
  • the display frequency of hot videos is controlled while the negative impact of cold videos on users is controlled, so as to more reasonably solve the E&E problem in the video recommendation system.
  • FIG. 5 is a structural block diagram of an apparatus for determining a recommended item provided by an embodiment of the present application.
  • the apparatus may be implemented by software and/or hardware, and may generally be integrated in a computer device, and the item to be recommended may be determined by executing a method for determining a recommended item .
  • the device includes the following modules.
  • the recall item sequence determination module 501 is configured to determine the recall item sequence corresponding to the current item collection according to the target user characteristic data corresponding to the target user and the recall model corresponding to the current item collection for each item collection in the item library, wherein,
  • the item library contains at least three item sets corresponding to different heat levels;
  • the scoring and sorting module 502 is configured to perform scoring and sorting on the items in the recalled item sequence, and determine the first candidate item sequence;
  • the score boosting module 503 is configured to perform score boosting processing on the target items corresponding to the first preset popularity level in the first candidate item sequence, to obtain a second candidate item sequence, wherein the first preset popularity level includes Minimum heat level;
  • the target recommended item determination module 504 is configured to determine a target recommended item sequence corresponding to the target user according to the second candidate item sequence.
  • the device for determining a recommended item determines, for each item set in the item library, a corresponding recall item sequence according to the target user feature data corresponding to the target user and the recall model corresponding to the current item set, wherein the item
  • the library contains at least three item sets corresponding to different heat levels, the items in the recall item sequence are scored and sorted, and the first candidate item sequence is determined, and the target corresponding to the first preset heat level in the first candidate item sequence is determined.
  • the items are subjected to score promotion processing to obtain a second candidate item sequence, wherein the first preset popularity level includes the lowest popularity level, and a target recommended item sequence corresponding to the target user is determined according to the second candidate item sequence.
  • the recall item sequence is determined according to the corresponding recall model for at least three item sets corresponding to different heat levels.
  • the recall model in the heat stage is used to determine the recall item sequence, which makes the recall item sequence more targeted and accurate. After the items in the recall item sequence are scored and sorted, the target items that contain the lowest heat level items are scored. It can effectively increase the display frequency of low-fat projects, and even if cold projects are distributed more reasonably.
  • FIG. 6 is a structural block diagram of a computer device according to an embodiment of the present application.
  • the computer device 600 includes a memory 601, a processor 602, and a computer program that is stored in the memory 601 and can run on the processor 602. When the processor 602 executes the computer program, the method for determining a recommended item provided by the embodiment of the present application is implemented. .
  • the embodiments of the present application further provide a storage medium containing computer-executable instructions, where the computer-executable instructions are used to execute the recommended item determination method provided by the embodiments of the present application when executed by a computer processor.
  • the recommended item determination device, device, and storage medium provided in the above embodiments can execute the recommended item determination method provided by any embodiment of the present application, and have corresponding functional modules and effects for executing the method.
  • the method for determining a recommended item provided by any embodiment of the present application.

Abstract

一种推荐项目确定方法、装置、设备及存储介质。其中，该推荐项目确定方法包括：针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定当前项目集合对应的召回项目序列，其中，项目库中包含至少三个对应不同热度等级的项目集合（101）；对召回项目序列中的项目进行评分排序，并确定第一候选项目序列（102）；对第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，第一预设热度等级包括最低热度等级（103）；根据第二候选项目序列确定与目标用户对应的目标推荐项目序列（104）。

