EP3845023A1 - Enabling wireless network personalization using zone of tolerance modeling and predictive analytics - Google Patents
Enabling wireless network personalization using zone of tolerance modeling and predictive analyticsInfo
- Publication number
- EP3845023A1 EP3845023A1 EP19856044.3A EP19856044A EP3845023A1 EP 3845023 A1 EP3845023 A1 EP 3845023A1 EP 19856044 A EP19856044 A EP 19856044A EP 3845023 A1 EP3845023 A1 EP 3845023A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- user
- data
- satisfaction
- network
- context
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 claims abstract description 131
- 238000010801 machine learning Methods 0.000 claims abstract description 95
- 230000008569 process Effects 0.000 claims abstract description 61
- 238000005457 optimization Methods 0.000 claims abstract description 35
- 238000004422 calculation algorithm Methods 0.000 claims description 59
- 230000000694 effects Effects 0.000 claims description 55
- 238000012517 data analytics Methods 0.000 claims description 16
- 238000004891 communication Methods 0.000 claims description 14
- 238000012544 monitoring process Methods 0.000 claims description 14
- 230000003542 behavioural effect Effects 0.000 claims description 12
- 230000006872 improvement Effects 0.000 claims description 3
- 230000004931 aggregating effect Effects 0.000 claims description 2
- 230000006399 behavior Effects 0.000 description 53
- 238000013461 design Methods 0.000 description 38
- 230000006870 function Effects 0.000 description 32
- 230000008901 benefit Effects 0.000 description 27
- 238000013459 approach Methods 0.000 description 26
- 238000002474 experimental method Methods 0.000 description 23
- 238000005259 measurement Methods 0.000 description 23
- 238000012549 training Methods 0.000 description 21
- 238000012360 testing method Methods 0.000 description 20
- 238000005516 engineering process Methods 0.000 description 18
- 230000001413 cellular effect Effects 0.000 description 16
- 238000013528 artificial neural network Methods 0.000 description 15
- 230000007423 decrease Effects 0.000 description 15
- 238000004519 manufacturing process Methods 0.000 description 15
- 238000004088 simulation Methods 0.000 description 15
- 238000013468 resource allocation Methods 0.000 description 14
- 238000010200 validation analysis Methods 0.000 description 14
- 230000000875 corresponding effect Effects 0.000 description 13
- 238000011161 development Methods 0.000 description 13
- 230000018109 developmental process Effects 0.000 description 13
- 238000007781 pre-processing Methods 0.000 description 13
- 230000002596 correlated effect Effects 0.000 description 12
- 210000002569 neuron Anatomy 0.000 description 12
- 230000008859 change Effects 0.000 description 11
- 230000001419 dependent effect Effects 0.000 description 11
- 238000012706 support-vector machine Methods 0.000 description 11
- 238000013480 data collection Methods 0.000 description 10
- 238000002790 cross-validation Methods 0.000 description 9
- 238000003066 decision tree Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 9
- 208000018910 keratinopathic ichthyosis Diseases 0.000 description 9
- 239000000203 mixture Substances 0.000 description 9
- 238000007637 random forest analysis Methods 0.000 description 9
- 210000004027 cell Anatomy 0.000 description 8
- 238000009472 formulation Methods 0.000 description 8
- 238000007726 management method Methods 0.000 description 8
- 230000009471 action Effects 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 5
- 230000003416 augmentation Effects 0.000 description 5
- 238000001914 filtration Methods 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 238000011160 research Methods 0.000 description 5
- 238000012800 visualization Methods 0.000 description 5
- 238000013506 data mapping Methods 0.000 description 4
- 230000003247 decreasing effect Effects 0.000 description 4
- 238000009826 distribution Methods 0.000 description 4
- 230000008451 emotion Effects 0.000 description 4
- 238000000513 principal component analysis Methods 0.000 description 4
- 230000009467 reduction Effects 0.000 description 4
- 230000002194 synthesizing effect Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 3
- 230000003190 augmentative effect Effects 0.000 description 3
- 230000001186 cumulative effect Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 230000002829 reductive effect Effects 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 230000001133 acceleration Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000002860 competitive effect Effects 0.000 description 2
- 230000008094 contradictory effect Effects 0.000 description 2
- 239000002360 explosive Substances 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000005304 joining Methods 0.000 description 2
- 230000015654 memory Effects 0.000 description 2
- 230000003278 mimic effect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- 238000012384 transportation and delivery Methods 0.000 description 2
- 101100161935 Caenorhabditis elegans act-4 gene Proteins 0.000 description 1
- 241000282412 Homo Species 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 230000010267 cellular communication Effects 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000013434 data augmentation Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000014509 gene expression Effects 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 230000036651 mood Effects 0.000 description 1
- 230000006855 networking Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000037081 physical activity Effects 0.000 description 1
- 230000002062 proliferating effect Effects 0.000 description 1
- 230000035755 proliferation Effects 0.000 description 1
- 238000012797 qualification Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000007493 shaping process Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 210000003813 thumb Anatomy 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W28/00—Network traffic management; Network resource management
- H04W28/16—Central resource management; Negotiation of resources or communication parameters, e.g. negotiating bandwidth or QoS [Quality of Service]
- H04W28/24—Negotiating SLA [Service Level Agreement]; Negotiating QoS [Quality of Service]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W24/00—Supervisory, monitoring or testing arrangements
- H04W24/02—Arrangements for optimising operational condition
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W72/00—Local resource management
- H04W72/50—Allocation or scheduling criteria for wireless resources
- H04W72/54—Allocation or scheduling criteria for wireless resources based on quality criteria
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/02—Services making use of location information
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/06—Selective distribution of broadcast services, e.g. multimedia broadcast multicast service [MBMS]; Services to user groups; One-way selective calling services
- H04W4/08—User group management
Definitions
- the subject application relates to telecommunication networks and more particularly, to a method and system for managing and allocating wireless network resources to optimize User satisfaction.
- Networks are expected to interconnect devices, humans and even things while maintaining a very good QoS.
- networks are expected to support a wider range of applications and use cases such as vehicular ad- hoc networks and virtual reality applications.
- Such applications require network services to be delivered with a variety of network performance characteristics (e.g., rate, latency, security, and quality of experience (QoE)) which raises significant technical challenges for service providers today.
- QoE quality of experience
- revenues associated with different services and application will widely vary which, in effect, will bring new business challenges.
- Current networks lack flexibility in balancing the implementation of cost-optimized and performance-optimized applications.
- the evolving fifth generation (5G) wireless networks are envisioned to cope with these rising challenges while maintaining a profitable business and high end-user QoE. But as will be explained, they fall short of providing a complete solution.
- Network Function Virtualization (NFV) technology is proposed for 5G and beyond networks to isolate the software and hardware aspects of networks in order to transform network functions from dedicated hardware appliances to software-based applications running on commercial off- the-shelf equipment.
- Software Defined Networks (SDN) along with NFV are considered as enablers for Network Slicing (NS) in 5G.
- SDN Software Defined Networks
- NS Network Slicing
- the concept of NS is proposed to allow operators to provide customized, reliable services with increased efficiency while reducing capital expenditure and the operating expenses of wireless networks.
- Each slice is associated with a set of resources including bandwidth and network topology.
- the differentiator of NS from the current QoS-based solution is its ability to provide an end-to- end virtual network for a given user. This level of flexibility cannot be offered by any of the current networks.
- 4G networks can discriminate between VoIP traffic from other traffic types such as web browsing.
- 4G networks are not able to differentiate and customize the same type of traffic (e.g., VoIP) initiated by different users.
- the aforementioned shortcomings of the current networks are addressed by the features offered by NS in 5G and beyond networks.
- Service Slicing can be also utilized to deal with different types of services with different QoS requirements. Since wireless networks resources are extremely limited, services are sliced based on their QoS requirements and network scenario. QoS Requirements associated with each slice are decided based on the service associated with the slice. For instance, it has been suggested to define a service utility function for each service according to the service- specific constraints and QoS requirements.
- the network described herein is more agile and flexible, and is able to micro-manage the resources within a slice and tailor them to the end user’s specific needs and requirements.
- user satisfaction behavior information is not available in current networks, the range of QoS to achieve the required average user satisfaction for all users in most situations is predetermined and is fixed. Based on the specified QoS range for a particular service, networks are optimized to increase the efficiency of the resources. But this results in an inefficient system.
- the described personalized networks provide a more efficient system, at the same time providing higher levels of User satisfaction.
- the improved network :
- Zone-of-Tolerance ZoT
- the proposed framework utilizes user satisfaction feedback to personalize the cellular network decisions and hence, micro-manages the available resources so that maximum user satisfaction is achieved with a minimum amount of resources. Saving resources is very valuable to the network since it can be utilized for more critical applications such as public safety and autonomous cars. In order to achieve the ultimate balance between network resources and user satisfaction, we answer the following questions:
- One aspect of the invention avoid networks over-provisioning by designing orthogonal networks which achieves the required user satisfaction levels using minimum resources, such as bandwidth and power.
- Embodiments of the invention include the following:
- Personalized wireless networks utilize the non-intrusive real-time user satisfaction feedback in order to personalize wireless networks decisions and hence, micromanage the available resources so that the required personalized satisfaction levels are achieved with the minimum allocated resources.
- Fig. 1 presents a block diagram of an exemplary communication network for effecting the invention.
- Fig. 2 presents a schematic diagram of a user satisfaction model and a visualization of an example illustrating the relationship between the zone of tolerance, D, user satisfaction and personalized network decisions.
