A cold start recommendation method and a thompson sampling recommendation method

Abstract

The application provides a cold start recommendation method and a Thompson sampling recommendation method. The cold start recommendation method comprises the following steps: obtaining to-be-recommended information and a characteristic attribute of the to-be-recommended information; performing dimension classification on the characteristic attribute of the to-be-recommended information to obtain characteristic information of at least two dimensions; obtaining known information of a target user, the target user being a user without behavior data within a first preset time; selecting a plurality of recommended information from to-be-recommended information corresponding to each dimension characteristic information based on the association between the known information and the characteristic information of the at least two dimensions, and generating a recommendation pool based on the plurality of recommended information. The application performs information recommendation on the target user in the cold start period from the recommended information corresponding to the plurality of dimension characteristic information, improves the information recommendation hit rate, and enables the target user to quickly pass through the cold start period.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of user behavior recommendation, in particular to a cold start recommendation method and a Thompson sampling recommendation method. BACKGROUND

[0002] A supply and demand platform is a public platform where users publish supply and demand information and third-party service agencies publish information. The existing supply and demand information matching mechanism is to recommend information to users by classifying information and adopting a Thompson sampling recommendation method. The principle of the Thompson sampling recommendation is that when a piece of information is triggered by a user's click, browse, contact, etc., an association between the user and the information is established, and the two parameters of the beta distribution, alpha and beta, are updated, and the same type of information is recommended and explored according to the information type.

[0003] In the period of cold start of new users without historical behavior data, a uniform distribution assumption is adopted, the two parameters of the beta distribution, alpha and beta, are set to 1, and each piece of information is regarded as a uniform distribution probability, and the recommendation probability is corrected through continuous recommendation and feedback.

[0004] However, in the initialization stage of new users, that is, the cold start stage, the recommended information lacks prior data, and the system needs to adopt a completely random mode to make a large number of recommendation attempts, so that the recommendation quality is difficult to guarantee, the user experience is poor, and further, the user loses interest and has no positive feedback, forming a vicious cycle. SUMMARY

[0005] Therefore, it is necessary to provide a cold start recommendation method and a Thompson sampling recommendation method to solve the technical problems of low information recommendation hit rate and poor user experience due to lack of user behavior data in the cold start period of users in the prior art.

[0006] To solve the above problems, on the one hand, the present application provides a cold start recommendation method, comprising:

[0007] obtaining a to-be-recommended information and a characteristic attribute of the to-be-recommended information;

[0008] dimensionally classifying the characteristic attribute of the to-be-recommended information to obtain characteristic information of at least two dimensions;

[0009] obtaining known information of a target user, the target user being a user without behavior data within a first preset time;

[0010] based on the association between the known information and the characteristic information of the at least two dimensions, selecting a plurality of recommended information from the to-be-recommended information corresponding to each dimension characteristic information, and generating a recommendation pool based on the plurality of recommended information;

[0011] Push the information in the recommendation pool to the target user based on a preset recommendation algorithm.

[0012] In some possible implementation manners, based on the relevance between the known information and the feature information of each dimension, a plurality of recommendation information is selected from the to-be-recommended information corresponding to each dimension feature information, and a recommendation pool is generated based on the plurality of recommendation information, including:

[0013] Based on the relevance between the known information and the feature information of any dimension in the at least two dimensions, corresponding recommendation information is selected from the to-be-recommended information corresponding to the feature information of the any dimension, respectively.

[0014] Based on all the recommendation information corresponding to the feature information of the at least two dimensions, a recommendation pool is generated.

[0015] In some possible implementation manners, the feature information is any one of position information, hotness value information, enterprise information, and classification information.

[0016] In some possible implementation manners, if the feature information is enterprise information, the feature attributes corresponding to the enterprise information are enterprise labels and information publishing time, and the known information is target user login time.