Description

推荐项目确定方法、装置、设备及存储介质
本申请要求在2020年07月09日提交中国专利局、申请号为202010658808.4的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，例如涉及推荐项目确定方法、装置、设备及存储介质。
背景技术
随着信息技术的快速发展，移动互联网的普及，信息过载已越来越成为人们生活中的挑战。推荐系统的出现，通过学习用户的历史行为，预测用户对其他内容的喜好程度，一方面方便了用户快速获取自己喜好的内容，另一方面满足了平台选择合适的用户展现自己的物品或资源等(可统称为项目)的需求，极大地缓解了用户与平台中间沟通的难题，促进了平台与用户的交互性。
长期来看，良好的推荐系统需要不断有新项目进来进行补充，从而淘汰掉历史的尾部项目，形成推荐系统的一个良性循环。但对于新项目来说，往往还没有用户行为或用户行为较少，推荐系统很难判断用户的偏好度，无法进行精准推荐，而盲目分发新项目，很可能伤害到用户，甚至造成用户流失。上述问题可概括为推荐系统的冷启动问题或推荐系统中的探索与利用(Exploration and Exploitation，E&E或EE)问题，而为解决冷启动问题所形成的推荐项目确定方案仍不够完善，需要改进。
发明内容
本申请提供了推荐项目确定方法、装置、设备及存储介质，可以优化推荐项目确定方案。
提供了一种推荐项目确定方法，该方法包括：
针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定当前项目集合对应的召回项目序列，其中，所述项目库中包含至少三个对应不同热度等级的项目集合；
对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列；
对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，所述第一预设热度等级包括最低热 度等级；
根据所述第二候选项目序列确定与所述目标用户对应的目标推荐项目序列。
还提供了一种推荐项目确定装置，该装置包括：
召回项目序列确定模块，设置为针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定当前项目集合对应的召回项目序列，其中，所述项目库中包含至少三个对应不同热度等级的项目集合；
评分排序模块，设置为对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列；
分数提升模块，设置为对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，所述第一预设热度等级包括最低热度等级；
目标推荐项目确定模块，设置为根据所述第二候选项目序列确定与所述目标用户对应的目标推荐项目序列。
还提供了一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如本申请实施例提供的推荐项目确定方法。
还提供了一种计算机可读存储介质，存储有计算机程序，该计算机程序被处理器执行时实现如本申请实施例提供的推荐项目确定方法。
附图说明
图1为本申请实施例提供的一种推荐项目确定方法的流程示意图；
图2为本申请实施例提供的又一种推荐项目确定方法的流程示意图；
图3为本申请实施例提供的一种推荐视频确定方法的流程示意图；
图4为本申请实施例提供的一种召回模型配置的示意图；
图5为本申请实施例提供的一种推荐项目确定装置的结构框图；
图6为本申请实施例提供的一种计算机设备的结构框图。
具体实施方式
下面结合附图和实施例对本申请进行说明。此处所描述的实施例仅仅用于解释本申请，而非对本申请的限定。为了便于描述，附图中仅示出了与本申请相关的部分而非全部结构。此外，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。
图1为本申请实施例提供的一种推荐项目确定方法的流程示意图，该方法可以由推荐项目确定装置执行，其中该装置可由软件和/或硬件实现，一般可集成在服务器等计算机设备中。如图1所示，该方法包括以下步骤。
步骤101、针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定当前项目集合对应的召回项目序列，其中，所述项目库中包含至少三个对应不同热度等级的项目集合。
示例性的，本申请实施例中的项目可以包括在互联网(可体现为多种平台)上发布的物品(如在线商城中的商品或二手交易网站上的物品等等)或资源(如短视频平台中的短视频或推荐引擎中的新闻等等)，项目的类型不做限定。一般的，用户可通过平台对应的客户端浏览平台推荐的项目，平台对应的后端服务器等可以根据用户的特点有针对性的进行项目推荐。推荐的时机可根据平台的实际情况进行设计，如用户打开平台时、用户刷新页面时、或用户切换页面时等等，又如还可以是定时推荐(如20秒推荐一次)等。
示例性的，目标用户可以理解为当前需要被推荐项目的用户。目标用户特征数据可以理解为能够体现目标用户自身特点的属性或行为等相关数据。用户特征数据例如可包括用户名称、用户性别、用户所在区域或用户职业等固有属性数据，还可包括用户等级、用户社交关系以及用户历史操作行为等其他相关数据。可选的，在本步骤之前，还可包括确定目标用户，并获取目标用户对应的目标用户特征数据。
示例性的，项目库可理解为平台中可用于推荐的项目构成的库，项目库的存在形式或存储位置等不做限定。热度等级可理解为项目对应的一种动态属性，随着时间的推移，一个项目对应的热度等级可以发生变化，对于一个新发布的项目来说，热度等级最低，经过一段时间后，用于评价热度等级的一些因素可能发生变化，该项目对应的热度等级就会发生变化。热度等级的评价标准可根据实际需求进行设定，例如，热度等级可基于项目发布时间、项目展现次数和/或频率、以及项目对应的用户行为数据量中的至少一个确定。举例而言，热度等级可从低至高依次分为起量阶段、扩散阶段、承接阶段和热门阶段，其中，承接阶段和热门阶段也可合并为一个阶段。根据热度等级的不同，项目库中的项目可归属于不同的项目集合，对于当前时刻的一个项目来说，所属的项目集 合是一定的，但对于后续的一个时刻来说，该项目可能因对应的热度等级发生变化而转入其他项目集合。这里为了便于说明，采用了集合的概念对项目进行描述，但在实际的存储空间内多个项目并不一定以集合的方式进行存储，一般会因当前所对应的热度等级不同而被输入至不同的召回模型，进而有机会进入相应的召回项目序列。
相关技术中，推荐系统一般只使用一种召回模型，导致在数据粒度和算法上都比较单一，无法提供更加精准以及个性化的推荐服务，例如对于热度较低的冷项目来说，很容易被淹没在大量项目内，无法得到合理的扩散。本申请实施例中，设置了多个热度等级，对于不同的热度等级，分别设置了对应的召回模型，也即对于不同热度等级的项目来说，所采用的召回模型可以是不同的，因此，能够更加有针对性的设计召回模型，提升项目召回的针对性和准确度，也使得进入召回项目序列的项目更加合理。多个项目集合分别对应的召回模型本申请实施例不做限定，可根据实际需求进行设置。
示例性的，可以将目标用户特征数据和当前项目集合对应的项目信息输入到对应的召回模型中，根据召回模型的输出结果确定对应的召回项目序列。其中，项目信息可包括多个项目分别对应的属性或特征等相关数据。召回项目序列中的项目的数量可根据实际需求进行设置，不同召回项目序列对应的项目数量可以相同也可以不同。可选的，每个召回项目序列中包含的项目数量为对应的设定数量，存在至少两个不相等的设定数量，也即，可存在至少两个包含项目数量不等的召回项目序列，实现更有针对性地召回。可选的，若召回项目序列中包含的项目数量相同，也可在后续操作中按照预设的配置比例等分别进行截取，截取步骤的执行位置不做限定。
步骤102、对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列。
示例性的，可对多个召回项目序列中的所有不同热度等级的项目进行统一的评分，并按照评分结果由高至低进行排序，按照一定的数量要求截取排名在前(如topA，其中，top指排序在前的，A指排名序号，topA指排名序号1至A)的项目，得到第一候选项目序列。
