WO2020211840A1 - 物料推荐方法和系统 - Google Patents

物料推荐方法和系统 Download PDF

Info

Publication number
WO2020211840A1
WO2020211840A1 PCT/CN2020/085347 CN2020085347W WO2020211840A1 WO 2020211840 A1 WO2020211840 A1 WO 2020211840A1 CN 2020085347 W CN2020085347 W CN 2020085347W WO 2020211840 A1 WO2020211840 A1 WO 2020211840A1
Authority
WO
WIPO (PCT)
Prior art keywords
recommendation
pool
level
recommended
materials
Prior art date
Application number
PCT/CN2020/085347
Other languages
English (en)
French (fr)
Inventor
杨闯亮
程晓澄
周开拓
Original Assignee
第四范式（北京）技术有限公司
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
Application filed by 第四范式（北京）技术有限公司 filed Critical 第四范式（北京）技术有限公司
Publication of WO2020211840A1 publication Critical patent/WO2020211840A1/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

Definitions

  • the present disclosure generally relates to a computer-implemented recommendation scheme, and more specifically, to a method and system for material recommendation.
  • the purpose of the present disclosure is to provide a material recommendation method and system to solve the poor recommendation effect caused by too many new materials or insufficient exploration flow in the prior art.
  • the present disclosure provides a material recommendation method.
  • the method includes: determining the number n of recommended pools, where n is an integer greater than or equal to 3; determining the display flow rate of a single material in each level of recommendation pool so that the The display flow increases from the first-level recommendation pool to the n-1th-level recommendation pool; for a specific recommendation request, select candidate materials from the n-level recommendation pool; recommend based on the selected candidate materials; and obtain the recommended materials Corresponding user behavior data, and the method also includes allocating materials in each recommendation pool in the following manner: putting the continuously generated new materials into the first-level recommendation pool; for the materials in the i-1th-level recommendation pool, Count the user behavior data corresponding to the material with the corresponding display traffic as a unit, and determine whether to remove the material from the i-1 level recommendation pool, keep it in the i-1 level recommendation pool or not based on the statistical results Move into the i-th recommended pool; and, for the materials in the n-th recommended pool, use the corresponding display flow as
  • the present disclosure provides a material recommendation system.
  • the system includes: a recommendation pool generating unit that determines the number of recommendation pools n, where n is an integer greater than or equal to 3; and a display flow determination unit that determines the recommended pool at each level Corresponding to the display flow of a single material, so that the display flow increases from the first-level recommendation pool to the n-1th-level recommendation pool; the candidate material selection unit selects candidate materials from the n-level recommendation pool for a specific recommendation request The recommendation unit, which makes recommendations based on the selected candidate materials; the data acquisition unit, which obtains user behavior data corresponding to the recommended materials; and the screening unit, which allocates materials in each recommendation pool in the following way: puts the continuously generated new materials into The first-level recommendation pool; for the materials in the i-1th-level recommendation pool, the user behavior data corresponding to the materials are counted with the corresponding display flow as a unit, and based on the statistical results, the materials are determined from the i-th -1 level recommendation pool moved out, retained in
  • Another aspect of the present disclosure provides a computer-readable storage medium, wherein when the computer instructions are executed by at least one computing device, the at least one computing device is caused to execute the material recommendation method as described above.
  • Another aspect of the present disclosure provides a system including at least one computing device and at least one storage device storing instructions, wherein the instructions when executed by the at least one computing device cause the at least one computing device to execute The recommended method of materials as described above.
  • the method and system for material recommendation can solve the problem of poor recommendation effect caused by too much new material generated or insufficient exploration flow by adding a multi-level recommendation pool. Furthermore, through the multi-level recommendation pool, high-quality content can be found faster with less traffic, the cold-start recommendation algorithm can be optimized, and the screening of high-quality content can be realized more quickly.
  • FIG. 1 is a flowchart showing a method of material recommendation according to an exemplary embodiment of the present disclosure
  • FIG. 2 is a block diagram showing a system for material recommendation according to an exemplary embodiment of the present disclosure.
  • FIG. 3 is a schematic diagram illustrating an environment in which the system in FIG. 2 is used for material recommendation according to an exemplary embodiment of the present disclosure.
  • FIG. 1 is a flowchart illustrating a method of material recommendation according to an exemplary embodiment of the present disclosure.
  • step S10 the number n of recommended pools can be determined, where n is an integer greater than or equal to 3. In other words, at least 3 recommended pools are generated.
  • any suitable method may be adopted to determine the number of recommended pools.
  • the number n of recommended pools may be determined based on at least one of the generation speed of the new material and the flow distribution.
  • the material can indicate any content or items that may be recommended, for example, blog posts, news information, forum posts, videos, short videos, music, pictures, funny paragraphs, etc.
  • the generation speed of new materials can be estimated by counting the total number of materials generated in a specific period. For example, if the number of new materials generated in a specific time period (for example, one day) is 10,000, n can be determined to be 3 to generate a 3-level recommendation pool; if the number of new materials generated during a specific time period (for example, one day) If the number is 100,000, n can be determined as 4, thereby generating a 4-level recommendation pool. It should be noted that the numerical values here are only examples and do not limit the scope of the present disclosure.
  • the number n of recommended pools may be determined based on traffic distribution.
  • the flow allocation refers to dividing the flow corresponding to the user request into the exploration flow and the utilization flow in the "exploration-utilization" approach.
  • exploit is a relatively definite interest for users, and mining should be used to cater to them; however, if users only use the known interests of users, users will soon be repeated and bored, so it is necessary to continuously explore users’ interest in new materials. interest of.
  • the former can obtain a stable effect, but it is not necessarily the best.
  • the latter may obtain a better effect, but it may also obtain an effect that is not as good as the previous method. Therefore, the distribution between exploratory flow and utilization flow needs to be balanced according to specific conditions.
  • the more the number of new materials the more exploratory traffic is needed to explore the new materials; on the other hand, the excessive exploratory traffic will affect the user experience.
  • the allocated exploration flow that is, the exploration flow in the "exploration-utilization" approach
  • relatively few recommendation pools can be set
  • a relative More recommended pool if the allocated exploration flow is relatively low
  • the recommendation pool from level 1 to level n-1 can correspond to the exploration traffic in the "exploration-utilization" approach, and the level n recommendation pool can correspond to the utilization in the "exploration-utilization” approach flow.
  • the exploration flow may satisfy at least one of the following conditions: the exploration flow is proportional to the quantity of new materials, and the exploration flow does not exceed one tenth of the total flow including the exploration flow and the utilization flow.
  • the exploration flow is proportional to the quantity of new materials
  • the exploration flow does not exceed one tenth of the total flow including the exploration flow and the utilization flow.
  • the exploration flow can be increased to 3.9 million, and the utilization flow can be increased to 46.1 million, that is, the exploration flow can be proportional to the number of new materials.
  • the exploration flow can be controlled within 5 million by adjusting other conditions, and the utilization flow is 45 million, that is, the exploration flow does not exceed one-tenth of the total flow. .
  • the two methods of determining the recommended pool number based on the rate of new material generation and determining the recommended pool number based on the total flow rate described above can also be combined to comprehensively determine the appropriate recommended pool number.
  • step S20 the display flow rate corresponding to a single recommended material in each level of recommendation pool can be determined, so that the display flow rate increases sequentially from the level 1 recommendation pool to the n level recommendation pool.
  • the display flow of each level of recommendation pool for a single recommended material can be determined, so that in the recommendation pool corresponding to the exploration flow, as the level grows, the corresponding display flow increases, and the recommendation pool corresponding to the utilization flow corresponds to Has the most impression traffic.
  • the display flow of a single recommended material corresponding to the first recommended pool can be 100, and the display flow of a single recommended material corresponding to the second recommended pool can be 4500.
  • the display flow rate of the third-level recommendation pool corresponding to a single recommended material can be greater than the display flow rate of the second-level recommendation pool corresponding to a single recommended material.
  • the exploration traffic is 3.25 million and the utilization traffic is 46.75 million.
  • the exploration flow rate is 3.9 million and the utilization flow rate is 46.1 million, that is, the exploration flow rate can be proportional to the number of new materials. It should be noted that although the number of new materials in the second-level recommended pool in the example is 5% of the first-level recommended pool, the percentage shown here is only for ease of description and does not limit the technical concept of the present disclosure. In other exemplary embodiments, other proportional relationships may be selected or not restricted by specific proportional relationships.
  • the display flow of a single recommended material corresponding to the first recommended pool can be 100
  • the display flow of a single recommended material corresponding to the second recommended pool can be 200
  • the third recommended pool The display flow rate corresponding to a single recommended material can be 4500
  • the display flow rate of a single recommended material corresponding to the level 4 recommendation pool can be greater than the display flow rate of a single recommended material corresponding to the level 3 recommendation pool.
  • the method in addition to the quantity of new materials, can also continuously obtain the total flow rate corresponding to the user request, and perform step S10 and step S20 again according to the change of the total flow rate and the generation speed of the new material, thereby Dynamically adjust the number n of recommended pools and the display flow rate of a single recommended material in each recommended pool.
  • the number of recommended pools is dynamically increased (for example, a first-level recommended pool is added) as at least one of the generation speed of new materials increases and the allocated exploration flow decreases, or as new At least one of the decrease in the production speed of the material and the increase in the allocated exploration flow dynamically reduces the number of recommended pools (for example, reduces the first-level recommended pool).
  • the number of recommended pools n can also be determined relatively fixedly based on experience or historically stable data.
  • the 3-level recommendation pool can be adjusted to the 4-level recommendation pool.
  • the display traffic of the first recommended pool can be 100
  • the second recommended pool The display traffic can be 4500, and still simply assume that the number of new materials in the second-level recommendation pool is 5% of the first-level recommendation pool, then the expected exploration traffic is 3.25 million and the utilization traffic is 46.75 million.
  • the level 3 recommendation pool can be adjusted to the level 4 recommendation pool, and in step S20, the level 1 recommendation pool can be adjusted.
  • the display traffic corresponding to a single recommended recommended material is adjusted to 100
  • the display traffic in the second-level recommendation pool can be adjusted to 300
  • the display traffic in the third-level recommendation pool can be adjusted to 4600.
  • the second-level recommendation pool The quantity of new materials is 5% of the level 1 recommendation pool, and the quantity of new materials in the level 3 recommendation pool is 5% of the quantity of new materials in the level 2 recommendation pool. In this way, after the adjustment, the total exploration flow required for the four-level recommendation pool is 2.53 million, so as to explore as many new materials as possible with less exploration flow.
  • the specific numerical values here are only examples for making the concept of the present disclosure clear, and do not limit the scope of the present disclosure. In the above-mentioned embodiment, only the situation where the quantity of new materials increases is shown. In contrast, when the quantity of new materials decreases, one level of recommendation pool can be reduced and the display flow corresponding to each level can be adjusted appropriately.
  • the specific method may be the same or similar to the above-mentioned embodiment, and the repeated description will be omitted here.
  • the total flow rate is unchanged.
  • the total number of recommended pool levels and the display traffic corresponding to each level can be adjusted with reference to the foregoing embodiment.
  • the situation that the real-time total flow is less than the expected total flow is the same or similar to the situation that the number of new materials increases. Therefore, the total number of recommended pool levels and the display flow corresponding to each level can be adjusted with reference to the above embodiments.
  • step S30 candidate materials can be selected from the n-level recommendation pool for a specific recommendation request.
  • At least one target recommendation pool may be determined among the n-level recommendation pools; and candidate materials may be selected from the determined at least one target recommendation pool.
  • the target recommendation pool may be a single recommendation pool or multiple recommendation pools.
  • the target recommendation pool is a single recommendation pool
  • a single target recommendation pool can be determined from the first-level recommendation pool to the n-1th-level recommendation pool
  • the nth level recommendation pool may be determined as the target recommendation pool.
  • a single target recommendation pool is determined from the level 1 recommendation pool to the n-1 level recommendation pool by at least one of the following methods: random selection The target recommendation pool is selected according to the characteristics of the user who issued the recommendation request and the target recommendation pool is selected according to preset rules.
  • a recommendation pool selected at random or selected according to a preset rule may be used as the target recommendation pool, or the target recommendation pool may also be selected in combination with the characteristics of the user (ie, the user requesting the recommendation) for which the material will be recommended.
  • the characteristics of the user who made the recommendation request may include, but are not limited to, the attributes of the user, such as the user’s gender, the user’s age, the user’s ID, the set of recommended content clicked by the user, and the recommendation clicked by the user.
  • the characteristics of the user may also include attribute information related to the environment, for example, characteristic information related to the environment in which at least one of the user and other related parties is located.
  • the attribute information related to the environment can be the browser version that displays the candidate content, the category of the terminal device that displays the candidate content (for example, desktop, tablet, smart phone), the model of the terminal device, weather, season, recent Hot events, etc.
  • the target recommendation pool is at least one recommendation pool (for example, one or more recommendation pools)
  • at least one target recommendation pool may be determined among the n-level recommendation pools according to one of the following items for a specific recommendation request: exploration traffic And use traffic usage, random selection method, preset rule selection method, and characteristics of the user who made the recommendation request.
  • exploration traffic and utilization traffic can indicate how much exploration traffic is used and how much utilization traffic is used, which can include the usage of at least one of the past period of time and the current one, or the estimated possible future Usage.
  • the step of selecting candidate materials from the determined at least one target recommendation pool may include: when the target recommendation pool is a single target recommendation pool, randomly selecting from the single target recommendation pool or selecting according to the attention index A predetermined number of candidate materials.
  • the step of selecting candidate materials from the determined at least one target recommendation pool may include: in a case where the target recommendation pool is a plurality of target recommendation pools, selecting from each target in at least one of the following selection methods: The total number of candidate materials selected in the recommended pool is a predetermined number: randomly selected and selected according to the attention indicators.
  • all materials in the n-level recommendation pool may be comprehensively considered to select candidate materials.
  • all n-level recommendation pools may be taken as a whole, and a total of a predetermined number of candidate materials are selected from the n-level recommendation pool according to at least one of the following selection methods: random selection and selection according to attention indicators.
  • the attention indicator refers to the basis for screening out candidate materials that will be recommended next.
  • the attention indicator can include but not limited to the ID, category, title, abstract, keywords, and the new material generated Geographical location, etc.
  • step S40 recommendation is made based on the selected candidate materials.
