How to Cope with Bias in Wellbeing AI? – Towards Fairness in Wellbeing AI by Personal and Long-term Evaluation Keiki Takadama The University of Electro-Communications keiki@inf.uec.ac.jp Abstract bias, (2) data bias, (3) sampling bias, (4) algorithmic bias, This paper focuses on the fairness in ML (machine learning) (5) interaction bias, (6) self-selection bias, and (7) second- meaning that output of ML should not be “biased” and aims order bias. The essential difference among them is summa- at clarifying bias in wellbeing AI. From the analysis of bias rized as follows. from the viewpoint of healthcare, the bias in wellbeing AI can be reduced by employing the personal and long-term evalua- (1) Activity bias tion while many biases in ML arise. To investigate an effec- This bias arises from the different number of active/silent tiveness of the personal and long-term evaluation, our previ- users. For example, only top 4% of Amazon users posted ous research conducted the human subject experiment by fo- cusing on sleep of aged persons in care house and found that the reviews, which means that we cannot receive mes- our wellbeing AI based on the personal and long-term evalu- sages from all users, i.e., they are the biased messages. ation succeed to extract knowledge for good sleep and to es- (2) Data bias timate mind change of an aged person from her sleep quality This bias arises from the different number of data. For change. example, the number of Westerner face pictures tends to be larger than that of Asian in the face pictures in dataset Introduction such as MS (Microsoft) celebrity dataset. (3) Sampling bias “Can ML (machine learning) provide a fair decision?”. To This bias arises from the fact where the sampled data is answer this question, this paper starts to explain one exam- not always followed by true distribution. For example, an ple. As you may know, Amazon developed the ML person- asthma patient rate in near highway tends to be higher nel recruitment system but stopped it in 2018 because than the rate in whole area. This means that the data in women do not tend to be recruited in comparison with men big city is different from the data in whole area due to the reason why most of input data for ML is men’s (4) Algorithmic bias data (Dastin, 2018). This example suggests an importance This bias arises from the different outcome caused by dif- of fairness in ML. In other words, the output by ML should ferent algorithms. For example, a search ranking by be fair or should not be “biased”. What should be noted here Google is different from the ranking by Bing. This means is that many healthcare systems based on ML (hereafter we that the behaviors of users are biased by the different call it as “wellbeing AI”) have also a rick of providing the search engine. biased outputs and such outputs are very critical for our (5) Interaction bias daily life. From this fact, this paper aims at investigating This bias arises from the different interaction according what kinds of biases arises in wellbeing AI and how such to web presentation. When focusing on the medicine list biases can be reduced to cope with them. For this issue, this on the pharmacy web site, for example, they are differ- paper starts to explain bias in general by focusing on the bias ently displayed, e.g., one by one or all image. In the one on the Internet and the bias in ML, and then clarifies bias in by one representation, users are hard to watch the less Wellbeing AI. prioritized medicines because they need to scroll the web page to find them. In the all image representation, on the other hand, users tend to watch the upper left image of Bias on the Internet medicine because we usually read the sentence from left According to Baeza-Yeate, the bias on the web on the Inter- to right and its line starts from upper to lower. Such dif- net is categorized as follows (Baeza-Yates 2018): (1) activity ferent representations causes bias of selecting medicines. ___________________________________ In T. Kido, K. Takadama (Eds.), Proceedings of the AAAI 2022 Spring Symposium “How Fair is Fair? Achieving Wellbeing AI”, Stanford University, Palo Alto, California, USA, March 21–23, 2022. Copyright © 2022 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). 