- Fig. 3 presents a process flow diagram for a big data-driven Al-based network personalization framework in an embodiment of the invention.
- Fig. 4 presents a schematic diagram showing QoSd, QoSp, D, and the corresponding user satisfaction for the two contexts (C1 and C2) associated with both personalized and non-personalized networks.
- Fig. 5 presents a graph of Total QoSN P, QoSPr, QoSd, and QoSN P - QoSPr for three users vs. time in hours.
- Fig. 6 presents a graph of Average user satisfaction for the three users vs. time in hours for the personalized and non-personalized networks.
- Fig. 7 presents a second schematic diagram of a User Zone of Tolerance (ZoT) model.
- ZoT User Zone of Tolerance
- Fig. 9 presents a Tree Data Generator (TG) model for a working professional persona in an embodiment of the invention.
- TG Tree Data Generator
- Fig. 10 presents a Hidden Markov model in an embodiment of the invention.
- Fig. 11 presents a graphic visualization of the problem of user satisfaction prediction from user context data.
- Fig. 13 presents an accuracy swarm plot for 10-folds cross-validation using DT, Knn, and RF algorithms.
- Fig. 14 presents a table of exemplary sample instances from the proposed synthetic dataset.
- Fig. 15 presents a block diagram showing the relationships between the features of time, location, speed, and activity, in an embodiment of the invention.
- Figs. 16a and 16b present a user speed heat map and the ring of locations for a weekday and a weekend day, respectively.
- Fig. 17 presents a graphic representation setting out the percentage of time the user spent at each location over each time period on a weekday.
- Fig. 18 presents an exemplary histogram of the recorded instances on a weekday for six activities at two locations.
- Fig. 20 presents a heat map of an instance count for the requested services and the corresponding applications over a week.
- Fig. 22 presents an exemplary accuracy swarm plot for 10-folds cross-validation using DT, Knn, and RF algorithms.
- Fig. 23 presents a process flow diagram of an exemplary data-driven persona prediction framework for personalized wireless networks.
- Fig. 24 presents a graph of cumulative EVR vs. the number of components.
- Fig. 25 presents a graph of the accuracy and accuracy variance of the proposed persona prediction framework vs. stack size using SVM.
- Fig. 26 presents a graph of the accuracy and accuracy variance of the proposed persona prediction framework vs. stack size using DNN.
- Fig. 27 presents a graph of the confidence and confidence variance of the persona prediction framework vs. stack size using DNN.
- Fig. 28 presents a process flow diagram of an exemplary big data-driven satisfaction prediction framework in an embodiment of the invention.
- Fig. 29 presents a schematic diagram of an exemplary data mapping for user location feature.
- Fig. 30 presents a histogram for user satisfaction classes of the WPP dataset in an embodiment of the invention.
- Fig. 31 presents an exemplary network structure of the proposed DNN model in an embodiment of the invention.
- Fig. 32 presents a graph comparing the accuracies of DNNs using different optimization methods.
- Fig. 33 presents a graph comparing the training accuracy of DNNs using different learning rates.
- Fig. 34 presents a graph comparing the DNN model performance for different training data sizes.
- Fig. 35 presents a graph comparing training and validation accuracy for the chosen DNN model.
- Table I sets out an exemplary set of simulation parameters.
- Table II sets out exemplary weekdays and weekends location rings for the four user personas.
- Table III presents an exemplary set of services and their associated demand rate for the considered set of applications.
- Table IV sets out exemplary features of the CH dataset.
- Table V sets out exemplary features of a WPP dataset in an embodiment of the invention.
- Table VI Simulation parameters for experimental results of Fig. 36 and 37.
- Personalized networks optimize two correlated and contradicting objectives in real-time: user satisfaction and resource utilization.
- wireless networks produce colossal amounts of data and most of this data is in real-time.
- a system that is capable of digesting these data to create relevant and meaningful decisions in real-time at the user level using machine learning (ML) and big data analytics is the ultimate solution to meeting the aforementioned objectives simultaneously.
- ML machine learning
- big data analytics is the ultimate solution to meeting the aforementioned objectives simultaneously.
- a wireless communications network will typically comprise various wireless devices such as smart phones and laptop computers, which can access a smart network (5G, Internet, etc.) via wireless base- stations and/or networks, Wi-Fi routers, WiLAN, metronet and other similar wireless devices and networks.
- the software which provides much of the functionality described herein will typically operate on one or more servers, either as part of the smart network, or connected to it.
- the description of the invention is set out in six Parts as follows:
- Part I provides an overview of process shown in Fig. 3.
- Parts II and III describe the data collection process identified by [2] and [3] (blocks [1] and [2])).
- Parts IV, V, and VI describe the processes identified by [4], [5] and [6] respectively.
- a system that is capable of digesting these data to create relevant and meaningful decisions in real time at the user level using machine learning (ML) and big data analytics is the ultimate solution to meeting the aforementioned objectives simultaneously.
- ML machine learning
- User feedback and context information awareness is needed for the personalized networks described in most embodiments of the invention. This is due to the fact that the inherent patterns and information in context and feedback data provide service providers with tangible data that can be utilized to make optimized and personalized decisions.
- User feedback collection can be done in real time or offline in a number of ways that can be either intrusive (e.g., surveys, feedback boxes) or non-intrusive which employ ML and artificial intelligence (Al).
- intrusive e.g., surveys, feedback boxes
- Non-intrusive which employ ML and artificial intelligence (Al).
- Al artificial intelligence
- wireless networks the utilization of user feedback from intrusive methods is discussed in [1 ], [2]
- the authors in [1 ] propose an approach called“user-in-the-loop” which utilizes real-time feedback to integrate spatial demand control to wireless networks where users are motivated to move to less congested areas.
- non-intrusive user feedback is widely discussed in the computational intelligence literature [7]— [9].
- the intrusive feedback collection methods do not represent all users because the majority of users wouldn't complain, they just change the provider.
- non-intrusive feedback collection methods enable more frequent feedback data collection which, consequently, increases the accuracy and relevance of networks decisions.
- This article proposes the utilization of context data along with non-intrusive user feedback data to personalize wireless networks. While the proposed personalization concept could potentially be applied to all wireless networks, we focus in this article specifically on wireless cellular networks as a use case. We introduce the concept of wireless network personalization through addressing the following four important questions:
- Service quality is defined as a comparison between subscribers’ expectations and service performance [10].
- Current networks are designed mostly to be a “Universal Fit,” where service providers deliver services with a quality appeal to all types of users.
- user expectations of service quality are not “One Size Fits All.”
- the way around this inefficiency is to tailor the network for each user's dynamic and context-dependent needs and expectations. This level of fine-grained network decision optimization will enable service providers to provide personalized, satisfactory services for the majority of users at a minimum cost.
- Personalized networks employ ML and big data analytics, which make real-time network decisions and actions possible through automation. Automation can be achieved by analyzing the enormous amounts of data produced by networks to identify relevant patterns and thereby predict context-dependent user needs and expectations.
- Network function virtualization (NFV) technology is proposed for 5G to isolate the software and hardware aspects of networks in order to transform network functions from dedicated hardware appliances into software-based applications.
- SDN software-defined networks
- NS network slicing
- the concept of NS is proposed to allow operators to provide customized, reliable services with increased efficiency while reducing capital expenditure and operating expenses of wireless networks.
- Each slice is associated with a set of resources, including bandwidth and network topology.
- network personalization provides an end-to-end virtual network tailored to each user-specific needs and expectations.
- Fig. 2 Drawing on concepts of service quality from business and marketing studies, our model of user satisfaction (S) in wireless networks is shown in Fig. 2.
- S user satisfaction
- Fig. 2 satisfaction is divided into four levels: A, B, C, and D.
- the division and number of satisfaction levels could vary depending on service providers’ preferences.
- QoS can be a vector with several elements, such as rate, reliability, latency, and jitter. Nonetheless, for simplicity, we assume that QoS is solely defined by rate.
- Our proposed user satisfaction model encompasses the following five main notions:
- QoSc/ the demanded QoS by the user, which represents the maximum QoS associated with the requested service.
- QoSp the provided QoS by the network.
- QoSai the adequate (minimum) QoS required to achieve a satisfaction level of i.
- D the difference between the QoS demanded by the user and the QoS provided by the network ( QoSd - QoSp).
- QoSp As shown in Fig. 2, as QoSp decreases, D increases and, consequently, satisfaction decreases. To keep user satisfaction at a certain level, QoSp should be within the ZoT associated with the targeted satisfaction level. It is important to note that QoS a( is what changes from one user to another, which consequently changes the width of the ZoT,. Moreover, demand is assumed to be dependent on the application and service type; hence, it is constant for all users requesting service of the same application.
- Fig. 2 we present a simple example to illustrate how ZoT, D, and user satisfaction are related.
- a service provider is trying to optimize the network such that a certain user in the network has a satisfaction level of B.
- the service provider is utilizing a big data-driven Al personalized network to predict ZOTB at different time slots (T1 to T6).
- the personalized network optimizes D during each time slot.
- the predicted user satisfaction level is D.
- the optimizer which is part of the personalized network, suggests increasing D fr to save resources (+).
- the optimizer suggests decreasing D U further by allocating more resources (-).
- the predicted user satisfaction for the provided D is level B, which is the targeted user satisfaction. Therefore, the optimizer suggests keeping D as is I (+ ⁇ -).
- the problem of extracting knowledge from this huge amount of data presents two subproblems: a big data problem and an Al problem.
- Al is defined as any process that senses the environment and takes actions to maximize the success probability of the targeted goal.
