Contextual Intelligence as a Driver of Services Innovation Teresa Lunt, VP PARC © 2013 PARC

Harnessing Context Content and context analytics infer relationships among information to construct a Personal Semantic Network for each individual Explicit Tags & Preferences … People Content Tasks Events Places Topics Extracted content & meta-data Sensed Location, Proximity, … Inferred activity & inter-relationships…. User model: Patterns, Interests, Behavior, Relationships, Personality… Use it to infer: Current Situation (Venue, Who I’m with, What I’m doing), and Likely Future 2 © 2013 PARC

Example: Psychological Profiles Online behavior is highly correlated with personality variables Extract and classify textual features from online behavior to build psychological profiles as indicators for future behavior Personality Traits Neuroticism Agreeableness Conscientiousness Extraversion Openness Perceived Stress Self-Assurance Overall Mood/Emotion Extraversion © 2013 PARC

Contextual Recommendations combine multiple factors to drive decisions * Distance w1 Personal Utility Score GPS + NFC EMAIL * Purchase History w2 Calendar The next step is to take these models of past behavior, along with an estimate of current and likely future situations, and to recommend a course of action. (or take the action directly, if the confidence of the recommendation is high enough). We’ve developed an extensible, general-purpose hybrid recommendation engine for decision support. Given any set of alternatives (coupons, movies, stores, activities, health choices, financial options, etc.), there is a set of scoring functions that query the user behavior model and current situation to score each of the alternatives based on what it knows. E.g., the distance function simply uses current and likely near-future distance to score the alternatives; the purchase history makes its score based on your history of purchasing in different product categories; etc. This can also call out to third party databases, like say a credit card company. The scores are combined in a linear weighted combination to generate the overall utility score – the weights are adjusted per individual based on the observance of the user outcome and the success/fail of the scoring function’s prediction in the observed outcomes. The weights can change dynamically. (incidentally, this hybrid recommendation is a patented technique. Parc also invented collaborative filtering, a common recommendation algorithm) Ranked Alternatives * Predicted Activity w3 Option 25 0.77 Option 16 0.64 Option 3 0.60 Option 49 0.57 Option 8 0.39 * Others … w4 Alternatives © 2013 PARC

Lean Service Innovation Digital Nurse Assistant: Contextual information Improve clinical decisions Reduce errors Increase patient time Improper Payment Amounts Fraud, Waste & Abuse: Health insurance Transportation Credit cards Social services Police Operations Enable evidence-based decisions Parking Enforcement: Shared real-time context for coordinating & documenting activities Sentiment Analysis & Topic Detection in Social Media: Customer intelligence Emerging event detection © 2013 PARC

Lean Service Innovation Call centers Mining audio & transcripts for learning Real-time analysis & intervention Intelligent call routing QA, training, call center analytics Transportation services Personalized rider services in metro area transit systems Participating vendors create offers Riders get real-time contextual recommendations Personalized services for the car Proactively infer user’s intent, anticipate needs, take action Smart information feeds & prioritized e-mail, Facebook, Twitter, texts, etc Interactive calendar assistant Understands when driver’s attention is available (e.g., at red light) © 2013 PARC

Thank You!