Self-adaptive distribution method and system for marketing bidding resources

Abstract

The invention relates to the technical field of Internet, in particular to a marketing bidding resource adaptive allocation method and system, and the method comprises the steps: receiving a bidding request, and carrying out the analysis of the bidding request, and obtaining an audience identifier, an audience network identifier and an access address identifier; calling a historical bidding log in a preset observation period, and aggregating to obtain a request scale, an occurrence time period, the number of times of short-time window requests and the diversity of access addresses; generating an anomaly judgment feature set which comprises a daily request proportion, occurrence duration, short-time burst and a diversity ratio; carrying out statistical estimation on the features to obtain a threshold set; comparing each identification feature with a threshold value, writing the identification features meeting abnormal conditions into a blacklist, marking other identification features as conventional, and removing the identification features according to an unappearing time limit; if a blacklist is hit in the bidding pre-processing stage, the request is judged to be a non-intention request and marketing bidding resource allocation is reduced or canceled, otherwise, the request is judged to be an intention request; and outputting a judgment and distribution result, and writing back the log to update the blacklist. According to the invention, the problem of unreasonable resource consumption and distribution caused by mixing of non-intentional bidding requests in the marketing bidding process can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet, in particular to a marketing bidding resource adaptive allocation method and system. BACKGROUND

[0002] In the programmatic advertising of real-time bidding (RTB), the bidding requests received by the demand-side platform usually follow the industry protocol and carry audience and media side information, including time, audience identification, network address, client identification, access address and domain name fields. Due to the existence of ad fraud and non-intentional traffic, part of the bidding requests are not triggered by real users, but present abnormal request scale, frequency or regularity of access behavior characteristics. In existing research, many methods focus on identifying abnormalities from the media or domain name dimension, but may ignore non-intentional bidding requests that are not from abnormal media groups, but generated by specific audience groups or even a single audience, thereby causing waste of bidding opportunities and computing resources. Therefore, a solution is needed that can identify requests in real time before bidding and be integrated with the bidding process. SUMMARY

[0003] In view of the above technical problems, the present application provides a marketing bidding resource adaptive allocation method and system, aiming to solve the problem of unreasonable resource consumption and allocation caused by non-intentional bidding requests mixed in the marketing bidding process, based on audience side identification to build access behavior statistics and abnormality determination rules, and offline generation and maintenance of audience identification blacklist; in the pre-bidding stage of real-time bidding, the arrived bidding request is matched with the blacklist of the previous period, and then the bidding request is divided into intentional request and non-intentional request, the non-intentional request is executed to reduce or cancel resource allocation, the intentional request is normally allocated resources and enters the bidding process, to realize adaptive scheduling and governance of bidding resources.

[0004] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0005] According to one aspect of the present application, a marketing bidding resource adaptive allocation method is provided, the method comprising: receiving a bidding request and parsing to obtain audience identification, audience network identification and access address identification of the bidding request; calling historical bidding logs corresponding to the audience identification and the audience network identification in a preset observation period, and respectively aggregating to obtain access behavior statistics, the access behavior statistics including daily request scale, time period level occurrence, request scale in a short time window and access address diversity; generating an abnormality determination feature set from the access behavior statistics, the abnormality determination feature set including at least daily request proportion feature, occurrence duration feature, short-time burst feature and access address diversity ratio feature; statistically estimate the abnormality determination feature set to determine a threshold set; compare the abnormality determination features of each of the audience identifier and the audience network identifier with the threshold set, mark the audience identifier and the audience network identifier as abnormal audience and write into a blacklist when a preset abnormality condition is met, mark the audience identifier and the audience network identifier as regular audience when the preset abnormality condition is not met, and set a retention policy for the blacklist to delete entries when a preset non-appearance deadline is exceeded; In the pre-bid processing stage, if the audience identifier or the audience network identifier of the real-time arrived bid request hits the blacklist, determine the bid request as a non-intentional request and reduce or cancel the marketing bid resource allocation to not participate in the bid, otherwise, determine the bid request as an intentional request and allocate the marketing bid resource to participate in the bid; output the determination result and the resource allocation result, and write back to the historical bid log to update the blacklist.

[0006] Further, the marketing bid resource includes one or more of budget resource, bid opportunity resource and bid calculation resource for participating in the bid; The audience identifier is used to distinguish between terminals or browser instances, and the audience identifier is updated when the terminal identifier information is changed. The audience network identifier is formed by combining network address information and client identifier information to maintain the distinction of the audience when the network address is reused. The bid request further includes publisher information, which includes domain name identifier and publisher identifier, the domain name identifier is obtained by parsing the access address identifier, and the publisher identifier is used to uniquely identify the media party or position party.

[0007] Further, the generation of the daily request proportion feature includes: statistically count the number of bid requests generated by each of the audience identifier or the audience network identifier in the daily time window in the daily request scale, and perform a ratio operation on the number of bid requests and the total number of bid requests in the same daily time window to obtain the daily request proportion; Compare the daily request proportion with the proportion threshold value, and determine that the preset abnormality condition is met when the daily request proportion exceeds the proportion threshold value matched with the corresponding identifier type.

[0008] Further, the generation of the appearance duration feature includes: divide the daily time window into a plurality of continuous time period units, and count the number of time periods in which bid requests appear for each of the audience identifier or the audience network identifier in the time period unit as the appearance duration; compare the appearance duration with the appearance duration threshold value, and determine that the preset abnormality condition is met when the appearance duration reaches or exceeds the appearance duration threshold value, wherein the time period unit is a time granularity unit of hour level or minute level.

[0009] Further, the generation of the short-time burst feature comprises: counting the number of bidding requests of each of the audience identifiers or the audience network identifiers within a sliding time window of a second level to obtain a short-time window request number; and aggregating the short-time window request numbers within the observation period to obtain at least one of a maximum burst value, an average burst value, or a burst quantile value; and determining that the preset abnormal condition is met when the maximum burst value, the average burst value, or the burst quantile value exceeds a burst threshold.

[0010] Further, the statistics of the access address diversity comprises: counting the number of different access address identifiers associated with each of the audience identifiers or the audience network identifiers within the observation period, and performing a ratio operation on the number of different access address identifiers and the daily request scale to obtain an access address diversity ratio; determining that the preset abnormal condition is met when the access address diversity ratio is lower than a diversity threshold; and / or determining that the preset abnormal condition is met when a corresponding relationship between the number of different access address identifiers and the number of bidding requests deviates from a preset monotonic relationship, wherein the preset monotonic relationship is used to represent a regular access behavior in which the number of different access address identifiers increases synchronously with the increase of the number of bidding requests.