  • the recommendation can be made for all selected candidate materials, or it can be recommended for only a part of the selected candidate materials.
  • the method can directly recommend materials to users, or recommend materials to users through related parties.
  • the selected candidate materials may be sorted, at least one candidate material suitable for recommendation is selected based on the sorting result, and the selected at least one candidate material is determined according to the display flow rate determined for the at least one target recommendation pool.
  • Materials are recommended.
  • the selected candidate materials may be sorted based on the machine learning model.
  • the recommended materials may be selected based on the sorting results based on at least one of the characteristics of the user who made the recommendation request and the attention indicators of the new materials. At least one candidate material.
  • the above-mentioned machine learning model may be a model that predicts the user's acceptance of candidate materials.
  • the method of training the machine learning model may include data splicing, feature extraction, and model training using preset machine learning algorithms (including parameter tuning and other processes) ) And other steps.
  • the machine learning algorithm used may be various machine learning methods such as neural network, support vector machine, decision tree, logistic regression, etc.
  • the machine learning model can be any type of machine learning model, for example, a logistic regression (LR) model, a support vector machine (SVM), a gradient boosting decision tree or a deep neural network, etc., but is not limited thereto.
  • LR logistic regression
  • SVM support vector machine
  • the exemplary embodiments of the present disclosure do not specifically limit specific machine learning algorithms.
  • other methods such as statistical algorithms can also be combined.
  • step S50 user behavior data corresponding to the recommended material can be obtained.
  • the user behavior data may indicate the user's acceptance of the recommended material.
  • the user behavior data involves at least one operation of clicking, forwarding, sharing, commenting, liking, and liking the recommended candidate material.
  • the user behavior data may be classified into at least one of positive behavior, negative behavior, and neutral behavior according to the meaning it represents.
  • user behavior data can be obtained in real time.
  • user behavior data can be obtained directly from users, or user behavior data can be obtained through other parties.
  • the method further includes a step S60 of allocating materials in each recommended pool.
  • step S60 of allocating materials in each recommended pool described here is not only executed after step S50, and further speaking, there is no strict execution sequence between step S10 and step S60.
  • step S60 is performed in the following manner: the continuously generated new materials are put into the level 1 recommendation pool; for the materials in the level i-1 recommendation pool, the statistics are calculated with the corresponding display flow rate as the unit.
  • Corresponding user behavior data determine whether to remove the material from the i-1 level recommendation pool, keep it in the i-1 level recommendation pool or move it into the i level recommendation pool; and, for the nth level For the materials in the recommendation pool, use the corresponding display flow as the unit to count the user behavior data corresponding to the materials, and based on the statistical results, determine whether to remove the materials from the nth level recommendation pool or keep them in the nth level recommendation Pool, where i is any integer greater than or equal to 2 and less than n.
  • the display traffic in the i-1th level recommendation pool is a predetermined value, for example, the predetermined value can be 300, then when the display traffic of the material reaches the predetermined value (for example, 300), count the users corresponding to the material Behavioral data.
  • the user's acceptance level of the material is obtained, and if the acceptance level is not higher than all acceptance levels in the i-2 level recommendation pool, the material is removed from the i-1 level Recommendation pool; in the case that the acceptance level is not lower than all acceptance levels in the i-th recommendation pool, the material is moved into the i-th recommendation pool; otherwise, the material is retained in the i-1 level recommendation Pool.
  • the user behavior data corresponding to the material is counted.
  • the user's acceptance level of the material is obtained based on the statistical result. In the case that the acceptance level is not higher than all acceptance levels in the n-1th level recommendation pool, the material is removed from the nth level. Level recommendation pool; otherwise, the material is retained in the nth level recommendation pool.
  • the numerical values shown here are only for ease of presentation and do not limit the scope of the concept of the present disclosure.
  • step S60 according to the user behavior data collected in step S50, the actual popularity (ie, acceptance) of the material at the time of recommendation can be measured, and at least one of the following operations can be performed: compare Popular materials are put into the next-level recommendation pool (for the situation from the first recommendation pool to the n-1 level recommendation pool), kept in the current recommendation pool, and the less popular materials are removed.
  • the display flow of a single recommended material in each level recommendation pool is increased from the first level recommendation pool to the nth level recommendation pool, so that only Explore new materials in the primary recommendation pool (for example, the first-level recommendation pool) with less display traffic, and initially determine their quality, and then filter out high-quality content into the next-level recommendation pool (for example, the second-level recommendation pool).
  • Level recommended pool the confidence level (for example, specific display traffic) can be gradually reached in the subsequent recommendation pool. Therefore, it can solve the poor recommendation effect caused by too much new materials generated or insufficient exploration flow. Further, through the specific multi-level recommendation pool, high-quality content can be found faster with less traffic.
  • recommendation algorithms include but are not limited to ⁇ -greedy algorithm (Epsilon-Greedy), Topmpson sampling, UCB algorithm (Upper Confidence Bound), etc.
  • FIG. 2 is a block diagram showing a system 10 for material recommendation according to an exemplary embodiment of the present disclosure.
  • the method shown in FIG. 1 may be executed by the system 10 shown in FIG. 2.
  • the system 10 may be a system that uses an improved discovery-utilization method to perform recommendations.
  • the system 10 may include: a recommendation pool generation unit 210, a display flow determination unit 220, a candidate material selection unit 230, a recommendation unit 240, a data acquisition unit 250, and a screening unit 260.
  • the recommended pool generating unit 210 may determine the number n of recommended pools, where n is an integer greater than or equal to 3. In other words, at least 3 recommended pools are generated.
  • any suitable method may be adopted to determine the number n of recommended pools to obtain n-level recommended pools R1, R2, ..., Rn.
  • the recommended pool generating unit 210 may determine the number n of recommended pools based on at least one of the generation speed of the new material and the flow distribution.
  • the material here may indicate any content or items that may be recommended, for example, blog posts, news information, forum posts, videos, short videos, music, pictures, funny paragraphs, etc.
  • the method of obtaining new materials can include obtaining materials directly/indirectly from the material producer, including but not limited to at least one of the following methods of obtaining: uploading by the author, publishing by the author, reprinting by others, searching the corresponding database, and from related media/media Get etc.
  • the recommendation pool generation unit 210 may estimate the generation speed of new materials by counting the total number of materials generated in a specific period. For example, if the number of new materials generated in a specific time period (for example, one day) is 10,000, n can be determined to be 3 to generate a 3-level recommendation pool; if the number of new materials generated during a specific time period (for example, one day) If the number is 100,000, n can be determined as 4, thereby generating a 4-level recommendation pool. It should be noted that the numerical values here are only examples and do not limit the scope of the present disclosure.
  • the recommended pool generating unit 210 may determine the number n of recommended pools based on the traffic distribution.
  • the flow allocation refers to dividing the flow corresponding to the user request into the exploration flow and the utilization flow in the "exploration-utilization" approach.
  • the "exploration-utilization" approach has been described in detail above, and will not be repeated here.
  • the recommendation pool generation unit 210 may set relatively few recommendation pools, and if the allocated exploration flow is relatively low, the recommendation pool generation unit 210 may set relatively more recommendations. Pool.
  • the first-level recommendation pool to the n-1th-level recommendation pool correspond to the exploration flow in the "exploration-utilization" approach
  • the nth-level recommendation pool corresponds to the utilization flow in the "exploration-utilization” approach.
  • the system 10 may divide the flow corresponding to the user request into the exploration flow and the utilization flow in the "exploration-utilization" mode.
  • the exploration flow may satisfy at least one of the following conditions: Proportional, exploration traffic does not exceed one-tenth of the total traffic including exploration traffic and utilization traffic.
  • the system 10 can combine the above two aspects. For example, suppose that the total flow is 50 million, and the expected amount of new materials generated on that day is 10,000. Then, in this embodiment, the exploration flow can be 3.25 million and the utilization flow can be 46.75 million.
  • the exploration flow can be increased to 3.9 million, and the utilization flow can be increased to 46.1 million, that is, the exploration flow can be proportional to the number of new materials.
  • the exploration flow can be controlled within 5 million by adjusting other conditions, and the utilization flow is 45 million, that is, the exploration flow does not exceed one-tenth of the total flow. .
  • the recommended pool generating unit 210 may also combine the above-described two methods of determining the number of recommended pools based on the new material generation speed and determining the number of recommended pools based on the total flow rate to comprehensively determine the appropriate number of recommended pools.
  • the display flow determination unit 220 may determine the display flow of a single recommended material in each level of recommendation pool, so that the display flow is sequentially increased from the level 1 recommendation pool R1 to the n level recommendation pool Rn.
  • the display flow determination unit 220 may determine the display flow of each level of recommendation pool for a single recommended material, so that in the recommendation pool corresponding to the exploration flow, as the level increases, the corresponding display flow increases, and corresponds to the utilization flow.
  • the recommended pool of has the largest display traffic.
  • the display flow determining unit 220 may determine that the display flow of a single recommended material corresponding to the first-level recommendation pool is 100, and the second-level recommendation pool The display flow corresponding to a single recommended material is 4500, and the display flow of a single recommended material corresponding to the third-level recommendation pool is greater than the display flow of a single recommended material corresponding to the second-level recommendation pool.
  • the display flow determination unit 220 may determine that the exploration flow is 3.25 million and the utilization flow is 46.75 million. In another embodiment, if the number of new materials is increased to 12,000 and other conditions remain unchanged, the display flow determination unit 220 can determine that the exploration flow is 3.9 million and the utilization flow is 46.1 million, that is, the exploration flow can be compared with that of the new material. The quantity is proportional.
  • the flow determination unit 220 can determine that the quantity of new materials in the second-level recommended pool is 5% of the first-level recommended pool, the percentage shown here is only for ease of description and is not limited The technical idea of the present disclosure. In other exemplary embodiments, the display flow determination unit 220 may determine other proportional relationships or not be restricted by specific proportional relationships.
  • the display flow determination unit 220 may determine that the display flow of a single recommended material corresponding to the first recommended pool is 100, and the second recommended pool corresponds to a single recommended material
  • the display flow of the third-level recommendation pool is 4,500 for a single recommended material.
  • the display flow of the fourth-level recommendation pool for a single recommended material can be greater than the display flow of the third-level recommendation pool for a single recommended material.
  • the system 10 can also continuously obtain the total flow corresponding to the user request, and the recommended pool generating unit 210 and the display flow determining unit 220 can be based on changes in the total flow and the new material The generation speed executes the operation again, thereby dynamically adjusting the number n of recommended pools and the display flow rate of a single recommended material in each recommended pool.
  • the recommended pool generating unit 210 dynamically increases the number of recommended pools (for example, increases the first-level recommended pool) , Or with at least one of a decrease in the generation speed of new materials and an increase in the distributed exploration flow, the recommended pool generation unit 210 dynamically reduces the number of recommended pools (for example, reduces the first-level recommended pool).
  • the recommendation pool generation unit 210 may also determine the number n of recommendation pools relatively fixedly based on experience or historically stable data.
  • the recommendation pool generating unit 210 may add one level of recommendation pool, and the display flow determination unit 220 may appropriately adjust the display flow corresponding to each level.
  • the recommendation pool generation unit 210 may adjust the level 3 recommendation pool to the level 4 recommendation pool.
  • the recommended pool generating unit 210 has generated a total of 3 recommended pools, where the display traffic determining unit 220 determines the display of the first-level recommended pool The flow is 100, the display flow of the second-level recommended pool is 4500, and still simply assume that the amount of new materials in the second-level recommended pool is 5% of the first-level recommended pool, then the exploratory traffic generated as expected is 3.25 million , The utilization flow is 46.75 million.
  • the number of new materials actually generated increases, for example, to 20,000, if the exploration traffic doubles to 6.5 million in direct proportion, it will exceed one-tenth of the total traffic and thus affect the user experience.
  • the recommended pool generating unit 210 may adjust the level 3 recommendation pool to the level 4 recommendation pool based on the number of new materials obtained and the total flow, and the display flow determination unit 220 may adjust the level 1
  • the recommended pool corresponds to the display flow rate of a single recommended recommended material adjusted to 100
  • the display flow rate in the second-level recommendation pool can be adjusted to 300
  • the display flow rate in the third-level recommendation pool can be adjusted to 4600.
  • the quantity of new materials in the recommended pool is 5% of the first-level recommended pool
  • the number of new materials in the third-level recommended pool is 5% of the new material in the second-level recommended pool.
  • the total exploration flow required for the four-level recommendation pool is 2.53 million, so as to explore as many new materials as possible with less exploration flow.
  • the specific numerical values here are only examples for making the concept of the present disclosure clear, and do not limit the scope of the present disclosure.
  • the recommended pool generating unit 210 can reduce the recommended pool by one level and the display flow determination unit 220 can appropriately adjust each The impression traffic corresponding to the level.
  • the operations of the execution recommendation pool generating unit 210 and the display flow determining unit 220 may be the same as or similar to the foregoing embodiment, and the repeated descriptions thereof will be omitted here.
  • the total flow rate is unchanged.
  • the total number of recommended pool levels and the display traffic corresponding to each level can be adjusted with reference to the foregoing embodiment. For example, the situation that the real-time total flow is less than the expected total flow of the day is the same or similar to the situation that the number of new materials increases. Therefore, the total number of recommended pool levels and the display flow corresponding to each level can be adjusted with reference to the above embodiments.
  • the candidate material selection unit 230 may select the candidate material CR from the n-level recommendation pool for a specific recommendation request.
  • the candidate material selection unit 230 may determine at least one target recommendation pool among the n-level recommendation pools for a specific recommendation request; and may select candidate materials from the determined at least one target recommendation pool.
  • the target recommendation pool may be a single recommendation pool or multiple recommendation pools.
  • the candidate material selection unit 230 may select from the first-level recommendation pool to the n-1th-level recommendation pool A single target recommendation pool is determined, and in a case where the specific recommendation request corresponds to the utilization flow, the candidate material selection unit 230 may determine the nth level recommendation pool as the target recommendation pool.
  • the candidate material selection unit 230 determines a single target from the first-level recommendation pool to the n-1th-level recommendation pool by at least one of the following methods when the specific recommendation request corresponds to the exploration flow Recommendation pool: randomly select the target recommendation pool, select the target recommendation pool according to the characteristics of the user who made the recommendation request, and select the target recommendation pool according to the preset rules.