[0017] Based on the relevance between the known information and the feature information of any dimension in the at least two dimensions, corresponding recommendation information is selected from the to-be-recommended information corresponding to the feature information of the any dimension, respectively, including:

[0018] Based on the target user login time and the information publishing time, the to-be-recommended information is sorted in the order of information publishing time.

[0019] At least one target enterprise label is selected from a preset enterprise label pool.

[0020] Based on the enterprise label, a plurality of recommendation information corresponding to the target enterprise label is sequentially selected from the to-be-recommended information in the order of sorting.

[0021] The number of the plurality of recommendation information is not greater than the expected selection number of the to-be-recommended information corresponding to the enterprise information.

[0022] In some possible implementation manners, if the feature information is position information, the feature attribute corresponding to the position information is information publisher geographic position, and the known information is target user geographic position.

[0023] Based on the relevance between the known information and the feature information of any dimension in the at least two dimensions, corresponding recommendation information is selected from the to-be-recommended information corresponding to the feature information of the any dimension, respectively, including:

[0024] determine a straight-line distance between the target user and the information publisher based on a preset distance algorithm, a geographic location of the information publisher, and a geographic location of the target user;

[0025] sort the to-be-recommended information according to the straight-line distance from near to far;

[0026] select a plurality of recommended information from the to-be-recommended information in order according to the sorting;

[0027] The number of the plurality of recommended information is not greater than a to-be-recommended information expected selection quantity corresponding to the location information.

[0028] In some possible implementation manners, if the feature information is hotness value information, the hotness value information includes information display quantity, information click browsing quantity, information click contact quantity, and information publishing time, and the known information includes target user login time;

[0029] based on the association between the known information and the feature information of any one of the at least two dimensions, select corresponding recommended information from to-be-recommended information corresponding to the feature information of any one of the at least two dimensions, including:

[0030] based on the target user login time and the information publishing time, select information with information publishing time within a second preset time range from the to-be-recommended information, and generate a preliminary selection pool based on to-be-recommended information with information publishing time within the second preset time range;

[0031] based on a preset hotness value algorithm and the information display quantity, the information click browsing quantity, and the information click contact quantity, determine the hotness value of the to-be-recommended information in the preliminary selection pool;

[0032] sort the to-be-recommended information in the preliminary selection pool according to the hotness value;

[0033] select a plurality of recommended information from the to-be-recommended information in the preliminary selection pool in order according to the sorting;

[0034] The number of the plurality of recommended information is not greater than a to-be-recommended information expected selection quantity corresponding to the hotness value information.

[0035] In some possible implementation manners, the method further includes:

[0036] based on a preset total to-be-recommended information expected selection quantity and a to-be-recommended information expected selection weight corresponding to feature information of any one single dimension, determine a to-be-recommended information expected selection quantity corresponding to the feature information of any one single dimension;

[0037] The expected number of to-be-recommended information corresponding to any single-dimension feature information is any one of the expected number of to-be-recommended information corresponding to enterprise information, the expected number of to-be-recommended information corresponding to location information, and the expected number of to-be-recommended information corresponding to the hotness value information.

[0038] In some possible implementation manners, before the known information of the target user is acquired, the method further includes:

[0039] Acquiring information browsed by the current user within a first preset time;

[0040] Determining whether the current user has logged into an account within the first preset time;

[0041] If the current user has not logged into the account within the first preset time, the current user is determined to be the target user;

[0042] If the current user has logged into the account within the first preset time, it is further determined whether the current user has browsed information within the preset time;

[0043] If the current user has browsed the information within the first preset time, the current user is determined to be a non-target user;

[0044] If the current user has not browsed the information within the first preset time, the current user is determined to be the target user.

