EP2005382A2 - Évaluation de la qualité du trafic vers des sites réseau au moyen de paramètres de trafic interdépendants - Google Patents

Évaluation de la qualité du trafic vers des sites réseau au moyen de paramètres de trafic interdépendants

Info

Publication number
EP2005382A2
EP2005382A2 EP07758955A EP07758955A EP2005382A2 EP 2005382 A2 EP2005382 A2 EP 2005382A2 EP 07758955 A EP07758955 A EP 07758955A EP 07758955 A EP07758955 A EP 07758955A EP 2005382 A2 EP2005382 A2 EP 2005382A2
Authority
EP
European Patent Office
Prior art keywords
agent
sessions
traffic
network site
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP07758955A
Other languages
German (de)
English (en)
Other versions
EP2005382A4 (fr
Inventor
Joel Berman
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Individual
Original Assignee
Individual
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US11/567,718 external-priority patent/US10567255B2/en
Application filed by Individual filed Critical Individual
Publication of EP2005382A2 publication Critical patent/EP2005382A2/fr
Publication of EP2005382A4 publication Critical patent/EP2005382A4/fr
Withdrawn legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising

Definitions

  • the disclosed technology relates to assessing the value of traffic associated with network sites.
  • An increasing number of companies, agencies, individuals, and other parties use online advertising to advertise to users of Internet or other network sites or services.
  • An advertiser purchases advertising space from an individual publisher or from an advertising network that distributes advertisements to one or more publishers.
  • a publisher or advertising network may charge the advertiser using one of several methods, including cost-per-click and cost-per-impression.
  • cost-per-click system an advertiser is charged based on the number of times that agents click on its advertisement.
  • An advertiser is not charged when a publisher displays an advertisement to an agent unless the agent clicks on the advertisement.
  • a cost-per-impression system an advertiser is charged based on the number of times a publisher displays its advertisement to an agent.
  • Click fraud or fraudulent clicks on advertisements, is an issue that concerns advertisers and publishers who use cost-per-click and other payment models.
  • impression fraud or displays of advertisements in situations where the advertisements will not make an impression on a human user, is an issue that concerns advertisers and publishers who use cost-per-impression and other payment models.
  • Click or impression fraud can take a number of forms, including clicks on an advertisement by or displays of an advertisement to competitors, web robots, or users with personal or political agendas.
  • an adware or clickware virus may install itself on a computer and generate clicks on or impressions of advertisements without the computer user's knowledge. Fraudulent clicks or impressions do not generate revenue or other value for an advertiser; however, the advertiser must pay for the clicks or impressions. Click or impression fraud therefore harms the advertiser by increasing advertising expense, and at the same time harms the publisher by lowering the perceived value of traffic the advertiser receives from the publisher.
  • click or impression fraud detection systems classify each click or impression in a binary manner as either "good” or “bad.” Publishers may use the results of click or impression fraud detection systems in a number of ways. In some cases, a publisher may subtract bad clicks or impressions from the total number of clicks or impressions, charging an advertiser for only good clicks or impressions.
  • Binary click or impression fraud detection systems have several drawbacks. A click or impression may not fall neatly into either the good or bad category, or it may be impossible to determine from the data set that represents the click or impression whether in fact the click or impression is good or bad.
  • a binary approach will therefore unfairly characterize those clicks or impressions that fall somewhere in between.
  • advertisers may have differing thresholds as to the type of traffic they are willing to accept.
  • One advertiser may consider a user simply viewing its web site as a valuable transaction; another advertiser may only consider a purchase to be a valuable transaction.
  • a binary system does not allow an advertiser to set a level that determines the quality of traffic for which it is willing to pay the publisher. Advertisers and publishers alike would therefore benefit from having a more accurate system of click or impression fraud detection in order to better assess the value of traffic to publisher sites.
  • Figure 1 is a block diagram of a representative facility for scoring the quality of network traffic and an environment in which the facility operates.
  • Figure 2 is a flow diagram of a method of computing the quality of network traffic.
  • Figure 3 is a flow diagram of a method of computing a correlation between a rule set that is used to assess the quality of traffic and a desired agent action.
  • Figure 4 is a block diagram of a data structure used to compute the correlation between each rule in the rule set used to assess the quality of traffic to a network site and a desired agent action.
  • Figure 5 is a flow diagram of a method of scoring an agent action based on a rule set.
  • Figure 6 is a flow diagram of a method of generating training set data.
  • Figure 7 is a flow diagram of a method of identifying correlated parameters that characterize traffic associated with network sites.
  • Figure 8 is a block diagram of a data structure used to identify correlated parameters that characterize traffic associated with network sites, the data structure depicted prior to processing data characterizing the traffic.
  • Figure 9 is a block diagram of a data structure used to identify correlated parameters that characterize traffic associated with network sites, the data structure depicted after processing data characterizing the traffic.
  • a software and/or hardware facility for scoring the quality of traffic to a site accessible via the Internet or other network is described.
  • the facility extracts session data, or information identifying an agent's interactions with a server, from one or more server logs or other data sources that are obtained from a publisher, advertiser, or third party.
  • the facility may obtain supplemental data from external data sources that assists in interpreting the agent's interactions with the server.
  • a multi-factor analysis in the form of a rule set is applied by the facility to the session data. The analysis of the session data identifies agent actions that are desirable to a publisher, advertiser, or third party.
  • Agent actions that are desirable to a publisher, advertiser, or third party include any activity that generates value for the publisher, advertiser, or third party, such as a click, a conversion (e.g., purchase), a submission of a form, bookmarking of the site, a rollover event, an impression, or other activity by the user.
  • the facility generates a relative score for each agent action or for an aggregate number of agent actions based on whether the agent action is desired by the publisher, advertiser, or third party. The score may be used to assess the quality of the traffic received by a network site. Lower scores are indicative of fraudulent, likely fraudulent, or otherwise non-productive traffic having little value, whereas higher scores are indicative of traffic having desirable characteristics and therefore greater value.
  • the score generated by the facility may be provided to the publisher or advertising network that published an advertisement.
  • the publisher or advertising network may use the score for a number of purposes. For example, a publisher or advertising network may elect to use a pricing method that charges an advertiser a variable amount based on the score of the traffic an advertisement receives.
  • the publisher or advertising network may decide not to charge for traffic that falls below a certain threshold. For example, certain sites desiring to market the high value of their traffic may elect to charge only for traffic having a score reflective of a low likelihood of fraud.
  • a publisher or advertising network that receives a score on a real-time basis may decide not to display advertisements to agents that have a score indicating that the agent poses a high risk of fraud.
