18 over time, as noted later. The preferences can relate to, by way of non-limiting examples, fuel brands, restaurants, hotels, vehicle service centers, car accessory brands or type, music/entertainment/social media, hobbies, retail stores, stocks, sports teams, preference for paid or free services and applications, and other information. User data 60 may also include interaction information such as prior sales or purchase information, call center interactions, social media activity and other information.

Still further, user data may include preferences and settings regarding notifications that the user would like to receive or not. These preferences and settings enable a user to determine, for each program or app, which may include vehicle system programs (e.g. notifications regarding fuel level, tire pressures, etc) and apps added to the vehicle or remote device by the user (generally referred to as apps hereafter), specific conditions for when and how notifications should be sent to the user. A user might choose to have no notifications delivered from one or more apps, or to receive notifications only when the vehicle is not moving, or when the vehicle is moving below a threshold speed, or when the vehicle is on a certain type of road (and not other types of roads, for example), or only after the vehicle is stopped and placed into a park mode, or based on time of day, or to provide audio notifications or other hands-free operations, and so on.

Next, external data 64 may be provided to and used in the analysis by the information system 10. External data 64 may include, by way of non-limiting examples, mobility services, insurance information, lease and other financial data, data from other, similar vehicles, data from third parties (e.g. sales, promotions, general information), information about the terrain and environment, map data including information about the geography, businesses, road and the like, traffic information, status of orders or deliveries requested by the user, and the like.

In use, a wide range of notifications and communications may be provided to a vehicle and the occupants of the vehicle. The notifications may relate to, by way of non-limiting examples, vehicle systems (e.g. fuel level alerts, low battery alerts, engine/oil/battery temperature alerts, and other warnings or vehicle indicators, application notifications specific to individual applications accessed through the control system (e.g. the IVI system) or a device paired to the vehicle IVI or control system, a navigation system or program (e.g. for traffic, accident, construction or road conditions, and route instructions), a publicly broadcast notice (e.g. an AMBER or dangerous person alert), a weather alert (e.g. for severe weather), calendar alerts (e.g. reminders and appointment information) as well as notices from a device paired to the vehicle (e.g. a smart phone and the various notifications and alerts that occur with smart phone use).

The various notifications can be categorized or rated based on content, priority or importance, and the context for the notification. In at least some implementations, the system includes an importance level classifier, a content classifier and a context-aware prioritization model or classifier.

The importance level classifier (which may be part of a personalized learning and customer notification model) may be arranged to provide a base importance rating which may be a rating applied relative to one or more thresholds and on any desired scale. By way of non-limiting examples, the importance ratings may be numeric (e.g. 1 to 10) or tiered like maximum, high, default, low, minimum. The highest ratings may be given to critical notifications that should be delivered to a user regardless of the current driving situation, and the lower ratings may be given to notifications that can be delayed and delivered during a driving situation deemed appropriate for delivery of the notification, as noted herein.

In at least some implementations, the importance level classifier includes a gradient boosting matrix which may use a tree-based learning algorithm like LightGBM. This system can be efficient, readily scalable and suitable for large datasets. In at least some implementations, the mathematical model is represented by a sum of decision trees F_i(x):

$F (x) = \sum_{i = 1}^{M} F_{i} (x)$

where M is the number of trees used in the algorithm. Each tree F_i(x) may be trained sequentially to correct the errors of the preceding trees, and build a robust system.

In at least some implementations, the rated importance of a notification can be adjusted based upon the context in which the notice is to be provided. For example, a notification that is deemed important and relevant to current conditions (e.g. real-time vehicle operating parameters) may have an adjusted importance rating that is greater than the base importance rating, and greater than an adjusted rating in instances in which the same notice is less relevant or not relevant to the current conditions. Thus, notifications may have a base rating and one or more condition specific ratings that are dynamically adjusted by the system. This can enable the system to avoid distracting a driver with a less important notification during a higher stress driving situation that is occurring now or is likely to occur in the near-term (e.g. within a threshold time period, that may be adjustable). The less important notification can then be delivered when the driving situation changes to an appropriate level.

In at least some implementations, the content classifier may categorize notifications by their content. Representative and non-limiting examples of categories for notifications may include, greetings, compliments, social media notices, management notices, speed limit indications or warnings, hard braking warnings, warnings about phone use while driving, weather or traffic warnings, emergency warnings, promotion notices, notifications that require a responsive action or no responsive action, and so on.

