Abstract: A digital medium environment includes an action processing application that performs actions including personalized recommendation. A learning algorithm operates on a sample-by-sample basis (e.g., each instance a user visits a web page) and recommends an optimistic action, such as an action found by maximizing an expected reward, or a base action, such as an action from a baseline policy with known expected reward, subject to a safety constraint. The safety constraint requires that the expected performance of playing optimistic actions is at least as good as a predetermined percentage of the known performance of playing base actions. Thus, the learning algorithm is conservative during exploratory early stages of learning, and does not play unsafe actions. Furthermore, since the learning algorithm is online and can learn with each sample, it converges quickly and is able to track time varying parameters better than learning algorithms that learn on a block basis.

Abstract: Sequential hypothesis testing in a digital medium environment is described using continuous data. To begin, a model is received that defines at least one data distribution. Testing data is also received that describes an effect of user interactions with the plurality of options of digital content on achieving an action using continuous non-binary data. Values of parameters of the model are then estimated for each option of the plurality of options based on the testing data. In one example. A variance estimate is then generated based on the estimated values of the parameters of the model for each option of the plurality of options. From this, a determination is made as to a decision boundary based on the variance estimate and an estimate for a mean value of each option of the plurality of options based on the testing data.

Abstract: Paired testing techniques in a digital medium environment are described. A testing system receives data that describes user interactions, e.g., with digital content or other items. The data is organized by the testing system as pairs of user exposures to the different item. Filtering is then performed based on these pairs by the testing system to remove “tied” pairs. Tied pair are pairs of user interactions that result in the same output for binary data (e.g., converted or did not convert) or are within a defined threshold amount for continuous non-binary data. The filtered pair data is then tested, e.g., until criteria of a stopping rule are met as part of sequential hypothesis testing. The testing, for instance, may be used to evaluate which item of digital marketing content exhibits a greater effect, if any, on conversion and control subsequent deployment of this digital marketing content as a result.

Abstract: Embodiments of the present invention are directed at providing a sequential multiple hypothesis testing system. In one embodiment, feedback is collected for hypothesis tests of a multiple hypothesis tests. Based on the collected feedback, a sequential p-value is calculated for each of the hypothesis tests utilizing a sequential statistic procedure that is designed to compare an alternative case with a base case for a respective hypothesis test. A sequential rejection procedure can then be applied to determine whether any of the hypothesis tests have concluded based on the respective p-value. A result of the determination can then be output to apprise a user of a state of the multiple hypothesis test. This process can then be repeated until a maximum sample size is reached, termination criterion is met, or all tests are concluded. Other embodiments may be described and/or claimed.

Abstract: Sequential hypothesis testing techniques are described, which involve testing sequences of increasingly larger number of samples until a winner is determined. In particular, sequential hypothesis testing techniques is based on whether a result of a statistic has reached statistical significance that defines a confidence level in the accuracy of the results. Sequential hypothesis testing also permits the user to “peek” into the test through use of a user interface (e.g., dashboard) to monitor the test in real time as it is being run. Real time output of this information in a user interface as a part of sequential hypothesis testing may be leveraged in a variety of ways. In a first example, a user may make changes as the test is run. In another example, flexible execution is also made possible in that the test may continue to run even if initial accuracy guarantees have been met.

Abstract: Sample size determination techniques in sequential hypothesis testing in a digital medium environment are described. The sample size may be determined before a test to define a number of samples (e.g., user interactions with digital marketing content) that are likely to be tested as part of the sequential hypothesis testing in order to achieve a result. The sample size may also be determined in real time to define a number of samples that likely remain for testing in order to achieve a result. The sample size may be determined in a variety of ways, such as through simulation, based on a gap between conversion rates for different options being tested, and so on.

Abstract: A digital medium environment is described to recommend advertisements using ranking functions. A ranking function is configured to compute a score by applying a user context vector associated with a user to individual ranking weight vectors associated with advertisements, and provide the advertisement with the highest score to the user. In order to learn the ranking weight vectors for the ranking function, training data is obtained that includes user interactions with advertisements during previous sessions as well as user context vectors. The ranking weight vectors for the ranking function associated with each advertisement can then be learned by controlling the score generated by the ranking function to be higher for positive interactions than the negative interactions. To do so, the ranking weight vectors may be learned by optimizing an area under the curve ranking loss (AUCL) for the ranking function.

Abstract: Systems and methods for selecting optimal policies that maximize expected return subject to given risk tolerance and confidence levels. In particular, methods and systems for selecting an optimal ad recommendation policy—based on user data, a set of ad recommendation policies, and risk thresholds—by sampling the user data and estimating gradients. The system or methods utilize the estimated gradients to select a good ad recommendation policy (an ad recommendation policy with high expected return) subject to the risk tolerance and confidence levels. To assist in selecting a risk-sensitive ad recommendation policy, a gradient-based algorithm is disclosed to find a near-optimal policy for conditional-value-at-risk (CVaR) risk-sensitive optimization.

Abstract: Risk quantification, policy search, and automated safe policy deployment techniques are described. In one or more implementations, techniques are utilized to determine safety of a policy, such as to express a level of confidence that a new policy will exhibit an increased measure of performance (e.g., interactions or conversions) over a currently deployed policy. In order to make this determination, reinforcement learning and concentration inequalities are utilized, which generate and bound confidence values regarding the measurement of performance of the policy and thus provide a statistical guarantee of this performance. These techniques are usable to quantify risk in deployment of a policy, select a policy for deployment based on estimated performance and a confidence level in this estimate (e.g.

Abstract: Risk quantification, policy search, and automated safe policy deployment techniques are described. In one or more implementations, techniques are utilized to determine safety of a policy, such as to express a level of confidence that a new policy will exhibit an increased measure of performance (e.g., interactions or conversions) over a currently deployed policy. In order to make this determination, reinforcement learning and concentration inequalities are utilized, which generate and bound confidence values regarding the measurement of performance of the policy and thus provide a statistical guarantee of this performance. These techniques are usable to quantify risk in deployment of a policy, select a policy for deployment based on estimated performance and a confidence level in this estimate (e.g.

Abstract: Risk quantification, policy search, and automated safe policy deployment techniques are described. In one or more implementations, techniques are utilized to determine safety of a policy, such as to express a level of confidence that a new policy will exhibit an increased measure of performance (e.g., interactions or conversions) over a currently deployed policy. In order to make this determination, reinforcement learning and concentration inequalities are utilized, which generate and bound confidence values regarding the measurement of performance of the policy and thus provide a statistical guarantee of this performance. These techniques are usable to quantify risk in deployment of a policy, select a policy for deployment based on estimated performance and a confidence level in this estimate (e.g.