步骤103、对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，所述第一预设热度等级包括最低热度等级。
示例性的，由于前述步骤中不同热度等级的项目混合在一起参与统一的评分排序，因此，对于热度等级较低的项目来说，一般得分会较低，排序位置也相对靠后，对包含最低热度等级在内的第一预设热度等级的项目对应的得分进 行提升处理，可以提升包含最低热度等级在内的第一预设热度等级的项目对用户展现概率或频率，使得新项目能够得到有效的扩散。在经过分数提升处理后，可重新按照分数进行排序，得到第二候选项目序列。
步骤104、根据所述第二候选项目序列确定与所述目标用户对应的目标推荐项目序列。
示例性的，可按照目标推荐项目序列对应的项目数量要求，从第二候选项目序列中截取出分数较高(如topB，其中，B小于A)的项目序列，得到与所述目标用户对应的目标推荐项目序列。
本申请实施例中提供的推荐项目确定方法，针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定对应的召回项目序列，其中，项目库中包含至少三个对应不同热度等级的项目集合，对召回项目序列中的项目进行评分排序，并确定第一候选项目序列，对第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，第一预设热度等级包括最低热度等级，根据第二候选项目序列确定与目标用户对应的目标推荐项目序列。通过采用上述技术方案，针对至少三个对应不同热度等级的项目集合分别根据对应的召回模型确定召回项目序列，可以对项目所处的热度阶段进行更细致的划分，从而利用更加符合项目所对应的热度阶段的召回模型来确定召回项目序列，使得召回项目序列更加有针对性以及更加准确，在对召回项目序列中的项目进行评分排序后，针对包含最低热度等级项目的目标项目进行分数提升处理，可有效提升低热度项目的展现频率，也即使冷项目得到更加合理的分发。
在一些实施例中，存在至少一个所对应的召回模型的数量为至少两个的项目集合，也即存在至少一个目标项目集合，所述目标项目集合所对应的召回模型的数量为至少两个；每个召回模型对应一个召回项目序列。这样设置的好处在于，对于同一个热度等级的项目来说，可以采用多样的召回模型(一个召回模型可理解为一个召回通道)进行召回，不同的召回模型可以有自己的特点，利用各自特有的优势来实现互相补充，共同起作用，有利于挖掘更多的优质内容，提高召回项目的多样性。可选的，所对应的召回模型的数量为至少两个的项目集合可包括对应最低热度等级的项目集合，最低热度等级的项目可以理解为处于推荐系统的起步阶段，一般用户交互数据较少，若采用单一的常规的召回模型，一般无法在人群中进行很好的扩散，采用多通道可以更好地展示冷启项目。
可选的，对应最低热度等级的项目集合所对应的召回模型中包括以下至少一个：基于社交关系的冷启动召回模型、基于地理位置的冷启动召回模型、基 于用户兴趣的冷启动召回模型以及基于个性化的冷启动召回模型。其中，多个召回模型的结构以及相关参数不做限定，可根据实际需求进行设置。下面对上述几种召回模型做示意性说明。
基于社交关系的冷启动召回模型，基本策略可以是将与目标用户存在预设社交关系(如好友关系，关注或被关注关系，或者同处于一个通信群组等)的项目发布者发布的新项目(对应最低热度等级)推荐给目标用户，对于这些项目发布者发布的新项目来说，可以采用第一预设策略进行排序，第一预设策略可基于与目标用户的亲密度以及目标用户是否正在关注对应的展位等因素来设定。示例性的，以短视频为例，可理解为通过社交关系或关注展位等进行起量，当目标用户在平台进行浏览时，平台可以根据目标用户的社交关系，按一定的配比，将目标用户的好友发布的新视频推荐给用户，当用户浏览关注展位时，平台也可以根据该关注展位视频对应的发布者与目标用户的亲密度向目标用户推荐项目。
基于地理位置的冷启动召回模型，基本策略可以是将所在地理位置与目标用户当前地理位置之间距离处于预设距离范围内的项目发布者发布的新项目(对应最低热度等级)推荐给目标用户。示例性的，以短视频为例，可理解为通过获取目标用户的地理位置，将目标用户的附近用户发布的新视频推荐给目标用户。
基于用户兴趣的冷启动召回模型，基本策略可以是基于目标用户的历史行为确定目标用户对不同项目发布者的兴趣程度，将兴趣程度高于预设兴趣阈值的项目发布者确定为感兴趣项目发布者，将与感兴趣项目发布者相似度满足预设相似度需求的项目发布者发布的新项目(对应最低热度等级)推荐给目标用户。示例性的，以短视频为例，构建拍客-用户兴趣打分矩阵，对打分矩阵进行分解，得到拍客间的相似程度，当目标用户再次来平台进行浏览时，根据目标用户的历史喜好的拍客，搜索最近邻未知拍客，将该拍客发布的新视频推荐给目标用户，实现视频冷启目的。
基于个性化的冷启动召回模型，基本策略可以是采用预设模型基于新项目(对应最低热度等级)的属性和其他用户的行为数据计算新项目的特征向量，用户访问平台时，用相同的预设模型计算目标用户的特征向量，再通过目标用户的特征向量和新项目的特征向量计算兴趣打分，根据得分向目标用户推荐新项目。示例性的，以短视频为例，将新视频的基本属性和少量的行为作为特征输入，经训练好的深度神经网络(Deep Neural Networks，DNN)模型计算新视频特征向量，目标用户访问平台时，用相同DNN模型计算目标用户特征向量，再计算目标用户特征向量和新视频特征向量的距离，进行兴趣打分，根据兴趣 分向目标用户推荐topC的新视频进行冷启。
可选的，对应次低热度等级的项目集合所对应的召回模型中包括以下至少一个：基于用户的协同过滤算法的召回模型和基于跟随动作的召回模型。对于次低热度等级的项目，虽然积累的用户行为不是很多，但相比最低热度等级项目来说，用户行为相对多一些，可以采用其他的更多参考用户行为的召回模型进行召回。其中，多个召回模型的结构以及相关参数不做限定，可根据实际需求进行设置。下面对上述两种召回模型做示意性说明。其中，所述次低热度等级的项目对应的用户行为多于所述最低热度等级的项目对应的用户行为。
基于用户的协同过滤算法的召回模型(User Collaboration Filter，UserCF)，基本思想是指众多的用户可以齐心协力，通过不断地和平台互动，使自己的推荐列表能够不断过滤掉自己不感兴趣的项目，从而越来越满足自己的需求。而基于用户可以是指通过分析用户对项目的行为(如浏览、收藏、点赞以及转发等)计算出哪些用户是兴趣相似的，然后把兴趣相似的用户所关注的项目进行相互推荐。基本策略可以是获取目标用户的特征向量，根据该特征向量计算与目标用户的相似度满足预设相似度要求的多个种子用户，以多个种子用户为种子，获取种子用户针对项目集合产生的行为数据，根据行为数据对项目集合中的多个项目进行打分排序，得到召回项目序列。示例性的，以短视频为例，通过DNN模型获取目标用户的特征向量，根据特征向量计算与目标用户最相似的topD用户，将这些相似用户作为种子，搜索种子用户的点赞、关注、分享、下载以及完播等行为的冷门视频(次低热度等级的项目)，根据不同行为进行加权打分，最终返回topE视频对用户进行推荐。
基于跟随动作的召回模型(follow action)，基本策略可以是获取目标用户的关注列表，将关注列表中用户发生正向行为的项目推荐给目标用户。示例性的，以短视频为例，获取用户的关注列表，将关注列表中用户发生正向行为的视频加权打分后推荐给用户，其中，正向行为例如可包括点赞、关注、分享、下载以及完播等行为。
在一些实施例中，对于其他热度等级的项目集合来说，可以采用主流的召回模型。例如，可包括youtube DNN模型以及基于项目的协同过滤算法的召回模型(ItemCF)等等。
在一些实施例中，所述对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列，包括：利用预设排序打分模型对所述召回项目序列中的项目进行评分，得到召回项目序列中的项目对应的排序分；根据所述排序分对所述召回项目序列中的项目进行由高至低的排序，得到第一序列；获取所述第一序列中排序在前的预设数量的项目，并根据所获取的项目确定第一候选项目 序列。其中，预设排序打分模型的数量和类型不做限定，例如可包括用于粗略排序的第一排序打分模型和用于精细排序的第二排序打分模型，排序打分模型例如可以是深度兴趣模型(Deep Interest Network，DIN)。
