Discusses recent results in the decision-making on machine learning adjusted for long-term social welfare

Author: Lydia T. Liu, Sarah Dean, Esther Rolf, Max Simchowitz, Moritz Hardt

Participation: Liu Tianci, Xiao Kun

Because the machine learning system is vulnerable to discriminatory behavior caused by the introduction of historical data, it is considered necessary to use fairness rules to restrict the behavior of the system in certain application scenarios and expect it to protect vulnerable groups and bring long-term benefits. Recently, Berkeley AI Institute published a blog to discuss the long-term impact of the static fairness criteria and found that the results are far from people's expectations. Related papers have been accepted by the ICML 2018 Conference.

Machine learning systems that are trained to "minimize prediction errors" often show discriminatory behavior based on sensiTIve characterisTIcs, and historical deviations in the data may be one of the reasons. . For a long time, in many application scenarios such as loans, employment, criminal justice, and advertising, machine learning has been criticized as "due to historical reasons, it has potentially hurt the neglected and vulnerable groups."

This article discusses the recent results of the researcher's decision to adjust the benefits of machine learning with the goal of long term social welfare. In general, machine learning models generate a score that summarizes information about individuals and then makes decisions about them. For example, a credit score summarizes a person's credit history and financial behavior to help the bank assess its credit rating. We use this loan scenario as an example throughout the text. Any user group has its specific distribution in the credit score, as shown in the following figure.

Credit score and repayment distribution

By defining a threshold, scores can be turned into decisions. For example, a person who scores above the lending threshold can get a loan, while a person who is below the lending threshold is rejected. This decision rule is called a threshold policy. The score can be understood as the estimated probability code of the loan default. For example, 90% of people with a credit score of 650 will repay their loans. As a result, the bank can estimate the expected return for a loan with a credit score of 650 for a user and, similarly, it can predict the expected return for a loan to all users with a credit score above 650 (or any given threshold).

2. Loan thresholds and results

Without considering other factors, the bank will try to maximize its total revenue. The benefit depends on the ratio between the amount of loan repayment that is recovered and the amount of loss that the loan defaulted. In the chart above, the loss-to-loss ratio is 1:4. Because the cost of losses is higher than the return, the bank will be more conservative in lending and raise the lending threshold. We refer to the total number of people above this threshold as the selecTIon rate.

Result curve

Loan decisions affect not only banking institutions but also individuals. In a breach of contract (lender cannot repay the loan), not only the bank lost the proceeds, but also the lender's credit score will be reduced. In a successful loan performance, the bank gains revenue while the credit score of the lender increases. In this example, the change ratio of a user's credit score is 1 (performance): -2 (default)

In the threshold strategy, the outcome is defined as the change expectation of a group's score, which can be parameterized as a function of the selection rate. This function is called the outcome curve. When the selection rate of a group changes, the result will also change. The results at these overall headcount levels will also depend on the probability of repayment (coded by the score), cost, and the benefit of individual loan decisions.

The figure above shows the results of a typical population curve. When there are enough individuals within a group to obtain a loan and successfully repay, their average credit score may increase. At this time, if the average score change is positive, the result of unconstrained gain maximization can be obtained. Deviation from income maximization, in order to provide loans to more people, the average score change will increase to the maximum. Call it the altruis TIc optimum. It is also possible to increase the selection rate to a certain value so that the average score change is lower than the average score change when the unconstrained gain is maximized, but it is still positive, that is, the area indicated by the yellow dot shade in the figure. The selection rate in this area is called relative harm. However, if there are too many users who are unable to repay the loan, the average score will decrease (the average score will change to negative) and enter the shaded area of ​​the red horizontal line.

4. Loan threshold and result curve

Multi-group situation

How does a given threshold strategy affect individuals in different groups? Two groups with different credit score distributions will have different results.

Assuming that the second group and the first group have different distributions of credit scores and fewer people in the group, they are understood to be historically disadvantaged groups. To represent it as a blue group, we hope to ensure that the bank’s loan policy will not unreasonably harm or deceive them.

Assume that banks can choose different thresholds for each group. Although this may face legal challenges, group-based thresholds cannot be avoided in order to prevent the possible differential results due to fixed threshold decisions.

5. Different groups of loan decisions

Naturally, there is a problem: what kind of threshold selection can be expected to improve in the distribution of the blue population score. As mentioned above, an unconstrained banking strategy maximizes returns and selects points where the balance of payments and loans are profitable. In fact, the maximum revenue threshold (credit score of 580) is the same in both groups.

Fairness criteria

Groups with different score distributions have different shapes of score curves (the upper half of the original text shows the real credit score data and the result curve of a simple result model). Another alternative to maximizing unconstrained returns is fairness constraints: Making certain groups of people equal in decision-making through certain objective functions. At present, various fairness standards have been put forward, appealing to intuition to protect vulnerable groups. Through the results model, we can formally answer: whether the fairness constraint really encourages more positive results.

A common fairness criterion, demographic parity, requires banks to give the same proportion of loans in both groups. With this requirement, banks continue to maximize their revenue as much as possible. Another criterion, equality of opportunity: The true positive rate in both groups is equal, requiring the bank to have the same proportion of loans to individuals in both groups who will repay the loan.

