Abstract: A system predicts user intent to take an action and delivers content items to the user that match that intent. A plurality of features or attributes for each tracking pixel in a set of tracking pixels can be acquired based on content items and landing pages associated with each tracking pixel. For example, features for a tracking pixel can be determined based on information associated with a content item that enabled a user to access a landing page from which the tracking pixel was fired or triggered. In this example, features for the tracking pixel can also be determined based on information associated with the landing page. The features for the tracking pixels can be utilized to train a machine learning model. The machine learning model can be trained to predict whether or not a particular user intends to produce a conversion (e.g., make a purchase).

Abstract: A social networking system provides content items to a user via a feed that may include one or more sponsored content items. Multiple sponsored content items may be included in a set that is presented in the feed via a scrollable content unit that presents a sponsored content item from the set and presents additional sponsored content items from the set when user interaction is received. To place sponsored content items in the feed, the social networking system scores a set of sponsored content items based on prior user interactions with content presented via scrollable content units and a bid amount of a sponsored content item in the set. The set of sponsored content items is ranked among other sponsored content items based on its score. If the set of sponsored content items is selected for inclusion in the feed, the social networking system orders the sponsored content items in the set for presentation via the scrollable content unit.

Abstract: An online system selecting content items for presentation to its users accounts for likelihoods of users performing actions associated with content items when selecting content items. The online system maintains models determining likelihoods of users performing various actions. If a content item is associated with an action that infrequently occurs, information for determining the model for the action is limited, so the online system increases a bid amount associated with the content item during a time interval to an amount based on a likelihood of the user performing a more frequently occurring alternative action and an average bid amount for the alternative action from content items previously presented to users. The online system also determines an amount based on the model for the action and the bid amount for during the time interval and stops increasing the bid amount when the rate of change has less than a threshold magnitude.

Abstract: A social networking system provides content items to a user via a feed that may include one or more sponsored content items. Multiple sponsored content items may be included in a set that is presented in the feed via a scrollable content unit that presents a sponsored content item from the set and presents additional sponsored content items from the set when user interaction is received. To place sponsored content items in the feed, the social networking system scores a set of sponsored content items based on prior user interactions with content presented via scrollable content units and a bid amount of a sponsored content item in the set. The set of sponsored content items is ranked among other sponsored content items based on its score. If the set of sponsored content items is selected for inclusion in the feed, the social networking system orders the sponsored content items in the set for presentation via the scrollable content unit.

Abstract: A social networking system provides content items to a user via a feed that may include one or more sponsored content items. Multiple sponsored content items may be included in a set that is presented in the feed via a scrollable content unit that presents a sponsored content item from the set and presents additional sponsored content items from the set when user interaction is received. To place sponsored content items in the feed, the social networking system scores a set of sponsored content items based on prior user interactions with content presented via scrollable content units and a bid amount of a sponsored content item in the set. The set of sponsored content items is ranked among other sponsored content items based on its score. If the set of sponsored content items is selected for inclusion in the feed, the social networking system orders the sponsored content items in the set for presentation via the scrollable content unit.

Abstract: An online system presenting content items to a user generates a model that predicts a latent metric describing user actions that occur at least a reasonable amount of time after presentation of content items. To determine the latent metric, the online system retrieves one or more models predicting likelihoods of the user performing various interactions when presented with the content items and determines weights associated with different retrieved models. Combining the weighted retrieved models generates a model for determining the latent metric. As the retrieved models are based on data accessible to the online system in less than the reasonable amount of time after presenting content items, weighing the retrieved models allows the online system to predict the latent metric describing user actions occurring after content items are presented. When selecting content items for the user, the online system accounts for the latent metric determined by the generated model.

Abstract: A system predicts user intent to take an action and delivers content items to the user that match that intent. A plurality of features or attributes for each tracking pixel in a set of tracking pixels can be acquired based on content items and landing pages associated with each tracking pixel. For example, features for a tracking pixel can be determined based on information associated with a content item that enabled a user to access a landing page from which the tracking pixel was fired or triggered. In this example, features for the tracking pixel can also be determined based on information associated with the landing page. The features for the tracking pixels can be utilized to train a machine learning model. The machine learning model can be trained to predict whether or not a particular user intends to produce a conversion (e.g., make a purchase).

