is it possible to use apache mahout without hadoop dependency?

Yes. Not all of Mahout depends on Hadoop, though much does. If you use a piece that depends on Hadoop, of course, you need Hadoop. But for example there is a substantial recommender engine code base that does not use Hadoop.

You can embed a local Hadoop cluster/worker in a Java program.

Definitely, yes. In the Mahout Recommender First-Timer FAQ they advise against starting out with a Hadoop-based implementation (unless you know you're going to be scaling past 100 million user preferences relatively quickly).

You can use the implementations of the Recommender interface in a pure-Java fashion relatively easily. Or place one in the servlet of your choice.

Technically, Mahout has a Maven dependency on Hadoop. But you can use recommenders without the Hadoop JARs easily. This is described in the first few chapters of Mahout in Action - you can download the sample source code and see how it's done - look at the file RecommenderIntro.java.

However, if you're using Maven, you would need to exclude Hadoop manually - the dependency would look like this:

<dependency>        <groupId>org.apache.mahout</groupId>        <artifactId>mahout-core</artifactId>        <exclusions>            <exclusion>                <groupId>org.apache.hadoop</groupId>                <artifactId>hadoop-core</artifactId>            </exclusion>        </exclusions></dependency>