[0011] Further, the determination of the threshold set comprises: constructing cumulative distribution statistics or quantile statistics on the values of the abnormality determination feature set within the observation period, and selecting a demarcation point in a tail region of the cumulative distribution statistics as a threshold, so that the audiences falling into a main body distribution region are determined as regular and the audiences falling into the tail region are determined as abnormal; and establishing the threshold set independently for the audience identifiers and the audience network identifiers respectively, so as to adapt to the differences in access behaviors of different identifier types.

[0012] Further, the updating of the blacklist adopts a combination of offline construction and online application, and specifically comprises: At the end of each of the observation periods, an abnormal audience list is generated based on the historical bidding logs of the observation period, and the abnormal audience list is added to the existing blacklists of the audience identification and the audience network identification to form a cumulative blacklist; at the time of addition, a latest occurrence time is recorded for each entry of the blacklist to implement the retention policy; and in the real-time processing stage, only the cumulative blacklist from the completed latest observation period is used to match the bidding request of the current period; when the bidding request is determined to be the non-intentional request, the determination result is written into a bidding response and a discard or bypass process is implemented, and when the bidding request is determined to be the intentional request, the bidding request is sent to a bidding process to achieve adaptive allocation of the marketing bidding resource and reduce real-time calculation overhead.

[0013] According to another aspect of the present application, a marketing bidding resource adaptive allocation system is provided, comprising: A bidding access module is configured to receive a bidding request and parse audience identification, audience network identification and access address identification of the bidding request; A log calling module is configured to call historical bidding logs corresponding to the audience identification and the audience network identification within a preset observation period, and aggregate access behavior statistics including daily request size, period-level occurrence, request size in a short time window and access address diversity, respectively; An abnormality judgment module is configured to generate an abnormality determination feature set from the access behavior statistics, the abnormality determination feature set at least including a daily request proportion feature, an occurrence duration feature, a short-time burst feature and an access address diversity ratio feature; A threshold definition module is configured to statistically estimate the abnormality determination feature set to determine a threshold set; A maintenance module is configured to compare the abnormality determination features of each of the audience identification and the audience network identification with the threshold set, mark as an abnormal audience and write into a blacklist if a preset abnormality condition is met, and mark as a regular audience if not met, and set a retention policy for the blacklist to be removed when an entry exceeds a preset non-occurrence deadline; A distribution control module is configured to, in a bidding pre-processing stage, if the audience identification or the audience network identification of the bidding request arriving in real time hits the blacklist, determine the bidding request as a non-intentional request and reduce or cancel marketing bidding resource allocation to not participate in bidding, otherwise determine as an intentional request and allocate the marketing bidding resource to participate in bidding; An output module is configured to output a determination result and a resource allocation result, and write back to the historical bidding logs to update the blacklist.

[0014] The technical scheme of the present application has the following beneficial effects: Compared with the processing mode of only relying on media side clustering or static rules, the present application starts from the audience perspective, realizes the identification and precipitation of abnormal audiences by extracting features reflecting abnormal access behaviors and setting corresponding threshold values and rules, thereby having the ability to identify both group abnormality and individual abnormality of two non-intentional sources; meanwhile, the present application can be integrated into the demand side real-time processing link as a plug-in process in the pre-bidding stage, realizes rapid discrimination through blacklist matching, marks and discards or bypasses non-intentional bidding requests, and only sends intentional requests into the bidding process, thereby reducing invalid bidding and invalid calculation, and improving the use efficiency of marketing bidding resources and overall delivery performance. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of a marketing bidding resource adaptive allocation method in an embodiment of the present application; Figure 2 A structural block diagram of a marketing bidding resource adaptive allocation system in an embodiment of the present application; Figure 3 A terminal device for implementing a marketing bidding resource adaptive allocation method in an embodiment of the present application; Figure 4 A computer readable storage medium storing a marketing bidding resource adaptive allocation method in an embodiment of the present application. DETAILED DESCRIPTION

[0016] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the implementations set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and fully convey the inventive aspects to those skilled in the art. The described features, structures, or characteristics can be combined in one or more implementations. In the following description, numerous specific details are provided to give a thorough understanding of implementations of the application. One skilled in the relevant art will recognize, however, that the

[0017] Furthermore, the accompanying drawings are merely illustrative of the invention. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0018] This invention provides an adaptive allocation method for marketing bidding resources. (Refer to...) Figure 1 The diagram shown is a flowchart illustrating an adaptive allocation method for marketing bidding resources according to an embodiment of the present invention. This method can be applied to electronic devices such as personal computers and servers. The method can be executed by a system, which can be implemented by software and / or hardware. Specifically, the method may include the following steps S101-S107: In step S101, a bidding request is received and parsed to obtain the audience identifier, audience network identifier, and access address identifier of the bidding request.

[0019] The marketing bidding resources include one or more of the following: budget resources, bidding opportunity resources, and bidding calculation resources for participating in the bidding; the audience identifier is used to distinguish between terminal or browser instances, and the audience identifier is updated when the terminal identifier information changes; the audience network identifier is formed by combining network address information and client identifier information to maintain audience differentiation even when the network address is reused; the bidding request also includes publisher information, which includes a domain name identifier and a publisher identifier, wherein the domain name identifier is obtained by resolving the access address identifier, and the publisher identifier is used to uniquely identify the media party or ad placement party.

[0020] After a bidding request enters the demand-side platform via the real-time bidding link, it first undergoes request access and protocol field parsing. Bidding requests follow an industry-standard request structure, carrying key information from both the user and media sides, including fields such as time, audience identifier, network address information, client identifier information, geographic location, access address, host information, and domain name. This request structure is defined by OpenRTB. Therefore, the receiving phase requires structured parsing and standardized extraction of the request, mapping the raw fields to a set of basic identifiers that can be directly used by subsequent rules and models, while retaining necessary context fields related to bidding calculations to support subsequent allocation and control of marketing bidding resources.

[0021] Specifically, the audience identification is used to characterize the browser or terminal instance, facilitating the aggregation of access behavior from the perspective of the audience. The audience identification can identify the browser instance by using the third-party Cookie, and the identification is updated when the Cookie is cleared, so it needs to be treated as a changeable identification when parsing: on the one hand, the current value is output as the audience identification of this request, and on the other hand, the same terminal can have a new audience identification in different time periods in the log or state, so as to avoid misjudging the update as a new device.

[0022] The audience network identification is used to improve the distinguishability in the case of network address multiplexing, and can be formed by combining network address information and client identification information, wherein the network address information corresponds to the IP address of the user machine, and the client identification information corresponds to the Useragent, which is used to represent the browser type and version, operating system and version, etc. Since the same IP may be shared by multiple users in a certain scenario, relying only on IP will reduce the distinguishability, so IP and Useragent should be spliced or jointly encoded when parsing to obtain a stable audience network identification output, providing a unified key value for subsequent access behavior statistics according to the audience dimension.