  • the candidate material selection unit 230 may select a recommendation pool selected randomly or according to preset rules as the target recommendation pool, or the candidate material selection unit 230 may also combine the user who will recommend the material (that is, the recommendation request User) characteristics to select the target recommendation pool.
  • the characteristics of the user who made the recommendation request may include, but are not limited to, the attributes of the user, such as the user’s gender, the user’s age, the user’s ID, the set of recommended content clicked by the user, and the recommendation clicked by the user.
  • the characteristics of the user may also include attribute information related to the environment, for example, characteristic information related to the environment in which at least one of the user and other related parties is located.
  • the attribute information related to the environment can be the browser version that displays the candidate content, the category of the terminal device that displays the candidate content (for example, desktop, tablet, smart phone), the model of the terminal device, weather, season, recent Hot events, etc.
  • the candidate material selection unit 230 may determine at least one target among the n-level recommendation pools according to one of the following for a specific recommendation request Recommendation pool: Explore the usage of traffic and utilization, random selection method, selection method of preset rules, and characteristics of the user who made the recommendation request.
  • the usage of exploration traffic and utilization traffic can indicate how much exploration traffic is used and how much utilization traffic is used, which can include the usage of at least one of the past period of time and the current one, or the estimated possible future Usage.
  • the candidate material selection unit 230 may select a predetermined number of candidate materials randomly from the single target recommendation pool or select a predetermined number of candidate materials according to the attention index when the target recommendation pool is a single target recommendation pool. In an exemplary embodiment, the candidate material selection unit 230 may select a predetermined number of candidate materials from each target recommendation pool according to at least one of the following selection methods when the target recommendation pool is a plurality of target recommendation pools. : Randomly selected and selected according to indicators of concern. In addition, for a specific recommendation request, the candidate material selection unit 230 may comprehensively consider all materials in the n-level recommendation pool to select candidate materials.
  • the candidate material selection unit 230 may take all n-level recommendation pools as a whole, and select a predetermined number of candidate materials from the n-level recommendation pool in at least one of the following selection methods: random selection and selection according to attention indicators.
  • the attention indicator refers to the basis for screening out candidate materials that will be recommended next.
  • the attention indicator can include but not limited to the ID, category, title, abstract, keywords, and the new material generated Geographical location, etc.
  • the recommendation unit 240 makes recommendations based on the selected candidate materials.
  • the recommendation unit 240 may recommend all selected candidate materials, or the recommendation unit 240 may only recommend a part of the selected candidate materials.
  • the recommendation unit 240 may directly recommend the material to the user, or may recommend the material to the user through a related party (not shown).
  • the recommendation unit 240 may sort the selected candidate materials, select at least one candidate material suitable for recommendation based on the sorting result, and compare the selected candidate materials according to the display flow determined for the at least one target recommendation pool. At least one candidate material is recommended.
  • the recommending unit 240 may sort the selected candidate materials based on a machine learning model.
  • the recommending unit 240 may sort based on at least one of the characteristics of the user who made the recommendation request and the attention index of the new material. As a result, at least one candidate material suitable for recommendation is selected.
  • the above-mentioned machine learning model may be a model that predicts the user's acceptance of candidate materials.
  • the method of training the machine learning model may include data splicing, feature extraction, and model training using preset machine learning algorithms (including parameter tuning and other processes) ) And other steps.
  • the machine learning algorithm used may be various machine learning methods such as neural network, support vector machine, decision tree, logistic regression, etc.
  • the machine learning model can be any type of machine learning model, for example, a logistic regression (LR) model, a support vector machine (SVM), a gradient boosting decision tree or a deep neural network, etc., but is not limited thereto.
  • LR logistic regression
  • SVM support vector machine
  • the exemplary embodiments of the present disclosure do not specifically limit specific machine learning algorithms.
  • other methods such as statistical algorithms can also be combined.
  • the data obtaining unit 250 may obtain user behavior data corresponding to the recommended material.
  • the user behavior data may indicate the user's acceptance of the recommended material.
  • the user behavior data involves at least one operation of clicking, forwarding, sharing, commenting, liking, and liking the recommended candidate material.
  • the user behavior data may be classified into at least one of positive behavior, negative behavior, and neutral behavior according to the meaning it represents.
  • the data acquisition unit 250 will continue to acquire the acceptance level of the user.
  • the data obtaining unit 250 may obtain user behavior data in real time.
  • the data obtaining unit 250 may directly obtain user behavior data from the user, or may obtain user behavior data through other parties.
  • the system 10 also includes a screening unit 260 that allocates materials in each recommended pool.
  • a screening unit 260 that allocates materials in each recommended pool.
  • the candidate material selection unit 230 selects candidate materials and recommends them by the recommendation unit 240, and the data acquisition unit 250 also acquires user behavior data. continuous.
  • the screening unit 260 allocates materials in each recommended pool in the following manner: put the continuously generated new materials into the first-level recommended pool; for the materials in the i-1th recommended pool, the corresponding display flow is The unit counts user behavior data corresponding to the material, and based on the statistical results, determines whether to remove the material from the i-1 level recommendation pool, keep it in the i-1 level recommendation pool or move it into the i level recommendation pool ; And, for the materials in the nth-level recommendation pool, use the corresponding display flow as a unit to count the user behavior data corresponding to the materials, and based on the statistical results, determine to remove the materials from the nth-level recommendation pool Still remain in the nth level recommendation pool, where i is any integer greater than or equal to 2 and less than n.
  • the screening unit 260 may first put it into the level 1 recommendation pool. Assuming that the display flow rate in the i-1th level recommendation pool is a predetermined value, for example, the predetermined value may be 300, then when the display flow rate of the material reaches the predetermined value (for example, 300), the screening unit 260 counts the material The corresponding user behavior data. Further, the screening unit 260 obtains the user's acceptance level of the material based on the statistical result, and if the acceptance level is not higher than all acceptance levels in the i-2th recommendation pool, remove the material from the i-th recommendation pool. -1 recommendation pool; if the acceptance level is not lower than all acceptance levels in the i-th recommendation pool, move the material to the i-th recommendation pool; otherwise, keep the material in the i-th recommendation pool Level 1 recommended pool.
  • the screening unit 260 counts the user behavior data corresponding to the material.
  • the screening unit 260 obtains the user's acceptance level of the material based on the statistical result, and if the acceptance level is not higher than all acceptance levels in the n-1th level recommendation pool, the material Remove the n-th recommended pool; otherwise, keep the material in the n-th recommended pool.
  • the numerical values shown here are only for ease of presentation and do not limit the scope of the concept of the present disclosure.
  • the actual popularity (ie, acceptance) of the material when it is recommended can be measured according to the user behavior data collected by the data acquisition unit 250, and at least one of the following operations can be performed: The more popular materials are put into the next level recommendation pool (for the i-1 level recommendation pool, where 2 ⁇ i ⁇ n), kept in the current recommendation pool, and the less popular materials are removed.
  • the recommendation pool generation unit 210 generates multi-level recommendation pools R1, R2,..., Rn, and the display flow determination unit 220 sets the display flow rate for a single recommended material in each level recommendation pool from The first-level recommendation pool R1 to the nth-level recommendation pool Rn are successively increased, so that the system 10 can explore new materials in the primary recommendation pool (for example, the first-level recommendation pool R1) with less display traffic. And preliminary determine its quality.
  • the system 10 can solve the problem of poor recommendation effect caused by too much new material generated or insufficient exploration flow, and can be found faster with less flow Quality content.
  • FIG. 3 is a schematic diagram illustrating an environment in which the system 10 in FIG. 1 is used for material recommendation according to an exemplary embodiment of the present disclosure. It should be noted that the scenes shown in the drawings are only examples, and must not be used to limit the exemplary embodiments of the present disclosure.
  • the environment shown in FIG. 3 may include a system 10 for material recommendation, a network 20, and user terminals 30 and 40.
  • the user terminal 30 and the user terminal 40 may respectively refer to multiple terminals.
  • the system 10 can be the system 10 described above with reference to Figures 1 to 2.
  • the system 10 can be deployed on the IT facilities of entities such as media and operators that handle material distribution, and can also be deployed on entities that specialize in providing recommendation services. IT facilities.
  • the network 20 may include routers, switches, servers, cloud servers, etc.
  • the user terminals 30 and 40 may include any type of electronic products that can access the network 20, such as cellular phones, smart phones, tablet computers, wearable devices, personal digital assistants (PDA), portable multimedia players (PMP), digital cameras, Music players, portable game consoles, navigation systems, digital TVs, 3D TVs, personal computers (PCs), household appliances, laptop computers, etc.
  • the user terminals 30 and 40 may also be desktop computers, workstation computers, or servers.
  • IP Internet Protocol
  • TCP Transmission Control Protocol
  • UDP User Datagram Protocol
  • RDMA remote direct memory access
  • the user terminal 30 may be a new material provider and upload the new material to at least one of the network 20 and the server in the network 20.
  • the user terminal 40 may send a recommendation request to at least one of the servers in the network 20 and the network 20, and then at least one of the servers in the network 20 and the network 20 forwards the recommendation request to the system 10.
  • the user terminal 40 The recommendation request can be sent directly to the system 10.
  • the system 10 may generate a recommendation list based on the materials stored in at least one of the server in itself, the network 20, and the network 20.
  • the method for generating the recommendation list can be the same as the method described above with reference to FIG. 1 and FIG. 2, and will not be repeated here.
  • the system 10 After the system 10 generates the recommendation list, it can provide the recommended materials to the user terminal 40 through the network 20 directly or via a third party.
  • the units included in the system for material recommendation according to an exemplary embodiment of the present disclosure may be respectively configured as software, hardware, firmware, or any combination of the foregoing items to perform specific functions.
  • these devices can correspond to dedicated integrated circuits, can also correspond to pure software codes, and can also correspond to modules combining software and hardware.
  • one or more functions implemented by these apparatuses may also be uniformly performed by components in physical physical devices (for example, processors, clients, or servers).
  • the computing device may include a storage component and a processor.
  • the storage component stores a set of computer-executable instructions, and when the set of computer-executable instructions is executed by the processor, the method for material recommendation is executed.
  • the computing device can be deployed in a server or a client, and can also be deployed on a node device in a distributed network environment.
  • the computing device may be a PC computer, a tablet device, a personal digital assistant, a smart phone, a web application, or other devices capable of executing the above set of instructions.
  • the computing device does not have to be a single computing device, and may also be any device or a collection of circuits that can execute the above-mentioned instructions (or instruction sets) individually or jointly.
  • the computing device may also be a part of an integrated control system or a system manager, or may be configured as a portable electronic device interconnected with a local or remote (e.g., via wireless transmission) interface.
  • the processor may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor.
  • the processor may also include analog processors, digital processors, microprocessors, multi-core processors, processor arrays, network processors, and the like.
  • Some operations described in the method for material recommendation according to an exemplary embodiment of the present disclosure can be implemented in software, some operations can be implemented in hardware, and in addition, can also be implemented in a combination of software and hardware. These operations.
  • the processor can run instructions or codes stored in one of the storage components, where the storage component can also store data. Instructions and data can also be sent and received via a network via a network interface device, wherein the network interface device can use any known transmission protocol.
  • the storage component can be integrated with the processor, for example, RAM or flash memory is arranged in an integrated circuit microprocessor or the like.
  • the storage component may include an independent device, such as an external disk drive, a storage array, or any other storage device that can be used by a database system.
  • the storage component and the processor may be operatively coupled, or may communicate with each other, for example, through an I/O port, a network connection, or the like, so that the processor can read files stored in the storage component.
  • the computing device may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, mouse, touch input device, etc.). All components of the computing device may be connected to each other via at least one of a bus and a network.
  • a video display such as a liquid crystal display
  • a user interaction interface such as a keyboard, mouse, touch input device, etc.
  • a system including at least one computing device and at least one storage device storing instructions, wherein when the instructions are executed by the at least one computing device, the at least one computing device causes the at least one computing device to perform the above Describe the method used for material recommendation.
  • the method for material recommendation shown in FIG. 1 can be executed by at least one computing device. Since the method for material recommendation has been described in detail in FIGS. 1 and 2 above, the content of this part of this disclosure will not be repeated.
  • the aforementioned system and computing device for material recommendation may be integrated in a server of a content operator, for example, may be integrated in a server of an application program that provides materials.
  • it can also be integrated in a third-party server to obtain a recommendation list that recommends new materials to users, and then the content operator will recommend to users based on the recommendation list (for example, the API provided by the third-party server interface).
  • the method for material recommendation may be implemented by a program recorded on a computer-readable storage medium.
  • a method for storing instructions may be provided.
  • a computer-readable storage medium wherein, when the instructions are executed by at least one computing device, the at least one computing device is caused to execute the method for material recommendation as described above.
  • the computer program in the above-mentioned computer-readable storage medium can be run in an environment deployed in computer equipment such as a client, a host, an agent device, a server, etc. It should be noted that the computer program can also be used to perform additional steps in addition to the above steps. Alternatively, more specific processing is performed when the above steps are performed. The content of these additional steps and further processing has been described with reference to FIG. 1 and FIG. 2, and will not be repeated here in order to avoid repetition.
  • system for material recommendation can completely rely on the operation of a computer program to achieve corresponding functions, that is, the functional architecture of each device and computer program corresponds to each step, so that the entire system can pass A special software package (for example, lib library) is called to realize the corresponding function.
  • a special software package for example, lib library
  • each unit included in the system for material recommendation may also be implemented by hardware, software, firmware, middleware, microcode, or any combination thereof.
  • the program code or code segment used to perform the corresponding operation can be stored in a computer-readable medium such as a storage medium, so that the processor can read and run the corresponding program Code or code segment to perform the corresponding operation.
  • the material recommendation method, system, and computer-readable storage medium provided by the present disclosure can solve the problem of poor recommendation effect caused by too much new material generated or insufficient exploration flow by adding a multi-level recommendation pool. Furthermore, through the multi-level recommendation pool, high-quality content can be found faster with less traffic, the cold-start recommendation algorithm can be optimized, and the screening of high-quality content can be realized more quickly.

