UbiComp'13 - 1

I could not write for long. I developed an agent system and run a field study. I finished writing my 9 month report and passed my first viva. We also submitted a conference paper. In the next couple of posts, I will write about papers which were published in UbiComp'13.

The first paper that I am going to mention is "The Collective Infrastructural Work of Electricity: Exploring Feedback in a Prepaid University Dorm in China" by Liu et al.

The paper presents findings of a field study. The study focuses on how students living in university dorms with a prepaid electricity system adopt a digital feedback system. In the study participants can use a web interface or a mobile application. The web interface enables users to compare their consumption with different dormitory apartments. Also it displays the amount of energy consumed and remained. The mobile application shows recent consumption trend and remaining energy balance. It also offers automatic low-balance reminder where users can set the threshold for the reminder.

The method used in the field study is explained briefly in the following.

- Participants are recruited through the University's bulletin system.

- No incentive was provided. Participation was voluntary basis.

- The evaluation of the system performed based on the log data and semi-structured interviews.

- Two rounds of interviews were conducted. The first round was conducted with 10 students two weeks into the deployment. The interview questions included when, where and why the participants used the system. The second round was conducted with 6 students after 7 months to understand long-term user behavior.

The high level structure of the paper:
-> Abstract
-> Introduction
    - Eco-feedback
    - Some related challenges and findings
    - The study, its aim and its contribution
-> Related Work
    - What have been done
    - What have not been done
    - Exceptional studies
    - How what you do is different
-> Background and Method
-> Findings
-> Collective Infrastructural Work
-> Discussion
-> Conclusion

Persuasive Social Actors

In this post, I will try to summarize the fifth chapter of the book "Persuasive Technology: Using Computers to Change What We Think and Do" by Fogg. In the chapter how computers can provide a range of social cues that trigger social reactions in their human users is explored. Computers are considered as social  actors that can persuade their users through rewarding the users with positive feedback, modeling a target behavior or attitude and providing social support.

Persuasion dynamics (social influence) are expressed as normative influence (peer pressure), social comparison, group polarization and social facilitation.

Normative social influence: is the influence of peer people that lead us to shape our attitudes, values or behaviors to conform them.

Social comparison (keep up with the Joneses): is based on the idea that people perform self-evaluations by comparing themselves to others in order to reduce uncertainties in their domain.

Group polarization: is the case when groups make a decision affected by its members' initial attitude, and the decision is more severe than initial tendency.

Social facilitation: is the inclination of people to perform better in simple tasks or the tasks they are good at doing when they are being watched or observed by others. However, if the tasks are not simple then performance might be affected adversely because of nervousness.

The writer introduces five social cues that make people to see computers as though they were living beings. These cues are physical, psychological, language, social dynamics and social roles, which are briefly explained with their influences as follows.

Physical Cues (e.g., face,eyes,body and movement)

It is possible to produce computing devices that can have physical attributes such as eyes, a mouth and movement. These attributes can be facilitated to convey social presence and persuade their human users. Also, physical attractiveness has a considerable impact in social influence. Hence, physically attractive device or interface might be more persuasive than unattractive ones. The attractiveness might create the halo effect, that leads users to believe that the product is also intelligent, capable, reliable and credible. However, the criteria for attractiveness changes from culture to culture,generation to generation and individual to individual.
- The Personality Study showed that people prefer computers whose 'personalities' match with them.
- The Affiliation Study showed that people who worked with a computer labeled as their teammate were more likely to conform the suggestions delivered by the computers.
Language
Computing technology can also convey social presence through simple interface elements like dialog boxes without physical attributes. Computing devices can utilize written or spoken language to convey social presence. Asking questions to perform a task, offering congratulations for completing a task, or reminding available software updates can lead people to infer that  the computing product is alive in some sense. One of the most effective persuasive uses of language is to offer praise. Offering praise by text, image or sound can lead users to be more open to persuasion.
Social Dynamics (unwritten rules for interacting with others)
In many cultures, there is a shared set of  patterns among people to interact with each other. These patterns construct social rules for interactions such as meeting people, taking turns and forming lines. Those who don't conform the rules pay a social price (e.g. being alienated).
- The Reciprocity Study, which is based on the principle that people will feel the need to reciprocate when computing technology has done a favor for them, showed that people worked with the same helpful computer on two tasks performed almost twice as much work for their computers on the second task as did the other participants.
Social Roles
Authority roles such as teacher, doctor, counselor and expert are played by people.  It is common to believe that authorities are intelligent and powerful. Computing products can act in these roles so that being in a position of authority will lead to self-inflicted influence. Computing technology that assumes roles of authority will have enhanced powers of persuasion. Designers should provide different social roles for their target audience, such roles might be a octor, servant or commander.