[0023] The access address identification is used to represent the accessed page or resource link corresponding to the bidding request, which is an important input of the media side identification. The media side related identification at least includes the access address, domain name and publisher identification, wherein the access address represents the link accessed by the bidding request, the domain name is the domain name part of the access address, and the publisher identification is a unique identification assigned to each publisher. Therefore, in the parsing stage, the access address field needs to be standardized and the access address identification needs to be generated, and the host information is extracted from the access address and further parsed to obtain the domain name identification, to form an aggregation key at the domain name level; the publisher identification field is directly read and output as the publisher identification, which is used to uniquely identify the media party or position party. Through the above parsing, a single bidding request is uniformly converted into structured results such as audience identification, audience network identification, access address identification, domain name identification and publisher identification.

[0024] In step S102, the historical bidding logs corresponding to the audience identification and the audience network identification are called in a preset observation period, and access behavior statistics are aggregated respectively, including daily request size, time period level occurrence, request size in a short time window, and access address diversity.

[0025] The statistics of the access address diversity include: counting the number of different access address identifiers associated with each of the audience identifiers or the audience network identifiers in the observation period, and performing a ratio operation on the number of different access address identifiers and the daily request size to obtain the access address diversity ratio; when the access address diversity ratio is lower than a diversity threshold, it is determined that the preset abnormal condition is met; and / or when the correspondence between the number of different access address identifiers and the number of bidding requests deviates from a preset monotonic relationship, it is determined that the preset abnormal condition is met, wherein the preset monotonic relationship is used to represent the regular access behavior that the number of different access address identifiers increases synchronously with the number of bidding requests.

[0026] Specifically, in a preset observation period, access behavior statistics can be retrieved and aggregated from a historical bidding log along two main lines of audience identifiers and audience network identifiers. The historical bidding log usually contains time, audience identifier, network side information (such as the combination of IP address and user agent), and access address information (such as URL, host, domain, etc.), so the requests of the same audience identifier or the same audience network identifier in the observation period can be merged and counted around these fields, thereby obtaining daily request size, period-level occurrence, request size in a short time window, and access address diversity statistics. The observation of access behavior can cover time granularities such as seconds, minutes, hours, days, etc., among which daily and hourly levels can depict the overall traffic and occurrence period distribution of a class of users in a day, and the second-level window is used to capture the intensive request mode in a very short time. These statistics are collectively used to represent whether the access behavior conforms to the regular mode.

[0027] In the aggregation of daily request size and period-level occurrence, the number of requests per day in the observation period can be counted to form the daily request size, and the number of hours in which requests occur in a day can be further counted to form the period-level occurrence. Regular access behavior usually shows that the daily number of requests is not abnormally high, and only appears in part of the hours in a day; relatively, if a certain audience generates an abnormally large number of requests in a single day, or continuously appears in almost all hours, it can be considered as an abnormal access tendency. The request size in a short time window can be counted in a fixed window of seconds or less, and regular access behavior usually generates only one to two requests in a second, while a higher frequency of requests in a very short time presents suspicious characteristics. The above-mentioned daily, hourly and second-level statistics can be directly calculated by time stamp in days, hours and seconds when aggregated, and used as input for subsequent abnormal condition judgment.

[0028] The statistics of access address diversity can be operated around the deduplication count and ratio of access address identification: for each audience identification or audience network identification, the number of different access address identifications associated with it is counted in each natural day of the observation period, and the ratio operation is performed with the request size of the day to obtain the access address diversity ratio. Correspondingly, the diversity ratio can be defined as ; wherein, represents a certain audience identification or audience network identification, represents the number of different access address identifications associated with the dth day, represents the number of requests on the dth day. Generally, the more the number of requests, the more the number of different access addresses will increase simultaneously, and the two present a regular relationship of increasing diversity with increasing requests; if there are points with a very high number of requests but a small number of different access addresses, they deviate from this regular relationship and can be used as an abnormal feature. At the same time, the diversity ratio should be maintained at a certain level in most cases, and if it is lower than the diversity threshold (for example, the threshold 0.05 determined by the statistical distribution), it is determined that the preset abnormal condition is met; or when and deviate from the preset monotonic relationship (manifested as high request size but no simultaneous increase in diversity), it is also determined that the preset abnormal condition is met. The above aggregation of access address diversity and request size and determination can be performed offline by integrating the audience identification with its historical bidding requests and access logs to form a storage result of access behavior statistics.

[0029] In step S103, an abnormality determination feature set is generated from the access behavior statistics, and the abnormality determination feature set at least includes a daily request proportion feature, an occurrence duration feature, a short-time burst feature, and an access address diversity ratio feature.

[0030] Specifically, in step S103, the generation of the abnormality determination feature set can be understood as: further structuring the access behavior statistics aggregated in the observation period into a set of feature values that can be used for rule determination, and providing input for subsequent threshold-based abnormality labeling. The characterization of access behavior usually revolves around three dimensions: the identification dimension is used to carry the statistical object, the measurement dimension is used to cover the traffic size, the behavior frequency and the behavior regularity, and the observation dimension covers different time granularities such as seconds, hours, days, etc., so that the behavior differences of the same identification at different scales are explicit. In implementation, a plurality of rule features extracted for audience identification and audience network identification can be uniformly represented as a feature vector, and the identification and historical bidding requests and access logs are integrated offline to form a reusable access behavior statistics base to support continuous calculation and update. Typical manifestations of abnormal behavior at the statistical level include: the request size at the day dimension is abnormally large, the time coverage within the day is too wide, the request suddenly increases within the second-level time, and the request size and the number of access addresses are imbalanced, etc.

[0031] The generation of the occurrence duration feature includes: dividing a day-level time window into a plurality of continuous time period units, counting the number of time periods in which the bidding request of each audience identification or audience network identification occurs in the time period unit as the occurrence duration; comparing the occurrence duration with an occurrence duration threshold value, and determining that the preset abnormal condition is met when the occurrence duration reaches or exceeds the occurrence duration threshold value, wherein the time period unit is a time granularity unit of hour level or minute level.