在一些实施例中，所述获取排序在前的预设数量的项目，并根据所获取的项目确定第一候选项目序列，包括：获取排序在前的预设数量的项目，并根据所获取的项目确定第二序列；将处于所述第一序列中且未处于所述第二序列中的第二预设热度等级的项目，添加至所述第二序列之后，得到第一候选项目序列，其中，所述第二预设热度等级包括最低热度等级。这样设置的好处在于，在选取第二序列时，热度等级较低的项目可能因排序分较低并未入选，但若经过后续的分数提升处理后，可能有机会进入第二候选项目序列，因此，可提高热度较低的项目的展现概率。其中，被添加的项目的添加顺序可参照第一序列中的顺序确定，也即对于被添加到第二序列中的第一项目和第二项目来说，若在第一序列中第一项目相对第二项目更靠前，则在第一候选项目序列中，第一项目相对第二项目依然更靠前。
在一些实施例中，所述对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，包括：针对所述第一候选项目序列中的对应第一预设热度等级的每个目标项目，基于当前目标项目的展现次数对所述当前目标项目对应的排序分进行提升处理，得到当前目标项目对应的新的排序分；重新根据新的排序分对所述第一候选项目序列中的项目进行排序，得到第二候选项目序列。这样设置的好处在于，可以基于目标项目的展现次数进行有针对性的分数提升处理。可选的，展现次数越小，对应的提升幅度越大。
在一些实施例中，所述基于当前目标项目的展现次数对所述当前目标项目对应的排序分进行提升处理，得到当前目标项目对应的新的排序分，包括：利用上置信界(Upper Confidence Bound，UCB)算法基于当前目标项目的展现次数对所述当前目标项目对应的排序分进行提升处理，得到当前目标项目对应的新的排序分，其中，所述UCB算法对应的第一项中包含所述排序分，所述UCB算法对应的第二项中包含所述展现次数的倒数。这样设置的好处在于，UCB算法中包含排序部分和提权部分，但排序部分算法比较简单，精度不足，采用本申请中设计的排序结果替换UCB算法中的提权部分，而利用UCB算法中的提权部分对目标项目的排序分进行提升处理，实现对UCB算法的改进，可以得到更理想的最终排序结果。
在一些实施例中，所述重新根据新的排序分对所述第一候选项目序列中的项目进行排序，得到第二候选项目序列，包括：重新根据新的排序分对所述第 一候选项目序列中的项目进行排序，得到第三序列；针对所述第三序列中包含的第三预设热度等级的目标项目进行首次展示位置调节和/或相邻展示位置间隔调节，得到第二候选项目序列，其中，所述第三预设热度等级包括最低热度等级。展示位置也可理解为在第二候选项目序列中的排序位置。这样设置的好处在于，在得到按照经过分数提升处理的新的排序分重新排序的第三序列后，可针对低热度等级的项目进行展示位置的控制，防止冷项目首次展示位置过于靠前或冷项目展示位置过于密集，引起用户的不适，降低对用户的负面影响。
在一些实施例中，所述对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列，包括：对所述召回项目序列中的项目进行相同项目融合处理和/或违规项目过滤处理，得到待排序项目序列；对所述待排序项目序列中的项目进行评分排序，并确定第一候选项目序列。其中，相同项目融合处理可理解为去重处理，当一个项目集合对应多个召回模型时，同一个项目可能存在于两个以上的召回项目序列中，可进行去重处理。另外，可以在排序之前对项目进行审核处理，过滤掉非合规项目，保证平台的健康环境。可选的，对于召回项目序列中出现次数大于或等于2次的项目，也可进行标识，在计算排序分时，可考虑对带有标识的项目适当提升排序分。
在一些实施例中，在所述根据所述第二候选项目序列确定与所述目标用户对应的目标推荐项目序列之后，还包括：将所述目标推荐项目序列中的项目推荐给所述目标用户。其中，可按照预设推荐策略进行推荐，预设推荐策略可根据实际需求设置，例如可包括推荐时间点以及在界面中的推荐位置等。
图2为本申请实施例提供的又一种推荐项目确定方法的流程示意图，如图2所示，该方法可包括以下步骤。
步骤201、获取目标用户对应的目标用户特征数据。
步骤202、针对项目库中的每个项目集合，根据目标用户特征数据和与当前项目集合对应的召回模型，确定当前项目集合对应的召回项目序列。
其中，项目库中包含至少三个对应不同热度等级的项目集合。热度等级基于项目发布时间、项目展现次数和/或频率、以及项目对应的用户行为数据量确定。存在至少一个所对应的召回模型的数量为至少两个的项目集合，每个召回模型分别对应一个召回项目序列，存在至少两个包含项目数量不等的召回项目序列。
步骤203、对召回项目序列中的项目进行相同项目融合处理和/或违规项目过滤处理，得到待排序项目序列。
步骤204、利用预设排序打分模型对待排序项目序列中的项目进行评分，得 到待排序项目序列中的项目对应的排序分。
示例性的，采用DIN模型来计算待排序项目序列中的每个项目对应的排序分。
步骤205、根据排序分对待排序项目序列中的项目进行由高至低的排序，得到第一序列。
步骤206、获取第一序列中排序在前的第一预设数量的项目，并根据所获取的项目确定第二序列。
步骤207、将处于第一序列中且未处于第二序列中的第二预设热度等级的项目，添加至第二序列之后，得到第一候选项目序列。
其中，所述第二预设热度等级包括最低热度等级。
步骤208、针对第一候选项目序列中的对应第一预设热度等级的每个目标项目，利用UCB算法基于当前目标项目的展现次数对当前目标项目对应的排序分进行提升处理，得到当前目标项目对应的新的排序分。
其中，所述第一预设热度等级包括最低热度等级。
示例性的，记新的排序分为score，原排序分为rank_score(也即步骤204中的排序分)，可基于如下变形后的UCB算法公式来计算score：
score＝rank_score+boost_score/T beta
其中，T表示项目的展现次数(也即当前目标项目的当前展现数)，boost_score为第一预设数值(可以是经验值或通过实验调出的固定值，一般取值范围为1到10)，beta为第二预设数值(可以是经验值或通过实验调出的固定值，一般取值范围为0到1)。
步骤209、重新根据排序分对第一候选项目序列中的项目进行排序，得到第三序列。
步骤210、针对第三序列中包含的第三预设热度等级的目标项目进行首次展示位置调节和/或相邻展示位置间隔调节，得到第二候选项目序列。
其中，所述第三预设热度等级包括最低热度等级。第一预设热度等级、第二预设热度等级和第三预设热度等级可以相同，也可不同。
步骤211、获取第二候选项目序列中排序在前的第二预设数量的项目，得到与目标用户对应的目标推荐项目序列。
步骤212、按照预设推荐策略将目标推荐项目序列中的项目推荐给目标用户。
本申请实施例中提供的推荐项目确定方法，针对至少三个对应不同热度等级的项目集合分别根据对应的召回模型确定召回项目序列，其中，一个或一些项目集合可以对应不同的召回模型，可以利用更加符合不同热度阶段的多样化的召回模型来确定多个召回项目序列，使得召回项目序列更加有针对性、更加准确以及更加丰富，对UCB算法进行改进，利用预设排序打分模型计算项目序列的排序分，并基于展现次数的倒数对热度等级较低的项目进行分数提升处理，并对低热度项目的展示位置进行有效控制，在有效提升低热度项目的展现频率的同时控制冷项目对用户产生的负面影响，从而更加合理的解决推荐系统中的E&E问题。
图3为本申请实施例提供的一种推荐视频确定方法的流程示意图，以短视频推荐场景为例进行说明，可基于视频发布时间、视频展现次数和/或频率、以及视频对应的用户行为数据量等将短视频对应的热度等级设定为起量阶段、扩散阶段、承接阶段和热门阶段。
示例性的，该方法可包括以下步骤。
步骤301、获取目标用户对应的目标用户特征数据。
步骤302、针对视频库中的每个视频集合，根据目标用户特征数据和与当前视频集合对应的召回模型，确定当前视频集合对应的召回视频序列。
其中，视频库中包含四个对应不同热度等级的视频集合。每个召回模型可以视为一个召回通道。