[0045] In some possible implementation manners, based on the preset recommendation algorithm, the information in the recommendation pool is pushed to the target user, and the information in the recommendation pool is pushed to the target user, and then the method further includes:

[0046] Determining whether the current information pushed to the target user triggers a user behavior

[0047] If the current information triggers the user behavior, the current information is removed from the recommendation pool;

[0048] If the current information does not trigger the user behavior, it is further determined whether a recommendation number of the current information reaches a preset recommendation number;

[0049] If the current user reaches the preset recommendation number, the current information is removed from the recommendation pool;

[0050] If the current user does not reach the preset recommendation number, the current information is returned to the recommendation pool;

[0051] Determining whether a remaining information number in the recommendation pool is less than a preset information number;

[0052] If the remaining information number in the recommendation pool is not less than the preset information number, the recommendation continues;

[0053] If the remaining information quantity in the recommendation pool is less than the preset information quantity, the cold start mode is exited.

[0054] In another aspect, the application also provides a Thompson sampling recommendation method, obtaining behavior data of a target user, the behavior data being behavior data generated after the target user is recommended information based on the cold start recommendation method of any one of the above;

[0055] Taking the behavior data of the target user as prior probability of a preset Thompson sampling algorithm;

[0056] Based on the preset Thompson sampling algorithm, the information to be recommended is recommended to the target user.

[0057] The beneficial effects of the above embodiments are that the cold start recommendation method provided by the application first obtains information to be recommended and characteristic attributes of the information to be recommended, then classifies the characteristic attributes of the information to be recommended in dimensions to obtain multi-dimensional characteristic information, then selects recommendation information from the information to be recommended corresponding to each dimensional characteristic information based on the relevance of known information and characteristic information of a target user, and generates a recommendation pool to recommend to the target user, thereby realizing selective recommendation of a target user in a cold start period from multiple recommendation dimensions when lacking user behavior data, improving the information recommendation hit rate, making the target user quickly pass through the cold start period, and simultaneously achieving uniform information recommendation by multi-dimensional recommendation, and avoiding too concentrated information recommendation.

[0058] Further, after the cold start, feedback of the target user to the information in the recommendation pool is considered, and the feedback behavior data is taken as prior probability of a subsequent recommendation algorithm, so that the target user is recommended again by the Thompson sampling algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0060] Figure 1 An embodiment flowchart of the cold start recommendation method provided by the application is shown in the figure.

[0061] Figure 2 An embodiment flowchart of the cold start recommendation method provided by the application is shown in the figure. Figure 1 An embodiment flowchart of the cold start recommendation method provided by the application is shown in the figure.

[0062] Figure 3 An embodiment flowchart of the cold start recommendation method provided by the application is shown in the figure. Figure 2 An embodiment flowchart of the cold start recommendation method provided by the application is shown in the figure.

[0063] Figure 4 For the present application Figure 2 An embodiment flowchart of step S201 when the feature information is location information;

[0064] Figure 5 For the present application Figure 2 An embodiment flowchart of step S201 when the feature information is heat value information;

[0065] Figure 6 An embodiment flowchart of the present application for judging a target user;

[0066] Figure 7 An embodiment flowchart of the recommendation pool information display logic provided by the present application

[0067] Figure 8 An embodiment flowchart of the Thompson sampling recommendation method provided by the present application. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0069] It should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the present application show the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowcharts can not be implemented in sequence, and the steps without logical context relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowcharts or one or more operations can be removed from the flowcharts by those skilled in the art under the guidance of the content of the present application. Some block diagrams shown in the drawings are functional entities, which do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0070] The terms "first", "second", etc. in the embodiments of the present application are only used for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the technical features defined with "first", "second" can explicitly or implicitly include at least one of the features. The association relationship of the associated objects described by "and / or" indicates that there can be three kinds of relationships, for example: A and / or B can represent the three cases of A alone, A and B together, and B alone.

[0071] Reference herein to "embodiments" means that a particular feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of the application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor does it necessarily refer to a separate or alternative embodiment in isolation from other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with each other.

[0072] For the convenience of understanding the present application, some special terms are explained first.

[0073] User cold start: refers to recommending items that may be of interest to new users. New users refer to users with little interaction data and lack of historical behavior data, so it is difficult to capture user interest preferences. The cold start problem is a challenging problem in the recommendation system and is also a problem that the industry and academia are highly concerned about.