[0032] The occurrence duration feature is used to characterize the coverage range of a certain identification within a day. Its generation can divide a day-level time window into continuous time period units, determine whether the bidding request of the identification occurs in each time period unit, and count the time period units in which the request occurs to obtain the occurrence duration. Generally, an ordinary identification only occurs in part of the day, and when an identification continuously occurs in a large range and for a long period of time within a day, it is often regarded as an abnormal signal. Therefore, the occurrence duration can be compared with an occurrence duration threshold value: when the occurrence duration reaches or exceeds the threshold value, the abnormal condition is met, which is used to capture the behavior mode of continuous online and wide coverage within a day. The threshold value can also be determined by statistical estimation of the occurrence duration distribution, so that the normal identification distribution area and the abnormal tail distribution area are distinguished.

[0033] The generation of the short-time burst feature includes: using a sliding time window of seconds, counting the bidding requests of each audience identifier or audience network identifier in the sliding time window to obtain a short-time window request number; and aggregating the short-time window request numbers in the observation period to obtain at least one of a maximum burst value, an average burst value, or a burst quantile value; when the maximum burst value, the average burst value, or the burst quantile value exceeds a burst threshold value, it is determined that the preset abnormal condition is met.

[0034] The short-time burst feature is used to describe the burst behavior on a second scale. The bidding requests of each identifier can be counted on a second time granularity to obtain the request number in each second window, and the second counts are aggregated in the observation period to form statistical quantities such as a maximum burst value, an average burst value, or a burst quantile value. When there are too many request numbers in a second window, it is usually difficult to be explained by normal manual operation, and thus can be used as an abnormality judgment basis. This rule can be embodied at the threshold comparison level as follows: the abnormal condition is met when the second count or the aggregated statistics exceeds the burst threshold value. The access address diversity ratio feature corresponds to the coupling relationship between the access address and the request size. The number of different access addresses associated with each identifier in the observation period is first counted, and then a ratio operation is performed with the number of bidding requests in the same period or the same day to form the access address diversity degree corresponding to each request. When the request size is large but the associated access address is small, it often reflects the abnormal characteristics of too concentrated access target and convergent behavior pattern. In addition, the experience rule that the number of different access addresses usually increases synchronously when the number of requests increases can also be used as a supplementary criterion for the abnormal condition. After integrating these features, the rule features can be organized into a vector form and used for threshold judgment. For example, the feature vector can be represented as: ; and the abnormal marking is completed under the input of the rule set and the threshold set: ; When any rule meets the corresponding threshold condition, it is marked as abnormal, otherwise it is marked as normal.

[0035] The generation of the daily request proportion feature includes: in the daily request size, counting the number of bidding requests generated by each audience identifier or audience network identifier in a daily time window, and performing a ratio operation on the number of bidding requests and the total number of bidding requests in the same daily time window to obtain a daily request proportion; comparing the daily request proportion with a proportion threshold value, when the daily request proportion exceeds the proportion threshold value matched with the corresponding identifier type, it is determined that the preset abnormal condition is met.

[0036] The daily request proportion feature is used to reflect the contribution degree of a certain identifier to the total bidding request in a daily time window. The generation process can count the number of requests for each identifier per day, count the total number of bidding requests in the same daily time window, and then perform a ratio operation on the two to obtain the proportion. The higher the proportion, the more likely it is that a single identifier continuously generates a large number of requests in a short period of time, thereby having suspiciousness. In order to meet the determinability, the proportion feature can be compared with a proportion threshold value: the setting of the threshold value can be derived from the statistical estimation of the proportion distribution of each identifier on the current day, so that most ordinary identifiers fall on one side of the threshold value, and a small number of tail identifiers fall on the other side, thereby realizing the separation of abnormal candidates. The essence of comparing this feature with the threshold value is to convert the daily request size from an absolute quantity to a relative quantity, and to reduce the interference brought by the overall traffic fluctuation of different dates.

[0037] In step S104, a statistical estimation is performed on the abnormal determination feature set to determine a threshold value set.

[0038] The determination of the threshold value set includes: constructing a cumulative distribution statistic or quantile statistic for the values of the abnormal determination feature set in the observation period, and selecting a demarcation point in the tail region of the cumulative distribution statistic as a threshold value, so that the audience falling into the main distribution region is judged to be regular and the audience falling into the tail region is judged to be abnormal; and the threshold value set is established independently for the audience identifier and the audience network identifier, respectively, to adapt to the differences in access behavior of different identifier types.

[0039] Specifically, in this step, the threshold value set is used to convert the abnormal determination feature set into a boundary condition that can be directly distinguished: only the rules or features are not enough to complete the abnormal / regular planning of the audience, and a corresponding specific threshold value needs to be configured for each rule; when the rule value of a certain audience in a given observation period meets any threshold value condition, it is marked as abnormal, otherwise it is regular. The threshold value is not fixed and unchangeable by experience, but is obtained by statistical estimation of the rule value in the observation period: the value distribution of each rule for the total audience in a time measurement window (for example, by day or by second) is made cumulative distribution statistic or quantile statistic, the tail region of the distribution is focused on, and the demarcation point of the tail region is taken as the threshold value, so that most of the audience in the main distribution region falls on the regular side, and a small number of audiences in the tail region falls on the abnormal side; and each rule threshold value is applied to the statistical value of the rule in a time period, and is calculated by a statistical estimation method.

[0040] Taking the first rule of statistics per day as an example, first count the number of requests of each audience identifier in a day, and further calculate the proportion of the total requests in the day. In the cumulative distribution statistics result, it can be observed that the proportion of most audiences is small, and a small number of audiences in the tail have a significantly larger proportion. Therefore, the threshold can be selected as the dividing point in the tail. The statistical result shows that more than 99% of audienceId has a proportion less than 0.03 in the total requests in the day, and a small number of audiences with a proportion of 0.03 or more are usually accompanied by abnormal behavior and extremely high request quantity, so 0.03 is selected as the threshold of audienceId. Similarly, the threshold of (IP, UA) is 0.02. The key point here is that the threshold comes from the dividing point of the tail of the distribution, rather than subjective judgment of individual audiences. At the same time, the determination of the threshold depends on the estimation of the cumulative distribution or quantile position of the rule value of all audiences in the observation period, so that most audiences naturally fall into the main distribution area.

