Patent Application
20080267207
1. Field of the Invention
The present invention relates to Internet traffic patterns. More particularly, the present invention relates to context-sensitive, self-adjusting models for targeting users based on traffic patterns.
2. Description of the Related Art
During typical Internet use, web users often utilize search engines or directories in order to find content relevant to what they would like to see. These search engines or directories often make money by placing advertising on the result pages of the users' searches, or otherwise profiting from the users' activity within the web pages operated by the search engine or directory. Hence, search engines and directories have a strong desire to steer users towards continuing to view web pages they operate even once the initial search has been completed.
In an embodiment, new traffic pattern data is received pertaining to a first time period having a first time increment. The traffic pattern data may be received as, or converted to, count information reflecting the probabilities that the user will select particular links services. An incremental table is accessed to determine stale traffic pattern count data, which is the traffic pattern count data in the incremental table over a second time period having the first time increment, wherein the second time period covers the oldest traffic pattern count data in the incremental table. Then a count table is updated to reflect removal of the stale traffic pattern count data, wherein the count table contains global count data at a higher level of granularity than the incremental table. The incremental table and count table are updated to reflect addition of the new traffic pattern data.
Reference will now be made in detail to specific embodiments of the invention including the best modes contemplated by the inventors for carrying out the invention. Examples of these specific embodiments are illustrated in the accompanying drawings. While the invention is described in conjunction with these specific embodiments, it will be understood that it is not intended to limit the invention to the described embodiments. On the contrary, it is intended to cover alternatives, modifications, and equivalents as may be included within the spirit and scope of the invention as defined by the appended claims. In the following description, specific details are set forth in order to provide a thorough understanding of the present invention. The present invention may be practiced without some or all of these specific details. In addition, well-known features may not have been described in detail to avoid unnecessarily obscuring the invention.
In an embodiment of the present invention, a search engine or directory recommends additional web pages or content within the control of the search engine or directory in response to user navigation. Depending on the context (e.g., currently viewed web page, prior web pages visited, links clicked upon, etc.), a user will exhibit different behavior and intentions. In order to leverage this context-sensitive user activity, a modeling process may be utilized. This modeling process is self-adjusting and adapts itself to a user's behavior. By doing so, the application is able to operate in real-time, with minimal computational cost.
This embodiment of the present invention may include three steps/components. The first allows for an input stream of modeling data. The second includes an incremental modeling process that includes the concept of a “time window.” The third involves the detection of model evolution. Each of these steps/components will now be described in more detail.
Input modeling data may be fed into an infrastructure of the embodiment of the invention in a streaming fashion. Data may be retrieved on a periodic basis (e.g., daily) and stored locally. Necessary data extraction, transformation, and load steps may be performed on the raw data in order to transform it for use by the incremental modeling process. The data may be received as raw data and then transformed into “counts” through a modeling process The modeling process utilizes the actual data regarding paths that the user has followed during web sessions along with information regarding other users' web experiences. The modeling process may then derive “counts” for a user/link combination that indicates the probability that the user will click on the particular list. In a simple example, this probability may simply reflect the overall actual number of times the user (and users deemed “similar” to the user) clicked on the link with respect to the total number of times the users clicked on any link, however embodiments are foreseen using more sophisticated modeling techniques.
The incremental modeling process is called “incremental” because it can be re-trained in an amount of time that is linearly proportional with the size of the delta defined as newly added data+expired data. Thus, the complexity of this process is on the order of that delta.
One possible implementation of this modeling process is a Naive Bayesian classifier that is extended to create a linearly incremental process. A Naive Bayesian classifier is a well known probability model that will not be described in more detail here. The embodiment using a naive Bayesian classifier is described in more detail later in this document.
In order to provide the “incremental” part of the modeling process, a “time window” is introduced. A time window is an interval that has a start date (and/or time) and an end date (and/or time). For example, a time window may start on Jun. 01, 2006 and end on Sep. 01, 2006. A “sliding time window” is a time window that has a fixed length, e.g. 30 days, but can slide along a timeline. In a typical case, a sliding time window moves forward one day each day, to reflect the advancement of time, and dynamically specifies a new range of time that may be used for the modeling process.