  • the rule set used by the facility to generate the score may be manually determined.
  • Rules used to determine the desirability of an agent action may include the physical location of the agent, the agent's browsing habits, search terms entered by the agent, rates charged to the advertiser for each agent action, the network topology from which the agent originated, and other characteristics. Rules may also be automatically determined by the facility based on the correlation between a trait of a given agent, advertiser, publisher, or other feature, and a measurement of the resulting fraudulent traffic associated with that trait. Different rule sets may be generated that are optimized for different environments Within each rule set, each rule may be weighted differently to generate an optimum combination of rules to comprise the rule set, and rules may be selectively removed if they do not perform well in an environment.
  • the facility generates training set data for use in scoring the quality of traffic to network sites.
  • the facility designs and executes one or more experimental advertising campaigns, each of which is likely to attract fraudulent or non-fraudulent agents.
  • the facility receives one or more server or other data logs corresponding to the session traffic generated by the advertising campaigns.
  • the facility standardizes the server logs and places the server log data in a uniform format.
  • the facility selects sessions from the standardized server log data that fall outside of a normal range. Within these selected sessions, clusters of sessions are further identified, and each cluster is characterized as representing fraudulent, non-fraudulent, or unknown traffic.
  • a false positive detection algorithm is applied to each cluster, eliminating those clusters falsely identified as indicating fraudulent or non-fraudulent traffic.
  • the facility adds clusters identified as indicating fraudulent or non-fraudulent traffic to the training set.
  • the training set data may be used as described herein to identify traffic parameters that are indicative of fraudulent traffic.
  • the facility scores the quality of traffic to a network site by evaluating various parameters associated with traffic to the site and determining combinations of traffic parameters that are indicative of the quality of traffic to the site.
  • the facility receives server log data or other data sources representing session traffic associated with one or more network sites.
  • the facility selects n rules that have been identified as statistically significant for the network site, each rule measuring one or more traffic parameters.
  • the facility applies each rule to the data associated with a session and generates a result for each rule.
  • the combination of all rule results for a particular session is referred to as a result vector.
  • the facility uses an association table having a plurality of rows, where each row represents a unique combination of rule results (i.e., a unique result vector).
  • the facility maps the generated result vector for each session to the corresponding row in the association table and records whether the transaction associated with the session is fraudulent or non-fraudulent.
  • a fraud odds ratio which represents how fraudulent sessions having a particular result vector are in comparison to sessions having different result vectors, is computed for each result vector.
  • the facility parses the association table to identify those result vectors having the highest fraud odds ratio.
  • the facility may optionally cross-validate the identified result vectors by applying the rules represented by a result vector to traffic that has not yet been processed by the facility. The facility may retain only those combinations of rules that produce consistent results, i.e., indicate similar proportions of fraudulent transactions.
  • the facility may match the generated result vectors of the traffic to the corresponding row or rows in the association table and to the associated fraud odds ratio. Thus, the facility can estimate whether and to what degree the new traffic is likely to represent fraudulent or non-fraudulent transactions.
  • FIG. 1 is a block diagram illustrating the components of a traffic scoring facility 100 and a representative environment in which the traffic scoring facility operates.
  • the traffic scoring facility 100 extracts data from publishers, advertisers, agents, and/or other parties to generate a relative score of the quality of traffic received by a site accessible by a network (hereinafter "a publisher site").
  • the generated score relates to the value of the traffic to that site as characterized by one or more parties associated with the site.
  • the publisher site may be a site that displays online advertisements, and the desired agent on that site may be a human user that would benefit from viewing advertisements, rather than robotic traffic or human users with fraudulent intent.
  • the score generated by the facility would thereby relate to the likelihood that the agent was a human user with non- fraudulent intent.
  • the publisher site includes a publisher server 105 and a server log 125.
  • the publisher server 105 serves one or more pages 120 containing content to an agent that interacts with the site.
  • An agent is any human, software entity, or device that interacts with the site, including, but not limited to, a human user, a human user with fraudulent intent, a robot, spyware, an adware or clickware virus, or improperly operating software.
  • the server log 125 stores session data associated with the agent's interactions with the publisher server 105.
  • the publisher site 105 may be any type of site accessible via a network such as a search engine site, a directory site, a news or other content site, a social networking site, a photo sharing or other service site, or any other online property.
  • a publisher site will typically include or utilize the services of an advertising service 165 to populate the publisher page 120 that it produces with one or more advertisements.
  • An advertiser 180 may purchase advertising placements on a publisher site in a variety of ways. For example, the advertiser may purchase keywords from the advertising service 165 that are relevant to the products and/or services that the advertiser offers.
  • the advertising service 165 generates a list 122 of advertisements. The list may be generated by matching the keywords of a search query or a content page to advertiser-keyword pairings in its advertising database 170.
  • the list 122 of advertisements generated by the advertising service 165 may be displayed on the publisher page 120, often above or alongside content provided by the publisher site.
  • an agent takes an action associated with an advertisement displayed on a publisher page 120
  • the agent is directed to an advertiser site 135.
  • the agent may interact with the site, such as by viewing content, purchasing products or services, and other activities.
  • data identifying an agent's actions is stored in a server log 140.
  • the data in the server log may be utilized to characterize the value of the agent's interaction with the advertiser site.
  • a record of the agent action is maintained by the advertising service 165 and the advertiser may be charged for the agent action. For example, if the advertising service relies on a cost- per-click model, the advertiser will be charged for an agent click on an advertisement. If the advertising service relies on an impression model, the advertiser will be charged for each time an advertisement is displayed to an agent.
  • a billing component 175 may deduct an amount from an advertiser's account equal to the number of agent actions multiplied by the value the advertiser has agreed to pay for each agent action.
  • limits may be set by the advertiser on the amount of its advertising spending, such as setting a maximum amount to be paid for each agent action or an aggregate amount to be paid within a certain time period.
  • traffic to the publisher server 105 over a network 130 is of value to the publisher and to advertisers that advertise on the publisher site.
  • human users 145 may interact with the publisher server 105 and take action associated with advertisements in which they have an interest, leading to the possibility of monetary or other gain for advertisers.