[0074] Beta distribution: a density function that is a conjugate prior distribution of Bernoulli distribution and binomial distribution, which has important applications in machine learning and mathematical statistics. In probability theory, Beta distribution, also known as B distribution, refers to a set of continuous probability distributions defined in the interval (0, 1).

[0075] Thompson sampling algorithm: also known as random probability matching algorithm, the reward expectation of each arm obeys a certain unknown distribution (in the present application, mainly discussing a multi-armed bandit machine in which each arm obeys Bernoulli distribution), as a kind of Bayesian multi-armed bandit machine problem algorithm, the main idea of Thompson sampling algorithm is to assume that the probability μ i with prior distribution Beta(alpha, beta), if a decision is executed successfully, the return is 1, otherwise it is 0. If decision i has been executed and is successful Q i (t) times, execute and fail P i (t) times, update the posterior distribution of the i-th decision variable to Beta(Q i (t)+1, P i (t)+1), and use the above result as the prior distribution of the decision in the next iteration. Then from Beta(Q i (t)+1, Pi (t)+1) is sampled i (t), is selected such that i(t) = arg i max theta i (t) is observed, and finally if the reward value is 1, then Q i (t) is incremented by 1, and otherwise P i (t) is incremented by 1. In short, Thompson selects deterministically according to the probability that each arm is the best arm. In a recommendation system, multiple recommendation strategies can be regarded as multi-armed bandits, and each recommendation strategy has a certain probability of generating user click or purchase behavior.

[0076] The present application provides a cold start recommendation method, which is described below.

[0077] Figure 1 An embodiment flow diagram of the cold start recommendation method provided by the present application is shown in FIG. 1. Figure 1 The cold start recommendation method includes the following steps.

[0078] S101, obtaining to-be-recommended information and characteristic attributes of the to-be-recommended information;

[0079] S102, dimensionally classifying the characteristic attributes of the to-be-recommended information to obtain characteristic information of at least two dimensions;

[0080] S103, obtaining known information of a target user, the target user being a user without behavior data within a first preset time;

[0081] S104, selecting a plurality of recommended information from to-be-recommended information corresponding to each dimension characteristic information based on the association between the known information and the characteristic information of the at least two dimensions, and generating a recommendation pool based on the plurality of recommended information;

[0082] S105, pushing the information in the recommendation pool to the target user based on a preset recommendation algorithm.

[0083] It should be noted that the to-be-recommended information refers to the information entity content published by the user, the characteristic attributes of the to-be-recommended information refer to the information marked by the platform for the to-be-recommended information after the user publishes the information, which is used to represent some information such as time, location, publisher, click and browse volume, etc. The preset recommendation algorithm is random recommendation, and the end condition of the preset recommendation algorithm is that all the information in the recommendation pool is recommended and displayed. When all the information in the recommendation pool for the target user is displayed, the cold start enters the original system Thompson sampling recommendation logic, and in the cold start recommendation process, the target user will give behavior feedback to the recommended information, including clicking, browsing, contacting, etc. These will be accumulated as the behavior data of the user.

[0084] It should be noted that the recommendation pool of the cold start exists in a copy for each target user, and once the copy of the recommendation pool for a target user is generated, any modification of the recommendation pool does not affect the copy.

[0085] Compared with the prior art, the cold start recommendation method provided by the embodiment of the application first acquires the to-be-recommended information and the characteristic attribute of the to-be-recommended information, then classifies the characteristic attribute of the to-be-recommended information in dimensions to obtain multi-dimensional characteristic information, and then selects the recommendation information from the to-be-recommended information corresponding to each dimensional characteristic information based on the association between the known information of the target user and the characteristic information, to generate a recommendation pool for the target user, thereby realizing the selective recommendation to the target user in the cold start period from multiple recommendation dimensions when the user behavior data is lacking, improving the hit rate of the information recommendation, and enabling the target user to quickly pass through the cold start period.