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

提供了一种物料推荐方法和系统，所述方法包括：确定推荐池的数量n，其中，n为大于或等于3的整数；确定每一级推荐池中对应单个物料的展示流量，使得展示流量从第1级推荐池到第n-1级推荐池依次递增；针对特定推荐请求，在n级推荐池之中选取候选物料；基于选取的候选物料进行推荐；以及获取与所推荐物料对应的用户行为数据，并且，方法还包括按照以下方式在各个推荐池中调配物料：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池。

Description

物料推荐方法和系统
本申请要求申请号为201910311757.5，申请日为2019年4月18日，名称为“物料推荐方法和系统”的中国专利申请的优先权，其中，上述申请公开的内容通过引用结合在本申请中。
技术领域
本公开总体说来涉及计算机实现的推荐方案，更具体地讲，涉及一种用于物料推荐的方法和系统。
背景技术
随着互联网技术的发展，产生了海量内容，向用户推荐其感兴趣的个性化内容会有助于极大提升用户的使用体验，也为内容提供者带来巨大的收益。通常，内容运营商每天都会产生大量的新物料。对于物料的冷启动可以使用探索-利用(explore&exploit)方法来执行，即，通过探索(explore)新物料，然后利用(exploit)探索收集到的行为数据(即，exploit算法)去计算优质内容，从而实现内容推荐。为确保用户体验，仅采用小部分流量用于探索新物料，其他大量的流量都采用exploit算法对优质内容做推荐。然而，如果新物料太多，则需要将大量流量做探索，这样将影响线上效果或者导致用户流失；另一方面，用于探索的流量太少又会导致新的物料探索不完，一些优质物料无法发现，同样会影响用户体验。
发明内容
本公开的目的在于提供一种用于物料推荐方法和系统，以解决现有技术中因新物料太多或探索流量不够而导致的推荐效果不佳。
本公开提供了一种物料推荐方法，所述方法包括：确定推荐池的数量n，其中，n为大于或等于3的整数；确定每一级推荐池中对应单个物料的展示流量，使得所述展示流量从第1级推荐池到第n-1级推荐池依次递增；针对特定推荐请求，在所述n级推荐池之中选取候选物料；基于选取的候选物料进行推荐；以及获取与所推荐物料对应的用户行为数据，并且，所述方法还包括按照以下方式在各个推荐池中调配物料：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池；并且，针对第n级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池，其中，i为大于或等于2且小于n的任一整数。
本公开提供了一种物料推荐系统，所述系统包括：推荐池生成单元，确定推荐池的数量n，其中，n为大于或等于3的整数；展示流量确定单元，确定每一级推荐池中对应 单个物料的展示流量，使得所述展示流量从第1级推荐池到第n-1级推荐池依次递增；候选物料选取单元，针对特定推荐请求，在所述n级推荐池之中选取候选物料；推荐单元，基于选取的候选物料进行推荐；数据获取单元，获取与所推荐物料对应的用户行为数据；以及筛选单元，按照以下方式在各个推荐池中调配物料：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池；并且，针对第n级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池，其中，i为大于或等于2且小于n的任一整数。
本公开的另一方面提供了一种计算机可读存储介质，其中，当所述计算机指令被至少一个计算装置运行时，促使所述至少一个计算装置执行如上所述的物料推荐方法。
本公开的另一方面提供了一种包括至少一个计算装置和至少一个存储指令的存储装置的系统，其中，所述指令在被所述至少一个计算装置运行时，促使所述至少一个计算装置执行如上所述的物料推荐方法。
根据本公开的一个或多个方面，用于物料推荐方法和系统通过增加多级推荐池，能够解决因产生的新物料太多或探索流量不够而导致的推荐效果不佳的问题。进一步地，通过多级推荐池，可以用更少的流量更快地发现优质内容，优化冷启动推荐算法，更快速地实现对优质内容的筛选。
附图说明
从下面结合附图对本公开实施例的详细描述中，本公开的这些和其他方面和优点将变得更加清楚并更容易理解，其中：
图1是示出根据本公开的示例性实施例的物料推荐的方法的流程图；
图2是示出根据本公开的示例性实施例的用于物料推荐的系统的框图；以及
图3是示出应用根据本公开的示例性实施例的利用图2中的系统进行物料推荐的环境的示意图。
具体实施方式
现将详细参照本公开的实施例，所述实施例的示例在附图中示出，其中，相同的标号始终指的是相同的部件。以下将通过参照附图来说明所述实施例，以便解释本公开。在此需要说明的是，在本公开中出现的“若干项之中的至少一项”均表示包含“该若干项中的任意一项”、“该若干项中的任意多项的组合”、“该若干项的全体”这三类并列的情况。例如“包括A和B之中的至少一个”即包括如下三种并列的情况：(1)包括A；(2)包括B；(3)包括A和B。又例如“执行步骤一和步骤二之中的至少一个”，即表示如下三种并列的情况：(1)执行步骤一；(2)执行步骤二；(3)执行步骤一和步骤二。
图1是示出根据本公开的示例性实施例的物料推荐的方法的流程图。
参照图1，在步骤S10中，可以确定推荐池的数量n，其中，n为大于或等于3的整数。也就是说，至少生成3级推荐池。
根据本公开的示例性实施，可采用任何适当的方式来确定推荐池的数量。作为示例， 可以基于新物料的生成速度和流量分配之中的至少一个来确定推荐池的数量n。这里，物料可指示任何有可能推荐的内容或项目，例如，博客文章、新闻资讯、论坛帖子、视频、短视频、音乐、图片、搞笑段子等。
这里，作为示例，可通过统计特定时段生成的总物料数量来估算新物料的生成速度。例如，若在特定时段(如，一日)产生的新物料数量为1万，则可以将n确定为3，从而生成3级推荐池；若在特定时段(如，一日)产生的新物料数量为10万，则可以将n确定为4，从而生成4级推荐池。需要说明的是，这里的数值仅是举例，并不限制本公开的范围。
根据本公开的示例性实施例，可以基于流量分配来确定推荐池的数量n。这里，流量分配是指将用户请求对应的流量划分为“探索-利用”方式中的探索流量和利用流量。其中，利用(exploit)是对用户比较确定的兴趣，要利用开采迎合；然而，若仅利用用户已知的兴趣使用，用户很快会重复和厌烦，因此需要不断探索(explore)用户对新物料的兴趣。前者可以获得稳定的效果，但是不一定是最优的，后者可能会得到更优的效果，但是也可能得到一个不如以前方法的效果。因此，探索流量与利用流量之间的分配需要根据具体情况进行平衡。一方面，新物料的数量越多，则需要更多的探索流量对新物料进行探索；另一方面，探索流量过多则会影响用户体验。
在实施例中，如果分配的探索流量(即，“探索-利用”方式中的探索流量)比较高，则可设置相对较少的推荐池，而如果分配的探索流量比较低，则可设置相对较多的推荐池。
推荐池中，可将第1级推荐池到第n-1级推荐池对应于“探索-利用”方式中的探索流量，并将第n级推荐池对应于“探索-利用”方式中的利用流量。
在具体的实施例中，探索流量可满足以下条件之中的至少一个：探索流量与新物料的数量成正比、探索流量不超过包括探索流量和利用流量的总流量的十分之一。上述两种方面可结合采用，例如，不妨假设总流量为5000万，当日预期产生的新物料数量为1万，那么，在本实施例中，探索流量可以为325万，利用流量可以为4675万。
在另一实施例中，若新物料数量增加至1.2万且其它条件不变，那么探索流量可以增加至390万，利用流量增加至4610万，即，探索流量可以与新物料的数量成正比。在又一实施例中，若新物料数量增加至2万，那么可以通过调整其它条件将探索流量控制在500万内，利用流量为4500万，即，探索流量不超过总流量的十分之一。
这里，还可将以上描述的基于新物料产生速度确定推荐池数量和基于总流量大小确定推荐池数量的这两种方式进行结合，综合确定适当的推荐池数量。
在步骤S20中，可以确定每一级推荐池中对应单个推荐物料的展示流量，使得所述展示流量从第1级推荐池到第n级推荐池依次递增。
这里，可确定每一级推荐池针对其中单个推荐物料的展示流量，使得与探索流量对应的推荐池中，随着级别的增长，对应的展示流量增长，而与利用流量对应的推荐池所对应的展示流量最大。
例如，在示例实施例中，假设已经确定总共有3级推荐池，第1级推荐池对应单个推荐物料的展示流量可以为100，第2级推荐池对应单个推荐物料的展示流量可以为4500，第3级推荐池对应单个推荐物料的展示流量可以大于第2级推荐池对应单个推荐物料的展示流量。具体说来，总流量若为5000万，当日预期产生的新物料数量为1万，且简单假设第2级推荐池中的新物料数量为第1级推荐池的5％，那么，在本实施例中，探索流量 为325万，利用流量为4675万。在另一实施例中，若新物料数量增加至1.2万且其它条件不变，那么探索流量为390万，利用流量为4610万，即，探索流量可以与新物料的数量成正比。需要说明的是，虽然示例中第2级推荐池中的新物料数量为第1级推荐池的5％，但是，这里示出的百分比仅是为了便于描述，不限制本公开的技术构思。在其它示例实施例中，可以选择其它的比例关系或者不受特定比例关系的限制。
又例如，在总共4级推荐池的示例中，第1级推荐池对应单个推荐物料的展示流量可以为100，第2级推荐池对应单个推荐物料的展示流量可以为200，第3级推荐池对应单个推荐物料的展示流量可以为4500，第4级推荐池对应单个推荐物料的展示流量可以大于第3级推荐池对应单个推荐物料的展示流量。
在示例实施例中，除了新物料的数量之外，所述方法还可以不断地得到用户请求对应的总流量，并根据总流量的变化和新物料的生成速度再次执行步骤S10和步骤S20，从而动态调整推荐池的级数n以及每一级推荐池中对应单个推荐物料的展示流量。例如，在实施例中，随着新物料的生成速度加快和分配的探索流量的减小之中的至少一个而动态地增加推荐池的数量(例如，增加一级推荐池)，或者随着新物料的生成速度降低和分配的探索流量的增大之中的至少一个而动态地减少推荐池的数量(例如，减少一级推荐池)。另外，还可以基于经验或者历史上看比较稳定的数据相对固定地确定推荐池的数量n。
例如，当新物料的数量增大时，可以增加一级推荐池并适当调整每一级对应的展示流量。如上所述的3级推荐池的示例中，可以将3级推荐池调整为4级推荐池。作为示例，总流量若为5000万，当日预期产生的新物料数量为1万，已经确定总共有3级推荐池，其中，第1级推荐池的展示流量可以为100，第2级推荐池的展示流量可以为4500，且仍简单假设第2级推荐池中的新物料数量为第1级推荐池的5％，那么，按照预期产生的探索流量为325万，利用流量为4675万。然而，当实际产生的新物料数增大，例如增大至2万时，若探索流量按照正比增加一倍至650万，则会超过了总流量的十分之一并因此影响用户体验。针对这一问题，在示例实施例中，可在步骤S10，基于获得的新物料数量和总流量，将3级推荐池调整为4级推荐池，并且在步骤S20，可以将第1级推荐池对应单个推荐推荐物料的展示流量调整为100，第2级推荐池中的展示流量可以调整为300，第3级推荐池中的展示流量可以调整为4600，仍简单假设第2级推荐池中的新物料数量为第1级推荐池的5％，第3级推荐池中的新物料数量为第2级推荐池中的新物料数量的5％。通过这样方式，调整后总共4级推荐池所需的探索流量为253万，从而以较少的探索流量探索尽可能多的新物料。需要说明的是，这里的具体数值仅是为使本公开构思清楚而进行的举例，并不限制本公开的范围。在上述实施例中，仅示出了新物料数量增大的情况，相对地，当新物料数量减少时，可以减少一级推荐池并适当调整每一级对应的展示流量。具体的方法可以与上述实施例相同或相似，在这里将省略其重复的描述。
另一方面，在上述实施例均假定总流量不变。相比之下，在总流量发生变化时可以参照上述实施例对推荐池的总级数和每一级对应的展示流量作出调整。例如，实时总流量小于预期总流量的情况与新物料数量增大的情况相同或相似，因此，可以参照上述实施例，对推荐池的总级数和每一级对应的展示流量作出调整。
接下来，在步骤S30中，可以针对特定推荐请求，在所述n级推荐池之中选取候选物料。
在示例实施例中，可以针对特定推荐请求，在所述n级推荐池之中确定至少一个目 标推荐池；以及从确定的至少一个目标推荐池中选取候选物料。这里，目标推荐池可以是单个推荐池，也可以是多个推荐池。
在目标推荐池为单个推荐池的情况下，作为示例，在所述特定推荐请求对应于探索流量的情况下，可从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池，并且，在所述特定推荐请求对应于利用流量的情况下，可将第n级推荐池确定为目标推荐池。具体说来，在所述特定推荐请求对应于探索流量的情况下，通过以下方式之中的至少一种从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池：随机选取目标推荐池、根据与发出推荐请求的用户的特性来选取目标推荐池和根据预设规则选取目标推荐池。通过上述方式，可将随机挑选或根据预设规则挑选的推荐池作为目标推荐池，或者，也可结合将针对其推荐物料的用户(即，推荐请求的用户)的特性来选取目标推荐池。
这里，作为示例，发出推荐请求的用户的特性可以包括但不限于用户自身的属性，例如，用户的性别、用户的年龄、用户的ID、用户点击过的推荐内容的集合、用户点击过的推荐内容发布方的集合等等。另外，所述用户的特性还可以包括与环境有关的属性信息，例如，与用户和其他关联方之中的至少一个所处环境相关的特征信息。作为示例，与环境有关的属性信息可以为展示候选内容的浏览器版本、展示候选内容的终端设备的类别(例如，台式机、平板电脑、智能手机)、终端设备的型号、天气、季节、近期热点事件等。
在目标推荐池为至少一个推荐池(例如，一个或多个推荐池)的情况下，可以针对特定推荐请求根据以下项之一在所述n级推荐池之中确定至少一个目标推荐池：探索流量与利用流量的使用情况、随机选取方式、预设规则选取方式和发出推荐请求的用户的特性。例如，探索流量与利用流量的使用情况可指示分别使用了多少探索流量和多少利用流量，这里既可以包括过去一段时间和当前之中的至少一个的使用情况，也可以包括预估的将来可能的使用情况。