  • a portion of the traffic to the publisher server 105 may come from sources that are not valuable to the publisher or advertisers.
  • robots 150 may arrive at a publisher site and click on advertisements in search of email addresses or other data. Such robotic traffic increases the load on the publisher site, while at the same time fails to generate valuable traffic for advertisers.
  • Other traffic 155 with fraudulent intentions can also reach the publisher site.
  • Fraudulent traffic 155 does not generate valuable traffic for an advertiser and may also be of little value to the publisher.
  • a traffic scoring facility 100 may be operated by, or provided as a service to, publishers or advertisers. Data from server logs or other sources may be submitted to the traffic scoring facility 100. Data may also be submitted directly from the publisher server 105 to the traffic scoring facility 100 in real time. As will be described in additional detail herein, the traffic scoring facility applies one or more rule sets stored in data store 160 to the received data to generate a score that, on a per agent action or on an aggregate agent action basis, characterizes the traffic. The score reflects the anticipated value of the traffic, i.e., "good" traffic that is likely to be of value will receive a higher score and "bad" traffic that is likely to be of little value will receive a lower score.
  • the traffic scoring facility 100 may provide the score to the billing component 175 of the advertising service.
  • the advertising service may utilize the score in a variety of ways. For example, the advertising service may elect to use a pricing method that charges an advertiser 180 a variable amount based on the score of the traffic its advertisement receives. Traffic having a higher score may be charged to the advertiser at a higher rate, while traffic having a lower score may be charged to the advertiser at a lower rate. Alternatively, a publisher may agree not to charge an advertiser for traffic that does not exceed a certain score. In this manner, an advertiser has greater control over the quality of traffic to which it advertises.
  • an advertiser may determine whether to continue using an advertising service 165 or a particular publisher based on the quality of traffic it receives. Based on the score of the traffic it receives, an advertiser may also decide to make adjustments to the keywords it purchases from the advertising service 165 or to its methodology for placing advertisements. In addition, an advertiser or publisher may use the score to assess the damage or loss of revenue resulting from low quality traffic.
  • the traffic scoring facility 100 utilizes one or more rule sets stored in data store 160 to rate the quality of traffic an advertisement receives.
  • the facility analyzes traffic in one or more environments and selects an optimum set of rules (an "environment rule set") that may be used to score traffic within each environment.
  • an environment rule set is applied to traffic from an environment to score traffic in that environment.
  • FIG. 2 is a flow chart of a process 182 for generating and validating rule sets that may be used to assess the value of traffic in various environments.
  • a global rule set is generated.
  • Each rule in the global rule set is an expression that receives as input one or more parameters associated with an agent's session.
  • the rule When the rule is applied by the facility to the input parameters, it produces a result that reflects the value of an agent's actions associated with that agent's session.
  • Rules may measure agent activities, including the speed with which an agent clicks on an advertisement (i.e., velocity), the time an agent spends on an advertiser's site, or the length or number of keywords an agent enters as search terms. Rules may also measure other characteristics of the agent.
  • one rule may score IP addresses and maintain a "blacklist" of IP addresses that generate low quality traffic.
  • the blacklist may contain IP addresses, scores generated by traffic originating from those IP addresses, and other details.
  • the facility may evaluate an IP address associated with a new agent by referencing the blacklist. Agents having IP addresses on the blacklist may be scored appropriately, and agents having an IP address in close proximity to IP addresses on the blacklist may be scored based on a function of the numerical distance between the agent's IP address and blacklisted IP addresses. Rules may also measure publisher and advertiser characteristics, including where a publisher resides within an advertising network hierarchy or the amount of money an advertiser is charged for each agent action associated with an advertisement.
  • a sophisticated rule related to a publisher distribution partner may measure whether more than a certain percentage (e.g., >80%) of the IP addresses associated with the publisher have multiple user agents associated with those IP addresses as measured on the day that a click from that distribution partner was received. Such a characteristic is often indicative of traffic with little or no value to an advertiser.
  • Each rule in the rule set may be phrased in a manner that allows the rule to be applied and the result to be expressed in binary form (i.e., "1" if the rule is satisfied or "0" if the rule is not satisfied).
  • a rule may be defined as a physical location of an agent, with a result of "1” if the physical location of the agent is in the United States and a result of "0" if the physical location of the agent is not in the United States.
  • the rule may be phrased in a manner such that the result may be expressed as having a value that varies from 0 to N (e.g., when N equals four, the result may take a value of 0, 1 , 2, 3, or 4).
  • the global rule set may be generated in a variety of ways.
  • the global rule set may be manually selected by an operator of the traffic scoring facility based on observed patterns of fraudulent traffic.
  • the facility may also automatically generate rules as it identifies environment features that correlate with fraudulent traffic.
  • the total number of rules in the global rule set may range from the tens to the hundreds or more.
  • the global rule set is not static; rather, it can be supplemented and modified over time. Ineffective rules may be removed from the global rule set, just as new rules may be added as they are generated or found beneficial.
  • the facility selects a traffic data set representative of traffic in a particular environment.
  • An environment may be any site, resource, or service having traffic that shares similar characteristics.
  • an environment may be a search web site that displays advertisements in conjunction with search results, an ecommerce web site that sells the products of an advertiser, or a content provider that offers a variety of services for others.
  • the traffic data set may be represented in the form of a server log, log file, or other data format that allows various parameters associated with the traffic to be evaluated.
  • a training set of data that is believed to be reflective of the actual data may be used.
  • a default training set of data may be used if no other agent action data is available.
  • the facility analyzes the global rule set to assess the ability of each rule in the global rule set to predict the likelihood of valuable traffic in the selected environment. That is, a determination is made as to the likelihood that a desired agent action will occur for each result permutation when a rule is applied to traffic associated with the environment.
  • Figure 3 is a flow chart of a process 200 that is implemented by the traffic scoring facility 100 to determine the correlation between each rule in the global rule set and a desired agent action.
  • the facility may rely on one or more server logs or other data sources created and maintained by an advertiser or publisher, as well as one or more sources of external data provided by third parties that may be used to further characterize or enhance the content in the server log.
  • the traffic scoring facility 100 receives a server log, which may be an advertiser server log 140 or a publisher server log 135.
  • a server log which may be an advertiser server log 140 or a publisher server log 135.
  • the facility determines whether there are other related server logs that may be used in the analysis. If other server logs exist, at a block 210 the related server log or logs are retrieved and reconciled with the first server log.