4 (6) Self-selection bias ⚫ Personal information This bias arises from the different number of users who Good information of others is not always good. For ex- are willing to participate or not. For example, many ques- ample, the knowledge of good sleep for a certain person tionnaires are returned from healthy persons, but not is not always useful for other persons. This indicates that from the non-healthy persons. This is because healthy that the personal data is very important in wellbeing. persons do not hesitate to tell their health information ⚫ Long-term evaluation without worry about it, while non-healthy persons do not Current evaluation of health is not enough because keep- want to tell their health information honestly due to their ing good health and better health (better life) are more worry about it. important than the current health. This indicates that the (7) Second-order bias long-term evaluation is very important in wellbeing. This bias arises from the original bias. After a biased in- From the viewpoint of the personal information and long- formation (in the high ranking) is spread, for example, term evaluation, the seven biases do not arise or can be re- active users post other messages related to this infor- duced as the following reasons. As shown in Figure 2, firstly, mation, and such messages are sampled in high possibil- the activity bias and self-selection bias do not arise because ity and increases its rank in search engine. This cycle am- the data comes from only one person. This indicates that the plifies the original bias. “single” user provides the “personal” data. Secondly, the data bias and sampling bias can be reduced if we can get Bias in Wellbeing AI long-term daily data. This is because such a kind of data is not heavily biased in comparison with the short-term daily To clarify the bias in Wellbeing AI for easy understanding, data due to the large number of data. Thirdly, an influence let start to simplify the bias from the viewpoint of ML. Ac- of the algorithmic bias and interaction is very small because cording to Mehrabi’s survey (Mehrabi et al. 2022), bias in only one person is affected. Finally, the second-order bias ML arises in the cycle of (i) users, (ii) data, and (iii) algo- can be reduced because other biases in this cycle are reduced rithm as shown in Fig. 1. The connection of the seven biases by the above reasons. From this analysis, the algorithm in on Web to the cycle from (i) to (iii) is summarized as follows. Wellbeing AI analyzes the “personal” data that comes from Firstly, the activity bias and self-selection bias arise in the the “single” user and provides the result to the user. This cycle from “user” to “data” because both biases are caused indicates that the “personal and long-term evaluation” (pre- by user and affect data. Secondly, the data bias and sampling cisely, the personal and long-term evaluation based on the bias arise in the cycle from “data” to “algorithm” because personal data) are important for fairness AI. both biases are found in data and affect algorithm. Thirdly, the algorithmic bias and interaction bias arise in the cycle from “algorithm” to “user” because both biases are caused by algorithms and affect user. Finally, the second-order bias also arises in this cycle, which is the same as the bias on the web. Figure 2: Bias in the cycle of Wellbeing AI Personal and long-term evaluation The goals of many examples of wellbeing AI are roughly classified into the following two categories. Figure 1: Bias in the cycle of ML ⚫ Keeping good health (not getting a disease) Since many patients such as dementia, diabetes, and sleep To consider the features of wellbeing in the cycle of aris- apnea syndrome (SAS) want to worsen their health, it is ing the bias of ML that connects with the bias on Web, the important to find something wrong for early detection. following features should be taken into consideration. 5 For this issue, the personal and long-term evaluation is suggested to change the time of having rehabilitation from needed for early detection. AM to PM. As a result, he could have a deep sleep. What ⚫ Better health (improving activities) should be noted here is that this knowledge is not always For better health, it is important to know (measure) the useful for other persons. For person B, when the aged person accumulated small progress and its change of activities of had rehabilitation in PM, he lost appetite due to tiredness of daily living (ADL) for aged persons, performance after rehabilitation and could not diner as usual. This caused him nap for office workers, and sleep quality for all ages. For not sleep very well because of hungry at night. For this prob- this issue, the personal and long-term evaluation is also lem, our Wellbeing AI suggested to change the time of hav- needed. ing rehabilitation from PM to AM. As a result, he could have Among them, this paper focuses on sleep of aged persons in a deep sleep. This results clearly show that the knowledge care house because sleep is significant for aged persons. For for good sleep is different among persons. example, aged person easily wakes up due to light sleep and Mind change estimation of aged person may wander in midnight. Sleep can also provide some mes- Figure 4 shows that the sleep quality of the aged diabetes sage of mind change of aged persons when their sleep qual- person before/after the great east Japan earthquake, where ity change from good to bad. This is because persons tend the blue dots indicate the deep sleep while the red dots indi- to have deep sleep without anxiety but change to light sleep cate the light sleep. The horizontal axis (f1) indicates the when