在示例实施例中，从确定的至少一个目标推荐池中选取候选物料的步骤可以包括：在目标推荐池为单个目标推荐池的情况下，从单个目标推荐池中随机选取或按照关注指标来选取预定数量的候选物料。在示例实施例中，从确定的至少一个目标推荐池中选取候选物料的步骤可以包括：在目标推荐池为多个目标推荐池的情况下，按照以下选取方式之中的至少一个分别从各个目标推荐池中选取总数为预定数量的候选物料：随机选取、按照关注指标选取。此外，针对特定推荐请求，可以综合考虑所述n级推荐池之中的所有物料来选取候选物料。例如，可以采用将全部n级推荐池作为整体，按照以下选取方式之中的至少一个从n级推荐池中选取总数为预定数量的候选物料：随机选取、按照关注指标选取。这里，关注指标是指用于筛选出接下来将有可能进行推荐的候选物料的依据，例如，关注指标可以包括但不限于新物料的ID、类别、标题、摘要、关键词、新物料产生的地理位置等。
在所述n级推荐池之中选取候选物料之后，在步骤S40中，基于选取的候选物料进行推荐。这里，可针对全部选取的候选物料进行推荐，或者，也可以仅针对一部分选取的候选物料进行推荐。在进行推荐时，所述方法可直接将物料推荐给用户，也可以通过相关方将物料推荐给用户。
在示例实施例中，可以对选取的候选物料进行排序，基于排序结果选出适合推荐的至少一个候选物料，按照针对所述至少一个目标推荐池所确定的展示流量，对选出的至少 一个候选物料进行推荐。可选地，可以基于机器学习模型对选取的候选物料进行排序，可选地，可根据发出推荐请求的用户的特性和新物料的关注指标之中的至少一个来基于排序结果选出适合推荐的至少一个候选物料。
作为示例，上述机器学习模型可以是预测用户对候选物料的接受情况的模型，训练机器学习模型的方法可以包括数据拼接、特征抽取、利用预设机器学习算法进行模型训练(包括参数调优等过程)等步骤。作为示例，所采用的机器学习算法可以是例如神经网络、支持向量机、决策树、逻辑回归等各种机器学习方法。并且，机器学习模型可以是任何类型的机器学习模型，例如，逻辑回归(LR)模型、支持向量机(SVM)、梯度提升决策树或深度神经网络等，但不限于此。应注意，本公开的示例性实施例对具体的机器学习算法并不进行特定限制。此外，还应注意，在训练和应用模型的过程中，还可结合统计算法等其他手段。
在步骤S50中，可以获取与所推荐物料对应的用户行为数据。这里，所述用户行为数据可表示用户针对其被推荐物料的接受程度。作为示例，所述用户行为数据涉及对推荐的候选物料进行点击、转发、分享、评论、点赞和点踩中的至少一种操作。这里，所述用户行为数据可根据其表示的含义而被划分为正向行为、负向行为和中性行为之中的至少一个。
由此可以看出，对于实际推荐的物料，所述方法会持续获取用户对其的接受程度。这里，可以实时获取用户行为数据。另外，可直接从用户获取用户行为数据，也可通过其他方来获取用户行为数据。
所述方法还包括在各个推荐池中调配物料的步骤S60。这里，应理解，新物料的生成是持续不断发生的，相应地，选取候选物料并对其进行推荐也是持续不断进行的，用户行为数据的获取也是持续的。因此，这里描述的在各个推荐池中调配物料的步骤S60并不仅在步骤S50后执行，进一步地说，步骤S10-步骤S60之间并不具有严格的执行顺序。具体说来，步骤S60按照以下方式执行：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池；并且，针对第n级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池，其中，i为大于或等于2且小于n的任一整数。
作为示例，每次有新的物料引入，可首先将其放入第1级推荐池。假设第i-1级推荐池中的展示流量为预定值，例如，预定值可以是300，那么当所述物料的展示流量达到预定值(例如，300)时，统计与所述物料对应的用户行为数据。进一步地，基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第i-2级推荐池中的所有接受程度的情况下，所述物料被移出第i-1级推荐池；在所述接受程度没有低于第i级推荐池中的所有接受程度的情况下，所述物料被移入第i级推荐池；否则，所述物料被保留在第i-1级推荐池。
在另一示例中，若对应于利用流量的第n级推荐池中的物料的展示流量达到预定值(例如，10000)时，则统计与所述物料对应的用户行为数据。在示例实施例中，基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第n-1级推荐池中的所有接受程度的情况下，所述物料被移出第n级推荐池；否则，所述物料被保留在第n级推 荐池。这里示出的数值仅是为了便于表述，并不限制本公开构思的范围。
综上，通过步骤S60，可以根据在步骤S50中收集的用户行为数据，衡量出物料在被推荐时的实际受欢迎程度(即，接受程度)，并执行以下操作之中的至少一个：将比较受欢迎的物料放入其下一级推荐池(针对第1级推荐池到第n-1级推荐池的情况)、保留在当前推荐池、将比较不受欢迎的物料去除掉。
在本公开提供的示例实施例中，通过使用多级推荐池，并使得每一级推荐池中对应单个推荐物料的展示流量从第1级推荐池到第n级推荐池依次递增，从而可以仅以较少的展示流量对位于初级推荐池(例如，第1级推荐池)中的新物料进行探索，并初步确定其质量，然后将优质内容筛选出来进入下一级推荐池(例如，第2级推荐池)。通过这种方式，可以在随后的推荐池中逐渐达到置信度(例如，特定的展示流量)。因此，可以解决因产生的新物料太多或探索流量不够而导致的推荐效果不佳。进一步地，通过所述特定的多级推荐池，可以用更少的流量更快地发现优质内容。
其外，如上描述的算法可以应用于各种推荐算法，作为示例，推荐算法包括但不限于ε-贪婪算法(Epsilon-Greedy)、Topmpson sampling，UCB算法(Upper Confidence Bound)等。
图2是示出根据本公开的示例性实施例的用于物料推荐的系统10的框图。这里，作为示例，可由图2所示的系统10来执行图1所示的方法。
如图2所示，系统10可以是使用改进的探索-利用方法来执行推荐的系统。系统10可以包括：推荐池生成单元210、展示流量确定单元220、候选物料选取单元230、推荐单元240、数据获取单元250和筛选单元260。
推荐池生成单元210可以确定推荐池的数量n，其中，n为大于或等于3的整数。也就是说，至少生成3级推荐池。
根据本公开的示例性实施，可以采用任何适当的方式来确定推荐池的数量n，以得到n级推荐池R1、R2、……、Rn。作为示例，推荐池生成单元210可以基于新物料的生成速度和流量分配之中的至少一个来确定推荐池的数量n。在示例实施例中，这里，物料可指示任何有可能推荐的内容或项目，例如，博客文章、新闻资讯、论坛帖子、视频、短视频、音乐、图片、搞笑段子等。新物料的获得方式可以包括直接/间接从物料生成方获取物料，包括但不限于以下获得方式之中的至少一个：经由作者上传、作者发布、他人转载、检索对应的数据库、从相关媒体/媒介获取等。
作为示例，推荐池生成单元210可通过统计特定时段生成的总物料数量来估算新物料的生成速度。例如，若在特定时段(如，一日)产生的新物料数量为1万，则可以将n确定为3，从而生成3级推荐池；若在特定时段(如，一日)产生的新物料数量为10万，则可以将n确定为4，从而生成4级推荐池。需要说明的是，这里的数值仅是举例，并不限制本公开的范围。
根据本公开的示例性实施例，推荐池生成单元210可以基于流量分配来确定推荐池的数量n。这里，流量分配是指将用户请求对应的流量划分为“探索-利用”方式中的探索流量和利用流量。关于“探索-利用”方式已经在上文中进行详细描述，在此不再赘述。
在实施例中，如果分配的探索流量比较高，则推荐池生成单元210可设置相对较少的推荐池，而如果分配的探索流量比较低，则推荐池生成单元210可设置相对较多的推荐池。
推荐池中，第1级推荐池到第n-1级推荐池对应于“探索-利用”方式中的探索流量，第n级推荐池对应于“探索-利用”方式中的利用流量。
系统10可将用户请求对应的流量划分为“探索-利用”方式中的探索流量和利用流量，在具体的实施例中，探索流量可满足以下条件之中的至少一个：与新物料的数量成正比、探索流量不超过包括探索流量和利用流量的总流量的十分之一。系统10可结合采用上述两方面。例如，不妨假设总流量为5000万，当日预期产生的新物料数量为1万，那么，在本实施例中，探索流量可以为325万，利用流量可以为4675万。在另一实施例中，若新物料数量增加至1.2万且其它条件不变，那么探索流量可以增加至390万，利用流量增加至4610万，即，探索流量可以与新物料的数量成正比。在又一实施例中，若新物料数量增加至2万，那么可以通过调整其它条件将探索流量控制在500万内，利用流量为4500万，即，探索流量不超过总流量的十分之一。
这里，推荐池生成单元210还可将以上描述的基于新物料产生速度确定推荐池数量和基于总流量大小确定推荐池数量的这两种方式进行结合，综合确定适当的推荐池数量。
展示流量确定单元220可以确定每一级推荐池中对应单个推荐物料的展示流量，使得所述展示流量从第1级推荐池R1到第n级推荐池Rn依次递增。
这里，展示流量确定单元220可确定每一级推荐池针对其中单个推荐物料的展示流量，使得与探索流量对应的推荐池中，随着级别的增长，对应的展示流量增长，而与利用流量对应的推荐池所对应的展示流量最大。
例如，在示例实施例中，假设推荐池生成单元210已经生成总共有3级推荐池，展示流量确定单元220可以确定第1级推荐池对应单个推荐物料的展示流量为100，第2级推荐池对应单个推荐物料的展示流量为4500，第3级推荐池对应单个推荐物料的展示流量大于第2级推荐池对应单个推荐物料的展示流量。具体说来，总流量若为5000万，当日预期产生的新物料数量为1万，且简单假设第2级推荐池中的新物料数量为第1级推荐池的5％，那么，在本实施例中，展示流量确定单元220可以确定探索流量为325万，利用流量为4675万。在另一实施例中，若新物料数量增加至1.2万且其它条件不变，那么展示流量确定单元220可以确定探索流量为390万，利用流量为4610万，即，探索流量可以与新物料的数量成正比。需要说明的是，虽然示例中展示流量确定单元220可以确定第2级推荐池中的新物料数量为第1级推荐池的5％，但是，这里示出的百分比仅是为了便于描述，不限制本公开的技术构思。在其它示例实施例中，展示流量确定单元220可以确定其它的比例关系或者不受特定比例关系的限制。
又例如，在推荐池生成单元210生成总共4级推荐池的示例中，展示流量确定单元220可以确定第1级推荐池对应单个推荐物料的展示流量为100，第2级推荐池对应单个推荐物料的展示流量为200，第3级推荐池对应单个推荐物料的展示流量为4500，第4级推荐池对应单个推荐物料的展示流量可以大于第3级推荐池对应单个推荐物料的展示流量。
在示例实施例中，除了新物料的数量之外，系统10还可以不断地得到用户请求对应的总流量，并且推荐池生成单元210和展示流量确定单元220可以根据总流量的变化和新物料的生成速度再次执行操作，从而动态调整推荐池的级数n以及每一级推荐池中对应单个推荐物料的展示流量。例如，在实施例中，随着新物料的生成速度加快和分配的探索流量的减小之中的至少一个，推荐池生成单元210动态地增加推荐池的数量(例如，增加一 级推荐池)，或者随着新物料的生成速度降低和分配的探索流量的增大之中的至少一个，推荐池生成单元210动态地减少推荐池的数量(例如，减少一级推荐池)。另外，推荐池生成单元210还可以基于经验或者历史上看比较稳定的数据相对固定地确定推荐池的数量n。
例如，当新物料的数量增大时，推荐池生成单元210可以增加一级推荐池，展示流量确定单元220可以适当调整每一级对应的展示流量。如上所述的3级推荐池的示例中，推荐池生成单元210可以将3级推荐池调整为4级推荐池。作为示例，总流量若为5000万，当日预期产生的新物料数量为1万，推荐池生成单元210已经生成总共有3级推荐池，其中，展示流量确定单元220确定第1级推荐池的展示流量为100，第2级推荐池的展示流量为4500，且仍简单假设第2级推荐池中的新物料数量为第1级推荐池的5％，那么，按照预期产生的探索流量为325万，利用流量为4675万。然而，当实际产生的新物料数增大，例如增大至2万时，若探索流量按照正比增加一倍至650万，则会超过了总流量的十分之一并因此影响用户体验。针对这一问题，在示例实施例中，推荐池生成单元210可以基于获得的新物料数量和总流量，将3级推荐池调整为4级推荐池，并且展示流量确定单元220可以将第1级推荐池对应单个推荐推荐物料的展示流量调整为100，第2级推荐池中的展示流量可以调整为300，第3级推荐池中的展示流量可以调整为4600，仍简单假设第2级推荐池中的新物料数量为第1级推荐池的5％，第3级推荐池中的新物料数量为第2级推荐池中的新物料数量的5％。通过这样方式，调整后总共4级推荐池所需的探索流量为253万，从而以较少的探索流量探索尽可能多的新物料。需要说明的是，这里的具体数值仅是为使本公开构思清楚而进行的举例，并不限制本公开的范围。在上述实施例中，仅示出了新物料数量增大的情况，相对地，当新物料数量减少时，推荐池生成单元210可以减少一级推荐池并且展示流量确定单元220可以适当调整每一级对应的展示流量。执行推荐池生成单元210和展示流量确定单元220的操作可以与上述实施例相同或相似，在这里将省略其重复的描述。
另一方面，在上述实施例均假定总流量不变。相比之下，在总流量发生变化时可以参照上述实施例对推荐池的总级数和每一级对应的展示流量作出调整。例如，实时总流量小于当日预期总流量的情况与新物料数量增大的情况相同或相似，因此，可以参照上述实施例，对推荐池的总级数和每一级对应的展示流量作出调整。
接下来，候选物料选取单元230可以针对特定推荐请求在所述n级推荐池之中选取候选物料CR。
在示例实施例中，候选物料选取单元230可以针对特定推荐请求在所述n级推荐池之中确定至少一个目标推荐池；并且可以从确定的至少一个目标推荐池中选取候选物料。这里，目标推荐池可以是单个推荐池，也可以是多个推荐池。
在目标推荐池为单个推荐池的情况下，作为示例，在所述特定推荐请求对应于探索流量的情况下，候选物料选取单元230可从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池，并且，在所述特定推荐请求对应于利用流量的情况下，候选物料选取单元230可将第n级推荐池确定为目标推荐池。具体说来，候选物料选取单元230在所述特定推荐请求对应于探索流量的情况下，通过以下方式之中的至少一种从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池：随机选取目标推荐池、根据与发出推荐请求的用户的特性来选取目标推荐池和根据预设规则选取目标推荐池。通过上述方式，候选物料选 取单元230可将随机挑选或根据预设规则挑选的推荐池作为目标推荐池，或者，候选物料选取单元230也可结合将针对其推荐物料的用户(即，推荐请求的用户)的特性来选取目标推荐池。