对于起量阶段，所对应的召回模型中包括基于社交关系的冷启动召回模型、基于地理位置的冷启动召回模型、基于用户兴趣的冷启动召回模型以及基于个性化的冷启动召回模型。起量阶段的这几个召回通道各有独特的优势，在资源冷启时互相补充，共同起作用，视频资源在推荐系统的起步阶段，由于没有足够的用户交互数据，一般不能通过单一的、常规的渠道在人群中进行很好的扩散。通过多通道、多形式地向用户展现冷启资源，有利于挖掘更多的优质内容，促进系统生态良好发展。同时，多通道冷启的方式提高了资源的多样性。
对于扩散阶段，所对应的召回模型中包括基于用户的协同过滤算法的召回模型和基于跟随动作的召回模型。当新视频进行一定起量阶段，已经具有了一些行为后，可通过上述通道进行扩散。
对于承接阶段和热门阶段，视频已经达到一定分发量或已经火起来，可以采用主流召回通道进行建库召回。
图4为本申请实施例提供的一种召回模型配置的示意图。如图4所示，起量阶段可以采用的召回模型包括基于社交关系(relationship)和跟随(follow) 展位的冷启动召回模型、基于地理位置的冷启动召回模型(nearby)、基于用户兴趣的冷启动召回模型(PosterCF)以及基于个性化的冷启动召回模型(DNN_cold)；扩散阶段可以采用的召回模型包括基于用户的协同过滤算法的召回模型(User Based Collaborative Filtering，UserCF)和基于跟随动作的召回模型(follow action)；承接阶段可以采用的召回模型包括DNN召回模型和基于DNN的冷启系统推荐召回模型(DNN EE)；热门阶段可以采用的召回模型包括基于热门视频组成的召回模型(hot，例如常见的热门排行榜)、基于项目的协同过滤算法的召回模型(ItemCF)和youtube DNN模型。上述仅为举例说明，在实际应用时，可采用其他召回模型替换其中的任意一个召回模型，也可采用相比于图4中更少或更多的召回模型，多个召回模型所对应的热度等级也可自由设置，本申请实施例均不做限定。
可选的，每个召回模型，也即每个召回通道，分别输出一个召回视频序列，每个召回视频序列的大小(也即包含的视频数量)可以相等。在进行后续步骤之前，可以按照每个通道配置的比例进行截断处理，得到处理后的召回视频序列。这样设置的好处在于，当一个召回模型对应的召回视频序列大小需要发生变化时，不需要对召回模型进行改动，直接调整对应的比例，即可实现快速准确的调整序列大小，避免出错。这种方式同样适用于视频以外的其他项目。
步骤303、对召回视频序列中的视频进行相同视频融合处理和违规视频过滤处理，得到待排序视频序列。
步骤304、利用预设排序打分模型对待排序视频序列中的视频进行评分，得到待排序视频序列中的视频对应的排序分。
步骤305、根据排序分对待排序视频序列中的视频进行由高至低的排序，得到第一序列。
步骤306、获取第一序列中排序在前的第一预设数量的视频，并根据所获取的视频确定第二序列。
示例性的，可获取topN的视频，得到第二序列，N例如可以是1000。
步骤307、将处于第一序列中且未处于第二序列中的起量阶段和扩散阶段的视频，添加至第二序列之后，得到第一候选视频序列。
步骤308、针对第一候选视频序列中的对应起量阶段和扩散阶段的每个目标视频，利用UCB算法基于当前目标视频的展现次数对当前目标视频对应的排序分进行提升处理，得到当前目标视频对应的新的排序分。
示例性的，记新的排序分为score，排序分为rank_score，可基于如下变形后的UCB算法公式来计算score：
score＝rank_score+boost_score/VV beta
其中，VV表示视频的展现次数(也即当前目标视频的当前展现数)，boost_score为第一预设数值(可以是经验值或通过实验调出的固定值，一般取值范围为1到10)，beta为第二预设数值(可以是经验值或通过实验调出的固定值，一般取值范围为0到1)。
步骤309、重新根据排序分对第一候选视频序列中的视频进行排序，得到第三序列。
步骤310、针对第三序列中包含的起量阶段和扩散阶段的目标视频进行首次展示位置调节和相邻展示位置间隔调节，得到第二候选视频序列。
其中，针对起量阶段和扩散阶段的目标视频的调整方式可以不同，例如，起量阶段的目标视频的首次展现位置与扩散阶段的目标视频相比可以更加靠后一些，起量阶段的目标视频的相邻展示位置间隔与扩散阶段的目标视频相比可以更加大一些等等。
步骤311、获取第二候选项目序列中排序在前的第二预设数量的项目，得到与目标用户对应的目标推荐视频序列。
示例性的，可获取topM的视频，得到目标推荐视频序列，M例如可以是20。
步骤312、按照预设推荐策略将目标推荐视频序列中的视频推荐给目标用户。
本申请实施例中提供的推荐视频确定方法，针对不同热度等级的视频集合分别根据对应的召回模型确定召回视频序列，每个热度等级均设置有多个召回模型，也即设置了多路召回通道，这些通道相互补充，又分管不同职能，多阶段多通道的架构，可以满足视频不同冷启程度的需求，共同完成冷视频流转分发，可以快速感知用户反馈，筛选出优质视频，提高冷启效率，也即当用户在平台上发布新视频时，可以快速得到扩散分发，从而打造出良好的生态体系，对UCB算法进行改进，利用预设排序打分模型计算视频序列的排序分，并结合UCB boost提权的方式，基于展现次数的倒数对热度等级较低的视频进行分数提升处理，并控制冷视频的下发密度减轻对用户的负面影响，在满足一定个性化需求的前提下，有效提升低热度视频的展现频率的同时控制冷视频对用户产生的负面影响，从而更加合理的解决视频推荐系统中的E&E问题。
图5为本申请实施例提供的一种推荐项目确定装置的结构框图，该装置可由软件和/或硬件实现，一般可集成在计算机设备中，可通过执行推荐项目确定方法来确定待推荐的项目。如图5所示，该装置包括以下模块。
召回项目序列确定模块501，设置为针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定当前项目集合对应的召回项目序列，其中，所述项目库中包含至少三个对应不同热度等级的项目集合；
评分排序模块502，设置为对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列；
分数提升模块503，设置为对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，所述第一预设热度等级包括最低热度等级；
目标推荐项目确定模块504，设置为根据所述第二候选项目序列确定与所述目标用户对应的目标推荐项目序列。
本申请实施例提供的推荐项目确定装置，针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定对应的召回项目序列，其中，项目库中包含至少三个对应不同热度等级的项目集合，对召回项目序列中的项目进行评分排序，并确定第一候选项目序列，对第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，第一预设热度等级包括最低热度等级，根据第二候选项目序列确定与目标用户对应的目标推荐项目序列。通过采用上述技术方案，针对至少三个对应不同热度等级的项目集合分别根据对应的召回模型确定召回项目序列，可以对项目所处的热度阶段进行更细致的划分，从而利用更加符合项目所对应的热度阶段的召回模型来确定召回项目序列，使得召回项目序列更加有针对性以及更加准确，在对召回项目序列中的项目进行评分排序后，针对包含最低热度等级项目的目标项目进行分数提升处理，可有效提升低热度项目的展现频率，也即使冷项目得到更加合理的分发。
本申请实施例提供了一种计算机设备，该计算机设备中可集成本申请实施例提供的推荐项目确定装置。图6为本申请实施例提供的一种计算机设备的结构框图。计算机设备600包括存储器601、处理器602及存储在存储器601上并可在处理器602上运行的计算机程序，所述处理器602执行所述计算机程序时实现本申请实施例提供的推荐项目确定方法。
本申请实施例还提供一种包含计算机可执行指令的存储介质，所述计算机可执行指令在由计算机处理器执行时用于执行本申请实施例提供的推荐项目确定方法。
上述实施例中提供的推荐项目确定装置、设备以及存储介质可执行本申请任意实施例所提供的推荐项目确定方法，具备执行该方法相应的功能模块和效果。未在上述实施例中详尽描述的技术细节，可参见本申请任意实施例所提供的推荐项目确定方法。