As the sliding time window changes, a linearly incremental algorithm may adjust (re-train) the model for the amount of time that is linear with the new data associated with the sliding time window, plus the data that is expired (moved out) from the previous sliding time window. For example, on a certain day the sliding time window of 30 days covers the entire period 9/1/2006 to 9/30/2006. On the next day, the sliding time window is moved forward one day to 9/2/2006 to 10/1/2006. The data of 10/1/2006, which can be referred to as D—10/1, is new to the sliding time window while the data of 0/1/2006, also known as D—9/1, may be expired by the sliding time window. A linearly incremental modeling process is able to retrain the model for 9/2/2006 to 10/1/2006 based on the existing model of 9/1/2006 to 9/30/2006 in an amount of time that is linear to (size(D—10/1)+size(D—9/1)).
In one embodiment of the present invention, a linear naive Bayesian process may be utilized. This process is based on conditional probability, where a user has a certain features vector. This vector may indicate certain characteristics or properties of the user. For example, a vector may indicate the user's gender, age, and a list of interests, such as (M, 25, retail/apparel, sports/basketball). Given a user with a certain feature vector, the process is able to calculate the probability that the user will click on a certain link or request another service.
The above probability may be computed by using the formula:
Given a set of features (FV1, FV2, . . . , FVn) and a set of prediction classes (counts, such as C1, C2, . . . , Cn), a Naive Bayesian model may be uniquely specified by a list of counts where count_i is a Naive Bayesian parameter. In other words, the counts may represent the probabilities of the corresponding actions occurring for the corresponding time periods (e.g., the probability of the user clicking on a certain link using a certain day's data). It should be noted that these “counts” may be represented either as absolute counts (e.g., overall number of clicks), or as probabilities (percentages or ratios).
The global count is kept for the complete data set at execution time. This data is kept in a count table, although how the data is actually stored is irrelevant. For purposes of this document, therefore, the term “table” shall be understood to mean any data structure capable of storing the underlying data.
In order to make the naive Bayesian model linearly incremental, another copy of the count data may also be kept at the finest granularity by which the corresponding sliding window is moved. For example, if the sliding window is to be moved one day at a time, then the count data is kept at a granularity of a day (or even finer granularity). This data is kept in an incremental table. Thus, the incremental table may contain, for example, all of the counts for a particular feature vector for the past month, organized into days).
Once the incremental modeling process has completed, the new model's evolution from the old model may be evaluated. This process determines how much the model of the new sliding time window is different from existing models based on earlier time periods. If the new model is different enough from the existing model (i.e., more than a certain threshold), then the new model may be used in determining targeting strategy for the user instead of the old model(s). This “model difference” is used as a factor in making this decision.
In one embodiment of the present invention, this model difference is the “click through rate lift ratio.” “Lift” refers to the amount of predicted benefit provided by the new model over a baseline model (typically a model that simply serves links randomly). The click through lift ratio is defined as the ratio of click through lift provided by the old model divided by the click through lift provided by the new model. In other words, for the most recent user behavior data set (D) the existing model (M_old) performs a click-through rate (CTR) lift of X% over the random-serving CTR while the new model (M_new) performs a CTR lift of Y%. The CTR lift ratio (R) is then defined as R=Y/X. In another embodiment this model difference may be defined as the ratio of revenue lift of the new model divided by the revenue lift of the old model.
A replacement threshold (T_r), is the level over which the new model will replace the old model. The definition of what qualifies as a “model difference” is adjustable to fit different types of business requirements, such as revenue, user acquisition rate, user retention rate. Thus, if the click through lift ratio is, for example, greater than a replacement threshold of.
It should also be noted that the present invention may be implemented on any computing platform and in any network topology in which analysis of path information is a useful functionality. For example and as illustrated in
While the invention has been particularly shown and described with reference to specific embodiments thereof, it will be understood by those skilled in the art that changes in the form and details of the disclosed embodiments may be made without departing from the spirit or scope of the invention. In addition, although various advantages, aspects, and objects of the present invention have been discussed herein with reference to various embodiments, it will be understood that the scope of the invention should not be limited by reference to such advantages, aspects, and objects. Rather, the scope of the invention should be determined with reference to the appended claims.