- supercomputers and distributed computing technologies are improving rapidly to the point that the use of big data analytics and prediction techniques for practical near real-time applications are currently possible.
- the proposed personalization frame-work collects information from the user environment and the network, predicts user needs and tolerance to service quality, and optimizes resource allocation to minimize cost and maintain certain user satisfaction levels.
- the proposed framework consists of three stages:
- the development stage is composed of the following modules all of which are implemented offline:
- Data mapping - Data from different users are mapped to shared space. Mapping user data enables ML models to capture correlations and inherent patterns. For instance, user location is recorded as GPS coordinates. However, generally, user satisfaction behavior is actually correlated to a particular type of location (e.g., home) rather than GPS coordinates.
- Cluster users into personas - A group of users who share similar user behavior and satisfaction patterns is referred to as a persona (see Part II). Associating users with pre-existing user personas will enable networks to provide highly personalized service with a minimal amount of data, thereby improving the efficiency of personalized networks. Nonetheless, at this stage, the network has no prior information on the number and types of unique personas implicitly available within the collected data. For this reason, at this early stage, unsupervised learning is used to cluster users into unique personas.
- Multi-phase persona and user satisfaction prediction the network has access to labeled context data with user persona and satisfaction levels.
- the processed data is used to build an ML model to predict user satisfaction levels for each user.
- the first phase is designed to output the personas probability vector.
- the second phase digest the personas probability vector along with the preprocessed labeled data in order to build a model capable of predicting the user satisfaction levels, for new and existing users, using a minimum amount of data.
- Part II we illustrate a satisfaction prediction example using different ML algorithms.
- the output ML model from the previous stage is integrated into the production environment to start making practical decisions based on new data.
- the production stage is where the network utilizes the trained ML models to achieve network personalization in real-time.
- the production stage is composed of the following modules:
- the first step is to continuously collect context information from users to predict personalized user satisfaction behavior.
- Data are collected from different sources, such as sensors and network data, and stored in a big data database (e.g. Hadoop distributed file system (HDFS)). Then, data are preprocessed using the same workflow used in the development stage.
- a big data database e.g. Hadoop distributed file system (HDFS)
- HDFS Hadoop distributed file system
- the next step is to use the multi-phase ML model trained in the development stage as an input to a multi-objective optimization problem.
- the optimization problem is formulated to micro-manage and optimize resources, and users’ satisfaction simultaneously based on each user’s QoS requirement and user satisfaction behavior.
- This optimization problem utilizes the ML model as its fitness function.
- the targeted satisfaction level for each user is decided by the network service provider and is fed as an input to the multi-objective optimization problem.
- the optimization problem outputs the optimum choice of D (D or i) which achieves the required satisfaction level using the minimum amount of resources.
- Resource allocation - The network utilizes A 0P t and other network parameters to allocate the best resource blocks (RB) that achieve the targeted user satisfaction level.
- the network After allocating resources to each user, the network records the QoSp along with user feedback (i.e., satisfaction).
- Tuning The measured user satisfaction is used to validate the predicted satisfaction levels. If user satisfaction was predicted correctly, the instance is fed to the database. Otherwise, the error is fed to the predictive model to relearn (i.e., concept drift). Relearning is used to improve the predictive model performance and to update the model with user behavioral changes that could occur over time.
- Online learning - Online learning is used to improve the predictive model proactively. Since network data become available in sequential order, batch learning techniques are not practical for real-time implementation. Instead, online learning techniques can dynamically adapt to new changes or patterns in user behavior and its relation to satisfaction
- the steps that need to be performed during the communication session should be assessed.
- the production process involves relatively fast operations such as data collection and using the ML models to performance predictions.
- the optimization of the resources should be done in near real-time. Although optimization could potentially require more time, meta-heuristic optimization provides a sufficiently good solution in a relatively short time.
- the development and deployment stages involve training, validating, and implementing the ML models. These are cumbersome, time-consuming, and involve heavy computation and processing. Nonetheless, since they are implemented offline, they should not affect the network proactivity.
- Fig. 4 illustrates two different contexts, C1 and C2.
- the value of D is illustrated by the length of the yellow bar and is depicted, for each context, in the yellow circles.
- the targeted user satisfaction for the considered user is assumed to be 5 and the QoSd is 5 M bits per second (bps).
- the non-personalized network allocated 3 Mbps, whereas the personalized network allocated 4 Mbps to reduce D from 2 to 1 Mbps.
- user satisfaction climbed from 1 for the non-personalized network to 5 for the personalized network.
- personalization enabled the network to predict that this user, during C2, would have a smaller ZoT5, and hence the minimum required QoSp is 4 Mbps. Accordingly, we can conclude the following: personalization can potentially increase user satisfaction to the desired level using a minimum amount of resources.
- the cellular network operator collects data about the users and stores it in a database.
- the collected data are of two types, real-time user satisfaction levels as well as context values, such as time, location, and application. Measurements are recorded at each measuring instant. The period between two measuring instances is referred to as a time slot (TS).
- TS time slot
- the operator collects data from the considered users with TS length of one second. Besides, the amount of resources used for each TS is recorded.
- Smin 4 satisfaction level of 4
- Table I Simulation parameters Dataset description - As shown in Fig. 3, user and network data are important requirements for personalized networks. Unfortunately, companies and institutions capable of collecting such data, particularly user data, do not publish them for privacy and confidentiality reasons. The way around this issue is to design and generate synthetic data that is flexible and has realistic characteristics.
- Parts II and III we proposed a synthetic dataset design to enable big data-driven wireless network personalization. The dataset is designed with four distinct user personas and it can be found in a publicly available GitHub repository [15]. The dataset is composed of context features along with their associated satisfaction values. The dataset in [15] is utilized to build the prototype for the proposed personalized network.
- the first premise of personalized networks is their ability to minimize the overall utilized resources at each instant. While resources in wireless networks are miscellaneous, in this paper, we confine resources to Bandwidth, which is proportional to QoSp in Mbps. The amount of saved resources is measured by calculating the difference between the QoSp provided by the non-personalized network (QoSNP) and QoSp provided by the personalized network (QoSPr) (i.e., QoSN P - QoSPr). In Fig. 5, we plot the total QoSNP, QoSPr, QoSd, and QoSNP - QoSPr for the three users vs. time in hours.
- QoSNP non-personalized network
- QoSPr personalized network
- the total amount of saved resources (QoSN P - QoSPr) fluctuates with time. Essentially, the network achieves the highest resources saving when the network attempts to maximize QoSp while the user has more tolerance to lower QoS. In this particular scenario, as shown in Fig. 5, the amount of saved resources was always greater than zero indicating that the personalized network was able to provide service with QoSPr ⁇ QoSNP; hence, it was able to save more resources (9703.8 Mbps over 24 hours) compared to the non-personalized network.
- the personalized network might suggest an increase in the provided resources to certain users (i.e., QoSPr > QoSNP) to push their satisfaction levels above the targeted minimum.
- QoSPr > QoSNP the provided resources for certain users
- this increase in the provided resources for low tolerance users is offset by the reduced amount of provided resources for high tolerance users.
- the extra amount of resources suggested to low tolerance users is the optimized minimum required to achieve targeted satisfaction.
- the second premise of personalized networks is their ability to maintain targeted satisfaction levels.
- Fig. 6 we plot the average user satisfaction for the three users vs. time in hours for both networks.
- the non-personalized network achieved higher satisfaction levels (an average of 4.87)
- the personalized network was able to maintain user satisfaction above the targeted level of 4 (an average of 4.31 ) and save resources, simultaneously.
- This article has proposed wireless network personalization as an enabler for resource micro-management based on users’ actual demands and needs.
- personalized networks utilize real-time non-intrusive user feedback coupled with context information to make fine-grained decisions that achieve higher user satisfaction levels using minimum resources.
- the user satisfaction model which is based on the notion of ZoT.
- this article focused on resource allocation, personalization can be employed to optimize various decisions in wireless networks, such as network failure detection and network security decisions.
- the technology and framework proposed for wireless networks can be applied to any network with users (e.g., wired networks) as well as other businesses and applications that require user feedback to improve service.
- wireless networks of the future will support personalized, fine-grained services and decisions by predicting user satisfaction in real-time using machine learning and big data analytics.
- Data- driven personalization will empower wireless networks to further optimize resources while maintaining user expectations of networks.
- acquiring data is necessary.
- datasets that comprise user behavior and corresponding user satisfaction information are generally not published due to privacy and confidentiality concerns.
- Part I One of the principal requirements of our proposed framework in Part I (and of any other personalization solution, for that matter) is to make wireless networks aware of the personalized experience of users and their satisfaction levels in real time.
- the real-time monitoring of user satisfaction levels by means of big data analytics and machine learning (ML) algorithms has been proposed for diverse applications, such as cloud gaming, healthcare, and smart cars [3]-[7]
- ZoT Zone of Tolerance
- the second key enabler for wireless network personalization is the availability of relevant datasets.
- a lack of published user behavior data labeled with ground truth user satisfaction information is holding back innovation into new approaches for personalizing wireless networks. Companies and institutions capable of collecting such data do not publish them for privacy and confidentiality reasons. The way around this issue is to design and generate synthetic data that is flexible and has realistic characteristics.
- synthetic data removes privacy and confidentiality concerns; therefore, it can be made publicly available for researchers.
- synthetic data can be generated in large volumes and with complex well- understood characteristics.
- Another fundamental advantage of synthetic datasets is the ability to redesign and change data structures as needed by varying certain input parameters. So instead of relying solely on real data over which we have little or no control, synthetic data provide us with great flexibility.