[0086] In order to improve the hit rate of the recommendation information and better expose the long-tail information while avoiding the concentration of the information recommendation, in some embodiments, the recommendation pool has a plurality of recommendation information corresponding to the characteristic information of at least two dimensions, and specifically, as shown in Figure 2 The step S104 includes:

[0087] S201, based on the association between the known information and the characteristic information of any one of the at least two dimensions, selecting the corresponding recommendation information from the to-be-recommended information corresponding to any one of the at least two dimensions;

[0088] S202, generating a recommendation pool based on all the recommendation information corresponding to the at least two dimensional characteristic information.

[0089] It should be noted that in other embodiments, the recommendation pool can also have only the recommendation information corresponding to the single-dimensional characteristic information.

[0090] In order to determine the characteristic information of the at least two dimensions, the characteristic information is any one of the position information, the hotness value information, the enterprise information and the classification information, and different recommendation information selection rules are used for different dimensional characteristic information.

[0091] Specifically, in one of the embodiments, if the characteristic information is the enterprise information, the characteristic attribute corresponding to the enterprise information is the enterprise label and the information publishing time, and the known information includes the login time of the target user, as shown in Figure 3 The step S201 includes:

[0092] S301, sorting the to-be-recommended information according to the publishing time in chronological order based on the login time of the target user and the information publishing time;

[0093] S302, selecting at least one target enterprise label from a preset enterprise label pool;

[0094] S303, selecting, in order, a number of recommendation information corresponding to the target enterprise label from the to-be-recommended information based on the enterprise label of the to-be-recommended information;

[0095] The number of the number of recommendation information is not greater than the expected selection quantity of the to-be-recommended information corresponding to the enterprise information.

[0096] It should be noted that the number of the number of recommendation information is less than the expected selection quantity of the to-be-recommended information corresponding to the enterprise information. In the case that the total amount of the to-be-recommended information corresponding to the target enterprise label is less than the expected selection quantity of the to-be-recommended information corresponding to the enterprise information, the to-be-recommended information corresponding to the target enterprise label is selected, and then the selection is stopped.

[0097] It can be understood that, for the characteristic information being classification information, the selection rule of the recommendation information is similar to the selection rule of the characteristic information being enterprise information, and the difference is that the characteristic attribute corresponding to the classification information is a classification label and an information publishing time.

[0098] In another embodiment, if the characteristic information is location information, the characteristic attribute corresponding to the location information is the geographic location of the information publisher, and the known information includes the geographic location of the target user, such as Figure 4 As shown in FIG. 2, step S201 includes:

[0099] S401, determining the straight-line distance between the target user and the information publisher based on a preset distance algorithm, the geographic location of the information publisher and the geographic location of the target user;

[0100] S402, sorting the to-be-recommended information according to the straight-line distance from near to far;

[0101] S403, selecting, in order, a number of recommendation information from the to-be-recommended information according to the sorting order;

[0102] The number of the number of recommendation information is not greater than the expected selection quantity of the to-be-recommended information corresponding to the location information.

[0103] It should be noted that the preset distance algorithm calculates the straight-line distance between the target user and the information publisher by calculating the latitude and longitude.

[0104] In order to improve the recommendation quality of the recommendation information corresponding to the location information, in some embodiments, after step S401, the method further includes: selecting information with a straight-line distance within a preset distance range from the to-be-recommended information. In this way, the information is preliminarily screened.