[0041] For the second rule of statistics per day, the number of hours in which each audience has requests in a day is counted, and the tail dividing point is selected as the threshold on its distribution. Statistics show that only a small number of audiences will have more than 5 hours of requests in a day. In actual behavior, 5, 6, and 7 hours are still reasonable, but more than 20 hours is an abnormal phenomenon, which can be explained by non-human access, so 20 is selected as the common threshold of audienceId and (IP, UA). For the third rule of statistics per second, the number of requests generated by each audience in 1 second is counted, and its distribution shows that only a small number of audiences will exceed 2 times. Under normal circumstances, there is usually only one request in 1 second, and a small amount of double-clicking can generate 2 requests and can be explained, but more than 2 times points to abnormality, so 3 is selected as the threshold of audienceId and (IP, UA). For the fourth rule, the threshold comes from the statistical distribution of the relationship between daily distinct URL and request quantity: the number of distinct URLs of each audience in a day is divided by the number of requests to obtain the statistical quantity of the rule, and the cumulative distribution statistics is performed on the statistical quantity, and the dividing point is selected in the tail. According to the statistical result, the dividing point is 0.05, and is applied to audienceId and (IP, UA) respectively. The summary result of the four rules and thresholds corresponds to form a threshold set, and the thresholds of the two types of audience identifiers are clearly distinguished to adapt to the differences in access behavior of different identifier types: such as the thresholds corresponding to the rules are audienceId: 0.03, 20, 3, 0.05; (IP, UA): 0.02, 20, 3, 0.05.

[0042] After the threshold set is determined, the abnormality determination feature set can be organized into a rule set R and a threshold set T according to the rule dimension, and the abnormality marking at the audience level is completed through threshold comparison, and is used for subsequent discrimination of requests: given a rule and its corresponding threshold , first build the audience classification function , wherein , mark the abnormal audience, 0 mark the regular audience; then get the request classification from , wherein }, represent non-intentional requests, 0 represents intentional requests. In this step, it needs to be emphasized that the threshold set is independently established for audienceId and (IP, UA) respectively: the distribution tail position of the same rule under different identification types may be different, so the threshold is not forced to be shared, but is selected as a dividing point in the tail based on the cumulative distribution or quantile statistics of the two sets of identification respectively, so as to obtain threshold sets that are independent of each other and can adapt to differences.

[0043] In step S105, the abnormality determination features of each of the audience identifiers and the audience network identifiers are compared with the threshold set, and those that meet the preset abnormality condition are marked as abnormal audiences and written into a blacklist, and those that do not meet the condition are marked as regular audiences, and a retention policy is set for the blacklist to delete entries that exceed a preset non-appearance period.

[0044] Specifically, in step S105, audience classification can be performed for each audience identifier and audience network identifier: first, map the abnormality determination feature set obtained by the identifier in the observation period to a set of data rule values, then compare each data rule value with its corresponding threshold; when any one data rule meets the corresponding threshold condition, the identifier is marked as an abnormal audience, otherwise it is marked as a regular audience. This comparison logic emphasizes the one-vote trigger principle, that is, as long as any threshold condition is met, it can enter the abnormal set, so as to quickly screen out abnormal audiences under the condition of unsupervised annotation through rule thresholds, and reduce the suspicious traffic size that needs to be processed in subsequent real-time links.

[0045] When the abnormal / regular marking is completed, the abnormal audiences are written into a blacklist for subsequent processing. The blacklist can be maintained separately according to the identification type, for example, AudienceId BlackList is maintained for audience identifiers, and IPUA BlackList is maintained for audience network identifiers; the blacklist is created at the beginning of the data, and is accumulated on a daily or observation period basis, and the abnormal identifiers identified in the current observation period are appended to the existing blacklist, thereby forming a reusable abnormal set across periods, so that the subsequent arriving requests can be quickly discriminated according to the blacklist without having to recalculate the full amount of features in real time.

[0046] To avoid the blacklist growing indefinitely and reduce the false positives for the identifiers that have been recovered or no longer exist, a retention policy needs to be set for the blacklist. The last time the identifier appeared can be recorded in the blacklist entry, and a preset non-appearance deadline is used as the removal condition: when a blacklist entry no longer appears in consecutive observation periods, it is removed from the blacklist; on the contrary, as long as the identifier appears again, its last appearance time is refreshed and it continues to be retained. In this way, the blacklist can cover the persistent abnormal audience while automatically cleaning up expired entries over time, keeping the size and timeliness of the blacklist within a controllable range; the same maintenance and removal mechanism can be used for the blacklist corresponding to the audience network identifier.

[0047] In step S106, in the pre-bidding processing stage, if the audience identifier or the audience network identifier of the real-time arrived bidding request hits the blacklist, the bidding request is determined as a non-intention request and the marketing bidding resource allocation is reduced or cancelled to not participate in bidding, otherwise it is determined as an intention request and the marketing bidding resource is allocated to participate in bidding.

[0048] Wherein, after the bidding request enters the real-time link of the demand side platform, it first goes through a pre-processing link before entering the auction, which is used for intention discrimination of the request and controls whether to enter the bidding process accordingly. The pre-processing can be embedded in the pre-bidding stage in the form of a plug-in program, and its core task is to divide the arrived bidding request into intention request and non-intention request: for a bidding request arrived on a certain day, read the audience identifier and audience network identifier therein, and match and verify them with the two types of blacklists that have been generated and can be directly used on the previous day respectively; if any identifier hits the blacklist, the bidding request is marked as a non-intention request, otherwise it is marked as an intention request. The marking result is written into the request reply information, which is used as the basis for processing in the subsequent link; when marked as a non-intention request, the request is discarded and does not enter the auction process; when marked as an intention request, the request is sent into the auction process to continue participating in bidding.

[0049] From the formalization point of view, this step can be regarded as a request layer classification driven by audience layer marking: first get the binary marking of whether the audience is abnormal, and then get the binary determination of whether the request is a non-intention from the marking. Corresponding to the content of step S105.

[0050] At the level of marketing bidding resource allocation, this step realizes adaptive control through whether to enter the auction: when the request is judged as a non-intentional request and discarded, no longer allocate the budget resources, bidding opportunity resources and bidding calculation resources required for participating in bidding for this request, so as to avoid consuming resources on invalid traffic; when the request is judged as an intentional request, allocate corresponding marketing bidding resources to it and enter the auction process, so as to realize resource inclination to requests that are more likely to produce effective bidding results. Since this step only needs to complete matching filtering based on the existing blacklist, without high calculation and complex configuration, and the additional time delay brought by filtering has little effect on the overall link, it is suitable for deployment in high-throughput real-time bidding systems and meets the scalability and low delay requirements.

[0051] In step S107, the judgment result and resource allocation result are output and written back to the historical bidding log to update the blacklist.

[0052] Among them, after the pre-processing link completes the intention judgment of the real-time bidding request, the judgment result and the corresponding resource allocation result need to be output to the outside, and the result needs to be put into the historical bidding log to form a traceable record. The output can be reflected in the bidding response or request reply field, that is, the marking information of this bidding request is written into the request reply; when marked as a non-intentional request, the resource allocation result is embodied as no longer consuming bidding calculation and bidding opportunity resources for this request and directly discarded; when marked as an intentional request, the resource allocation result is embodied as allowing to enter the auction process and continue to participate in subsequent bidding processing. The output written into the request reply and triggering the processing path of discarding or sending into the auction is the final judgment and scheduling result given by the real-time link to the outside.

