Signal Processing - January 2017 - 71

Suggestion, Notifications,
and Alerts Run Time

Offline Inference and Rule Recipe Authoring
Personal Profile Data
Device Sensor Signals
Web Services Signals

Collect and
Process

Inference and
Learn

Aggregate

Deploy and
Publish

User

Rule Recipe
Authoring

Figure 3. The proactive assistance system architecture.

suggestions are essential for the proactive agent to learn and
adapt to the user.

Notifications and alerts
In the notifications/alerts category, the system allows the user
to set rules to define the triggers for certain actions. If the triggering condition evaluates to TRUE, the action is executed.
The rules are defined over a set of signals. These signals are
produced by an information channel that can be evaluated by
the proactive agent. These channels represent many types of
information, such as date/time and location, as well as constantly updated data feeds generated by various web services,
which include weather, sports, news, finance, and entertainment. For example, one can create rules to obtain an alert when
the Seattle Seahawks score a touchdown. A user can set a rule
to be reminded of his or her mother's birthday. It is also possible to combine these signals to formulate more complex triggering rules. For example, a specific flight departure time, a
user's physical location, and a commute time to the airport can
all be used to trigger an alert that reminds the user that it is
time to leave for the airport. Once the trigger rule is set, the
agent monitors the signals from the corresponding information
channels to evaluate the rule. If the rule evaluates to TRUE, the
agent takes an action. The actions are communicated to the
user in a target device-specific manner, which could be a proactive entity card, SMS, or even a phone call. This type of proactive agent programming falls under the if-then recipes [60],
in which simple rules allow users to control many aspects of
their digital life.

Inference and suggestions
In the suggestions category, the agent infers the user's habits
and routines by reasoning over his or her past behavior and
makes a personalized recommendation to the user with the
goal of furthering the user's interest. For example, knowing
that the user watched comedy movies featuring a particular
actor in the past, the agent may suggest a new comedy movie
featuring that same actor in the future. The agent can also
suggest new experiences based on the logical sequencing of

different yet related (through time or location) events. For
instance, if the user made a restaurant reservation in a metropolitan city downtown, the agent may suggest nearby parking
places. Through inference, the agent can learn certain facts
about the user, by reasoning over the user's whereabouts and
movement patterns through time and location. For example,
the user's home and work location could be inferred by joining GPS data with time over several weeks. If the user is
spending most or all of his or her time between 9 a.m. and
5 p.m. during weekdays at a specific location over several
weeks, that is likely to be the user's work location. Likewise,
the user's commute hours between home and work could also
be inferred from combining home and work location with the
GPS data during the likely morning and evening commute
hours over several weeks. This inference is used to proactively show the traffic commute cards around the time the user
typically commutes to/from work (or home).
The key questions here are determining the type of suggestion and when to do it, because there is an associated cost with
the suggestion (if the action, relevance, or timing is wrong). To
get around the cold start problem (if the agent does not have
access to the user's past activity through a feedback loop or the
user is accessing the PDA for the first time), the user is also
given the ability to teach the agent his or her interests from a
precompiled list of topics, including news, sports, finance, technology, dining, and entertainment. The decision for taking a
proactive action is driven by a machine-learned model, given the
costs and benefits as constraints. The machine-learned model
combines a set of information in the user's profile, demographic
and content-based profiles, and online user behavior signals
(such as click through, dwell time, and dismissal), along with
the user's recent relevant activity (e.g., similar content searches),
which are captured in the history variable h in (1). This is used
to model whether a specific user u will like the specific suggested entity e. Standard machine-learning techniques, such as
maximum entropy models [35], gradient boosted decision trees
[55], and deep learning techniques [48], are used to incorporate
both user-specific online and offline signals to estimate the
probability that the user is expected to like the suggested entity

IEEE Signal Processing Magazine

January 2017

Table of Contents for the Digital Edition of Signal Processing - January 2017