  • the traffic scoring facility determines whether there is supplemental external data that may assist in interpreting a server log. Supplemental external data may include information from an advertiser about whether a conversion (e.g., purchase) or other desired agent action occurred, information from a publisher such as statistical information on advertisement performance, information relating to a geographical mapping of an IP address, and other types of information. If there is supplemental external data, the processing proceeds to a block 220 where the external data is retrieved.
  • a server log retrieved from a publisher may not contain information about whether a conversion occurred, because a conversion occurs after an agent is transferred from the publisher site to an advertiser site. This data may only be available from an advertiser.
  • data from the server log may be enhanced by data retrieved from the advertiser about whether a conversion occurred.
  • the IP address associated with a particular agent is known, the IP address may be correlated with a geographical database to determine the rough geographical area from which the agent is accessing the site.
  • Many types of supplemental external data can be used to enhance the data received from a server log.
  • a session may be defined as one or more entries in the server log or other data source indicative of an agent's interaction with a network site.
  • a session may contain no clicks, one click, or multiple clicks that occur as an agent interacts with the network site, and a session may span one or more visits to the network site within a period of time (e.g., within an hour, within 24 hours).
  • the facility may therefore apply one or more algorithms.
  • an agent may be identified by the first three bytes of its IP address in combination with the User Agent ID identified by the server log.
  • an agent may be identified by a combination of the User Agent ID identified by the server log, the last two parts of the domain name associated with the IP address (or the first two bytes of the IP address if no domain name can be found), and, if there are two or more agent actions, the elapsed time between the agent actions.
  • a desired agent action is an action taken by an agent that generates value for the advertiser, such as monetary or other gain. Desired agent actions may be defined by the party generating the traffic score or by the party that will be receiving the traffic score. For example, if the desired action to be taken by a agent is the purchase of a product or service, a desired agent action may be defined as paying for the product or service. As another example, if the desired action to be taken by an agent is the viewing of a particular media clip, a desired agent action may be defined as the downloading of media content to an agent during a session.
  • a desired agent action may be a click on an advertisement that is displayed to the agent.
  • a desired agent action may be globally recognized as beneficial by a large group of parties (e.g., the purchase of a product from a retailer or a human user), or more narrowly recognized as beneficial to a single party (e.g., the viewing of a trailer for a new movie at the site of the movie studio that is producing the movie).
  • FIG. 4 is a block diagram of a representative table 300 in which the results may be stored.
  • Each row 310a, 31 Ob 1 ...31 On in the table corresponds to one rule in the global rule set (i.e., from rulei to rule n ).
  • the first four columns in the table are used to record the result as each rule is applied to the session data.
  • Columns 315a and 315b correspond to sessions that resulted in desired agent actions.
  • Column 315a reflects desired agent actions that occur when the result of the rule is "0" (i.e., the rule was not satisfied).
  • Column 315b reflects desired agent actions that occur when the result of the rule is "1" (i.e., the rule was satisfied).
  • a count is maintained in each column, and is incremented to reflect a result that falls within that column.
  • Columns 320a and 320b correspond to sessions that did not result in desired agent actions.
  • Column 320a reflects sessions where a desired agent action does not occur and the result of the rule is "0" (i.e., the rule was not satisfied).
  • Column 320b reflects sessions where a desired agent action does not occur and the result of the rule is "1" (i.e., the rule was satisfied).
  • Figure 4 shows the results of rulei for 10,000 agent sessions stored in row 310a. While only four columns are depicted in table 300 to reflect the results of applying a rule, those skilled in the art will appreciate that the number of columns may be any number that reflects the number of potential results for a rule. For example, a rule with three results (“0,” “1 ,” or "2") may require six columns - three devoted to the sessions that result in desired agent actions, and three devoted to sessions that do not result in desired agent actions.
  • Figure 4 depicts a table whose contents and organization are designed to make it more comprehensible to the reader, those skilled in the art will appreciate that the actual data structure used by the facility to store this information may differ from the table shown.
  • the table may be organized in a different manner, may contain more or less information than shown, may be compressed and/or encrypted, and may otherwise be optimized in a variety of ways.
  • a test is made by the facility to determine if there are any more rules in the global rule set to apply. If additional rules exist, processing loops to block 255 to process the next rule in the list. If additional rules do not exist, processing continues to a decision block 275.
  • the facility determines if additional sessions remain in the server log to be processed. If additional sessions exist that are to be processed, the facility returns to block 245 to select the next session in the log. The process defined by blocks 245 through 270 is thereby repeated as many times as necessary to process the session information of all agent sessions as reflected in the server log or other data log. If no additional sessions remain to be processed at block 275, processing continues to a block 280.
  • the rule set is applied only to those entries in a server log that correspond to a click or impression.
  • An entry in a server log that corresponds to a click or impression may be referred to as an atomic session.
  • each rule in the rule set is applied only to each atomic session. This application of the rule set is particularly advantageous in the case of scoring clicks on or impressions of advertisements.
  • the facility calculates a correlation coefficient that is representative of the likelihood that the session parameters tested by the rule will indicate a desired agent action.
  • a correlation coefficient is representative of the likelihood that the session parameters tested by the rule will indicate a desired agent action.
  • an algorithm is applied to the session results stored in table 300.
  • a statistically significant number of agent actions must have been measured that resulted in both sessions in which a rule was satisfied and sessions in which a rule was not satisfied in order for a correlation coefficient to be calculated. If there are fewer than a certain number of agent actions (e.g., 50) where the rule is satisfied or, alternatively, fewer than a certain number of agent actions where the rule is not satisfied, the correlation coefficient is set to zero.
  • a rule for a correlation coefficient to be calculated, a rule must result in a statistically significant number of both desired agent actions and non-desired agent actions. If either all agent actions are valuable, alternatively, no agent actions are valuable, the correlation is set to zero. In other words, if a desired agent action either always occurs or never occurs, the rule may not be a good predictor of a desired agent action in the selected environment. If a sufficient number of agent actions are measured from both a rule satisfaction and a desired agent action standpoint, then a correlation coefficient may be calculated for that rule.
  • val_and_rule_satisfied is the percentage of the total number of agent actions in which a desired agent action occurs and the rule is satisfied (i.e., the percentage of agent actions that fall in column 315b);