Claims (14)

  1. 一种推荐项目确定方法，包括：
    针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定所述当前项目集合对应的召回项目序列，其中，所述项目库中包含至少三个对应不同热度等级的项目集合；
    对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列；
    对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，所述第一预设热度等级包括最低热度等级；
    根据所述第二候选项目序列确定与所述目标用户对应的目标推荐项目序列。
  2. 根据权利要求1所述的方法，其中，所述热度等级基于项目发布时间、项目展现次数、项目展现频率、以及项目对应的用户行为数据量中的至少一个确定。
  3. 根据权利要求1所述的方法，其中，存在至少一个目标项目集合，所述目标项目集合所对应的召回模型的数量为至少两个；
    每个召回模型对应一个召回项目序列。
  4. 根据权利要求3所述的方法，其中，对应所述最低热度等级的项目集合所对应的召回模型中包括以下至少一个：
    基于社交关系的冷启动召回模型、基于地理位置的冷启动召回模型、基于用户兴趣的冷启动召回模型以及基于个性化的冷启动召回模型。
  5. 根据权利要求3所述的方法，其中，对应次低热度等级的项目集合所对应的召回模型中包括以下至少一个：
    基于用户的协同过滤算法的召回模型和基于跟随动作的召回模型，
    其中，所述次低热度等级的项目对应的用户行为多于所述最低热度等级的项目对应的用户行为。
  6. 根据权利要求1所述的方法，其中，所述对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列，包括：
    利用预设排序打分模型对所述召回项目序列中的项目进行评分，得到所述召回项目序列中的项目对应的排序分；
    根据所述排序分对所述召回项目序列中的项目进行由高至低的排序，得到第一序列；
    获取所述第一序列中排序在前的预设数量的项目，并根据所获取的项目确定所述第一候选项目序列。
  7. 根据权利要求6所述的方法，其中，所述获取所述第一序列中排序在前的预设数量的项目，并根据所获取的项目确定所述第一候选项目序列，包括：
    获取所述排序在前的预设数量的项目，并根据所述所获取的项目确定第二序列；
    将处于所述第一序列中且未处于所述第二序列中的第二预设热度等级的项目，添加至所述第二序列之后，得到所述第一候选项目序列，其中，所述第二预设热度等级包括最低热度等级。
  8. 根据权利要求6所述的方法，其中，所述对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，包括：
    针对所述第一候选项目序列中的对应所述第一预设热度等级的每个目标项目，基于当前目标项目的展现次数对所述当前目标项目对应的排序分进行提升处理，得到所述当前目标项目对应的新的排序分；
    重新根据新的排序分对所述第一候选项目序列中的项目进行排序，得到所述第二候选项目序列。
  9. 根据权利要求8所述的方法，其中，所述基于当前目标项目的展现次数对所述当前目标项目对应的排序分进行提升处理，得到所述当前目标项目对应的新的排序分，包括：
    利用上置信界UCB算法基于所述当前目标项目的展现次数对所述当前目标项目对应的排序分进行提升处理，得到所述当前目标项目对应的新的排序分，其中，所述UCB算法对应的第一项中包含所述排序分，所述UCB算法对应的第二项中包含所述展现次数的倒数。
  10. 根据权利要求8所述的方法，其中，所述重新根据新的排序分对所述第一候选项目序列中的项目进行排序，得到所述第二候选项目序列，包括：
    重新根据新的排序分对所述第一候选项目序列中的项目进行排序，得到第三序列；
    针对所述第三序列中包含的第三预设热度等级的目标项目进行以下至少一种调节：首次展示位置调节、或相邻展示位置间隔调节，得到所述第二候选项目序列，其中，所述第三预设热度等级包括最低热度等级。
  11. 根据权利要求1所述的方法，其中，所述对所述召回项目序列中的项 目进行评分排序，并确定第一候选项目序列，包括：
    对所述召回项目序列中的项目进行以下至少一种处理：相同项目融合处理、或违规项目过滤处理，得到待排序项目序列；
    对所述待排序项目序列中的项目进行评分排序，并确定所述第一候选项目序列。
  12. 一种推荐项目确定装置，包括：
    召回项目序列确定模块，设置为针对项目库中的每个项目集合，根据目标用户对应的目标用户特征数据和与当前项目集合对应的召回模型，确定所述当前项目集合对应的召回项目序列，其中，所述项目库中包含至少三个对应不同热度等级的项目集合；
    评分排序模块，设置为对所述召回项目序列中的项目进行评分排序，并确定第一候选项目序列；
    分数提升模块，设置为对所述第一候选项目序列中的对应第一预设热度等级的目标项目进行分数提升处理，得到第二候选项目序列，其中，所述第一预设热度等级包括最低热度等级；
    目标推荐项目确定模块，设置为根据所述第二候选项目序列确定与所述目标用户对应的目标推荐项目序列。
  13. 一种计算机设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，其中，所述处理器执行所述计算机程序时实现如权利要求1-11中任一项所述的推荐项目确定方法。
  14. 一种计算机可读存储介质，存储有计算机程序，其中，所述计算机程序被处理器执行时实现如权利要求1-11中任一项所述的推荐项目确定方法。
PCT/CN2021/095567 2020-07-09 2021-05-24 推荐项目确定方法、装置、设备及存储介质 WO2022007520A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010658808.4A CN111859126B (zh) 2020-07-09 推荐项目确定方法、装置、设备及存储介质
CN202010658808.4 2020-07-09

