Beyond Classification Tasks - Data Mining with Decision Trees: Theory and Applications - page 98

Training the models
(Generating Attributes driven HMM trees)

Intention prediction

Training

Intention prediction model builder

Sequences
and Users
attributes
of goal 1

Sequences
and Users
attributes
of goal 2

Sequences
and Users
attributes
of goal i

...

As equence
of actions

User's
attributes

Attributes driven HMM tree generator

Tree for goal
1

Tree for goal
2

Tree for goal
i

...

Intention
prediction
method

Pruning Attributes driven HMM trees

Final Tree for
goal 1

Final Tree for
goal 2

Final Tree for
goal i

Predicted
goal for the
sequence

...

Fig. 8.2 Overall diagram of the attribute-driven HMM tree-based method for intention
prediction.

We assume that L is an algorithm for training HMMs. Particularly,
the model λ in each of the tree's nodes is trained using the Baum-Welch
algorithm and the probability p λ ( G = g j | s q ,u q ) is estimated using the

forward-backward algorithm as we show below.
Figure 8.2 presents the HMM training process schematically. The left
side in Figure 8.2 specifies the generation of the models used for intention
prediction based on an attribute-driven HMM tree for each possible goal.
The input for this process is a set of user action sequences divided into
goals and user attributes. The output of the process is a tree for each
goal based on different usage behavior of users in attaining this goal, and
user attributes. The right side of the figure presents the prediction process.
The inputs for the process are: the goal trees generated during the model
generation phase, a specific user session (i.e. a sequence of actions), and the
attributes of a user. The output is the predicted goal that the specific user
was trying to accomplish in this sequence (i.e. the goal with the highest
probability estimation).