Abstract: An online system receives multiple candidate components for including in content items to be presented to online system users. Upon identifying an opportunity to present content to a subject user of the online system, the online system dynamically generates an optimal content item for presentation to the subject user that includes one or more candidate components. Candidate components included in the optimal content item are selected by predicting an affinity of the subject user for each candidate component. The affinity of the subject user for a candidate component may be predicted using a machine-learned model that is trained using historical performance information about content items including the candidate component that were presented to viewing users of the online system having at least a threshold measure of similarity to the subject user. Components of content items used to train the model may be selected using a heuristic (e.g., Thompson sampling).

Abstract: An advertising (“ad”) system allows users to specify a budget for an advertisement (“ad”) campaign including ad requests (“ads”), and the ad system automatically determines bids on a per-impression basis to pace the ad spend according to the budget. The ad system computes a “regret” metric for the ad campaign, which is the total amount over the course of the ad campaign that the advertiser had to pay for presenting an ad above the ideal bid (e.g., the bid that would have captured at least a threshold number of the lowest priced impressions during the ad campaign's life while spending the budget). The ad system may use the regret metric to indicate the performance of the ad campaign, e.g., as feedback for the advertiser.

Abstract: An online system receives a sponsored content item including a maximum amount of compensation for accessing the content, a budget, and a tracking mechanism identifying an action. When an opportunity to present sponsored content to a user eligible to be presented with the sponsored content item is identified, the online system determines a likelihood of the user performing the action identified by the tracking mechanism and an average likelihood of other users performing the action identified by the tracking mechanism. Based on the determined likelihood and the average likelihood, the online system determines a subsidy value. Additionally, the online system generates a penalty value inversely proportional to a number of the identified action that have been identified. The online system increases a bid amount by the subsidy value decreases the bid amount by the penalty value to determine whether to present the sponsored content item to the user.

Abstract: An online system determines an estimated conversion rate for sponsored content items placed on content publishers and on the online system. The estimated conversion rate can be determined by a machine learning model trained using data describing content campaigns, content publishers, and online system users. This data is collected by the online system from content publishers and/or content campaigns that report conversion rates to the online system. By determining a ratio of estimated conversion rates with third party content on the content publisher against those on the online system, the online system can determine a publisher quality score for that content publisher. The online system uses the publisher quality score to normalize third party value contributions toward placing sponsored content on content publishers and the online system. Thus, disparities in the intrinsic value across publishers are diminished as third party value contributions are normalized based on the publisher conversion rates.

Abstract: A system predicts user intent to take an action and delivers content items to the user that match that intent. A plurality of features or attributes for each tracking pixel in a set of tracking pixels can be acquired based on content items and landing pages associated with each tracking pixel. For example, features for a tracking pixel can be determined based on information associated with a content item that enabled a user to access a landing page from which the tracking pixel was fired or triggered. In this example, features for the tracking pixel can also be determined based on information associated with the landing page. The features for the tracking pixels can be utilized to train a machine learning model. The machine learning model can be trained to predict whether or not a particular user intends to produce a conversion (e.g., make a purchase).

Abstract: A content delivery system adjusts bids across publishing channels to account for historical ratios that content was targeted to a content item's audience. A content campaign for users of a content delivery system is devised for two or more sponsored content providers. Targeting criteria for the content item is used to define an audience for the content item, and a sample group from that audience is chosen. The ratio of content impressions among the content providers is identified for the sample group among prior content presentations to these users. For the content item, based on the current ratio of presenting content across a secondary publishing channel and a benchmark publishing channel, a channel control factor is adjusted for the secondary publishing channel based on a numerical comparison of these two ratios. This adjusted channel control factor adjusts the bid price per impression for displaying content by the content campaign.