[0053] In order to form a closed loop update, the judgment result and the resource allocation result need to be written back to the historical bidding log as the input of subsequent offline statistics and blacklist update. In addition to retaining the basic fields of the bidding request, the historical bidding log can also attach the judgment mark of this time, the processing action, and the identification information used for correlation statistics, so as to support the summary calculation of the audience level on a daily or observation period basis, and precipitate the abnormal audience identified into the blacklist. The actual update method can adopt daily rolling accumulation: the blacklist is established at the beginning of the data, and the abnormal audience identified in each subsequent period is added to the blacklist, while the latest occurrence time is maintained, so as to perform elimination when the entry has not appeared for a long time, so as to keep the effectiveness and controllable scale of the blacklist.

[0054] The history bidding log after the write-back can also be used to derive auxiliary information to enhance the governance effect, for example, after marking the non-intended requests in real time, extracting the access address and domain name and other information from these requests and aggregating them to form a candidate set for subsequent analysis and monitoring, which is used to observe the coverage and concentration of non-intended traffic, and to provide a basis for policy optimization. Such derived statistics do not change the decision logic of this step, but rely on the write-back log to reuse the output results, thereby promoting the iterative updating of the blacklist and filtering effect with the period.

[0055] In an embodiment, the updating of the blacklist adopts a combination of offline construction and online application, specifically including: After the end of each observation period, an abnormal audience list is generated based on the history bidding log of the observation period, and the abnormal audience list is added to the existing blacklist of the audience identifier and the blacklist of the audience network identifier to form a cumulative blacklist; the latest occurrence time is recorded for each entry of the blacklist during the addition to perform the retention policy; and only the cumulative blacklist from the completed latest observation period is used in the real-time processing stage to match the bidding request of the current period; when the bidding request is determined to be the non-intended request, the determination result is written into the bidding response and the discard or bypass processing is performed, and when the bidding request is determined to be the intended request, the bidding request is sent to the bidding process to realize the adaptive allocation of the marketing bidding resources and reduce the real-time calculation overhead.

[0056] Specifically, the key of the embodiment is to complete the high-cost statistics and identification in the offline stage after the end of the period, and to compress the things that need to be done in the real-time stage into a fast matching and shunting, so that the abnormal audience set can be continuously updated, and the complex calculation is not put into the bidding main link. That is, after the end of each observation period, the aggregation result of the audience dimension is extracted from the history bidding log precipitated in the period, and the abnormal audience list is obtained according to the aggregation result, and then the abnormal audience list is added to the existing blacklist to form a cumulative blacklist; the blacklist can be maintained according to the audience identifier and the audience network identifier respectively, and the two types of blacklists are established at the beginning of the data, and then they are added by period to make the abnormal audience reusable across periods and gradually cover the long-term abnormal subjects.

[0057] To ensure the timeliness of the blacklist and avoid unbounded growth, the latest occurrence time is recorded for each blacklist entry during the addition, which is used to perform the retention policy subsequently: when a certain entry does not appear again for a plurality of consecutive periods, it is considered to be expired and is removed from the blacklist; when the entry appears again, the latest occurrence time is updated and the entry is continued to be retained. In this way, the blacklist can maintain coverage of the continuously abnormal audience and automatically clean up the historical entries that no longer appear, ensuring that the online matching set size is controllable and more close to the current traffic pattern.

[0058] The real-time stage only uses the accumulated blacklist formed by the last completed observation period to match the bidding request arrived in the current period, without the need for online side to recalculate the access behavior statistics or re-estimate the threshold. In implementation, the matching logic can be embedded in the plug-in processing flow of the pre-bidding stage: the audience identifier and audience network identifier are extracted from the real-time bidding request as the search key, and two types of blacklists are queried respectively; if hit, the request is marked as a non-intention request, the marking result is written into the request reply information, and the marked result is directly discarded or bypassed to make the request not enter the auction process; if not hit, the request is regarded as an intention request and sent to the auction process for bidding. This way takes the blacklist output by the offline stage as a lightweight index for online decision, and only key-value matching and shunting actions are performed on the online side, so the additional delay is very small, and the bidding calculation and bidding opportunity consumption caused by invalid requests can be significantly reduced, achieving the purpose of adaptive allocation of marketing bidding resources and reducing real-time calculation overhead.

[0059] Based on the same idea, as shown in Figure 2 , a marketing bidding resource adaptive allocation system is provided, comprising: A bidding access module 201 is configured to receive a bidding request and parse the audience identifier, audience network identifier and access address identifier of the bidding request; A log retrieval module 202 is configured to retrieve historical bidding logs corresponding to the audience identifier and the audience network identifier within a preset observation period, and aggregate the access behavior statistics respectively, wherein the access behavior statistics include daily request size, time period level occurrence, request size in a short time window and access address diversity; An anomaly judgment module 203 is configured to generate an anomaly judgment feature set from the access behavior statistics, wherein the anomaly judgment feature set at least includes a daily request proportion feature, an occurrence duration feature, a short-time burst feature and an access address diversity ratio feature; A threshold definition module 204 is configured to statistically estimate the anomaly judgment feature set to determine a threshold set; A maintenance module 205 is configured to compare the anomaly judgment features of each audience identifier and audience network identifier with the threshold set, mark as an abnormal audience and write into a blacklist if a preset anomaly condition is met, and mark as a regular audience if not met, and set a retention policy for the blacklist to be deleted when the entry exceeds a preset non-appearance period; A distribution control module 206 is configured to, in the pre-bidding processing stage, if the audience identifier or the audience network identifier of the real-time arrived bidding request hits the blacklist, determine the bidding request as a non-intention request and reduce or cancel the allocation of marketing bidding resources to participate in bidding, otherwise, determine the bidding request as an intention request and allocate the marketing bidding resources to participate in bidding; The output module 207 is configured to output the determination result and the resource allocation result, and write back to the historical bidding log to update the blacklist.

[0060] Compared with the processing mode of relying on media side clustering or static rules only, the system starts from the perspective of the audience, extracts features reflecting abnormal access behavior, sets corresponding threshold and rules, realizes the identification and sedimentation of abnormal audience, and has the ability to identify both group abnormality and individual abnormality. Meanwhile, the system can be integrated into the demand side real-time processing link as a plug-in in the pre-bidding stage, realize rapid identification through blacklist matching, mark and discard or bypass non-intentional bidding requests, and only send intentional requests into the bidding process, so as to reduce invalid bidding and invalid calculation, and improve the use efficiency of marketing bidding resources and the overall delivery performance.