[0105] In yet another embodiment, if the characteristic information is the heat value information, the characteristic attribute corresponding to the heat value information is the information display quantity, the information click browsing quantity, the information click contact quantity and the information publishing time, and the known information includes the login time of the target user, such as Figure 5As shown, step S201 includes:

[0106] S501, based on the target user login time and the information publishing time, selecting information with the information publishing time within the second preset time range from the to-be-recommended information, and generating a preliminary selection pool based on the to-be-recommended information with the information publishing time within the second preset time range;

[0107] S502, based on the preset heat value algorithm and the information display amount, the information click browsing amount, and the information click contact amount, determining the heat value of the to-be-recommended information in the preliminary selection pool;

[0108] S503, sorting the to-be-recommended information in the preliminary selection pool according to the heat value;

[0109] S504, selecting a number of recommended information from the to-be-recommended information in order according to the sorting;

[0110] The number of the recommended information is not greater than the expected selection number of the to-be-recommended information corresponding to the heat value information.

[0111] It should be noted that the preset heat value algorithm takes the display amount, the click browsing amount, and the click contact amount as the target, and performs linear superposition calculation according to different weights.

[0112] It should be noted that the characteristic attributes of the to-be-recommended information are classified into enterprise information, classification information, geographic location, heat value, and other dimension characteristic information according to the dimensions, and the data initialization is established depending on the previous step. Specifically, after the information is published, the enterprise user publishing the information is tagged and profiled according to the enterprise scale, business scope, registered location, survival time, office location, and other information, i.e., enterprise information. Then, the published information is classified, such as office building for rent, raw material purchase or supply, i.e., classification information. In addition, further labels are added to the information, such as geographic location, i.e., location information, and display amount, click browsing amount, and click contact amount, i.e., heat value information. In addition to the above, the browsing, contact, and dwell time data of the information are counted by date, and the information publishing time is classified as auxiliary information into each dimension characteristic information.

[0113] It should be further noted that in other embodiments, the characteristic attributes of the to-be-recommended information can be further classified into other dimensions based on actual application requirements, or further subdivided. The characteristic information of at least two dimensions is not limited to enterprise information, classification information, geographic location, and heat value, but can also be other dimensions related to information profiling.

[0114] It should be noted that the to-be-recommended information corresponding to each dimension characteristic information is independently calculated, and the system automatically removes duplicates when generating the recommendation pool of the target user.

[0115] In order to facilitate the guiding display of the recommended information, and to meet the strategic layout of the platform recommended information, in some embodiments, the proportion of the recommended information corresponding to the plurality of recommended feature information of the recommendation pool can be adjusted, and specifically, the method further comprises:

[0116] Based on the preset expected selection amount of the to-be-recommended information and the expected selection weight of the to-be-recommended information corresponding to any single-dimension feature information, the expected selection quantity of the to-be-recommended information corresponding to any single-dimension feature information is determined.

[0117] Among them, the expected selection quantity of the to-be-recommended information corresponding to any single-dimension feature information is any one of the expected selection quantity of the to-be-recommended information corresponding to the enterprise information, the expected selection quantity of the to-be-recommended information corresponding to the location information and the expected selection quantity of the to-be-recommended information corresponding to the hotness value information.

[0118] It should be noted that the total amount of information in the recommendation pool and the weight of each dimension information can be customized according to actual needs, such as being set according to different business promotion in different periods of the platform, or being set according to different time and different customer group needs.

[0119] In order to obtain better recommendation effect, in some embodiments of the present application, the judgment condition of the target user is limited, and before step S101, as shown in Figure 6 The method further comprises:

[0120] S601, obtaining the login and browsing information of the current user within a first preset time;

[0121] S602, judging whether the current user has logged in an account within the first preset time;

[0122] S603, if the current user has not logged in an account within the first preset time, the current user is determined as a target user;

[0123] S604, if the current user has logged in an account within the first preset time, it is further judged whether the current user has browsed information within the preset time;

[0124] S605, if the current user has browsed information within the first preset time, the current user is determined as a non-target user;

[0125] S606, if the current user has not browsed information within the first preset time, the current user is determined as a target user.

[0126] It should be noted that if the user is determined as a target user, cold start recommendation is performed, and if the user is determined as a non-target user, i.e. an old user, the original Thompson sampling recommendation logic is performed.