Abstract: An advertising campaign received by an online system has a specified budget, a specified duration, and includes multiple advertisement requests that each include advertisements for presentation to users of the online system. An ad request included in the advertising campaign is associated with a frequency limit specifying a maximum number of times an advertisement from the ad request is shown to a user during the specified duration. When selecting advertisements for presentation to a user, the online system determines an adjustment value for the ad request's bid amount based on a number of times the advertisement from the ad request has been presented to the user, an amount of the duration that has lapsed, and the frequency limit associated with the ad request. The online system modifies the ad request's bid amount by the adjustment value and uses the modified bid amount when selecting advertisements for presentation to the user.

Abstract: An online system receives content items from publishing users for presentation to other users. When selecting content for presentation to users, the online system accounts for amounts of compensation from publishing users when presenting content items. To prevent publishing users from exploiting content selection by the online system to obtain disproportionate presentation of their content items relative to other publishing users, the online system generates an estimated amount of revenue from various publishing users from presenting their content items. The online system compares an amount of compensation received from a publishing user to the estimated amount of revenue from the publishing user, and generates clusters of content items from the publishing user for review if the amount of compensation is at least a threshold amount less than the estimated amount of revenue.

Abstract: An online system determines a change in of revenue received when a budget for presenting content received from a publishing user is increased. Based on the change in revenue to the online system, the online system determines a subsidy that the online system can provide to one or more publishing users that increases revenue received by the online system from the publishing users, while accounting for the subsidy. To determine the change in revenue for determining the subsidy, the online system selects a subset of publishing users, increases budgets for content presentation provided by the subset of publishing users, and determines changes in revenue to the online system from the increased budgets.

Abstract: An online system presents different content items to different sets of users to evaluate how changes to content or changes to the online system affect user interaction with the content items or presentation of the content items. However, if the online system receives compensation for presenting different content items, the online system may receive a disproportionate amount of compensation for presenting one of the content items that improves user interaction. To prevent such disproportionate allocation of compensation between presentation of different contents items, the online system allocates sets of users to whom different content items are eligible to be presented to maintain a specified budget for presenting the different content items. The online system also differently allocates users across sets to whom different content items are eligible for presentation to prevent biasing of users from presentation of other different content items to users done in parallel.

Abstract: An online system receives multiple candidate components for including in content items to be presented to online system users. Upon identifying an opportunity to present content to a subject user of the online system, the online system dynamically generates an optimal content item for presentation to the subject user that includes one or more candidate components. Candidate components included in the optimal content item are selected by predicting an affinity of the subject user for each candidate component. The affinity of the subject user for a candidate component may be predicted using a machine-learned model that is trained using historical performance information about content items including the candidate component that were presented to viewing users of the online system having at least a threshold measure of similarity to the subject user. Components of content items used to train the model may be selected using a heuristic (e.g., Thompson sampling).

Abstract: An online system selecting content items for presentation to its users accounts for likelihoods of users performing actions associated with content items when selecting content items. The online system maintains models determining likelihoods of users performing various actions. If a content item is associated with an action that infrequently occurs, information for determining the model for the action is limited, so the online system increases a bid amount associated with the content item during a time interval to an amount based on a likelihood of the user performing a more frequently occurring alternative action and an average bid amount for the alternative action from content items previously presented to users. The online system also determines an amount based on the model for the action and the bid amount for during the time interval and stops increasing the bid amount when the rate of change has less than a threshold magnitude.

Abstract: An online system presenting advertisement content determines a bid amount for an advertisement for each new impression opportunity based on a budget for an advertising campaign provided by the advertiser, pacing bid amounts to spend the budget over the course of the advertising campaign. The online system applies an additional constraint that limits a cost metric for the advertising campaign such as an observed CPM (cost per thousand impressions) to a multiple of an average CPM for a target audience for presentation of advertisements of the advertising campaign. To compute the average CPM for the target audience, the online system samples users in the target audience and retrieves an average CPM for each online system user, and averages the retrieved per-user average CPMs.