[0061] The specific details of each module / unit in the system have been described in detail in the method part, and the undisclosed details can be referred to the method part, and thus will not be described again.

[0062] Based on the same idea, the embodiments of the present specification also provide a marketing bidding resource adaptive allocation device, as shown in Figure 3 .

[0063] The marketing bidding resource adaptive allocation device can be a terminal device or a server provided by the above embodiments.

[0064] The marketing competitive bidding resource adaptive allocation device can have a large difference due to configuration or performance, and can include one or more processors 301, memories 302, and buses. The memories 302 can store one or more applications or data. The memories 302 can include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) and / or a cache memory, for example, a plug-in mobile hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, and the like, which are equipped on an electronic device, and can further include a read-only storage unit. The applications stored in the memories 302 can include one or more program modules (not shown in the figure), and such program modules include, but are not limited to, an operating system, one or more applications, other program modules, and program data, each of which or a combination thereof can include an implementation of a network environment. Further, the processor 301 can be configured to communicate with the memories 302 to execute a series of computer-executable instructions in the memories 302 on the marketing competitive bidding resource adaptive allocation device. The marketing competitive bidding resource adaptive allocation device can further include one or more power supplies 303, one or more wired or wireless network interfaces 304, one or more I / O interfaces (input / output interfaces) 305, one or more external devices 306 (such as a keyboard) for communication, and can further communicate with one or more devices that enable a user to interact with the device, and / or any device that enables the device to communicate with one or more other computing devices (such as a router, a network switch, and the like). Such communication can be performed through the I / O interface 305. Further, the device can also communicate with one or more networks (such as a local area network (LAN)) through the wired or wireless interface 304.

[0065] Figure 3 Only the marketing competitive bidding resource adaptive allocation device with components is shown, and those skilled in the art can understand that, Figure 3 The structure shown does not constitute a limitation on the marketing competitive bidding resource adaptive allocation device, and can include fewer or more components than shown, or combine certain components, or different component arrangements.

[0066] In particular, in this embodiment, the marketing bidding resource adaptive allocation device comprises a memory and one or more programs, wherein one or more programs are stored in the memory, and the one or more programs can comprise one or more modules, and each module can comprise a series of computer executable instructions in the marketing bidding resource adaptive allocation device, and the one or more programs configured to be executed by one or more processors include computer executable instructions for: receiving a bidding request and parsing audience identification, audience network identification and access address identification of the bidding request; calling historical bidding logs corresponding to the audience identification and the audience network identification in a preset observation period, and respectively aggregating access behavior statistics including daily request size, time period level occurrence, request size in a short time window and access address diversity; generating an abnormality judgment feature set from the access behavior statistics, the abnormality judgment feature set at least including daily request proportion feature, occurrence time length feature, short-time burst feature and access address diversity ratio feature; statistically estimating the abnormality judgment feature set to determine a threshold set; comparing the abnormality judgment features of each of the audience identification and the audience network identification with the threshold set, marking as an abnormal audience and writing into a blacklist if a preset abnormality condition is met, and marking as a regular audience if not met, and setting a retention strategy for the blacklist to be deleted when an entry exceeds a preset non-appearance period; in a bidding pre-processing stage, if the audience identification or the audience network identification of the bidding request arriving in real time hits the blacklist, the bidding request is determined as a non-intention request and the marketing bidding resource allocation is reduced or cancelled to not participate in bidding, otherwise it is determined as an intention request and the marketing bidding resource is allocated to participate in bidding; outputting a determination result and a resource allocation result, and writing back to the historical bidding log to update the blacklist.

[0067] Based on the same idea, the exemplary embodiments of the present application also provide a computer readable storage medium having a program product stored thereon, which can implement the above-mentioned method of the present application. In some possible embodiments, various aspects of the present application can also be implemented in the form of a program product, which includes program code for causing a terminal device to execute the steps described in the above-mentioned “exemplary method” section according to various exemplary embodiments of the present application when the program product is run on the terminal device.

[0068] Reference Figure 4As shown, a program 400 for implementing the above-described method according to an exemplary embodiment of the present application is described, which can take a portable compact disc read-only memory (CD-ROM) and include a program code, and can be executed on a terminal device, such as a personal computer. However, the program product of the present application is not limited thereto, and in the present document, the readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0069] The program product can take any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium, for example, can be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the above. More specific examples (a non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0070] The computer readable signal medium can include a data signal propagated in baseband or propagated as a carrier wave, in which the readable program code is embodied. Such propagated data signal can take various forms, including but not limited to electro-magnetic, optical or any suitable combination thereof. The readable signal medium can also be any readable medium that is not a readable storage medium and that can transmit, propagate or transport for use by or in connection with an instruction execution system, apparatus, or device program code.

[0071] The program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, CSS, HTML, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device or entirely on the remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, such as through the Internet using an Internet service provider.

[0072] Those skilled in the art can easily understand from the above description of the embodiments that the example embodiments described herein can be implemented by software or by software in combination with necessary hardware. Therefore, the technical solutions according to the embodiments of the present application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.) or on a network, and includes a number of instructions to enable a computing device (which can be a personal computer, a server, a terminal system, or a network device, etc.) to execute the method according to the example embodiments of the present application.

[0073] In addition, the above-described figures are only schematic illustrations of the processes included in the method according to the example embodiments of the present application, and are not intended to be limiting. It is easy to understand that the processes shown in the above-described figures do not indicate or limit the time sequence of the processes. In addition, it is also easy to understand that the processes can be executed synchronously or asynchronously, for example, in a plurality of modules.

[0074] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, such a division is not mandatory. Indeed, according to the example embodiments of the present application, the features and functionalities of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functionalities of one module or unit described above can be further divided into embodied by a plurality of modules or units.

[0075] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of the application following, in general, the principles of the application and including such departures from the present disclosure as come within known use or customary practice within the art. The specification and examples are to be regarded as exemplary only, and the true scope and spirit of the application is indicated by the appended claims.