- synthetic data generation methods enable the generated datasets to cover most of the data space needed to generate meaningful results, and hence, provide us with deeper insight and stronger conclusions.
- the quality of the generated data can be controlled using various design techniques [9] For all its benefits, the advantages of synthetic data come at a price.
- Real data are usually dirty and contain various types of errors; hence, it can be foreseen that the creation of data with characteristics similar to real-world data is not straightforward. Synthetic data also needs to reflect errors, distributions, and patterns that exist in real data. Finally, an additional validation step might be necessary to ensure that the conclusions drawn from synthetic data extend to real-world applications.
- Fig. 7 Drawing on concepts of service quality from business and marketing studies, our model of user satisfaction in wireless networks is shown in Fig. 7 (note that Fig. 7 is much the same as Fig. 2, except that Fig. 2 uses a model satisfaction with levels A, B, C and D).
- Fig. 7 is much the same as Fig. 2, except that Fig. 2 uses a model satisfaction with levels A, B, C and D).
- satisfaction is divided into six discrete levels: 0, 1 , 2, 3, 4, 5.
- Our proposed user satisfaction model encompasses the following five main notions:
- QoSp the provided QoS by the network.
- QoSai the adequate (minimum) QoS required to achieve a satisfaction level of i.
- A the difference between the QoS demanded by the user and the QoS provided by the network (QoSd - QoSp).
- QoSp As shown in Fig. 7, as QoSp decreases, A increases and, consequently, satisfaction decreases. To keep user satisfaction at a certain level, QoSp should be within the ZoT associated with the targeted satisfaction level. It is important to note that QoS a , is what changes from one user to another, which consequently changes the width of the ZoT / . Moreover, QoSd is assumed to be dependent on the application and service type; hence, it is constant for all users requesting service of the same application.
- Context is a combination of several variables that influence user satisfaction, such as time, location, speed, and activity.
- Context variable modeling is addressed in Section III.
- users will have a consistent satisfaction behavior which depends solely on A.
- the following equation models user satisfaction (S) vs. A at a certain context:
- the tree data generator is a structure that specifies the relationships and correlation between context variables in user behavior generated datasets. This structure defines the rules, patterns, and dependencies that the generated data need to follow.
- Fig. 9 we illustrate a sample TG structure.
- Each TG is a collection of several nodes, where each node has a value and could have a child node. Nodes that do not have child nodes are called leaf nodes, whereas nodes with child nodes are called parent nodes.
- the first node in TGs is called the key attribute node (see Fig. 9). Since context information datasets are collected over time, the key attribute node will always be time.
- One of the main advantages of using TG to generate context data is its ability to exclude impossible or unwanted variable combinations.
- TG nodes are of two types: rule-based nodes and Hidden Markov Model-based (HMM) nodes.
- the values of rule-based nodes are generated on the basis of rules specified by TG inputs.
- the TG in Fig. 9 has a rule-based node called weekdays (WD).
- WD node has two possible values: weekday and weekend. In order to determine whether a day is a weekday (working day) or a weekend day, a rule is integrated into this node.
- HMM nodes by contrast, compute the state sequence conditional probabilities and perform a weighted random choice based on the computed probabilities.
- Fig. 10 illustrates the decision process of HMM nodes, where Xi represents a hidden state sequence, A is the state transition probabilities, and B is the observation probability matrix.
- the matrix B ⁇ b,, ⁇ is N x M with where an observation ( O f ) of an HMM node at time t is the sequence of values recorded at the predecessor parent nodes and M is the number of observation symbols.
- Identical TG input parameters mean that people belonging to the same persona have identical A and B matrixes, which indicates that they will have similar distributions and patterns within their data.
- Table II lists the four personas and their corresponding ring of locations for both weekdays and weekends. Locations grouped in one bracket indicate that the user can go to either one with a certain probability.
- the four datasets generated for the four personas are available in [10].
- Table II Weekdays and weekends location rings for the four user personas.
- Integrating errors into synthetic datasets is a fundamental requirement for creating datasets with realistic characteristics. Different datasets have different types of errors depending on several factors, such as the utilized data collection method and data type.
- One strategy that can be used to add errors to datasets is the integration of real data measurements with inherent errors and noise.
- we integrate real data from the sensor measurements dataset available in [11] The authors in [11] carried out an experiment to collect data for a group of 30 volunteers with ages ranging from 19 to 48 years in order to obtain a dataset of phone sensor data labeled with ground truth activity labels. The activity labels were the following: standing, sitting, laying down, walking, walking downstairs, and walking upstairs.
- Sensor data are collected from sensors, such as body acceleration, gravity acceleration, and body angular speed sensors.
- HA[Activity][i ⁇ represents the value in dataset HA, at column Activity and record number i):
- HAL Create a lookup table for HA dataset called HAL.
- HAL has three columns: Activity, Indexes, and Number of Records (NoR). Each row is called a record.
- One record is created for each consecutive set of HA records that have the same activity label.
- the indexes i.e., row numbers
- Algorithm 1 describes the process of creating the lookup table HAL.
- Algorithm 1 Dataset lookup table generation
- Algorithm 2 Real sensor data augmentation
- TR 4 The UCL's total number of records
- rand 4 random number between 1 and the length of AF
- Fig. 11 we visualize the problem of user satisfaction prediction from user context data. As shown in Fig. 11 , all the variables affecting user satisfaction prediction accuracy at a certain context are visualized as a plane. Each plane is associated with a certain context and user satisfaction behavior (i.e., mapper). User satisfaction behavior in each context is dependent on all context variables in the context plane. Missing variables in the context plane will create gaps, which will add noise to the associated user satisfaction mapper as a result and therefore hinder the accurate prediction of user satisfaction. It is worth noting that the accuracy of user satisfaction prediction is highly correlated with the number of missing variables and the amount of missing information.
- DT tet al.
- Knn tet al.
- the accuracies for the three algorithms are labeled DT-a, Knn-a, and RF-a. • A dataset with noisy satisfaction and real sensor measurements augmentation. The accuracies for the three algorithms are labeled DT-na, Knn-na, and RF-na.
- Fig. 13 depicts the 10-folds cross-validation prediction accuracies for the four experiments and for each ML algorithm. Since there are six satisfaction levels, the random choice accuracy level is 0.166. It can be seen that the best accuracies were achieved by the first experiment, which is expected since this experiment was performed on error-free data. By contrast, the fourth experiment had the worst accuracies of all experiments due to the errors of the added satisfaction uncertainty and the real sensor measurements. It is worth noting that the predictors’ performance in experiments three and four rely on the amount of noise added to the satisfaction mappers (i.e., the value of s).
- Wireless network personalization by means of machine learning and big data analytics is a tremendously promising research area.
- new research has been limited by a lack of published user behavior data with ground truth satisfaction labels due to privacy concerns and other technical limitations.
- error generation and augmentation strategies have been discussed.
- sample user satisfaction prediction experiments have been conducted and the effect of error integration on the prediction accuracies has been discussed.
- Wireless network personalization is proposed as a dynamic context-aware approach to maintaining the targeted personalized satisfaction levels with minimum resources.
- Wireless network personalization has two key enablers: measuring and predicting user satisfaction in real-time, and datasets that have both context and user satisfaction information.
- ZoT Zone of Tolerance
- wireless network personalization Due to the advantages of high levels of user satisfaction, and as a result of the fierce competition in the telecom industry, user satisfaction has started to attract attention from industry-based and academic communities. Superior levels of user satisfaction have proved to be a predictor of higher revenue, increased customer loyalty, and higher efficiency. Since user satisfaction is the ultimate goal, network overprovisioning can be avoided by designing enhanced context-aware networks that are capable of achieving the required user satisfaction levels using minimum resources. Such an approach is referred to as wireless network personalization. It utilizes non-intrusive real-time user satisfaction feedback to personalize wireless network decisions and thereby micromanage resources so that the required user satisfaction levels are achieved with a minimum allocation of resources.
- Fig. 7 Drawing on concepts of service quality from business and marketing studies, our model of user satisfaction in wireless networks is shown in Fig. 7 [7], We propose dividing user satisfaction into levels where each level is associated with a certain range of QoS. In Fig. 7, satisfaction is divided into 6 discrete levels: 0, 1 , 2, 3, 4, 5. The number of satisfaction levels and how these are divided could vary depending on service provider preferences.
- QoS can be a vector with several elements, such as rate, reliability, latency, and jitter. For simplicity’s sake, however, we assume here that QoS is solely defined by rate.
- Our proposed user satisfaction model encompasses the following five main notions:
- QoS p the provided QoS by the network.
- QoSa the adequate (minimum) QoS required to achieve a satisfaction level of i.
- D the difference between the QoS demanded by the user and the QoS provided by the network (QoSd - QoSp ).
- QoS p As shown in Fig. 7, as QoSp decreases, D increases and, as a result, satisfaction decreases. To keep user satisfaction at a certain level, QoS p should be within the ZoT associated with the targeted satisfaction level. It is important to note that QoSa , is what changes from one user to another, which consequently changes the width of the ZoT,. Moreover, QoS d is assumed to be dependent on the application and service type; hence, it is constant for all users requesting service of the same application. Also, QoS d 3 QoS p ; therefore, D > 0. III. DATASET GENERATION MODEL
- Wireless network personalization is an application that utilizes context data along with user satisfaction data to predict the real-time user satisfaction level in a non-intrusive manner, using it to make personalized network decisions (see Part I).