  • val is the percentage of the total number of agent actions in which a desired agent action occurs (i.e., the percentage of agent actions that fall in either column 315a or 315b);
  • rule_satisfied is the percentage of the total number of agent actions in which the rule is satisfied (i.e., the percentage of agent actions that fall in either column 315b or 320b);
  • non_val is the percentage of the total number of agent actions in which a desired agent action does not occur (i.e., the percentage of agent actions that fall in either column 320a or 320b);
  • rule_ ⁇ ot_satisfied is the percentage of the total number of agent actions in which the rule is satisfied (i.e., the percentage of agent actions that fall in either column 315a or 320a).
  • This correlation equation may be applied in circumstances when the result of the rule is binary (i.e., "0” or "1") and the occurrence of a desired user action is binary (i.e., a user action either did or did not occur). In other cases, an appropriately modified or different correlation equation may be used.
  • the result of the correlation coefficient calculation for each rule is stored in column 340 of table 300.
  • correlation coefficients may be used rather than regression coefficients (computed through logistic regression) since correlation coefficients offer a more robust methodology. That is, the logistic regression approach is a poor choice when the dependent variables are highly correlated.
  • the correlation approach described herein can be viewed as a ridge regression where regression coefficients are constrained and must have the same sign as the rule correlations.
  • the facility determines whether the correlation coefficient is statistically significant for the selected environment. The facility makes this determination by selecting only those rules that have a correlation coefficient that is in excess of a certain percentage. For example, the facility may keep only those rules that have a correlation coefficient in excess of 75% as part of a rule set that is associated with the environment (hereinafter, the "environment rule set"). If analysis of the correlation coefficient indicates that it is not statistically significant for the selected environment, the rule may be omitted from the environment rule set.
  • the final column 345 in the table 300 contains a flag for each rule that may be set to a value of "0" or "1.”
  • the result is stored in column 345. If a rule has a statistically significant correlation, the rule flag value in column 345 will be set to "1.” If the rule does not have a statistically significant correlation, the rule flag value in column 345 will be set to "0.” Only those rules with a rule flag value of "1 ,” i.e., those that are statistically significant, will be used in determining a score of traffic in the selected environment. Rules that perform well for a particular environment are therefore maintained in the rule set for that environment, whereas rules that do not perform well for a particular environment are discarded.
  • the environment rule set for the particular environment including the correlation coefficients for each of the rules in the environment rule set, is stored by the facility so that it may be used again without having to re-execute the process indicated by Figures 3 and 4.
  • the facility proceeds to a block 192 to determine whether there are any additional environments for which an environment rule set should be generated. If there are additional environments, processing loops to block 186 to process the next environment.
  • the facility will have generated and stored a rule set for each identified environment. In this way, a unique rule set may be generated for each publisher, advertiser, or other characteristic, or any combination thereof. Each rule set will reflect the predictive parameters of the traffic received by that environment.
  • FIG. 5 is a flow diagram illustrating a process 500 of computing a score for a new agent action associated with an agent's interaction with a network site.
  • a traffic data set is received by the facility reflecting an agent's interaction with a particular environment.
  • the traffic data set may take the form of a server log, log file, or other form of data that allows various parameters associated with the traffic to be evaluated.
  • the received traffic data set may be reconciled with any supplemental external data or any other traffic data sets that are related to the received traffic data set.
  • the facility identifies the environment with which the traffic data set is associated.
  • the environment may be identified based on the publisher, advertiser, agent, and/or other feature of the environment.
  • the stored environment rule set for the identified environment is selected by the facility. If no environment rule set is available for the identified environment, or if the environment cannot be reliably identified by the facility, a default rule set may be used by the facility.
  • the facility applies the environment rule set to each session within the traffic data set to generate a session score.
  • the facility applies an algorithm that results in the summation of a function of the correlation coefficients over all rules in the environment rule set.
  • each of the rules in the rule set is weighted equally. In some embodiments, it may be desirable to weight some rules more or less than others if the rules are deemed better or worse, respectively, than others at predicting whether a desired agent action will occur. In such a case, each rule may be weighted differently, e.g., by multiplying the rule by a coefficient prior to performing the summation.
  • a weighted algorithm may be represented by the following equation (3): n score ⁇ Y u K j f ⁇ c j )r j
  • K j is the weight applied to the rule
  • f(Cj) is the correlation coefficient for each rule
  • ⁇ is the result of the rule.
  • Each rule may be weighted by a value chosen from a small set of pre-selected values and may be optimized for a particular environment.
  • the facility may aggregate all session scores to generate a score for all or portions of the traffic data set.
  • An aggregate score may therefore be calculated for all traffic received by a publisher or advertising network, or it may calculated for a more limited environment defined by a single advertiser, publisher affiliate, or other group. In this manner, the quality of traffic may be determined globally for a publisher or advertiser network, or it may be determined on a per publisher affiliate, advertiser, or other group basis.
  • the score may optionally be normalized to place it into a form that may be more readily understood and used by advertisers and publishers.
  • the facility optionally normalizes the score so that the score will fall within a standard range. For example, the score may be normalized so that it always falls within a range from 300 to 800. When normalized in this fashion, each incremental change in the score may be more readily acted upon since the scale of the score has greater meaning.
  • a certain point change in a normalized score may correspond to a certain change in the proportion of agent actions that are valuable to an advertiser (e.g., the proportion of agent actions that are valuable to an advertiser is reduced by 50%).
  • the score may also be converted into a non-numeric scale, such as an A+, A, B+, etc. scale, that may facilitate a greater understanding of and simplify the use of the score.
  • FIG. 6 is a flow diagram of a process 600 of generating training set data that may be used to help score the quality of traffic to network sites.
  • the facility designs experimental advertising campaigns, each of which is designed to attract either fraudulent or non-fraudulent agents.
  • An experimental advertising campaign that is likely to attract fraudulent agents may involve purchasing advertising space on a third- or fourth-tier search engine or distribution partner site. These sites are often interested in increasing their revenue, even if it comes at the expense of the advertiser. They may engage in generating fraudulent agent actions associated with advertisements, such as through use of an automated robot agent.
  • An experimental advertising campaign that is likely to attract non- fraudulent agents may be a campaign in which agent actions associated with advertisements do not lead to generation of revenue. Because no payment is made, the incentive for fraud is low.
  • the experimental advertising campaigns may be developed by human or automated means.