这里，作为示例，发出推荐请求的用户的特性可以包括但不限于用户自身的属性，例如，用户的性别、用户的年龄、用户的ID、用户点击过的推荐内容的集合、用户点击过的推荐内容发布方的集合等等。另外，所述用户的特性还可以包括与环境有关的属性信息，例如，与用户和其他关联方之中的至少一个所处环境相关的特征信息。作为示例，与环境有关的属性信息可以为展示候选内容的浏览器版本、展示候选内容的终端设备的类别(例如，台式机、平板电脑、智能手机)、终端设备的型号、天气、季节、近期热点事件等。
在目标推荐池为至少一个推荐池(例如，一个或多个推荐池)的情况下，候选物料选取单元230可以针对特定推荐请求根据以下项之一在所述n级推荐池之中确定至少一个目标推荐池：探索流量与利用流量的使用情况、随机选取方式、预设规则选取方式和发出推荐请求的用户的特性。例如，探索流量与利用流量的使用情况可指示分别使用了多少探索流量和多少利用流量，这里既可以包括过去一段时间和当前之中的至少一个的使用情况，也可以包括预估的将来可能的使用情况。
在示例实施例中，候选物料选取单元230可在目标推荐池为单个目标推荐池的情况下，从单个目标推荐池中随机选取或按照关注指标来选取预定数量的候选物料。在示例实施例中，候选物料选取单元230可在目标推荐池为多个目标推荐池的情况下，按照以下选取方式之中的至少一个分别从各个目标推荐池中选取总数为预定数量的候选物料：随机选取、按照关注指标选取。此外，针对特定推荐请求，候选物料选取单元230可以综合考虑所述n级推荐池之中的所有物料来选取候选物料。例如，候选物料选取单元230可以采用将全部n级推荐池作为整体，按照以下选取方式之中的至少一个从n级推荐池中选取总数为预定数量的候选物料：随机选取、按照关注指标选取。这里，关注指标是指用于筛选出接下来将有可能进行推荐的候选物料的依据，例如，关注指标可以包括但不限于新物料的ID、类别、标题、摘要、关键词、新物料产生的地理位置等。
在候选物料选取单元230从所述n级推荐池之中选取候选物料之后，推荐单元240基于选取的候选物料进行推荐。这里，推荐单元240可针对全部选取的候选物料进行推荐，或者，推荐单元240也可以仅针对一部分选取的候选物料进行推荐。在进行推荐时，推荐单元240可直接将物料推荐给用户，也可以通过相关方(未示出)将物料推荐给用户。
在示例实施例中，推荐单元240可以对选取的候选物料进行排序，基于排序结果选出适合推荐的至少一个候选物料，按照针对所述至少一个目标推荐池所确定的展示流量，对选出的至少一个候选物料进行推荐。可选地，推荐单元240可以基于机器学习模型对选取的候选物料进行排序，可选地，推荐单元240可根据发出推荐请求的用户的特性和新物料的关注指标之中的至少一个来基于排序结果选出适合推荐的至少一个候选物料。
作为示例，上述机器学习模型可以是预测用户对候选物料的接受情况的模型，训练机器学习模型的方法可以包括数据拼接、特征抽取、利用预设机器学习算法进行模型训练(包括参数调优等过程)等步骤。作为示例，所采用的机器学习算法可以是例如神经网络、支持向量机、决策树、逻辑回归等各种机器学习方法。并且，机器学习模型可以是任何类型的机器学习模型，例如，逻辑回归(LR)模型、支持向量机(SVM)、梯度提升决策 树或深度神经网络等，但不限于此。应注意，本公开的示例性实施例对具体的机器学习算法并不进行特定限制。此外，还应注意，在训练和应用模型的过程中，还可结合统计算法等其他手段。
数据获取单元250可以获取与所推荐物料对应的用户行为数据。这里，所述用户行为数据可表示用户针对其被推荐物料的接受程度。作为示例，所述用户行为数据涉及对推荐的候选物料进行点击、转发、分享、评论、点赞和点踩中的至少一种操作。这里，所述用户行为数据可根据其表示的含义而被划分为正向行为、负向行为和中性行为之中的至少一个。
由此可以看出，对于实际推荐的物料，数据获取单元250会持续获取用户对其的接受程度。这里，数据获取单元250可以实时获取用户行为数据。另外，数据获取单元250可直接从用户获取用户行为数据，也可通过其他方来获取用户行为数据。
系统10还包括在各个推荐池中调配物料的筛选单元260。这里，应理解，新物料的生成是持续不断发生的，相应地，候选物料选取单元230选取候选物料并由推荐单元240对其进行推荐也是持续不断进行的，数据获取单元250获取用户行为数据也是持续的。具体说来，筛选单元260按照以下方式在各个推荐池中调配物料：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池；并且，针对第n级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池，其中，i为大于或等于2且小于n的任一整数。
作为示例，每次有新的物料引入，筛选单元260可首先将其放入第1级推荐池。假设第i-1级推荐池中的展示流量为预定值，例如，预定值可以是300，那么当所述物料的展示流量达到预定值(例如，300)时，筛选单元260统计与所述物料对应的用户行为数据。进一步地，筛选单元260基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第i-2级推荐池中的所有接受程度的情况下，将所述物料移出第i-1级推荐池；在所述接受程度没有低于第i级推荐池中的所有接受程度的情况下，将所述物料移入第i级推荐池；否则，将所述物料保留在第i-1级推荐池。
在另一示例中，若对应于利用流量的第n级推荐池中的物料的展示流量达到预定值(例如，10000)时，则筛选单元260统计与所述物料对应的用户行为数据。在示例实施例中，筛选单元260基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第n-1级推荐池中的所有接受程度的情况下，将所述物料移出第n级推荐池；否则，将所述物料保留在第n级推荐池。这里示出的数值仅是为了便于表述，并不限制本公开构思的范围。
综上，通过筛选单元260，可以根据数据获取单元250收集的用户行为数据，衡量出物料在被推荐时的实际受欢迎程度(即，接受程度)，并执行以下操作之中的至少一个：将比较受欢迎的物料放入其下一级推荐池(针对第i-1级推荐池的情况，其中2≤i≤n)、保留在当前推荐池、将比较不受欢迎的物料去除掉。
在本公开提供的示例实施例中，推荐池生成单元210生成多级推荐池R1、R2、……、Rn，展示流量确定单元220设定每一级推荐池中针对单个推荐物料的展示流量从第1级推 荐池R1到第n级推荐池Rn依次递增，从而使得系统10可以仅以较少的展示流量对位于初级推荐池(例如，第1级推荐池R1)中的新物料进行探索，并初步确定其质量。基于与参照图1描述的方法类似的本公开构思，系统10可以解决因产生的新物料太多或探索流量不够而导致的推荐效果不佳的问题，并且可以用更少的流量更快地发现优质内容。
图3是示出应用根据本公开的示例性实施例的利用图1中的系统10进行物料推荐的环境的示意图。应注意该附图中示出的场景仅为示例，而绝不可用于限制本公开的示例性实施例。
图3所示的环境可以包括用于物料推荐的系统10、网络20以及用户终端30和40。这里，应注意，用户终端30和用户终端40分别可指代多个终端。
其中，系统10可以是上面参照图1至图2描述的系统10，该系统10可以部署在经营物料分发的媒体、运营商等实体的IT设施上，也可部署在专门提供推荐服务的实体的IT设施上。网络20可以包括路由、交换机、服务器、云服务器等。用户终端30和40可以包括可以访问网络20的任何类型的电子产品，诸如蜂窝电话、智能电话、平板计算机、可穿戴装置、个人数字助理(PDA)、便携式多媒体播放器(PMP)、数字相机、音乐播放器、便携式游戏控制台、导航系统、数字电视、3D电视、个人计算机(PC)、家用电器、膝上型计算机等。用户终端30和40还可以是台式计算机、工作站计算机或服务器。用户终端30和40通过以太网协议、基于互联网协议(IP)的协议、基于传输控制协议(TCP)的协议、基于用户数据报协议(UDP)的协议、基于远程直接内存访问(RDMA)协议的协议以及基于NVMe-oF协议的协议或它们的组合来访问网络20和网络20中的服务器之中的至少一个。
进一步地将，用户终端30可以是新物料提供方，并将新物料上传至网络20和网络20中的服务器之中的至少一个。用户终端40可以向网络20和网络20中的服务器之中的至少一个发出推荐请求，随后网络20和网络20中的服务器之中的至少一个向系统10转发推荐请求，可选地，用户终端40可以直接向系统10发出推荐请求。
在系统10收到推荐请求后，可以基于存储于自身内部、网络20和网络20中的服务器之中的至少一个中的物料来生成推荐列表。生成推荐列表的方法可以与上面参照图1和图2描述的方法相同，在此不再赘述。
系统10生成推荐列表之后，可直接或经由第三方，通过网络20向用户终端40提供推荐的物料。
根据本公开示例性实施例的用于物料推荐的系统所包括的各单元可被分别配置为执行特定功能的软件、硬件、固件或上述项的任意组合。例如，这些装置可对应于专用的集成电路，也可对应于纯粹的软件代码，还可对应于软件与硬件相结合的模块。此外，这些装置所实现的一个或多个功能也可由物理实体设备(例如，处理器、客户端或服务器等)中的组件来统一执行。
在本公开示例性实施例中还提出一种用于物料推荐的计算装置。例如，该计算装置可包括存储部件和处理器，存储部件中存储有计算机可执行指令集合，当所述计算机可执行指令集合被所述处理器执行时，执行用于物料推荐的方法。
所述计算装置可以部署在服务器或客户端中，也可以部署在分布式网络环境中的节点装置上。此外，所述计算装置可以是PC计算机、平板装置、个人数字助理、智能手机、web应用或其他能够执行上述指令集合的装置。
这里，所述计算装置并非必须是单个的计算装置，还可以是任何能够单独或联合执行上述指令(或指令集)的装置或电路的集合体。计算装置还可以是集成控制系统或系统管理器的一部分，或者可被配置为与本地或远程(例如，经由无线传输)以接口互联的便携式电子装置。
在所述计算装置中，处理器可包括中央处理器(CPU)、图形处理器(GPU)、可编程逻辑装置、专用处理器系统、微控制器或微处理器。作为示例而非限制，处理器还可包括模拟处理器、数字处理器、微处理器、多核处理器、处理器阵列、网络处理器等。
根据本公开示例性实施例的用于物料推荐的方法中所描述的某些操作可通过软件方式来实现，某些操作可通过硬件方式来实现，此外，还可通过软硬件结合的方式来实现这些操作。
处理器可运行存储在存储部件之一中的指令或代码，其中，所述存储部件还可以存储数据。指令和数据还可经由网络接口装置而通过网络被发送和接收，其中，所述网络接口装置可采用任何已知的传输协议。
存储部件可与处理器集成为一体，例如，将RAM或闪存布置在集成电路微处理器等之内。此外，存储部件可包括独立的装置，诸如，外部盘驱动、存储阵列或任何数据库系统可使用的其他存储装置。存储部件和处理器可在操作上进行耦合，或者可例如通过I/O端口、网络连接等互相通信，使得处理器能够读取存储在存储部件中的文件。
此外，所述计算装置还可包括视频显示器(诸如，液晶显示器)和用户交互接口(诸如，键盘、鼠标、触摸输入装置等)。计算装置的所有组件可经由总线和网络之中的至少一个而彼此连接。
根据本公开示例性实施例的用于物料推荐的方法所涉及的操作可被描述为各种互联或耦合的功能块或功能示图。然而，这些功能块或功能示图可被均等地集成为单个的逻辑装置或按照非确切的边界进行操作。
例如，如上所述，提供一种包括至少一个计算装置和至少一个存储指令的存储装置的系统，其中，所述指令在被所述至少一个计算装置运行时，促使所述至少一个计算装置执行如上描述的用于物料推荐的方法。
也就是说，可由至少一个计算装置来执行图1所示的用于物料推荐的方法。由于上述在图1和图2中已经对用于物料推荐的方法进行了详细介绍，本公开对此部分的内容不再赘述。
可选地，上述用于物料推荐的系统和计算装置可被集成在内容运营方的服务器中，例如，可被集成在提供物料的应用程序的服务器中。除此之外，也可以可被集成在第三方服务器中以获得将新物料推荐给用户的推荐列表，再由内容运营方来根据推荐列表向用户进行推荐(例如，由第三方服务器提供的API接口)。
应理解，根据本公开示例性实施例的用于物料推荐的方法可通过记录在计算可读存储介质上的程序来实现，例如，根据本公开的示例性实施例，可提供一种存储指令的计算机可读存储介质，其中，当所述指令被至少一个计算装置运行时，促使所述至少一个计算装置执行如上描述的用于物料推荐的方法。
上述计算机可读存储介质中的计算机程序可在诸如客户端、主机、代理装置、服务器等计算机设备中部署的环境中运行，应注意，所述计算机程序还可用于执行除了上述步骤以外的附加步骤或者在执行上述步骤时执行更为具体的处理，这些附加步骤和进一步处 理的内容已经参照图1和图2进行了描述，这里为了避免重复将不再进行赘述。
应注意，根据本公开示例性实施例的用于物料推荐的系统可完全依赖计算机程序的运行来实现相应的功能，即，各个装置与计算机程序的功能架构中与各步骤相应，使得整个系统通过专门的软件包(例如，lib库)而被调用，以实现相应的功能。
另一方面，根据本公开示例性实施例的用于物料推荐的系统所包括的各个单元也可以通过硬件、软件、固件、中间件、微代码或其任意组合来实现。当以软件、固件、中间件或微代码实现时，用于执行相应操作的程序代码或者代码段可以存储在诸如存储介质的计算机可读介质中，使得处理器可通过读取并运行相应的程序代码或者代码段来执行相应的操作。
以上描述了本公开的各示例性实施例，应理解，上述描述仅是示例性的，并非穷尽性的，本公开不限于所披露的各示例性实施例。在不偏离本公开的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。因此，本公开的保护范围应该以权利要求的范围为准。
工业实用性
本公开提供的物料推荐方法、系统、计算机可读存储介质通过增加多级推荐池，能够解决因产生的新物料太多或探索流量不够而导致的推荐效果不佳的问题。进一步地，通过多级推荐池，可以用更少的流量更快地发现优质内容，优化冷启动推荐算法，更快速地实现对优质内容的筛选。