- context information datasets exist, large scale context information datasets, as well as datasets that have both context data and the corresponding satisfaction values, are not publicly available.
- the proposed dataset structure consists of user context data and the corresponding satisfaction values.
- Context data are composed of a set of context variables, such as time and location.
- the measurements recorded in the proposed dataset are generated for one user over one year.
- the model used to design each context variable is also presented.
- a context variable could be a measured or engineered feature. Measured features, such as time, day, location, and speed are directly obtained from sensors. By contrast, engineered features, such as activity, are created on the basis of the measured features. Measured and engineered features are recorded at each measuring instant. The period between two measuring instants is referred to as a Time Slot (TS).
- TS Time Slot
- Fig. 14 illustrates sample instances from the proposed synthetic dataset.
- Fig. 15 illustrates these relations between four examples of features. Nodes represent features and links represent relations between features.
- Date, Time of the day, and Day of the week are among the measured features added to the dataset.
- days of the week are classified into weekdays and weekends to create the Classified days engineered feature.
- the time of the day is classified into 7 periods to create the Time period engineered feature. Time periods are assigned as follows: Early Morning: 04:00AM-05:59AM, Morning: 06:00AM-10:59AM, Mid-day: 1 1 :00AM-13:59PM, Afternoon: 14:00PM-17:59PM, Night: 18:00PM-23:59PM, After midnight: 00:00AM-03:59AM.
- the user considered for this dataset is assumed to live in Ottawa, Canada. Instead of recording the user’s GPS location, we engineer the user Location feature, which is created by dividing the Ottawa area into a 100 * 100 grid with a distinct ID for each square. The area covered by each square is 91 * 10 ⁇ 5 km 2 and the distinct location ID is recorded to the dataset under the Location feature. Usually, most people follow location patterns in their daily life. Although these patterns do not occur with 100% probability, for the purpose of this dataset, we assume that the user has a certain and distinct location pattern for weekdays and weekends. In this paper, the set of roads and destinations that constitute a location pattern is called the ring of locations. Roads and destinations are added to the dataset under the Location name feature.
- Fig. 16a and Fig. 16b depict the weekday and weekend rings of locations for the considered user, respectively.
- Fig. 17 in order to illustrate the correlation between the time period and location features, we plot the percentage of time the user spent at each location over each time period on a weekday. From Fig. 17, it can be deduced that the user spends most of his ⁇ her time on weekdays at work and at home.
- Fig. 16a and Fig. 16b illustrate the speed heat map in km/hr for each location square using data generated for two days, a weekday and a weekend day, respectively.
- User speed in km/hr is classified into four ranges: High, Medium, Low, and Zero speed to create the Speed range engineered feature.
- N service requests in a day with at most one request within a TS.
- the arrival of a request during a certain instant is recorded to the dataset using a binary variable called Request arrived (see Fig. 14).
- Request arrival is dependent on time (t) in hours, the day of the week (d), location (I), and speed (s) in km/hr.
- a non-homogeneous Poisson process with arrival rate A r (t, d, I, s) is defined as a counting process N(t) for t > 0.
- the probability of a request arrival in an interval d is given by:
- a r is modeled as a function of the context variables f, d, I, and s.
- the application feature is also correlated with other context variables, such as time and location, and it is recorded to the dataset under the Application feature.
- Table III The set of services and their associated demand rate for the considered set of applications.
- Fig. 20 depicts the heatmap of the instance count for services requested by the user and their corresponding application over a one-week period. As shown in Fig. 20, the most popular service for this user is WhatsApp video.
- the Demand rate (i.e., QoS d ) feature represents the rate requirement associated with the service requested by the user.
- Table III lists the service demand rate requirement data aggregated from several sources.
- the rate assigned to the user by the network is recorded to the dataset under the Given rate (i.e., QoSp) feature.
- Given rate is designed in such a way that it is correlated to some other context variables,
- RECTIFIED SHEET (RULE 91.1) such as time and location to model the effect of cell congestion.
- A is computed and recorded to the dataset under the Delta feature.
- a and y have fixed values and user satisfaction model is a function of A.
- S is solely a function of A.
- ZoTs increase with the decrease of A for all satisfaction values.
- each context is associated with a certain combination of A and y.
- Fig. 22 depicts the 10-folds cross-validation prediction accuracies for each ML algorithm. Since there are six satisfaction levels, the random choice accuracy level is 0.166. Fig. 22 shows that the best performance is achieved using RF (ensemble) algorithm with an average accuracy of 0.884 compared to 0.85 for DT and Knn. It is worth mentioning that the predictors’ performance should be directly related to the number of relevant context variables available to the ML algorithms. In a future article, we will address the effect of insufficient context variables and the effect of noise introduced by sensors on the predictors’ performance.
- Wireless networks are designed to satisfy a minimum QoS requirement in order to assure that users receive a satisfactory service at any network condition. Integrating personalization into wireless networks will enable further optimization of the available resources such that actual user demand can be matched with a personalized QoS offered by the network. This level of micro-management can be achieved through the utilization of Al and big data analytics to predict and thereby optimize user satisfaction in different contexts. User satisfaction prediction necessitates the continuous measurement and tracking of user satisfaction in wireless networks, which is why the ZoT model was proposed. Based on the ZoT model, a synthesized context-based dataset was modeled along with the corresponding user satisfaction values. Eventually, an exemplary user satisfaction prediction experiment conducted on the proposed synthetic dataset showed that RF has a superior performance compared to DT and Knn.
- Wireless network personalization is an emerging technology that has considerable potential to achieve the ultimate balance between resource allocation and user satisfaction.
- One of the main enablers of personalized networks is the continuous monitoring and prediction of dynamic user satisfaction levels in various contexts. Accurate satisfaction prediction requires a lot of data, and unfortunately, data and the process of acquiring it are expensive. A closer look at user behavior and satisfaction levels reveal that certain users share certain similarities. A group of users who share similar user behavior and satisfaction patterns is referred to as a persona. Associating users with pre-existing user personas will enable networks to provide highly personalized service with a minimal amount of data, thereby improving the efficiency of personalized networks.
- ZoT Zone of Tolerance
- Part V we proposed a data-driven framework to measure and predict user satisfaction values in wireless networks. This framework describes the process of acquiring, analyzing, and modeling user satisfaction information using machine learning (ML), particularly deep learning techniques. Another step that supports the proposed framework in Part V and makes it more efficient and robust is the concept of user personas (see Part II).
- ML machine learning
- a persona in a personalized wireless network is a set of user context patterns and associated user satisfaction behavior characteristics that are shared by a group of people.
- Another notable advantage of integrating user persona modeling is that it enables personalized wireless networks to provide personalized services to new users joining the network about whom there is not yet enough information concerning their preferences or behavior in the network. This can be done by simply associating the new user with a pre-existing user persona.
- the user persona designs proposed in the literature can be categorized into two main categories.
- the first design category involves personas with fixed characteristics designed by domain experts. For instance, the author in [12] proposed a statistical technique to create user personas based on user needs and preferences.
- the second category involves flexible personas with dynamic characteristics created using data-driven real-time analytics, such as the solutions proposed in [13] and [14]
- a satisfaction behavior for a user represents his/her satisfaction patterns for various levels of service performance in a specific context. Integrating the capabilities of identifying these behavioral and satisfaction patterns is key for enabling personalized wireless networks.
- Fig. 23 we propose a data-driven framework for implementing user persona prediction in personalized wireless networks. The proposed framework consists of three main processes: development, deployment, and production. In this section, we discuss these three processes and the steps involved in each.
- the development process goes through multiple stages before the output model is deployed in the network.
- the steps involved in the development process are.
- WP working professional
- HS high school student
- US university student
- HM homemaker
- Data mapping In order to capture correlations between users belonging to the same persona, data should be mapped to shared space. For instance, user location is recorded as GPS coordinates. However,
- RECTIFIED SHEET (RULE 91.1) generally, user satisfaction behavior and persona types are actually correlated to a particular type of location (e.g., home) rather than to GPS coordinates. Therefore, the features in the CH dataset are mapped to the same shared space.
- Preprocessing The next stage is to preprocess the data, which consists of the following steps.
- the CH dataset was recorded for a period of one year. Also, the resolution (i.e., the time period between two records) of the CH dataset is one second. For the purpose of this paper, we selected a part of the dataset where the filtered dataset spans a four month period. The total number of records in the dataset is 31 ,536,000. Naturally, users do not use the network during all measuring instances. Therefore, since user satisfaction behavior is an important part of user persona design and prediction, we filtered out records that had no service requests from the network and that did not carry any satisfaction information. The dataset was filtered on the basis of the Request arrived feature, which is set to 1 when a user requests a service from the network and, otherwise, is set to 0. The filtered number of records is 38,166.
- PCA principal component analysis
- PCA identifies the hyperplane that lies closest to the data and then projects the data onto it.
- the hyperplane is chosen such that it preserves the highest variance.
- the PCA algorithm needs the number of principal components to be fed as input. The number of components should be as small as possible while maintaining a reasonable variance in the data.
- a useful piece of information that can be used to find the optimal number of components is the explained variance ratio (EVR) of each principal component.
- EVR describes the percentage of the dataset’s variance that lies along the axis of each component.
- Fig. 24 we plot the cumulative EVR as a function of the number of components. For the purpose of our model, we aimed to preserve at least 98% variance using the minimum number of components. As shown in Fig. 24, 98% cumulative EVR was achieved using at least 50 principal components.
- the next step is to deploy the model into production.
- Production machines are continuously running operating systems with servers that are highly optimized to meet expected load and demand.