* It should be kept in mind that unnecessary social cues might cause distracted and annoyed users.

AgentSwitch

This post will examine the AgentSwitch system which provides personalized recommendations on how much users can save by changing their energy tariff and shifting their deferrable loads to off-peak times. The system is based on uSwitch API, an energy data store, and a set of algorithms for consumption prediction and appliance disaggregation. More information about the work can be found in the following papers.

- "AgentSwitch: Towards Smart Energy Tariff Selection" by Ramchurn et al. 

- "Recommending Energy Tariffs and Load Shifting Based on Smart Household Usage Profiling" by Fisher et al.

Here, I will point out some of the findings derived from the evaluation of the system. Please read the second paper stated above for details about the evaluation process. The some of the findings are:

- a certain level of percieved accuracy of the predictions is important to further engage with and trust information provided by the system. What I understood from this is that it is required for a recommendation system to provide some evidences for its predictions in order to maintain the trust that the system is well-functioning.

- the threshold that potential savings have to exceed to motivate the users could be lowered by some degree of automation. Notifications, limiting factors and controllability are significant elements for the automation. Semi-autonomous systems can reduce the threshold for potential savings as they will ease the hassle that actual users need to deal with.

- current monetary benefits does not sufficiently motivate people to change their behavior, monetary benefits can be supported by game-like rewards or effective environmental benefits for better motivation. More persuasion dynamics need to be applied into this domain so as to increase the motivation and evaluate their impacts. I will talk about persuasion dynamics in my next post.

Some interesting questions for intelligent interface community:

- how to present uncertainty in intelligent UIs?

- how the balance between user control and autonomy can be achieved in a flexible way without overwheling the users with requests and undesired system actions.

The application of AgentSwitch is available here.

A Study of Combined Interventions

This post will briefly indicate the main points made in the paper "The effect of tailored information, goal setting and tailored feedback on household energy use, energy-related behaviors, and behavioral antecedents" by Abrahamse et al.

The study examines the effects of the combination of three interventions. The three intervention types used in the study are information, goal setting and feedback. The households are provided information about energy-related problems and individualised energy conservation tips. In terms of goal setting, the experimental groups are asked to save 5% both as an individual and as a group. As for feedback, the experimental groups   received feedback about energy savings. The hypotheses being tested in the study are as follows.

The households exposed to the combination of interventions

H1A - would save more direct (gas, electricity and fuel) and indirect energy (i.e. disposal of goods),

H1B - would adopt more energy-saving behaviours,

H1C - would acquire higher level of knowledge than the control group who had not received any information and feedback. Further, the households who received group goal and group feedback are expected to

H2A - save more direct and inderect energy,

H2B - adopt more energy-related behaviours than the households who only received personal goal and feedback.

The paper proposed the results that the households exposed to the interventions saved significantly more direct energy than the control group. As for indirect energy, the difference is not significant. However, the experimental groups reduced their indirect energy use whilst the control group increased indirect energy use. Further, the experimental groups showed that they obtained more knowledge and formed more energy-saving behaviours than the control group. Moreover, the group goal and group feedback do not have an additional positive effect according to the study. 

Intervention Strategies

In this post I will summarise the review paper "A review of intervention studies aimed at household energy conversation" by Abrahamse et al.

Interventions are divided into two categories: antecedent interventions and consequence interventions. The latter intervention type is based on the presumption that the availability of positive or negative outcomes will influence behaviour. The former intervention strategies aim to influence the underlying determinants of behaviour, which in turn are assumed to affect behaviour.

Antecedent interventions: Commitment, Goal Setting, Information and Modeling.

- Commitment: An oral or written pledge or promise to change behaviour (e.g. conserve energy).

- Goal Setting: A goal can be set by experimenters or by the households (e.g. reduce energy use by %5).

- Information: General or specific information about energy-related problems or solutions are used to encourage energy conservation behaviours (e.g. workshops, mass media campaigns and home audits).

- Modeling: This method is based on providing examples of recommended behaviours which are understandable, relevant, meaningful and rewarding to the households.

Consequence interventions: Feedback and Reward.

- Feedback: It consists of providing households with the information about energy consumption or energy savings (e.g. daily feedback, comparative feedback).

- Reward: Monetary rewards are used as motivators for energy conservation.

The effectiveness of the interventions is increased when antecedent and consequence interventions are combined and used together (e.g. goal setting with feedback).

Event Detection

This post summarises the paper "Event Detection for Non Intrusive Load Monitoring" by Anderson et al.

The main aim of the paper is to introduce a set of metrics for evaluating the event detection algorithms in NILM and to provide a dataset which enables others to test their event detection algorithms. The paper also briefly discusses the methods used for event detection in the NILM (for more information about NIALM please see Oliver's blog). According to the paper, there are three main categories which aim to detect explicit change-points in a time-series signal. The categories are listed as follows.