[0127] In order to ensure good target user behavior feedback effect while improving the recommendation pool display efficiency, in some embodiments, the display and elimination conditions of the information in the recommendation pool are limited, such as Figure 7 As shown in FIG. 10, in step S105, the information in the recommendation pool is pushed to the target user, and then the following steps are further included:

[0128] S701, judging whether the current information pushed to the target user triggers user behavior or not;

[0129] S702, if the current information triggers user behavior, the current information is eliminated from the recommendation pool;

[0130] S703, if the current information does not trigger user behavior, further judging whether the recommendation times of the current information reach the preset recommendation times or not;

[0131] S704, if the current user reaches the preset recommendation times, the current information is eliminated from the recommendation pool;

[0132] S705, if the current user does not reach the preset recommendation times, the current information is put back into the recommendation pool;

[0133] S706, judging whether the number of the remaining information in the recommendation pool is less than the preset information number or not;

[0134] S707, if the number of the remaining information is not less than the preset information number, the recommendation continues;

[0135] S708, if the number of the remaining information is less than the preset information number, the cold start mode is exited.

[0136] Wherein, the information triggering user behavior refers to the user's clicking and browsing behavior on the displayed information.

[0137] Since the above-mentioned cold start recommendation algorithm is used in cooperation with the Thompson recommendation algorithm, the embodiment of the present application further provides a Thompson sampling recommendation method, which is applied after the target user exits the cold start mode, as shown in FIG. 11, which includes: Figure 8

[0138] S801, obtaining the behavior data of the target user, which is the behavior data generated after the target user is recommended information by the above-mentioned cold start recommendation method;

[0139] S802, taking the behavior data of the target user as the prior probability of the preset Thompson sampling algorithm;

[0140] S803, recommending the information to be recommended to the target user based on the preset Thompson sampling algorithm.

[0141] ​It should be noted that the target user will give behavior feedback to the recommended information in the cold start recommendation process, including clicking, browsing, contacting and other actions, which will be accumulated as behavior data. The target user starts to use the cold start recommendation method to recommend information, and when the information in the recommendation pool is displayed, it is switched to the Thompson sampling recommendation method.

[0142] The cold start recommendation method and the Thompson sampling recommendation method provided by the present application are described in detail above, and the principles and implementation modes of the present application are described by applying specific examples. The above description of the embodiments is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description should not be understood as a limitation of the present application.

Claims

1. A cold start recommendation method, characterized in that, include: Obtain the information to be recommended and its feature attributes; The feature attributes of the information to be recommended are classified into at least two dimensions to obtain feature information, which is enterprise information, and the feature attributes corresponding to the enterprise information are enterprise tag and information publication time. Obtain known information about the target user, wherein the target user is a user who has no behavioral data within a first preset time period, and the known information is the target user's login time; Based on the correlation between the known information and the feature information of any one of the at least two dimensions, the corresponding recommended information is selected from the information to be recommended corresponding to the feature information of each dimension. A recommendation pool is generated based on all recommendation information corresponding to the at least two dimensions of feature information; Based on a preset recommendation algorithm, information from the recommendation pool is pushed to the target user; Based on the correlation between the known information and the feature information of any one of the at least two dimensions, corresponding recommended information is selected from the information to be recommended corresponding to the feature information of each dimension, including: Based on the target user's login time and the information's publication time, the information to be recommended is sorted in chronological order of publication time. Select at least one target company tag from the preset company tag pool; Based on the enterprise tags of the information to be recommended, several recommended information corresponding to the target enterprise tags are selected sequentially from the information to be recommended in order of sorting. The number of recommended information items is no greater than the expected number of recommended information items to be selected corresponding to the enterprise information.

2. The cold start recommendation method according to claim 1, characterized in that: The feature information can also be any one of location information, heat value information, and classification information.