- Collect data stack for each user In this stage, the network collects data from new users joining the network in order to predict their persona type. As mentioned earlier, a user persona describes users with similar user behavior and satisfaction patterns over time. Therefore, in order for the predictor to achieve accurate results, it should be able to decide based on more than one data record. For this reason, the proposed predictor is designed to congest a stack of data prior to deciding on the predicted user persona. To this end, we used random statistical sampling without replacement in order to build data stacks for different users. Each stack was sampled using the entire population of the test data associated with the ground truth persona. For each persona, we sampled 500 stacks of data for 500 users. Hence, the total number of testing stacks was 2000 stacks with a stack size of B. The effect of changing B on the performance of the proposed framework will be discussed in Section IV.
- Compute confidence After predicting the persona label for each data record, the next stage in the proposed framework is to assess the confidence of the predictor.
- the confidence is computed using two steps.
- the first step is to compute the histogram of the predicted labels listed in ⁇ 4 for each user / ' .
- the second step is to compute the confidence of the prediction by passing H, through a softmax layer defined by
- y is the one-hot encoded vector associated with the persona. For instance y for persona number 2 is [0, 1 , 0, 0]
- the proposed persona prediction framework is supported by a validation stage.
- the validation stage checks whether the confidence of the predicted persona is greater than a specified threshold C/, ⁇ As shown in Fig. 23, if Q,P 3 C / . « , P is passed as the final prediction result. However, if c,,p ⁇ (3 ⁇ 4 «, the network is directed to increase stack size, and hence collect more data from the user.
- the validated predictions are fed back to the deployed model.
- the deployed model continuously learns from the arriving samples in a process known as online learning.
- Second hidden layer 1024 neurons.
- prediction confidence plays a vital role in our proposed framework. Instructing the network to act on the basis of a falsely predicted persona will dramatically affect user experience and satisfaction levels. Since the network has no prior definite knowledge of the persona of new users, the framework is designed to assess the confidence levels of the predictions and consider them only if they pass a certain threshold.
- Fig. 27 we plot the DNN predictor confidence vs. stack size. It can be seen that, similar to the ML model accuracy, the confidence levels increase with the increase in stack size. Besides, the choice of optimum confidence threshold comes with a tradeoff. The higher the threshold, the greater the amount of data required for the predictions to pass. Interestingly, although the variance of the predictors’ accuracy is highly dependent on stack size, the variance in confidence levels for the four personas does not significantly change with the increase in stack size.
- wireless networks will be expected to support a wider range of applications and use cases, such as vehicular ad-hoc networks and virtual reality applications.
- Such applications require network services to be delivered with a variety of network performance characteristics (e.g., rate, latency, security, and quality of experience (QoE)), which poses fundamental technical challenges for the management of user experience.
- QoE quality of experience
- Enabling wireless networks to understand and characterize the relationship between network performance and user experience will empower networks to make more personalized decisions (e.g., configurations) and optimized actions (e.g., resource allocation).
- Personalizing wireless networks is the cornerstone of optimum resource allocation and user experience management.
- Part I we proposed a data-driven Al-based wireless network personalization framework, which enables networks to micromanage resources and make fine-grained personalized decisions based on dynamically changing user needs and expectations.
- One of the main enables of the proposed framework in Part I is user satisfaction measurement and monitoring.
- DNNs automate feature extraction from data that has complex structures and correlations, which thus reduce expensive human-dependent tasks that hinder automation and real-time network operations.
- data collected from wireless networks is increasingly large and heterogeneous and arrives in different formats and speeds from different sources [4]
- DNNs are considered to be one of the best tools for learning useful patterns for complex and colossal wireless network data.
- Intrusive collection methods require users to actively interact with the system to record satisfaction information.
- non-intrusive user satisfaction collection methods employ ML and Al to predict personalized user satisfaction without the need to disturb users.
- intrusive collection methods include surveys and feedback boxes.
- wireless networks the utilization of user feedback from intrusive methods is discussed in [5], [6].
- the authors in [5] propose an approach called”user-in-the-loop” which utilizes real-time feedback to integrate spatial demand control to wireless networks where users are motivated to move to less congested areas.
- the authors in [6] propose a data-guided resource allocation approach where offline feedback data (e.g., network measurements and user complaints) is employed to improve the average user experience.
- offline feedback data e.g., network measurements and user complaints
- the intrusive feedback collection methods do not represent all users because the majority of users wouldn't complain, they just change the provider.
- non-intrusive feedback collection methods enable more frequent feedback data collection which, consequently, increases the accuracy and relevance of networks decisions. Therefore, non-intrusive feedback methods are considered more practical compared to intrusive feedback. Nonetheless, due to the lack of data and the immaturity of the technology required to acquire and utilize non-intrusive user satisfaction feedback in wireless networks, it is not a common discussion topic in the literature and is limited to but a few applications [7] Nevertheless, non-intrusive user feedback has been proposed to make relevant automotive decisions in many applications, such as cloud gaming, healthcare, and human-computer interaction [8]-[10] In this paper, the proposed framework will enable non-intrusive user satisfaction feedback collection in wireless networks.
- Input features are a set of network KPIs [15], [16].
- Input features are a combination of network parameters and measurements aggregated directly from sensors.
- the authors in [17] proposed monitoring user QoE using a set of network parameters and EEG signals collected from sensors placed near the brain.
- Fig. 28 illustrates the complete framework we envision for predicting user satisfaction in wireless networks. This framework is designed to autonomously predict future personalized user satisfaction values in real-time for each user in the network in a non-intrusive manner.
- the proposed framework consists of the following four processes:
- the first step is to acquire context data, which can be done by monitoring sensors, aggregating and analyzing collected data, and predicting missing and future context information.
- KPI data is collected from the network and aggregated with context information.
- the authors in [19] provide some details on the process of acquiring and monitoring user context data.
- using all types of network parameters in the user satisfaction prediction problem is very popular in the literature.
- user experience is directly influenced by the network KPIs, and due to the fact that, as far as the user is concerned, the changes in network parameters directly affects KPIs, our framework suggests considering KPIs as the only network variables.
- the second step is to map data from different users to shared space.
- Mapping user data is a valuable step as it enables ML models to capture correlations and inherent patterns.
- Fig. 29 we illustrate an example of location feature mapping to a common space.
- User location is acquired from GPS sensors and its recorded as unique coordinates.
- user satisfaction behavior is actually correlated to a particular type of location (e.g., home) rather than GPS coordinates.
- the recorded coordinates should be classified into locations types.
- feature values GPS coordinates
- the common feature space is obtained.
- samples belong to four location classes.
- the color indicates the class of an instance.
- Data preprocessing - The third step is to retransform data and extract useful features. This step is important because it contributes to the reduction of noise and irrelevant data, which can degrade the predictive model’s performance.
- the second process in the framework is labeling context data by the actual user satisfaction values.
- the proposed process to capture actual user satisfaction is as follows:
- Emotions recognition The first step is to predict and measure user emotions and feelings.
- the detection of user emotions in a non-intrusive manner is widely discussed in computational intelligence literature.
- Emotions can be monitored using different types of data input, such as images and video [20], speech and sound [21], [22], body language [23]- [25], and other commercial sensors [26], [27]
- the next step is to estimate the correlation between measured user feelings and the performance of the service being used by the user. This step is necessary due to the fact that detected feelings are mostly not related to provided service performance.
- Processes 1 and 2 are designed to collect and label user data in an automated way in order to make the framework scalable and increase network intelligence and ability to detect users’ actual needs and demands.
- the network has enough labeled data to build a user satisfaction prediction model for each user with a good performance.
- ML algorithms There are several ML algorithms that can be employed and different techniques to tune the models depending on the type of data being utilized.
- the fourth process in our proposed framework is to predict future unlabeled user satisfaction values for each user from the context-satisfaction datasets aggregated by the network and stored in the database.
- the steps that need to be performed during the communication session should be assessed.
- the first process involves relatively fast operations, such as data collection and preprocessing.
- the second and third processes involve cumbersome, time-consuming, and complex operations, such as training, validating, and implementing the ML models. Nonetheless, since they are implemented offline, they should not affect the network pro-activity.
- the fourth process is operating during the communication session and it involves fast operations, such as using the ML models to performance predictions. Therefore, implementing this process in real-time would not raise practicality concerns. In this paper, we implement the third and fourth process in this framework.
- ZoT Zone of Tolerance
- D the difference between the KPIs demanded by the network and those provided by it.
- the ZoT model is designed to be personalized to each user and to dynamically change with the context in order to reflect the real characteristics of user behavior in the network.
- satisfaction values are modeled as discrete values (0-5) which makes the process of user satisfaction monitoring and prediction easier, stable, and more accurate.
- the rest of the paper discusses the process of building a DNN model in order to predict user satisfaction from context data, which are the third and fourth processes in our framework.
- Table V shows the features of the WPP dataset and an example of their values (the dataset in [30] has other features, such as real sensor measurements. However, for the purpose of this paper, we consider only the features listed in Table V).
- the WPP dataset was recorded for a period of one year with one record for each second.
- the total number of records is 31 ,536,000.
- the dataset has an indicator feature called Request arrived which is set to 1 when the user requests a service from the network.
- the filtered number of records is 38,166.
- MinMax scaling which is typically done via the following equation:
- Neural networks do not accept categorical values. Hence, encoding categorical values is needed prior to feeding the data into the model. In this paper, we encode categorical values using one-hot encoding.
- NNs One of the most important preprocessing steps for NNs is to check whether we have an unbalanced training data. Training NNs using unbalanced data will result in ignoring the classes with smaller representation in the dataset, which creates a biased predictor.