In at least some implementations, the content classifier may include a gradient boosting matrix which may use a tree-based learning algorithm like LightGBM, to classify the notifications into multiple categories. The content classifications and importance ratings can be combined into a matrix of model, with an importance level determined for each of the content classifications, if desired. In this way, the same rules may be applied for all notifications in a category so similarly categorized notifications are managed in the same way, in at least some implementations.

The mathematical model for multi-class classification may be done by extending a binary classification model to handle multiple classes. In an example in which f_k(x) is the predicted probability that the notification belongs to class k, where k ranges over all content categories, the final predicted class is then the one with the highest probability, from the following model, in which K is the total number of content categories:

$P (y = k ❘ x = \frac{\exp (f_{k} (x))}{\sum_{j = 1}^{K} \exp (f_{k} (x))}$

Next, the context-aware prioritization model can be used to dynamically adjust the importance rating of different notifications, recognizing that certain notifications are more relevant and/or important in some situations than other situations. This may be done by content classification, for each individual notification type or both, as desired. For example, weather-based notifications can be very relevant to only vehicles in or traveling near the area for which the notification is issued, and may have little or no importance to other vehicles. Similarly, traffic condition notifications for areas that a vehicle has already passed through may have little or no importance to that vehicle, but may be very relevant to vehicles traveling toward the area for which the traffic notification is to be issued. In addition to location-based relevance, time may be considered where some notifications like restaurant promotions or the like may be more relevant during times when people normally eat meals than at other times.

In at least some implementations, the system may consider and receive data regarding a number of context features. For example, representative but non-limiting context features include: driving patterns like vehicle speed, acceleration, deceleration, and other driving behavior features; time of day or date or month and the like; location like exact vehicle location obtained via GPS or area or region in which the vehicle is used; current weather and traffic conditions; user preferences by which a user can determine what notifications they want/need and priority levels for the notifications including context specific priority levels.

In at least some implementations, the context-aware prioritization model uses one or both of a rule-based approach and a reinforcement learning approach. For a rule-based approach, mathematical models involve defining a set of rules that dictate how the priority or importance should be adjusted based on various contextual features. For example, Adjusted Priority=Base Priority+Weather Rule+Traffic Rule+Location Rule+Time Rule+ . . . . With this approach, one or more rules may be established for various factors, and then those factors may be utilized to adjust the priority or importance rating as a function of the current conditions for the vehicle, that is, the current context of the vehicle.

For a reinforcement learning approach, the model involves defining a reward function that guides the learning process, where model may include determining an estimated value of taking a first action in a first state and then compare the estimated value to the reward that occurs from taking the first action in the first state. The mathematical model includes the definition of state representations, actions, and the reward function. For example:

$Q (s, a) \leftarrow Q (s, a) + α \cdot (R (s, a) + γ \cdot \max_{a^{'}} Q (s^{'}, a^{'}) - Q (s, a))$

Where:

- Q(s, a) is the estimated value of taking action a in state s,
- R(s, a) is the immediate reward of taking action a in state s,
- α is the learning rate,
- γ is the discount factor, and
- s′ is the next state.

Collectively, the various algorithms and mathematical formulas and analyses noted herein may be part of the personalized learning and customer notification model. The model may provide a comprehensive solution to notification analysis and management of the notifications.

User interaction with the notification management system can occur via a remote device (e.g. phone, tablet, computer and via an internet interface) or the IVI system 18, and in particular, a head-unit or main console thereof which may include one or more display screens 20 and the user interface 21. The user interface 21 may include one or more inputs that may be provided in one or more forms, such as but not limited to, touch responsive portions of a display, one or more manually actuated inputs (e.g. dials, buttons or switches), and/or audio inputs including a microphone via which verbal inputs can be given by a user.

The IVI system 18 may display various items and options that may be selected by a user. By way of some non-limiting examples, the items and options may include menu options of vehicle settings and preference menus (for control of heating and cooling options, audio video settings and preferences, door lock functionality, performance settings (sport, eco, etc) and various other settings), program icons displayed for included or embedded apps that may be selected by a user and run by the system, such as via a web portal or application programming interface (API). As noted herein, the system may include programs or “apps” or “web widgets” that may relate to a wide range of tasks and features, such as but not limited to, navigation, audio/video, social media, interaction with paired devices, text messaging, phone use, shopping, restaurants, reviews (e.g. Yelp), and an app store via which apps may be downloaded or updated.