close

Filter

Adaptive Personalization Manual / Version 2207

Table Of Contents

4.5 Working With Scoring

Scoring is a simple means to abstract an individual user's behavior on a website. In general, the idea is to assign scores to certain observable events and to combine these scores with the user's current scores whenever the events are observed.

Note

Note

Please note that profiling user behavior as well as taking automated decisions based on this profiling may be restricted by applicable privacy law and may require specific user consent to be obtained and may depend on how you implement the Adaptive Personalization module for your project.

Example

Assume the pages on a website are tagged with keywords and you want to keep track of how often the user visits pages tagged with a specific keyword. In this scenario, a visit on a page is an observable event, and the scores are the counters associated with each keyword. Whenever the user visits a page, the scores of all associated keywords are incremented by 1.

CoreMedia Adaptive Personalization supports scoring via the ScoringContext. It manages a set of scores and uses a ScoringStrategy to update scores if events are observed.

Scoring classes

Figure 4.9. Scoring classes


CoreMedia Adaptive Personalization comes with a set of predefined scoring strategies:

  • CountScoring This strategy simply counts the occurrence of events. That is, for each supplied event, the corresponding score is incremented. This strategy can be used to implement the keyword scenario described above.

  • PercentageFromMaxScoring This strategy weights each score by its percentage of the maximum score value. For each event, a score of the corresponding key is maintained and incremented by 1 whenever the event is observed.

  • PercentageFromTotalScoring This strategy weights each score by its percentage of the sum of all scores. For each event, a score of the corresponding key is maintained and incremented by 1 whenever the event is observed.

Going back to the keyword example, assume that page A is tagged with keywords 'foo' and 'bar', and page B is tagged with 'bar'. Further, assume that a new user visits page A once and page B twice. Here is the table of scores that result from applying the different strategies:

Strategy foo bar
CountScoring 1 3
PercentageFromMaxScoring 1/3 3/3
PercentageFromTotalScoring 1/4 3/4

Table 4.5. Example results


In addition to the application specific scores, two general scores are maintained by all three strategies:

  • __max__ contains the maximum score of all scores maintained by the context

  • __total__ contains the current total of all scores maintained by the context

The scoring strategies are interchangeable, that is, if you start with one you can reconfigure your system later to use a different one without loosing any data.

Configuring a ContextSource to use the ScoringContext

ScoringContext provides its own ContextCoDec implementation in the static inner class ScoringContext$CoDec. The codec can be used in any source that accepts a ContextCoDec or a ContextFactory. Because the ScoringContext requires a ScoringStrategy, you must inject the strategy you want to use for all decoded and created contexts into the codec.

Here is an example of how to configure a CookieSource to use a ScoringContext with the PercentageFromTotalScoring strategy:

<bean id="scoringCookie" 
 class="com.coremedia.personalization.context.collector.CookieSource"
 type="singleton">
  <property name="contextCoDec">
    <bean class="com.coremedia.personalization.scoring.
                  ScoringContext$CoDec">
      <property name="strategy">
        <bean class="com.coremedia.personalization.scoring.
                       PercentageFromTotalScoring"/>
      </property>
    </bean>
  </property>
  <property name="contextName" value="scoringContext"/>
</bean>
Writing your own ScoringStrategy

Writing your own scoring strategy is as simple as implementing the ScoringStrategy interface. Keep in mind that your implementation must be thread-safe because it is typically shared by several ScoringContext instances. Ideally, you simply do not use any modifiable state that is shared among threads.

In typical scenarios, processing events is far more frequent than reading scores. Thus, it's sensible to perform costly updates lazily only when scores are requested. To this end, your strategy may implement the ScoreValueTransformer interface. If a strategy implements this interface, the ScoreValueTransformer#transform method is called by the ScoreContext#getScore method and its result returned as the score. The supplied strategies PercentageFromMaxScoring and PercentageFromTotalScoring use this to perform the normalization of values only at the time of access.

The third interface that is relevant to scoring is MergeStrategy: A ScoringStrategy that allows merging of two sets of scores should implement this interface. Merging of scores is useful if you want to combine data from different context. A typical scenario is as follows: A user logs into your site and his scoring context is persisted in a database. Later, the user returns to the site and browses without logging in, thus new scores are collected. Then, with the user logging in, the formerly persisted data becomes available and can now be merged with the scores collected while the user was anonymous.

If your ScoringStrategy implements the MergeStrategy interface, a ScoringContext using your strategy will be able to perform the mergeWith operation.

Using a ScoringContext to track Keyword Clicks

CoreMedia Adaptive Personalization provides the KeywordInterceptor for the common use case in which you want to count the keywords associated with the pages a user clicks on. The KeywordInterceptor intercepts a CAE request after the controller but before the view is rendered and attempts to extract keywords from the 'self' bean in the model that is to be supplied to the view dispatcher. These keywords are sent as events to the configured ScoringContext. See the respective Javadoc for details.

Search Results

Table Of Contents
warning

Your Internet Explorer is no longer supported.

Please use Mozilla Firefox, Google Chrome, or Microsoft Edge.