- Fig. 30 we plot the histogram for the user satisfaction classes of the considered WPP dataset. The histogram of the classes shows that this dataset is highly imbalanced. In order to address this issue, we oversampled our dataset using the SMOTE algorithm [31].
- the dataset is split into 70% training set, 25% validation set, and 5% test set.
- the validation set will be used to tune the DNN model and the test set will be used to perform the hold-out test to make sure that the model generalizes well on new testing data.
- Stratified sampling is used to ensure that the statistics of satisfaction labels are similar for both the training and testing sets.
- our network architecture consists of 6 layers.
- the first layer is the input data layer.
- Layers 2 to 5 are the hidden layers, and layer 6 is the output layer with 6 nodes. The number of nodes per hidden layer will be tuned in the following section for the purpose of improving the prediction performance (i.e., classification accuracy) of the model.
- the output of each neuron is computed through two steps. The first step is to calculate the weighted linear combination of the inputs. Then, the second step is to calculate the nonlinear transformation of the output of the first step. Updating the DNN parameters is divided into two processes:
- each layer is updated using the following equation: where / is the layer, w ,/ is the weight of the connection from neuron (/, M) to neuron (k, I ), b' k is the bias of the (k, I) unit, and f( ) is a nonlinear activation function.
- ReLU rectified linear unit
- the output score is computed from the first layer to the sixth layer. Since we have a multi-class classification problem, we use the categorical cross-entropy as our cost function, which is given by
- the gradient is updated as follows:
- the implementation of the experiments in this paper were done in python.
- the DNN model was built using the Tensor-Flow library.
- TensorFlow is a high-performance computational framework with a highly flexible structure. Originally developed by Google’s engineers, TensorFlow comes with strong support for machine learning and especially deep learning algorithms. In addition, the Sklearn library was used for preprocessing the data, whereas seaborn and matplotlib were used for visualization purposes.
- the first step is to choose the best DNN architecture that yields the best performance.
- Second hidden layer (Layer 3): 512 neurons.
- Adagrad updates the model parameters according to the following rule: where G f, « is a diagonal matrix.
- G f, « is a diagonal matrix.
- Adam adaptive moment estimation
- L1 and L2 are the most popular types of regularization strategies, and they update the general cost function by integrating another term to reduce the values of the weight matrices.
- Dropout is also a very interesting regularization technique and it is widely used in the field of deep learning [34] Dropout randomly picks some nodes and removes them and all the incoming and outgoing connections associated with it. For our DNN design, we choose to integrate dropout into each layer. The dropout rate was tuned to 41 % for each layer.
- Fig. 34 we plot the accuracy of our DNN model vs. different training sizes. It can be seen that too few examples will result in low test accuracy. Moreover, Fig. 34 illustrates an increasing trend in test accuracy, which means that as more information becomes available to the wireless network, the performance of the user satisfaction predictor improves. 2) Performance results
- Fig. 35 illustrates the change in training accuracy and validation accuracy with the increasing number of epochs. From Fig. 35, it can be noted that the training and validation accuracies increase gradually with the increasing number of epochs. At 120 epochs, the training accuracy is 92%, whereas the validation accuracy reaches around 82%.
- we perform a hold-out test to check if the predictor is able to achieve comparable accuracy using the test set. The accuracy of the hold-out test is 0.81 %, which indicates the DNN design generalizes on new data.
- CV test we perform a CV test.
- RSPN personalized wireless cellular networks
- resource scheduling problems for cellular networks are modeled as an optimization problem with a single objective which is to maximize throughput, spectral efficiency, or fairness under certain constraints.
- the objective could be a trade-off between throughput/spectral efficiency and fairness [1]
- MOO problems are used to model optimization problems with more than one conflicting objective functions.
- RSPN is defined as a biobjective optimization problem that maximizes two conflicting objective functions: the total D for all users and the average satisfaction for all users.
- MOO problems are solved by finding the set of mutually nondominant solutions called the Pareto front.
- Pareto front solution set there is no solution better than the other considering that all solutions trade off the conflicting objective functions.
- the network can make granular personalized resource allocation decisions for each user to ensure that the required satisfaction level is achieved with the minimum cost (i.e. , resources).
- ⁇ ⁇ 1 , 2, ...., B ⁇ eNBs available in the network.
- k is constant for all b e S.
- the RB allocation indicator is denoted by binary decision variable e ⁇ 0,1 ⁇ , where x ⁇ n) _ ⁇ 1 > if RB n is assigned to « (i , 6)
- a uiW [ (i U(i 6)) , a (2, U(l b)) , , , «((,3 ⁇ 4)) ].
- a uab)) is the value of feature J for user u (i:bj .
- the inputs for the machine learning satisfaction prediction engine are user ID, context features A u(i b) , and user demand D u i,b) .
- the first objective function in (9a) aims at maximizing the average u b ( u b ) in order to maximize resourcesaving in the network.
- the second objective function in (10a) aims at maximizing the average satisfaction for all users. Both objective functions contradict with each other; hence, the solution set is expected to be a Pareto front where the optimum points trade-off both objectives.
- constraint (9b) ensures that, during each scheduling instant, each RB is being used by no more than one user.
- the second constraint (9c) prevents each eNB from allocating a total power more than the budget power PTM ax .
- the third constraint (9d) limits the rate provided to each user to values less than the demanded rate fRu b ,b.
- the forth constraint (9e) ensures that the allocated power for each user ' s a positive value.
- the second objective function (10a) has five constraint. Constraints (10c), (10d), (10e), and (1 Of) are similar to the constraints associated with the objective function in (9a).
- the last constraint (10g) maintains the satisfaction lower bound for each user. This constraint is added to differentiate between users target satisfaction and hence, provide a wider service level range and customizable prices.
- the cell within a cellular network that covers Ottawa, Canada.
- the cell has one eNB and it is connected to three active users moving within its coverage area.
- the area of the cell is divided into a k * k grid.
- the cellular network environment is simulated using the parameters listed in Table VI.
- the cellular network operator collects data about the users and stores it in a database.
- the collected data are of two types, real-time user satisfaction levels as well as context values, such as time, location, and application. Measurements are recorded at each measuring instant.
- the period between two measuring instances is referred to as a time slot (TS).
- the operator collects data from the considered users with TS length of one second. Besides, the amount of resources used for each TS is recorded.
- Fig. 36 a-d illustrate the average u b (Au b ) vs. the average satisfaction for all users during a single simulation frame (frame A).
- Fig. 36 a-d illustrate the optimum Pareto front solution for the algorithms NSGAII, NSGAIII, SPEA2, and EpsMOEA.
- S m/ n 1.
- RECTIFIED SHEET (RULE 91.1) other points that achieve that satisfaction value using less resources (i.e., higher A Ub ).
- they could choose any S min ; hence any of the Pareto front solutions to achieve the required average satisfaction using the minimum resources.
- Fig. 37 a-d illustrate the optimum Pareto front solutions at a different simulation frame (frame B) for the algorithms NSGAII, NSGAIII, SPEA2, and EpsMOEA.
- the Pareto front solution set is different at different frames. This due to the fact that the solutions depend on several factors including the satisfaction behavior for each user, context, and the cellular network environment.
- embodiments of the invention provide network operators with improved flexibility of operation in terms of personalized user satisfaction (rather than one averaged satisfaction value good for most users) and the amount of consumed resources.
- Personalized networks are able to efficiently exploit network resources in order to achieve the ultimate balance between user experience and profit.
- Personalized networks can enable operators to attract more customers with a various service price preferences. In other words, user satisfaction can be personalized to each user based on the service price charged by the operator.
- the technology and framework proposed for wireless network can be applied to any network with users (e.g., wired network and WiFi). In addition, it can be applied to other businesses and applications that require user feedback to improve the service. However, for the purpose of our work, we engineered the features in our ZoT model to fit wireless networks.
- Embodiments of the invention can be integrated to the current networks at no extra cost in terms of infrastructure. However, there may be extra computation power needed to operate the network. In addition, user feedback should be capture, quantified and labeled based on the proposed ZoT mode. Many companies could have interest in this technology especially in the telecom industry, including, for example: incumbent network operators, incumbent network equipment vendors and manufacturers, new entrants to wireless and networking systems, and companies in the big data and data analytics industry.
- the goal of the multi-objective optimizer is to provide service providers with a Pareto-front solution. This solution trades off satisfaction with the amount of recourse required.
- the optimizer provides service providers with the flexibility to choose a personalized user satisfaction target for each user and find the minimum amount of resources to achieve this target.
- the teachings described herein need not be limited to wireless networks or even to communication networks.
- the framework and functionality of extracting and/or predicting context and user satisfaction in an unobtrusive way, in real time, can be applied in a broad range of applications. At the very least, these applications would include connected and autonomous vehicles, battlefield (tactical) scenarios, gaming and wearables (AV/RV, haptics, etc.)
- AV/RV battlefield (tactical) scenarios
- AV/RV haptics, etc.
- teachings set out herein can be applied to any context in which application delivery can be fined-tuned in real-time based on a number of factors including the end user's likes/dislikes, ambience, circumstances, mood, how much she/he is willing to pay, etc.
- the method steps of the invention may be embodied in sets of executable machine code stored in a variety of formats such as object code or source code.
- Such code may be described generally as programming code, software, or a computer program for simplification.
- the executable machine code or portions of the code may be integrated with the code of other programs, implemented as subroutines, plug-ins, add- ons, software agents, by external program calls, in firmware or by other techniques as known in the art.