- Expert Heuristics: The detection of step-like changes present in the power consumption signal with the use of the standard deviation of the fixed-size set of examples. 

- Probabilistic Models: As an example, the Generalized Likelihood Ratio(GLR) is introduced. It requires the parameters including the window length , the power change and a threshold so as to calculate decision statistic.

- Matched-Filters: The matching of a known signal (mask) with an unknown one in order to detect the mask in the unknown signal.

The event detection metrics introduced for event-based NILM approach are True Positive Rate, True Positive Percentage, Total Power Change, Average Power Change and  Score Function. In order to show the validity of the metrics, the paper uses a modified generalized likelihood ratio(GLR). The GLR is tested with a  fully-labeled dataset for BLUED dataset which is available here.

The paper concludes that the total power change metric outperforms the other three metrics which are True Positive Rate,  True Positive Percentage and Average Power Change.

Agent vs Direct Manipulation

In this post, I will write some points made in the article "Designing for Human-Agent Interaction" by Lewis.

There are two main interface metaphors: agent and environment. While agent metaphor uses the computer as a mediator which responds to user requests,  environment metaphor enables the user to interact with the task domain directly. Agents are designed to automate repetitive, poorly specified or complex precesses by bridging the gap between a user's goals and actions.

The figure above taken from the article shows the possible automations that can occur in the execution or evaluation phases defined by Norman. The issue of the automation is that if the execution side is automated the user may fail to monitor effects of actions and cannot regulate the incorrect ongoing processes. On the other hand, if the evaluation side is automated the user might be unable to observe the results of actions. 

* I will keep updating this post.

Occupancy Prediction

In this post, I will try to guess the hours in which property is occupied by tenants. The prediction will be performed through the use of different arbitrary thresholds in electricity consumption data. The main aim is to find time intervals, which tenants are at home, with high accuracy. For privacy issues, the owner of the data is not represented.

11 Jan 2013, Friday.

Power-consumption hours: 00.00 - 01.00, 05.00 - 08.00, 17.00 - 24.00

12 Jan 2013, Saturday.

Power-consumption hours: 00.00 - 01.00 , 05.00 - 08.00, 11.00 - 14.00, 20.00 - 21.00, 23.00 - 24.00

13 Jan 2013, Sunday.

Power-consumption hours: 00.00 - 02.00 , 08.00 - 13.00, 20.00 - 22.00

14 Jan 2013, Monday.

Power-consumption hours: 06.00 - 22.00

15 Jan 2013, Tuesday.

Power-consumption hours: 00.00 - 01.00 , 06.00 - 08.00, 18.00 - 24.00

16 Jan 2013, Wednesday.

Power-consumption hours: 00.00 - 01.00 , 06.00 - 08.00, 17.00 - 24.00

17 Jan 2013, Thursday.

Power-consumption hours: 00.00 - 01.00 , 05.00 - 08.00, 17.00 - 24.00

Please note that the hours stated above represent the time intervals in which tenants consume observable energy. Hence, we might assume that at least one occupant is at home during these time intervals. 

MatPlotLib

MatPlotLib is a plotting library which is created by John Hunter. It offers an useful object-oriented API for Python programmers. For further information please see the link : http://matplotlib.org/

A sample graph which shows the historical electricity use in a day. 

Team Energy

I completed the implementation of TeamEnergy which is a web application allowing multiple users to coordinate their energy consumption activities. The project is uploaded to the bitbucket which provides unlimited private repositories for online users. The steps followed for the process of uploading can be found in the link below.

https://confluence.atlassian.com/display/BITBUCKET/Set+up+Git+and+Mercurial

The articles I read are as follows:

- A Large-Scale Study on Predicting and Contextualizing Building Energy Usage 

In this paper, a data-driven approach is represented to model energy consumption in buildings. The monthly electricity and gass bills are used as data which is collected by a utility for several years. Both parametric and non-parametric learning methods are utilised to model energy consumptions related to the features of the buildings. Based on these models, two end-user system is offered. One is for utilities or authorised institutions, and it visualises energy consumption for each unit. The other system enables consumers to enter their own data like the features of the building and energy usage, and the system, in return, provides a comparision of similar buildings. 

* In order to compare energy consumptions of buildings, electricity and gas usages are combined to obtain equivalent units through the use of conversation factor (1/3) 29.3 therms/kwh of the US. We can use this conversion in our home energy management system.

- The Smart Thermostat: Using Occupancy Sensors to Save Energy in Homes

Basically, this paper presents a smart thermostat which adjusts set points according to occupancy patterns that are generated through several sensors.  

Finally, I finished reading the book "Plans and Situated Actions" by Suchman.