  • the campaigns are executed by the facility. Advertisements are posted on network sites according to the constraints of the advertising campaigns.
  • an agent takes an action associated with an advertisement displayed on a network site, the agent is directed to an advertiser site. Once at the advertiser site, the agent may interact with the site, such as by viewing content, purchasing products or services, and other activities.
  • session data, or data identifying an agent's actions is stored in a server log.
  • the facility receives server or other data logs at a block 615.
  • the facility standardizes these server logs and places the data contained in the server logs into a uniform format.
  • Summary statistics may be extracted from the standardized server logs. For example, data may be summarized based on different session characteristics, such as IP address, agent, distribution partner, or other session characteristic.
  • the facility selects sessions from the standardized server log data with summary statistics that fall outside the normal range.
  • the facility may select sessions that falls below a certain percentile (e.g., first percentile or fifth percentile), above a certain percentile (e.g., ninety-ninth percentile or ninety-fifth percentile), or both.
  • the selected sessions represents the most or least fraudulent agent actions.
  • Clusters may be determined by constructing one or more n-dimensional diagrams, where n is a desired number of rules.
  • each rule is an expression that receives as input one or more parameters associated with an agent's session, such as IP address, time of day, and other parameters.
  • the rule is applied by the facility to the input parameters, it produces a result that reflects the value of an agent's actions associated with that agent's session.
  • the facility may construct a 2-dimensional diagram, where Rule 1 is mapped on the X-axis and Rule 2 is mapped on the Y-axis.
  • Rule 1 is mapped on the X-axis
  • Rule 2 is mapped on the Y-axis.
  • the selected sessions are mapped on the 2-dimensional diagram, and those sessions whose mapped points are clustered together in the diagram will be selected.
  • a clustering analysis may be performed for one or more permutations of rules. For instance, in the above example, additional 1 -dimensional diagrams may be constructed to evaluate clusters defined by each of Rule 1 and Rule 2. If clusters defined by three rules are desired, a 3-dimensional diagram may be constructed to evaluate clusters defined by Rule 1, Rule 2, and Rule 3.
  • Additional 2-dimensional diagrams may be constructed to evaluate clusters defined by Rule 2 and Rule 3 and by Rule 1 and Rule 3. Once clusters are selected, at a block 635 each cluster is characterized as representing fraudulent traffic, non-fraudulent traffic, or unknown.
  • the classification of clusters may be made by a human or it may be automated. For those clusters in which it is unknown whether the cluster indicates fraudulent or non-fraudulent traffic, further analysis may be performed. This analysis may be performed by a human or it may be automated.
  • One or more algorithms may be used by the facility to detect clusters at a block 630. Each algorithm may return, along with the detected cluster, a probability that the cluster indicates fraudulent or non-fraudulent traffic. For example, an algorithm may specify 90% certainty that an identified cluster represents fraudulent activity. The facility selects the cluster that is identified by the algorithm with the highest degree of certainty. For example, if algorithm 1 detects a first cluster and specifies 90% certainty that the cluster indicates fraudulent traffic and algorithm 2 detects a second cluster and specifies 80% certainty that the cluster indicates fraudulent traffic, the facility will select the first cluster because of the higher confidence that the first cluster indicates fraudulent traffic.
  • a false positive detection algorithm may be applied by a human or it may be automated.
  • a false positive detection algorithm may be applied first by an automated program and then by a human to those clusters unresolved by the automated program.
  • the goal of the false positive detection algorithm is to eliminate clusters improperly determined to indicate fraudulent or non-fraudulent traffic.
  • the algorithm may review session statistics, raw server log data, and other data sources, including data provided by third parties.
  • Those clusters that are approved by the false positive detection algorithm are added to the training set at a block 645.
  • the goal is to populate the training set with data that the facility has a high degree of confidence indicates fraudulent or non-fraudulent traffic. While the facility has been described to classify data in a binary form, e.g., fraudulent or non-fraudulent, one skilled in the art will appreciate that the facility may classify data in a non-binary form. For example, data may be classified in several categories, such as highly fraudulent, moderately fraudulent, moderately non-fraudulent, and highly non-fraudulent. Data could also be classified by a numerical score or according to one or more other classification schemes.
  • the training set data may be provided to the previously described traffic scoring system in order to train the system to better identify fraudulent traffic.
  • the training set data enables the traffic scoring system to evaluate which traffic parameters are better indicators of fraudulent traffic.
  • FIG. 7 is a flow chart of a process 700 for identifying combinations of parameters that characterize the quality of traffic associated with network sites.
  • the facility receives server log data or other data representing the traffic associated with one or more network sites.
  • the facility may also receive data from other sources indicative of an agent's interaction with a network site.
  • the server log data and any other data received may be provided to the facility in its original format, or it may be standardized so that the data obtained from various sources is presented in a uniform format.
  • summarized statistics may be extracted from the standardized data and provided to the facility. For example, data may be summarized based on different session characteristics, such as IP address, agent, distribution partner, or other session characteristic.
  • the facility selects n rules that are to be evaluated in order to identify combinations of rules that are indicative of traffic quality.
  • the n rules selected may be one or more of those rules identified as statistically significant, as described above, for the particular network site.
  • the facility produces an association table 800, as depicted in Figure 8.
  • the association table contains n + 4 columns, where n is the number of rules being evaluated. Columns 805a, 805b...805n are each associated with one of the rules being evaluated. Column 815 contains a count of the number of fraudulent transactions, column 820 contains a count of the number of non- fraudulent transactions, column 825 contains a count of the total number of transactions, and column 830 contains a calculated fraud odds ratio.
  • the table will be comprised of x n rows 810a, 810b...81 Om, where x is the number of values each rule may take and n is the number of rules being evaluated. Each row in the table corresponds to a unique combination of rule results.
  • Figure 8 depicts a table where x is 2, e.g., when applied each rule may have a result of either 0 or 1.
  • the first row 810a may contain values of [0, 0 0, 0, 0]
  • the next row 810b may contain values of [0, 0, ... , 0, 0, 1]
  • the last row 810m containing values of [1, 1 , ... , 1 , 1 , 1]. All combinations of rule results are thereby represented in the table. While Figure 8 depicts a table whose contents and organization are designed to make it more comprehensible to the reader, those skilled in the art will appreciate that the actual data structure used by the facility to store this information may differ from the table shown. For example, the table may be organized in a different manner, may contain more or less information than shown, may be compressed and/or encrypted, and may otherwise be optimized in a variety of ways.
  • the facility selects data representative of a single session, as defined above, from the server log data or other data.
  • the facility applies each of the n rules to the session data and generates a result for each rule.