Claims (38)

  1. 一种包括至少一个计算装置和至少一个存储指令的存储装置的系统，其中，所述指令在被所述至少一个计算装置运行时，促使所述至少一个计算装置执行用于物料推荐的以下步骤，包括：
    确定推荐池的数量n，其中，n为大于或等于3的整数；
    确定每一级推荐池中对应单个物料的展示流量，使得所述展示流量从第1级推荐池到第n-1级推荐池依次递增；
    针对特定推荐请求，在所述n级推荐池之中选取候选物料；
    基于选取的候选物料进行推荐；以及
    获取与所推荐物料对应的用户行为数据，
    并且，所述指令在被所述至少一个计算装置运行时，促使所述至少一个计算装置还执行以下步骤：按照以下方式在各个推荐池中调配物料：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池；并且，针对第n级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池，其中，i为大于或等于2且小于n的任一整数。
  2. 如权利要求1所述的系统，其中，确定推荐池的数量n的步骤包括：基于新物料的生成速度和流量分配之中的至少一个来确定推荐池的数量n，其中，第1级推荐池到第n-1级推荐池对应于“探索-利用”方式中的探索流量，第n级推荐池对应于“探索-利用”方式中的利用流量。
  3. 如权利要求2所述的系统，其中，针对特定推荐请求，在所述n级推荐池之中选取候选物料的步骤包括：针对特定推荐请求，在所述n级推荐池之中确定至少一个目标推荐池；以及从确定的至少一个目标推荐池中选取候选物料。
  4. 如权利要求2所述的系统，其中，针对特定推荐请求，在所述n级推荐池之中选取候选物料的步骤包括：针对特定推荐请求，综合考虑所述n级推荐池之中的所有物料来选取候选物料。
  5. 如权利要求1所述的系统，其中，基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池的步骤包括：
    基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第i-2级推荐池中的所有接受程度的情况下，所述物料被移出第i-1级推荐池；在所述接受程度没有低于第i级推荐池中的所有接受程度的情况下，所述物料被移入第i级推荐池；否则，所述物料被保留在第i-1级推荐池。
  6. 如权利要求1所述的系统，其中，基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池的步骤包括：
    基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第n-1级推荐池中的所有接受程度的情况下，所述物料被移出第n级推荐池；否则，所述物料被保 留在第n级推荐池。
  7. 如权利要求2所述的系统，其中，基于新物料的生成速度和流量分配之中的至少一个来确定推荐池的数量n的步骤包括：基于新物料的生成速度和流量分配之中的至少一个来动态地确定推荐池的数量n。
  8. 如权利要求3所述的系统，其中，针对特定推荐请求，在所述n级推荐池之中确定至少一个目标推荐池的步骤包括：
    在所述特定推荐请求对应于探索流量的情况下，从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池，并且，
    在所述特定推荐请求对应于利用流量的情况下，将第n级推荐池确定为目标推荐池。
  9. 如权利要求8所述的系统，其中，在所述特定推荐请求对应于探索流量的情况下，通过以下方式之中的至少一种从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池：随机选取目标推荐池、根据与发出推荐请求的用户的特性来选取目标推荐池和根据预设规则选取目标推荐池。
  10. 如权利要求3所述的系统，其中，针对特定推荐请求，在所述n级推荐池之中确定至少一个目标推荐池的步骤包括：针对特定推荐请求，根据以下项之一在所述n级推荐池之中确定至少一个目标推荐池：探索流量与利用流量的使用情况、随机选取方式、预设规则选取方式和发出推荐请求的用户的特性。
  11. 如权利要求3所述的系统，其中，从确定的至少一个目标推荐池中选取候选物料的步骤包括：
    在目标推荐池为单个目标推荐池的情况下，从单个目标推荐池中随机选取或按照关注指标来选取预定数量的候选物料。
  12. 如权利要求3所述的系统，其中，从确定的至少一个目标推荐池中选取候选物料的步骤包括：
    在目标推荐池为多个目标推荐池的情况下，按照以下选取方式之中的至少一个分别从各个目标推荐池中选取总数为预定数量的候选物料：随机选取、按照关注指标选取。
  13. 如权利要求1所述的系统，其中，基于选取的候选物料进行推荐包括：
    对选取的候选物料进行排序，基于排序结果选出适合推荐的至少一个候选物料，以进行推荐。
  14. 如权利要求13所述的系统，其中，基于机器学习模型对选取的候选物料进行排序。
  15. 如权利要求13所述的系统，其中，根据发出推荐请求的用户的特性和新物料的关注指标之中的至少一个来基于排序结果选出适合推荐的至少一个候选物料。
  16. 如权利要求2所述的系统，其中，探索流量满足以下条件之中的至少一个：
    探索流量与新物料的数量成正比、探索流量不超过总流量的十分之一。
  17. 如权利要求1所述的系统，其中，所述用户行为数据涉及对推荐的候选物料进行点击、转发、分享、评论、点赞和点踩中的至少一种操作。
  18. 如权利要求1所述的系统，其中，所述物料包括以下项之中的至少一项：博客文章、新闻资讯、论坛帖子、视频、短视频、音乐、图片、搞笑段子。
  19. 一种物料推荐系统，所述系统包括：
    推荐池生成单元，确定推荐池的数量n，其中，n为大于或等于3的整数；
    展示流量确定单元，确定每一级推荐池中对应单个物料的展示流量，使得所述展示流量从第1级推荐池到第n-1级推荐池依次递增；
    候选物料选取单元，针对特定推荐请求，在所述n级推荐池之中选取候选物料；
    推荐单元，基于选取的候选物料进行推荐；
    数据获取单元，获取与所推荐物料对应的用户行为数据；以及
    筛选单元，按照以下方式在各个推荐池中调配物料：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池；并且，针对第n级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池，其中，i为大于或等于2且小于n的任一整数。
  20. 一种由至少一个计算装置执行的物料推荐方法，包括：
    确定推荐池的数量n，其中，n为大于或等于3的整数；
    确定每一级推荐池中对应单个物料的展示流量，使得所述展示流量从第1级推荐池到第n-1级推荐池依次递增；
    针对特定推荐请求，在所述n级推荐池之中选取候选物料；
    基于选取的候选物料进行推荐；以及
    获取与所推荐物料对应的用户行为数据，
    并且，所述方法还包括按照以下方式在各个推荐池中调配物料：将持续产生的新物料放入第1级推荐池；针对第i-1级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池；并且，针对第n级推荐池中的物料，以对应的展示流量为单位来统计与所述物料对应的用户行为数据，并基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池，其中，i为大于或等于2且小于n的任一整数。
  21. 如权利要求20所述的物料推荐方法，其中，确定推荐池的数量n的步骤包括：基于新物料的生成速度和流量分配之中的至少一个来确定推荐池的数量n，其中，第1级推荐池到第n-1级推荐池对应于“探索-利用”方式中的探索流量，第n级推荐池对应于“探索-利用”方式中的利用流量。
  22. 如权利要求21所述的物料推荐方法，其中，针对特定推荐请求，在所述n级推荐池之中选取候选物料的步骤包括：针对特定推荐请求，在所述n级推荐池之中确定至少一个目标推荐池；以及从确定的至少一个目标推荐池中选取候选物料。
  23. 如权利要求21所述的物料推荐方法，其中，针对特定推荐请求，在所述n级推荐池之中选取候选物料的步骤包括：针对特定推荐请求，综合考虑所述n级推荐池之中的所有物料来选取候选物料。
  24. 如权利要求20所述的物料推荐方法，其中，基于统计结果来确定将所述物料从第i-1级推荐池中移出、保留在第i-1级推荐池还是移入第i级推荐池的步骤包括：
    基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第i-2级推荐池中的所有接受程度的情况下，所述物料被移出第i-1级推荐池；在所述接受程度没有 低于第i级推荐池中的所有接受程度的情况下，所述物料被移入第i级推荐池；否则，所述物料被保留在第i-1级推荐池。
  25. 如权利要求20所述的物料推荐方法，其中，基于统计结果来确定将所述物料从第n级推荐池中移出还是保留在第n级推荐池的步骤包括：
    基于统计结果得到用户对所述物料的接受程度，在所述接受程度没有高于第n-1级推荐池中的所有接受程度的情况下，所述物料被移出第n级推荐池；否则，所述物料被保留在第n级推荐池。
  26. 如权利要求21所述的物料推荐方法，其中，基于新物料的生成速度和流量分配之中的至少一个来确定推荐池的数量n的步骤包括：基于新物料的生成速度和流量分配之中的至少一个来动态地确定推荐池的数量n。
  27. 如权利要求22所述的物料推荐方法，其中，针对特定推荐请求，在所述n级推荐池之中确定至少一个目标推荐池的步骤包括：
    在所述特定推荐请求对应于探索流量的情况下，从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池，并且，
    在所述特定推荐请求对应于利用流量的情况下，将第n级推荐池确定为目标推荐池。
  28. 如权利要求27所述的物料推荐方法，其中，在所述特定推荐请求对应于探索流量的情况下，通过以下方式之中的至少一种从第1级推荐池到第n-1级推荐池之中确定单个目标推荐池：随机选取目标推荐池、根据与发出推荐请求的用户的特性来选取目标推荐池和根据预设规则选取目标推荐池。
  29. 如权利要求22所述的物料推荐方法，其中，针对特定推荐请求，在所述n级推荐池之中确定至少一个目标推荐池的步骤包括：针对特定推荐请求，根据以下项之一在所述n级推荐池之中确定至少一个目标推荐池：探索流量与利用流量的使用情况、随机选取方式、预设规则选取方式和发出推荐请求的用户的特性。
  30. 如权利要求22所述的物料推荐方法，其中，从确定的至少一个目标推荐池中选取候选物料的步骤包括：
    在目标推荐池为单个目标推荐池的情况下，从单个目标推荐池中随机选取或按照关注指标来选取预定数量的候选物料。
  31. 如权利要求22所述的物料推荐方法，其中，从确定的至少一个目标推荐池中选取候选物料的步骤包括：
    在目标推荐池为多个目标推荐池的情况下，按照以下选取方式之中的至少一个分别从各个目标推荐池中选取总数为预定数量的候选物料：随机选取、按照关注指标选取。
  32. 如权利要求20所述的物料推荐方法，其中，基于选取的候选物料进行推荐包括：
    对选取的候选物料进行排序，基于排序结果选出适合推荐的至少一个候选物料，以进行推荐。
  33. 如权利要求32所述的物料推荐方法，其中，基于机器学习模型对选取的候选物料进行排序。
  34. 如权利要求32所述的物料推荐方法，其中，根据发出推荐请求的用户的特性和新物料的关注指标之中的至少一个来基于排序结果选出适合推荐的至少一个候选物料。
  35. 如权利要求21所述的物料推荐方法，其中，探索流量满足以下条件之中的至少一个：探索流量与新物料的数量成正比、探索流量不超过总流量的十分之一。
  36. 如权利要求20所述的物料推荐方法，其中，所述用户行为数据涉及对推荐的候选物料进行点击、转发、分享、评论、点赞和点踩中的至少一种操作。
  37. 如权利要求20所述的物料推荐方法，其中，所述物料包括以下项之中的至少一项：博客文章、新闻资讯、论坛帖子、视频、短视频、音乐、图片、搞笑段子。
  38. 一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机指令，当所述计算机指令被至少一个计算装置运行时，促使所述至少一个计算装置执行如权利要求20至37中任一项所述的物料推荐方法。
PCT/CN2020/085347 2019-04-18 2020-04-17 物料推荐方法和系统 WO2020211840A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910311757.5 2019-04-18
CN201910311757.5A CN111831891B (zh) 2019-04-18 2019-04-18 物料推荐方法和系统