- the embodiments of the invention may be executed by a computer processor or similar device programmed in the manner of method steps, or may be executed by an electronic system which is provided with means for executing these steps.
- an electronic memory medium such as computer diskettes, hard drives, thumb drives, CD-Roms, Random Access Memory (RAM), Read Only Memory (ROM) or similar computer software storage media known in the art, may be programmed to execute such method steps.
- electronic signals representing these method steps may also be transmitted via a communication network.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Computer Networks & Wireless Communication (AREA)
- Artificial Intelligence (AREA)
- Signal Processing (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US201862724195P | 2018-08-29 | 2018-08-29 | |
PCT/CA2019/051197 WO2020041883A1 (en) | 2018-08-29 | 2019-08-29 | Enabling wireless network personalization using zone of tolerance modeling and predictive analytics |
Publications (2)
Publication Number | Publication Date |
---|---|
EP3845023A1 true EP3845023A1 (en) | 2021-07-07 |
EP3845023A4 EP3845023A4 (en) | 2022-06-15 |
Family
ID=69644788
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
EP19856044.3A Pending EP3845023A4 (en) | 2018-08-29 | 2019-08-29 | Enabling wireless network personalization using zone of tolerance modeling and predictive analytics |
Country Status (3)
Country | Link |
---|---|
EP (1) | EP3845023A4 (en) |
CA (1) | CA3111030A1 (en) |
WO (1) | WO2020041883A1 (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11736973B2 (en) | 2018-08-29 | 2023-08-22 | Carleton University | Enabling wireless network personalization using zone of tolerance modeling and predictive analytics |
CN111553482B (en) * | 2020-04-09 | 2023-08-08 | 哈尔滨工业大学 | Machine learning model super-parameter tuning method |
CN112446097B (en) * | 2020-12-03 | 2023-08-29 | 武汉第二船舶设计研究所(中国船舶重工集团公司第七一九研究所) | Multi-objective optimization method for volume and load of steam generator |
CN114004358B (en) * | 2021-12-29 | 2022-06-14 | 粤港澳大湾区数字经济研究院(福田) | Deep learning model training method |
WO2024035399A1 (en) * | 2022-08-10 | 2024-02-15 | Rakuten Mobile, Inc. | A system, method, device, and program for interface management of ai/ml applications over radio access parameters |
CN117236788B (en) * | 2023-10-30 | 2024-03-26 | 水利部珠江水利委员会技术咨询(广州)有限公司 | Water resource scheduling optimization method and system based on artificial intelligence |
Family Cites Families (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6775233B1 (en) | 1999-08-31 | 2004-08-10 | Lucent Technologies Inc. | Rate processor sharing method and apparatus for scheduling data transmissions in a CDMA wireless communication system |
US6785227B1 (en) | 2000-05-11 | 2004-08-31 | Lucent Technologies Inc. | Method and apparatus for multi-layer resource management in wireless communications systems |
US7512545B2 (en) | 2004-01-29 | 2009-03-31 | At&T Intellectual Property I, L.P. | Method, software and system for developing interactive call center agent personas |
US8751305B2 (en) | 2010-05-24 | 2014-06-10 | 140 Proof, Inc. | Targeting users based on persona data |
JP5320358B2 (en) * | 2010-08-24 | 2013-10-23 | 株式会社エヌ・ティ・ティ・ドコモ | Wireless communication system, base station, and wireless resource allocation method |
CN102662764B (en) * | 2012-04-25 | 2015-07-29 | 梁宏斌 | A kind of dynamic cloud computational resource optimizing distribution method based on SMDP |
CN102932308B (en) * | 2012-11-02 | 2016-07-27 | 重庆邮电大学 | Mixed service Scene dispatch method and dispatching patcher in a kind of OFDM relay system |
WO2014089765A1 (en) | 2012-12-12 | 2014-06-19 | 华为技术有限公司 | Method, device and system for service scheduling and service transfer rate controlling |
US20140229210A1 (en) * | 2013-02-14 | 2014-08-14 | Futurewei Technologies, Inc. | System and Method for Network Resource Allocation Considering User Experience, Satisfaction and Operator Interest |
-
2019
- 2019-08-29 WO PCT/CA2019/051197 patent/WO2020041883A1/en unknown
- 2019-08-29 CA CA3111030A patent/CA3111030A1/en active Pending
- 2019-08-29 EP EP19856044.3A patent/EP3845023A4/en active Pending
Also Published As
Publication number | Publication date |
---|---|
WO2020041883A1 (en) | 2020-03-05 |
CA3111030A1 (en) | 2020-03-05 |
EP3845023A4 (en) | 2022-06-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11736973B2 (en) | Enabling wireless network personalization using zone of tolerance modeling and predictive analytics | |
Gao et al. | The cloud-edge-based dynamic reconfiguration to service workflow for mobile ecommerce environments: a QoS prediction perspective | |
Shakarami et al. | A survey on the computation offloading approaches in mobile edge computing: A machine learning-based perspective | |
Iftikhar et al. | AI-based fog and edge computing: A systematic review, taxonomy and future directions | |
EP3845023A1 (en) | Enabling wireless network personalization using zone of tolerance modeling and predictive analytics | |
Maia et al. | An improved multi-objective genetic algorithm with heuristic initialization for service placement and load distribution in edge computing | |
Carvalho et al. | Computation offloading in Edge Computing environments using Artificial Intelligence techniques | |
Chen et al. | Exploiting Web service geographical neighborhood for collaborative QoS prediction | |
Alkurd et al. | Big-data-driven and AI-based framework to enable personalization in wireless networks | |
Abosaif et al. | Quality of service-aware service selection algorithms for the internet of things environment: A review paper | |
Violos et al. | Predicting visitor distribution for large events in smart cities | |
Wang et al. | A task recommendation framework for heterogeneous mobile crowdsensing | |
Jalodia et al. | A Deep Neural Network-Based Multi-Label Classifier for SLA Violation Prediction in a Latency Sensitive NFV Application | |
Jin et al. | Mobility-aware and privacy-protecting qos optimization in mobile edge networks | |
da Silva et al. | Online machine learning for auto-scaling in the edge computing | |
Mayhoub et al. | A Review of Client Selection Methods in Federated Learning | |
CA3126091A1 (en) | Enabling wireless network personalization using zone of tolerance modeling and ai-enabled optimization | |
Sountharrajan et al. | On-the-go network establishment of iot devices to meet the need of processing big data using machine learning algorithms | |
Ismaili-Alaoui et al. | Towards smart incident management under human resource constraints for an iot-bpm hybrid architecture | |
Ometov et al. | On applicability of imagery-based CNN to computational offloading location selection | |
Alkurd et al. | User persona in personalized wireless networks: A big data-driven prediction framework | |
Robertazzi et al. | Machine learning in networking | |
Vaidhyanathan | Data-driven self-adaptive architecting using machine learning | |
Priyanga et al. | Mobile app usage pattern prediction using hierarchical flexi-ensemble clustering (HFEC) for mobile service rating | |
Ouyang et al. | Mining Over Air: Wireless Communication Networks Analytics |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
17P | Request for examination filed |
Effective date: 20210329 |
|
AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
DAV | Request for validation of the european patent (deleted) | ||
DAX | Request for extension of the european patent (deleted) | ||
A4 | Supplementary search report drawn up and despatched |
Effective date: 20220513 |
|
RIC1 | Information provided on ipc code assigned before grant |
Ipc: H04W 4/08 20090101ALN20220509BHEP Ipc: H04W 4/02 20180101ALN20220509BHEP Ipc: G06N 3/04 20060101ALI20220509BHEP Ipc: G06N 3/08 20060101ALI20220509BHEP Ipc: H04W 28/24 20090101ALI20220509BHEP Ipc: H04W 24/02 20090101ALI20220509BHEP Ipc: H04W 4/00 20180101ALI20220509BHEP Ipc: G06N 20/10 20190101ALI20220509BHEP Ipc: H04W 4/38 20180101ALI20220509BHEP Ipc: G06N 20/00 20190101ALI20220509BHEP Ipc: H04W 72/08 20090101AFI20220509BHEP |
|
REG | Reference to a national code |
Ref country code: DE Ref legal event code: R079 Free format text: PREVIOUS MAIN CLASS: H04W0072080000 Ipc: G06N0020000000 |
|
GRAP | Despatch of communication of intention to grant a patent |
Free format text: ORIGINAL CODE: EPIDOSNIGR1 |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: GRANT OF PATENT IS INTENDED |
|
RIC1 | Information provided on ipc code assigned before grant |
Ipc: H04W 4/08 20090101ALN20240305BHEP Ipc: H04W 4/02 20180101ALN20240305BHEP Ipc: H04W 72/54 20230101ALI20240305BHEP Ipc: G06N 3/045 20230101ALI20240305BHEP Ipc: G06N 3/084 20230101ALI20240305BHEP Ipc: G06N 3/082 20230101ALI20240305BHEP Ipc: G06N 3/04 20060101ALI20240305BHEP Ipc: G06N 3/08 20060101ALI20240305BHEP Ipc: H04W 28/24 20090101ALI20240305BHEP Ipc: H04W 24/02 20090101ALI20240305BHEP Ipc: H04W 4/00 20180101ALI20240305BHEP Ipc: G06N 20/10 20190101ALI20240305BHEP Ipc: H04W 4/38 20180101ALI20240305BHEP Ipc: G06N 20/00 20190101AFI20240305BHEP |
|
INTG | Intention to grant announced |
Effective date: 20240319 |