  • the combination of all rule results for a particular session is referred to as a result vector.
  • the result vector is matched to the corresponding row in the association table. For example, if the result of each rule when applied to the session data is 0, the facility will match the session result vector to row 810a of table 800.
  • the facility records whether a transaction associated with the session was fraudulent or non-fraudulent.
  • the facility increments the count in column 815 of the appropriate row (e.g., row 810a in the above example). If the transaction was non-fraudulent or had indicia of being non- fraudulent, the facility increments the count in column 820 of the appropriate row. In either case, the facility also increments the count in column 825 of the appropriate row to denote the transaction.
  • the facility determines if additional sessions remain in the server log or other data source to be processed.
  • the facility may process all of the sessions in the server log or it may process fewer than all of the sessions in the server log. For example, the facility may reserve a certain percentage (e.g., 50%) of the sessions in order to validate, at a later time, the sessions that have been processed. If additional sessions exist that are to be processed, the facility returns to block 720 to select the next session.
  • the process defined by blocks 720 to 730 is thereby repeated as many times as necessary to process the session information of all agent sessions as reflected in the server log or other data source. If no additional sessions remain to be processed at block 735, processing continues to a block 740.
  • Figure 9 is an example of an association table 900 that may be generated by the facility after 15,200 sessions have been processed.
  • a fraud odds ratio is computed for each row of the association table (i.e., for each result vector).
  • the fraud odds ratio represents an estimate of how fraudulent sessions having a particular result vector are in comparison to sessions having different result vectors.
  • the fraud odds ratio may be computed by equation (4):
  • fraud_total y non _ fraud _ total
  • fraudulent_result is the number of fraudulent transactions for a particular result vector
  • non_fraudulent_result is the number of non-fraudulent transactions for a particular result vector
  • fraud_total is the number of fraudulent transactions for all result vectors
  • non_fraud_total is the number of non-fraudulent transactions for all result vectors.
  • Fraud odds ratios computed according to equation (4) are provided in column 930 of table 900. While a particular algorithm is disclosed for calculating the fraud odds ratio, one skilled in the art will appreciate that a variety of different algorithms may be used to determine which result vectors are characterized by the greatest proportion of fraudulent transactions.
  • each row in the association table may be viewed as a final node in a decision tree, where each rule represented in the row corresponds to a node in the decision tree.
  • the association table provides a method to generate the final decision tree nodes without having to generate the entire decision trees.
  • the facility parses the association table in order to identify clusters or subsets of the n rules.
  • a large number of subsets e.g., 5,000 may be identified according to an optimization algorithm, such as simulated annealing or Markov chain Monte Carlo (MCMC) methods.
  • MCMC Markov chain Monte Carlo
  • the facility may condense the association table initially created into an optimized association table.
  • the facility may eliminate from the initial association table those result vectors that were generated by fewer than a certain number (e.g., 100) of transactions.
  • the facility may remove those result vectors for which there is a high variance in fraud odds ratio, which indicates an inconsistent result.
  • the facility may create several (e.g., three) occurrences of the initial association table for each cluster or subset of rules identified at block 745.
  • Each occurrence of the table may correspond to a different characteristic (e.g., date) of the sessions or to a different subset of a training set.
  • the sessions that comprise each occurrence of the table may also be randomly selected.
  • the facility may calculate the fraud odds ratio for each result vector in each occurrence of the table. Those rows that have result vectors in which there is a high variance in the fraud odds ratio between table occurrences may be eliminated. Thus, only those result vectors that are consistent, or robust, are maintained by the facility in the optimized association table.
  • the filtering step may also be viewed as removing unstable, or unreliable, final nodes from a decision tree. Those nodes that produce variable results depending on session characteristic or subset of rules, and are thus not reliable predictors of fraudulent or non-fraudulent agent actions, are removed by the facility. The facility retains only those nodes that produce consistent results.
  • the facility selects those result vectors with the highest fraud odds ratio, further refining the optimized association table.
  • These result vectors represent combinations of traffic parameters that are characterized by the greatest proportion of fraudulent transactions.
  • the facility may choose to select a number (e.g., 200) of the result vectors having the highest fraud odds ratio in the association table.
  • the facility may choose to all result vectors having a fraud odds ratio greater than a certain number (e.g., 25).
  • the selected result vectors comprise the optimized association table; all other result vectors are discarded.
  • the facility may optionally perform cross-validation.
  • the facility may apply the combinations of rules selected at block 745 to the sessions that were not previously processed by the facility. Those rules that produce results that are consistent with the results of the previously processed sessions (i.e., indicate similar proportions of fraudulent agent actions) may be retained by the facility, while those rules that produce results that are inconsistent with the results of the previously processed sessions (i.e., indicate varying proportions of fraudulent agent actions) may be discarded by the facility.
  • the facility may generate result vectors for the traffic and match the generated result vectors to the corresponding row or rows in the optimized association table. In this way, the facility can estimate whether and to what degree the traffic likely resulted in fraudulent or non-fraudulent agent actions. If there is no row in the optimized association table that matches the result of the new traffic, a correlation formula, such as equation (2) or (3), above, or a regression formula may be applied to the traffic in order to estimate whether and to what degree the traffic resulted in fraudulent or non- fraudulent agent actions.
  • a correlation formula such as equation (2) or (3), above, or a regression formula may be applied to the traffic in order to estimate whether and to what degree the traffic resulted in fraudulent or non- fraudulent agent actions.
  • the result vector of the new traffic does may not produce a reliable estimate of whether and to what degree the traffic resulted in fraudulent agent actions.
  • the estimate of whether and to what degree new traffic resulted in fraudulent or non-fraudulent agent actions should be consistent whether the traffic is mapped to the optimized association table or, alternatively, a correlation or regression formula is applied to the traffic.
  • a session, a group of sessions, or a traffic data set as a whole may be scored according to an equation that is a function of the fraud odds ratio.
  • the score for a traffic data set may be the average score of all sessions identified within the traffic data set.
  • the score that is calculated is relative to the value of the traffic; higher scores indicate that the traffic resulted in more valuable (or less fraudulent) agent action(s), whereas lower scores indicate that the traffic resulted in less valuable (or more fraudulent) agent action(s). Calculating traffic scores in this manner results in a more refined estimate of traffic value that is of greater benefit to publishers and advertisers.
  • the score may optionally be normalized, as described above, to place it into a form that may be more readily understood and used by advertisers and publishers.