Publications (1)

Publication Number Publication Date
WO2022007520A1 true WO2022007520A1 (zh) 2022-01-13

Family

ID=73152660

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/095567 WO2022007520A1 (zh) 2020-07-09 2021-05-24 推荐项目确定方法、装置、设备及存储介质

Country Status (1)

Country Link
WO (1) WO2022007520A1 (zh)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114302242A (zh) * 2022-01-25 2022-04-08 聚好看科技股份有限公司 一种媒资推荐方法、显示设备及服务器
CN114564496A (zh) * 2022-03-01 2022-05-31 北京有竹居网络技术有限公司 一种内容推荐方法及装置
CN114676851A (zh) * 2022-04-08 2022-06-28 中国科学技术大学 召回和排序模型的联合训练方法、设备和存储介质
CN114785852A (zh) * 2022-04-07 2022-07-22 北京达佳互联信息技术有限公司 一种推送内容确定方法、装置、电子设备及存储介质
CN115033782A (zh) * 2022-05-18 2022-09-09 百度在线网络技术（北京）有限公司 推荐对象的方法、机器学习模型的训练方法、装置和设备
CN115062184A (zh) * 2022-06-29 2022-09-16 四川长虹电器股份有限公司 一种语音召回场景下的影片排序方法
CN115150668A (zh) * 2022-06-13 2022-10-04 咪咕文化科技有限公司 视频提示方法、装置、设备以及计算机存储介质
CN115474070A (zh) * 2022-08-10 2022-12-13 武汉斗鱼鱼乐网络科技有限公司 一种展示新内容的方法、装置、介质及设备
CN115563394A (zh) * 2022-11-24 2023-01-03 腾讯科技（深圳）有限公司 搜索召回方法、召回模型训练方法、装置和计算机设备
CN115587261A (zh) * 2022-12-09 2023-01-10 思创数码科技股份有限公司 政务资源目录推荐方法及系统
CN117455579A (zh) * 2023-12-25 2024-01-26 深圳市智岩科技有限公司 商品推荐干预方法、装置以及介质和设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160212494A1 (en) * 2013-09-06 2016-07-21 Beijing Qihoo Technology Company Limited Video recommendation method and device
CN110019945A (zh) * 2017-12-28 2019-07-16 飞狐信息技术（天津）有限公司 视频推荐方法及装置、存储介质及电子设备
CN110659388A (zh) * 2019-10-10 2020-01-07 北京奇艺世纪科技有限公司 一种待推荐信息筛选方法、装置、电子设备及存储介质
CN111008278A (zh) * 2019-11-22 2020-04-14 厦门美柚股份有限公司 内容推荐方法及装置
CN111339416A (zh) * 2020-02-25 2020-06-26 咪咕文化科技有限公司 热度召回方法、装置、电子设备及存储介质
CN111859126A (zh) * 2020-07-09 2020-10-30 有半岛(北京)信息科技有限公司 推荐项目确定方法、装置、设备及存储介质