[0076] It should be understood that the present application is not limited to the precise structures described hereinabove and illustrated in the accompanying drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims

1. A method for adaptive allocation of marketing bidding resources, characterized in that, The method includes: Receive bidding requests and parse them to obtain the audience identifier, audience network identifier, and access address identifier of the bidding request; Within a preset observation period, historical bidding logs corresponding to the audience identifier and the audience network identifier are retrieved and aggregated to obtain access behavior statistics. The access behavior statistics include daily request scale, time period occurrence, request scale within a short window, and access address diversity. An anomaly detection feature set is generated from the access behavior statistics. The anomaly detection feature set includes at least the daily request ratio feature, the occurrence duration feature, the short-term burst feature, and the access address diversity ratio feature. Statistical estimation is performed on the anomaly detection feature set to determine the threshold set; The abnormal judgment features of each audience identifier and the audience network identifier are compared with the threshold set. Those that meet the preset abnormal conditions are marked as abnormal audiences and written into the blacklist. Those that do not meet the conditions are marked as normal audiences. A retention policy is set for the blacklist to remove entries when they exceed a preset period of non-appearance. During the pre-bidding processing stage, if the audience identifier or audience network identifier of the bidding request that arrives in real time matches the blacklist, the bidding request is determined to be an unintentional request and the allocation of marketing bidding resources is reduced or canceled so as not to participate in the bidding; otherwise, it is determined to be an intentional request and the marketing bidding resources are allocated to participate in the bidding. Output the judgment result and resource allocation result, and write them back to the historical bidding log to update the blacklist.

2. The adaptive allocation method for marketing bidding resources according to claim 1, characterized in that, The marketing bidding resources include one or more of the following: budget resources, bidding opportunity resources, and bidding calculation resources; The audience identifier is used to distinguish between terminal or browser instances, and the audience identifier is updated when the terminal identifier information changes; the audience network identifier is formed by combining network address information and client identifier information to maintain the distinguishability of the audience even when the network address is reused; the bidding request also includes publisher information, which includes a domain name identifier and a publisher identifier. The domain name identifier is obtained by resolving the access address identifier, and the publisher identifier is used to uniquely identify the media party or the ad space provider.

3. The adaptive allocation method for marketing bidding resources according to claim 1, characterized in that, The generation of the daily request percentage feature includes: Within the daily request scale, the number of bidding requests generated by each audience identifier or audience network identifier within the daily time window is counted, and the ratio of the number of bidding requests to the total number of bidding requests within the same daily time window is calculated to obtain the daily request percentage. The daily request percentage is compared with the percentage threshold. When the daily request percentage exceeds the percentage threshold that matches the corresponding identifier type, it is determined that the preset abnormal condition is met.

4. The adaptive allocation method for marketing bidding resources according to claim 1, characterized in that, The generation of the occurrence duration feature includes: The daily time window is divided into multiple consecutive time period units. The number of time periods in which each audience identifier or audience network identifier has a bidding request within the time period unit is counted as the occurrence duration. The occurrence duration is compared with the occurrence duration threshold. When the occurrence duration reaches or exceeds the occurrence duration threshold, it is determined that the preset abnormal condition is met. The time period unit is a time granularity unit at the hour or minute level.

5. The adaptive allocation method for marketing bidding resources according to claim 1, characterized in that, The generation of the short-term burst feature includes: A sliding time window at the second level is used to count the bidding requests of each audience identifier or audience network identifier within the sliding time window to obtain the number of short-time window requests; and the number of short-time window requests is aggregated within the observation period to obtain at least one of the maximum burst value, the average burst value, or the burst percentile value; when the maximum burst value, the average burst value, or the burst percentile value exceeds the burst threshold, it is determined that the preset abnormal condition is met.

6. The adaptive allocation method for marketing bidding resources according to claim 1, characterized in that, The statistics on the diversity of access addresses include: The number of different access address identifiers associated with each audience identifier or audience network identifier during the observation period is counted, and the access address diversity ratio is obtained by calculating the ratio between the number of different access address identifiers and the daily request scale. When the access address diversity ratio is lower than the diversity threshold, the preset abnormal condition is determined to be met; and / or, when the correspondence between the number of different access address identifiers and the number of bidding requests deviates from the preset monotonic relationship, the preset abnormal condition is determined to be met, wherein the preset monotonic relationship is used to characterize the normal access behavior in which the number of different access address identifiers increases synchronously with the increase of the number of bidding requests.

7. The adaptive allocation method for marketing bidding resources according to claim 1, characterized in that, The determination of the threshold set includes: A cumulative distribution or quantile statistic is constructed for the values ​​of the anomaly determination feature set within the observation period, and a boundary point is selected as a threshold in the tail region of the cumulative distribution statistic, so that the audience falling into the main distribution region is judged as normal and the audience falling into the tail region is judged as abnormal; and independent threshold sets are established for the audience identifier and the audience network identifier respectively, to adapt to the differences in access behavior of different identifier types.

8. The adaptive allocation method for marketing bidding resources according to claim 1, characterized in that, The blacklist is updated using a combination of offline building and online application, specifically including: After each observation period ends, an abnormal audience list is generated based on the historical bidding logs of the observation period, and the abnormal audience list is appended to the existing blacklists of the audience identifier and the audience network identifier to form a cumulative blacklist. During appending, the most recent occurrence time of each entry in the blacklist is recorded to implement the retention policy. In the real-time processing phase, the bidding requests of the current period are matched only using the cumulative blacklist from the most recent completed observation period. When the bidding request is determined to be a non-intent request, the determination result is written to the bidding response and discarded or bypassed. When the bidding request is determined to be an intent request, the bidding request is sent to the bidding process to achieve adaptive allocation of the marketing bidding resources and reduce real-time computing overhead.

9. A marketing bidding resource adaptive allocation system, characterized in that, include: The bidding access module is used to receive bidding requests and parse them to obtain the audience identifier, audience network identifier, and access address identifier of the bidding request. The log retrieval module is used to retrieve historical bidding logs corresponding to the audience identifier and the audience network identifier within a preset observation period, and aggregate them to obtain access behavior statistics. The access behavior statistics include daily request scale, time period occurrence, request scale within a short window, and access address diversity. The anomaly detection module is used to generate an anomaly detection feature set from the access behavior statistics. The anomaly detection feature set includes at least the daily request ratio feature, occurrence duration feature, short-term burst feature, and access address diversity ratio feature. The threshold definition module is used to perform statistical estimation on the anomaly detection feature set to determine the threshold set; The maintenance module is used to compare the anomaly determination features of each audience identifier and the audience network identifier with the threshold set, mark those that meet the preset anomaly conditions as abnormal audiences and write them into the blacklist, mark those that do not meet the conditions as normal audiences, and set a retention policy for the blacklist to remove entries when they exceed a preset period of non-occurrence. The allocation control module is used to determine, during the pre-bidding processing stage, if the audience identifier or audience network identifier of the bidding request that arrives in real time matches the blacklist, then the bidding request is determined to be a non-intent request and the allocation of marketing bidding resources is reduced or canceled so that it does not participate in the bidding; otherwise, it is determined to be an intent request and the marketing bidding resources are allocated to participate in the bidding. The output module is used to output the judgment result and resource allocation result, and write them back to the historical bidding log to update the blacklist.

Images