Publications (1)

Publication Number Publication Date
WO2020211840A1 true WO2020211840A1 (zh) 2020-10-22

Family

ID=72837033

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/085347 WO2020211840A1 (zh) 2019-04-18 2020-04-17 物料推荐方法和系统

Country Status (2)

Country Link
CN (1) CN111831891B (zh)
WO (1) WO2020211840A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112435091A (zh) * 2020-11-23 2021-03-02 百果园技术(新加坡)有限公司 推荐内容的选择方法、装置、设备及存储介质
CN115017340A (zh) * 2022-04-29 2022-09-06 北京达佳互联信息技术有限公司 多媒体资源生成方法、装置、电子设备及存储介质

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113570421A (zh) * 2021-09-23 2021-10-29 枣庄职业学院 一种基于大数据分析的电商营销系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140123179A1 (en) * 2012-10-30 2014-05-01 Haohong Wang Zoomable content recommendation system
CN103810030A (zh) * 2014-02-20 2014-05-21 北京奇虎科技有限公司 一种基于移动终端应用市场的应用推荐方法、装置及系统
CN108809751A (zh) * 2018-04-23 2018-11-13 微梦创科网络科技（中国）有限公司 一种策略测试方法及装置
CN109284445A (zh) * 2018-12-11 2019-01-29 北京达佳互联信息技术有限公司 网络资源的推荐方法、装置、服务器及存储介质
CN109614301A (zh) * 2018-11-19 2019-04-12 微梦创科网络科技（中国）有限公司 一种信息的评估方法和装置

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7479333B2 (en) * 2004-12-13 2009-01-20 Hyteon, Inc. Fuel cell stack with multiple groups of cells and flow passes
CN104156483A (zh) * 2014-08-26 2014-11-19 天津大学 基于蚁群算法的地理信息多目标帕累托解的可视化方法
CN106897912B (zh) * 2017-01-18 2020-08-28 太仓中科信息技术研究院 基于epsilon反馈算法的推荐系统冷启动方法
CN108520034B (zh) * 2018-03-29 2021-01-05 阿里巴巴（中国）有限公司 应用推荐方法、装置及计算机设备
CN108694647B (zh) * 2018-05-11 2021-04-23 北京三快在线科技有限公司 一种商户推荐理由的挖掘方法及装置，电子设备

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140123179A1 (en) * 2012-10-30 2014-05-01 Haohong Wang Zoomable content recommendation system
CN103810030A (zh) * 2014-02-20 2014-05-21 北京奇虎科技有限公司 一种基于移动终端应用市场的应用推荐方法、装置及系统
CN108809751A (zh) * 2018-04-23 2018-11-13 微梦创科网络科技（中国）有限公司 一种策略测试方法及装置
CN109614301A (zh) * 2018-11-19 2019-04-12 微梦创科网络科技（中国）有限公司 一种信息的评估方法和装置
CN109284445A (zh) * 2018-12-11 2019-01-29 北京达佳互联信息技术有限公司 网络资源的推荐方法、装置、服务器及存储介质

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112435091A (zh) * 2020-11-23 2021-03-02 百果园技术(新加坡)有限公司 推荐内容的选择方法、装置、设备及存储介质
CN112435091B (zh) * 2020-11-23 2024-03-29 百果园技术(新加坡)有限公司 推荐内容的选择方法、装置、设备及存储介质
CN115017340A (zh) * 2022-04-29 2022-09-06 北京达佳互联信息技术有限公司 多媒体资源生成方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN111831891A (zh) 2020-10-27
CN111831891B (zh) 2024-08-13

Similar Documents

Publication Publication Date Title
US10397359B2 (en) Streaming media cache for media streaming service
US11386137B2 (en) Dynamic feedback in a recommendation system
TWI702844B (zh) 用戶特徵的生成方法、裝置、設備及儲存介質
RU2720899C2 (ru) Способ и система для определения зависящих от пользователя пропорций содержимого для рекомендации
US10783151B1 (en) Popularity-based content feed management system
CN106528813B (zh) 一种多媒体推荐方法和装置
WO2020211840A1 (zh) 物料推荐方法和系统
US10904599B2 (en) Predicting digital personas for digital-content recommendations using a machine-learning-based persona classifier
WO2020011068A1 (zh) 用于执行机器学习过程的方法和系统
US10380649B2 (en) System and method for logistic matrix factorization of implicit feedback data, and application to media environments
CN109062919B (zh) 一种基于深度强化学习的内容推荐方法及装置
US10311362B1 (en) Identification of trending content using social network activity and user interests
WO2020238502A1 (zh) 物品推荐方法及装置、电子设备及存储介质
CN110413867B (zh) 用于内容推荐的方法及系统
CN112989146B (zh) 向目标用户推荐资源的方法、装置、设备、介质和程序产品
US11914651B2 (en) Classification and selective filtering of a data set
US9530024B2 (en) Recommendation system for protecting user privacy
EP3326084A1 (en) Systems, apparatuses, methods and computer-readable medium for automatically generating playlists based on playlists of other users
US10943259B1 (en) Expansion of high performing placement criteria
JP7549669B2 (ja) ビデオグループパッケージを精緻化するためのユーザインターフェース
US9959322B1 (en) Ranking channels in search
KR20210075904A (ko) 타겟팅 및 기타 설정을 개선하기 위한 콘텐츠 제공자 추천
CN112218114B (zh) 视频缓存控制方法、装置和计算机可读存储介质
WO2018205642A1 (zh) 视频收益计算建模装置、方法和视频推荐装置、方法和服务器及存储介质
CN112269942B (zh) 一种推荐对象的方法、装置、系统及电子设备

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: 20790714

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: 20790714

Country of ref document: EP

Kind code of ref document: A1