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160212494A1 (en) * 2013-09-06 2016-07-21 Beijing Qihoo Technology Company Limited Video recommendation method and device
CN110019945A (zh) * 2017-12-28 2019-07-16 飞狐信息技术（天津）有限公司 视频推荐方法及装置、存储介质及电子设备
CN110659388A (zh) * 2019-10-10 2020-01-07 北京奇艺世纪科技有限公司 一种待推荐信息筛选方法、装置、电子设备及存储介质
CN111008278A (zh) * 2019-11-22 2020-04-14 厦门美柚股份有限公司 内容推荐方法及装置
CN111339416A (zh) * 2020-02-25 2020-06-26 咪咕文化科技有限公司 热度召回方法、装置、电子设备及存储介质
CN111859126A (zh) * 2020-07-09 2020-10-30 有半岛(北京)信息科技有限公司 推荐项目确定方法、装置、设备及存储介质

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114302242A (zh) * 2022-01-25 2022-04-08 聚好看科技股份有限公司 一种媒资推荐方法、显示设备及服务器
CN114302242B (zh) * 2022-01-25 2023-10-31 聚好看科技股份有限公司 一种媒资推荐方法、显示设备及服务器
CN114564496B (zh) * 2022-03-01 2023-09-19 北京有竹居网络技术有限公司 一种内容推荐方法及装置
CN114564496A (zh) * 2022-03-01 2022-05-31 北京有竹居网络技术有限公司 一种内容推荐方法及装置
CN114785852A (zh) * 2022-04-07 2022-07-22 北京达佳互联信息技术有限公司 一种推送内容确定方法、装置、电子设备及存储介质
CN114785852B (zh) * 2022-04-07 2024-04-09 北京达佳互联信息技术有限公司 一种推送内容确定方法、装置、电子设备及存储介质
CN114676851A (zh) * 2022-04-08 2022-06-28 中国科学技术大学 召回和排序模型的联合训练方法、设备和存储介质
CN114676851B (zh) * 2022-04-08 2024-03-29 中国科学技术大学 召回和排序模型的联合训练方法、设备和存储介质
CN115033782A (zh) * 2022-05-18 2022-09-09 百度在线网络技术（北京）有限公司 推荐对象的方法、机器学习模型的训练方法、装置和设备
CN115150668A (zh) * 2022-06-13 2022-10-04 咪咕文化科技有限公司 视频提示方法、装置、设备以及计算机存储介质
CN115062184A (zh) * 2022-06-29 2022-09-16 四川长虹电器股份有限公司 一种语音召回场景下的影片排序方法
CN115474070A (zh) * 2022-08-10 2022-12-13 武汉斗鱼鱼乐网络科技有限公司 一种展示新内容的方法、装置、介质及设备
CN115563394B (zh) * 2022-11-24 2023-03-28 腾讯科技（深圳）有限公司 搜索召回方法、召回模型训练方法、装置和计算机设备
CN115563394A (zh) * 2022-11-24 2023-01-03 腾讯科技（深圳）有限公司 搜索召回方法、召回模型训练方法、装置和计算机设备
CN115587261A (zh) * 2022-12-09 2023-01-10 思创数码科技股份有限公司 政务资源目录推荐方法及系统
CN117455579A (zh) * 2023-12-25 2024-01-26 深圳市智岩科技有限公司 商品推荐干预方法、装置以及介质和设备
CN117455579B (zh) * 2023-12-25 2024-04-09 深圳市智岩科技有限公司 商品推荐干预方法、装置以及介质和设备

Also Published As

Publication number Publication date
CN111859126A (zh) 2020-10-30

Similar Documents

Publication Publication Date Title
WO2022007520A1 (zh) 推荐项目确定方法、装置、设备及存储介质
US11263217B2 (en) Method of and system for determining user-specific proportions of content for recommendation
US9723044B1 (en) Stream of content for a channel
EP4080889A1 (en) Anchor information pushing method and apparatus, computer device, and storage medium
US10318538B2 (en) Systems, methods, and apparatuses for implementing an interface to view and explore socially relevant concepts of an entity graph
CN106302085B (zh) 即时通讯群组的推荐方法及系统
US20210382609A1 (en) Method and device for displaying multimedia resource
US20160232244A1 (en) Methods, systems, and media for personalizing computerized services based on mood and/or behavior information from multiple data sources
WO2017071251A1 (zh) 信息推送方法和装置
US9672530B2 (en) Supporting voting-based campaigns in search
US20150195624A1 (en) Media-Related Trends
KR20160059486A (ko) 연속적인 소셜 커뮤니케이션을 위한 시스템 및 방법
WO2016109567A1 (en) Bookmarking search results
US20230036644A1 (en) Method and system for exploring a personal interest space
WO2013184957A1 (en) Systems and methods of classifying content items
TW201447797A (zh) 內容個人化之多相排序方法和系統
WO2020238502A1 (zh) 物品推荐方法及装置、电子设备及存储介质
US11379528B2 (en) Tag-based, user-directed media recommendations
US20150081660A1 (en) Identifying Gaps In Search Results
US20150287092A1 (en) Social networking consumer product organization and presentation application
US20150142778A1 (en) Presenting Previously Selected Search Results
CN113778286B (zh) 应用控件显示方法、装置、存储介质及电子设备
CN110188123A (zh) 用户匹配方法及设备
CN111859126B (zh) 推荐项目确定方法、装置、设备及存储介质
US20240152512A1 (en) Machine learning for dynamic information retrieval in a cold start setting

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21838420

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21838420

Country of ref document: EP

Kind code of ref document: A1