3. The cold start recommendation method according to claim 2, characterized in that, If the feature information is location information, the feature attribute corresponding to the location information is the geographical location of the information publisher, and the known information is the geographical location of the target user; Based on the correlation between the known information and the feature information of any one of the at least two dimensions, corresponding recommended information is selected from the information to be recommended corresponding to the feature information of each dimension, including: Based on a preset distance algorithm, the geographical location of the information publisher, and the geographical location of the target user, the straight-line distance between the target user and the information publisher is determined; The information to be recommended is sorted from closest to furthest according to the straight-line distance. Select several recommended information items from the list of items to be recommended in order of sorting; The number of the recommended information items is no greater than the expected number of recommended information items to be selected corresponding to the location information.

4. The cold start recommendation method according to claim 2, characterized in that, If the feature information is popularity value information, the feature attributes corresponding to the popularity value information are the number of information displays, the number of information clicks and views, the number of information clicks and contacts, and the information publication time. The known information includes the target user's login time. Based on the correlation between the known information and the feature information of any one of the at least two dimensions, corresponding recommended information is selected from the information to be recommended corresponding to the feature information of each dimension, including: Based on the target user's login time and the information release time, information with a release time within a second preset time range is selected from the information to be recommended, and an initial selection pool is generated based on the information to be recommended within the second preset time range. Based on the preset popularity value algorithm and the information display volume, information click-through volume, and information click-contact volume, the popularity value of the information to be recommended in the initial selection pool is determined. The information to be recommended in the initial selection pool is sorted according to the popularity value. Select a number of recommended information items from the initial pool of recommended information in order of sorting. The number of recommended information items is no greater than the expected number of recommended information items to be selected corresponding to the popularity value information.

5. A cold start recommendation method according to any one of claims 2-4, characterized in that, The method further includes: Based on the preset total amount of recommended information to be selected and the weight of recommended information to be selected corresponding to any single-dimensional feature information, the expected number of recommended information to be selected corresponding to the single-dimensional feature information is determined. Wherein, the expected number of recommended information corresponding to any single-dimensional feature information is any one of the following: the expected number of recommended information corresponding to enterprise information, the expected number of recommended information corresponding to location information, and the expected number of recommended information corresponding to popularity value information.

6. The cold start recommendation method according to claim 1, characterized in that, Before obtaining known information about the target user, the method further includes: Obtain the login and browsing information of the current user within a first preset time period; Determine whether the current user has logged into the account within a first preset time period; If the current user has not logged into their account within the first preset time period, then the current user is identified as the target user; If the current user has logged into their account within the first preset time period, then it will be further determined whether the current user has browsed information within the preset time period; If the current user has viewed information within the first preset time period, the current user is determined to be a non-target user; If the current user has not viewed any information within the first preset time period, then the current user is identified as the target user.

7. The cold start recommendation method according to claim 1, characterized in that, Based on a pre-defined recommendation algorithm, information from the recommendation pool is pushed to the target user, and this process also includes: Determine whether the current information pushed to the target user triggers user behavior; If the current information triggers user behavior, then the current information will be removed from the recommendation pool; If the current information does not trigger user behavior, then it is further determined whether the number of times the current information has been recommended has reached the preset number of recommendations; If the current user reaches the preset number of recommendations, the current information will be removed from the recommendation pool; If the current user has not reached the preset number of recommendations, the current information will be returned to the recommendation pool. Determine whether the number of remaining information in the recommendation pool is less than the preset number of information; If the number of remaining information in the recommendation pool is not less than the preset number of information, then continue recommending; If the number of remaining information in the recommendation pool is less than the preset number of information, exit the cold start mode.

8. A Thompson sampling recommendation method, characterized in that, include: Acquire the behavioral data of the target user, wherein the behavioral data is the behavioral data generated after recommending information to the target user based on the cold start recommendation method as described in any one of claims 1-7; The behavioral data of the target user is used as the prior probability of the preset Thompson sampling algorithm; Based on the preset Thompson sampling algorithm, the information to be recommended is recommended to the target user.

Images