Landscapes

  • Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Engineering & Computer Science (AREA)
  • Finance (AREA)
  • General Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Marketing (AREA)
  • General Physics & Mathematics (AREA)
  • Economics (AREA)
  • Theoretical Computer Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Game Theory and Decision Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)

Abstract

L'invention concerne un équipement logiciel et/ou matériel servant à évaluer la qualité du trafic vers un site accessible par l'intermédiaire d'Internet ou d'un autre réseau. Cet équipement peut générer des données d'ensemble d'apprentissage et utiliser ces données pour identifier des paramètres représentatifs d'un trafic frauduleux vers un site et réduire l'effet de ce trafic frauduleux sur des annonceurs et des éditeurs. L'équipement selon l'invention peut évaluer la qualité du trafic vers un site par évaluation de divers paramètres associés au trafic et détermination de combinaisons de paramètres représentatifs de la qualité du trafic vers le site. Le trafic vers un site peut être évalué d'après ces combinaisons de paramètres associées à une ou plusieurs sessions. Des notes d'évaluation faibles sont représentatives d'un trafic présentant peu de valeur pour un éditeur, un annonceur ou une tierce partie; des notes d'évaluation élevées sont représentatives d'un trafic présentant de la valeur.
EP07758955A 2006-03-20 2007-03-20 Évaluation de la qualité du trafic vers des sites réseau au moyen de paramètres de trafic interdépendants Withdrawn EP2005382A4 (fr)

Applications Claiming Priority (5)

Application Number Priority Date Filing Date Title
US78390606P 2006-03-20 2006-03-20
US78417406P 2006-03-21 2006-03-21
US11/567,718 US10567255B2 (en) 2005-12-06 2006-12-06 Method and system for scoring quality of traffic to network sites
PCT/US2006/061704 WO2007067935A2 (fr) 2005-12-06 2006-12-06 Procede et systeme de notation de la qualite du trafic vers des sites de reseau
PCT/US2007/064454 WO2007109694A2 (fr) 2006-03-20 2007-03-20 Évaluation de la qualité du trafic vers des sites réseau au moyen de paramètres de trafic interdépendants

Publications (2)

Publication Number Publication Date
EP2005382A2 true EP2005382A2 (fr) 2008-12-24
EP2005382A4 EP2005382A4 (fr) 2012-01-11

Family

ID=38523269

Family Applications (1)

Application Number Title Priority Date Filing Date
EP07758955A Withdrawn EP2005382A4 (fr) 2006-03-20 2007-03-20 Évaluation de la qualité du trafic vers des sites réseau au moyen de paramètres de trafic interdépendants

Country Status (2)

Country Link
EP (1) EP2005382A4 (fr)
WO (1) WO2007109694A2 (fr)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10410245B2 (en) * 2013-05-15 2019-09-10 OpenX Technologies, Inc. System and methods for using a revenue value index to score impressions for users for advertisement placement

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000045264A1 (fr) * 1999-01-29 2000-08-03 Webtrends Corporation Procede et appareil permettant de compter les visiteurs d'un serveur web
US20050097204A1 (en) * 2003-09-23 2005-05-05 Horowitz Russell C. Performance-based online advertising system and method

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060015390A1 (en) * 2000-10-26 2006-01-19 Vikas Rijsinghani System and method for identifying and approaching browsers most likely to transact business based upon real-time data mining
US7401013B2 (en) * 2001-10-15 2008-07-15 Lockheed Martin Corporation Method to optimize test data
US20040148386A1 (en) * 2003-01-24 2004-07-29 Dennis Bushmitch Dynamic CC/PP-based profile generation framework for network conditions assessment
US7584287B2 (en) * 2004-03-16 2009-09-01 Emergency,24, Inc. Method for detecting fraudulent internet traffic

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2000045264A1 (fr) * 1999-01-29 2000-08-03 Webtrends Corporation Procede et appareil permettant de compter les visiteurs d'un serveur web
US20050097204A1 (en) * 2003-09-23 2005-05-05 Horowitz Russell C. Performance-based online advertising system and method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of WO2007109694A2 *

Also Published As

Publication number Publication date
WO2007109694A3 (fr) 2007-12-27
WO2007109694A2 (fr) 2007-09-27
EP2005382A4 (fr) 2012-01-11

Similar Documents

Publication Publication Date Title
US11627064B2 (en) Method and system for scoring quality of traffic to network sites
US11790396B2 (en) Preservation of scores of the quality of traffic to network sites across clients and over time
US10497034B2 (en) Auto adaptive anomaly detection system for streams
US20210035152A1 (en) Predicting the effectiveness of a marketing campaign prior to deployment
US20080270154A1 (en) System for scoring click traffic
US9152977B2 (en) Click fraud detection
US20080103887A1 (en) Selecting advertisements based on consumer transactions
US20100082400A1 (en) Scoring clicks for click fraud prevention
CN112581291B (zh) 风险测评异动检测方法、装置、设备及存储介质
Zheng et al. A scalable purchase intention prediction system using extreme gradient boosting machines with browsing content entropy
CN101449284A (zh) 使用互相关的流量参数对至网络站点的流量质量进行的评分
Lin et al. Predicting outcomes of active sessions using multi-action motifs
WO2007109694A2 (fr) Évaluation de la qualité du trafic vers des sites réseau au moyen de paramètres de trafic interdépendants
KR101166616B1 (ko) 클릭 트래픽을 스코어링하기 위한 시스템
Duan et al. Building and managing reputation in the environment of chinese e-commerce: A case study on taobao
CN116402565B (zh) 一种基于大数据分析的商品推荐方法及系统
CN115705382A (zh) 信息推送方法及其装置、存储介质
Chen et al. Mobile ad fraud: Empirical patterns in publisher and advertising campaign data
CN118037460A (zh) 一种保险产品的评测方法及系统

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 20081020

AK Designated contracting states

Kind code of ref document: A2

Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IS IT LI LT LU LV MC MT NL PL PT RO SE SI SK TR

A4 Supplementary search report drawn up and despatched

Effective date: 20111208

RIC1 Information provided on ipc code assigned before grant

Ipc: G06Q 30/00 20120101AFI20111202BHEP

DAX Request for extension of the european patent (deleted)
STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 20120707