they are worry about something. To tackle these issues, achievement degree of what an aged person wants to do, our previous research developed the wellbeing AI system while the vertical axis (f2) indicates the achievement degree for the first issue to extract knowledge for good sleep of what a care worker wants to do for an aged person. From (Takadama et al. 2015) and for the second issue to estimate this figure, the blue dots are located at the right side while mind change of aged person from sleep (Takadama 2013). the red dots are located at the left side before the earthquake. These researches took the approach of the personal and This is because an aged person tended to have a deep sleep long-term evaluation (in detail, we investigated the data of when she could achieve the activities (such as eating as the individual persons in one year). Technically, we devel- usual) because of not being worry about anything while she oped the sleep quality estimation system based on vital vi- tended to have a light sleep when she is hard to achieve the bration data from pressure sensor (Harada et al. 2016). activities (such as less eating as usual) because of being Knowledge extraction for good sleep worry about something. In our experiment, the daily activity and sleep quality are What should be noted here is that the blue and red dots recorded every day. In one day, many activities as such were mixed after the earthquake, which had a possibility of meals, rehabilitation, and bathing, are scored in the integer the message of something mind changes of aged person. For values. For example, when eating full amount of meal, the this issue, our Wellbeing AI estimated that amount of break- score is 3. When no rehabilitation, then the score is 0. In fast should change from full to medium and the time of hav- addition to the dairy activity, the sleep quality (i.e., the ratio ing rehabilitation should change from none to AM. After of deep sleep) is estimated by our sleep stage estimation to these changes, she could have a usual sleep. To verity these classify whether the deep or light sleep. suggestions, care workers asked to her and she said that she Figure 3 shows the knowledge for good sleep. For person was not willing to eat a full amount of breakfast due to news A, when the aged person had rehabilitation in AM, he be- of death of many people by tsunami caused by the earth- came to be tired and mostly took a nap. This caused him not quake. Regarding the rehabilitation, she did not like it and sleep very well at night. For this problem, our Wellbeing AI was often absent from rehabilitation. After the earthquake, she noticed that many people killed by tsunami could not extend their life while she could extend it by having rehabil- itation to tackle her diabetes. This changes her mind to be willing to exercise. Figure 3: Knowledge for good sleep Figure 4: Light and deep sleep before/after earthquake 6 Conclusions To explore the answer to the question of how we should cope with bias in Well-being AI, this paper started to focuses on bias on the Internet and bias in ML and analyzed the bias in wellbeing AI after connecting biases on the Internet and bias in the ML. From this analysis, the bias in wellbeing AI can be reduced by employing the personal and long-term evaluation while many biases in ML arise. This paper dis- cussed the fairness in Well-being AI from the viewpoint of the personal and long-term evaluation and found that our wellbeing AI based on the personal and long-term evalua- tion showed its potential by extracting knowledge for good sleep and estimating mind change of aged person from her sleep quality change. Future work includes an investigation of other domains. References Dastin, J. 2018. “Amazon scraps secret AI recruiting tool that showed bias against women”, Reuters, Oct. 11. Baeza-Yates, R. 2018. “Bias on the Web.” Communications of the ACM, Vol. 61, No. 6, pp. 54-61. Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., and Galstyan, A. 2022. “A Survey on Bias and Fairness in Machine Learning.” ACM Computing Surveys, Vol. 54, Issue 6, Article No. 115, pp. 1–35. Takadama, K. and Nakata, M. 2015. “Extracting Both Generalized and Specialized Knowledge by XCS using Attribute Tracking and Feedback,” 2015 IEEE Congress on Evolutionary Computation (CEC2015), pp. 3034-3041. Takadama, K. 2013. “Towards a Care Support System that Can Guess The Way Aged Persons Feel,” The AAAI 2013 Spring Sym- posia, Data Driven Wellness: From Self-Tracking to Behavior Change, AAAI (The Association for the Advancement of Artificial Intelligence), pp. 45-50. Harada, T., Uwano, F., Komine, T., Tajima, Y., Kawashima, T., Morishima, M., and Takadama, K. 2016. “Real-time Sleep Stage Estimation from Biological Data with Trigonometric Function Re- gression Model,” The AAAI 2016 Spring Symposia, Well-Being Computing: AI Meets Health and Happiness Science, AAAI (The Association for the Advancement of Artificial Intelligence), pp. 348-353. 7