Although these criteria are reasonable from the perspective of requiring fair decision-making fairness, they mostly ignore these future effects on group results. Figure 6 shows this by comparing strategic outcomes that maximize returns, demographic equality, and equal opportunity. Look at the changes in bank returns and credit scores for each loan strategy. Compared with maximizing the revenue strategy, demographic equality and equal opportunity have reduced the bank's revenue, but have it achieved a blue group result compared to the maximizing gains? Although compared to Altruistic Optimization, the Maximized Yield Strategy has too low a loan to blue groups, but the equal opportunity strategy (as compared to Altruistic Optimization) has too many loans, demographic equality is excessive loans, and it reaches relative damage areas. .

6. Simulation of loan decisions under constraints

If the objective of the fairness rule is to "promote or equitably promote the well-being of all groups in the long run," what has just been shown indicates that under certain circumstances, the fairness rule actually violates this objective. In other words, the fairness constraint will further reduce the existing welfare in the vulnerable groups. Establishing an accurate model to predict the effects of the strategy on the group's results may ease the unexpected damage caused by introducing fairness constraints.

Reflections on the "Fair" Machine Learning Results

The researchers proposed a perspective on the "fairness" of machine learning based on long-term results. Without a detailed model of delayed results, it is impossible to predict the impact of the fairness criterion as it is added to the classification system. However, if there is an accurate results model, the positive results can be optimized in a more direct way than the existing fairness criteria. Specifically, the resulting curve gives a deviation from the maximizing revenue strategy to improve the results most directly.

The resulting model is a concrete method of introducing domain knowledge in the classification process, and it can be in good agreement with many studies that point out that “fairness” in machine learning has background-sensitive characteristics. The resulting curve provides an interpretable visual tool for this application-specific trade-off process.

For more details, please read the original paper. This article will appear at the 35th ICML Conference this year. This study is only a preliminary exploration of the "result model can ease the impact of machine learning algorithms on social unintended consequences." Researchers believe that in the future, as machine learning algorithms will affect more people's lives, there will be more research work to ensure the long-term fairness of these algorithms.

Paper: Delayed Impact of Fair Machine Learning

Paper address: https://arxiv.org/pdf/1803.04383.pdf

Photocell Timer Switch Socket

24HR Electronic timer socket with photocell.

â‘  Light intensity setting

â‘¡ Light intensity detection

â‘¢ Countdown Timer ON & OFF

â‘£ 4 MODES:

Set > Detected: When the light intensity detection value is less than the set value, switch ON or OFF.

Detected > Set: When the light intensity detection value is greater than the set value, switch ON or OFF

ON : Always ON

OFF : Always OFF

NOTED:

1. The light intensity displayed by this machine is not the standard light intensity value (Lux), only the relative light intensity value.

2. The light intensity value is affected by the placement position and direction. Please determine the position first and then set it according to the actual light intensity detected. If you change the position or change the orientation, you need to reset the light intensity setting value suitable for the new position.

3. This product has built-in rechargeable battery. If it is not connected to AC for a long time, you need to connect the power supply to charge until the LCD can display normally.

MANUAL OPERATION

1. Press [UP" or [DOWN" to set the LUX value.

2. Press the [SET" key to start setting, and the P1 settable items will be flashed.

3. Press [UP" or [DOWN" to adjust the value.

4. Press [SET" key again to exit setting or enter next setting for countdown timer.

5. Repeat the [SET" key to start setting, and the P2 & P3 settable items will be flashed.

6. Press the [FUN" key to switch the working state in the following:

Set > Detected -> Detected > Set -> ON -> OFF

Set > Detected: Automatically switches when the detected ambient light intensity is darker than the set value

Detected >Set: Automatically switch when the detected ambient light intensity is brighter than the set value

When the brightness meets the setting conditions, the countdown starts as below:

Note:when the countdown is ON, the detected value is not displayed.

When the brightness does not meet the setting conditions, the countdown stops and waits:

After the countdown ON is reduced to 0, the countdown OFF starts immediately and the power is OFF.

After the countdown OFF is reduced to 0:

A. If the light intensity meets the set conditions, a new round of countdown will be started;

B. If the light intensity does not meet the set conditions, keep the power off and wait for the light to meet the conditions before turning on automatically.

NOTE:

1. If the power is cut off while the countdown is running, the countdown will be terminated immediately and the relay output will be off. After the power is turned on again, a new round of brightness detection will start.

2. Modifying the brightness value in the countdown operation will not affect the current countdown operation. After the off time of the current countdown, the new brightness setting value will take effect.

3. In the countdown on operation, change the setting value of the countdown on, this countdown will still be timed according to the original setting value; the new setting value will take effect when the next countdown on starts.

4. In the countdown off operation, change the setting value of countdown off, this countdown will still be timed according to the original setting value; the new setting value will take effect when the next countdown off is started.

NOTE: the brightness setting value, countdown ON or countdown OFF, any one of which is equal to 0, cannot be switched ON or OFF automatically.

Manual Control

When ON or OFF is displayed, it means that the power supply remains ON or OFF, as shown in the figure below:

Power Detection and Standby Mode

With AC power supply, the icon lights up and works normally.

When there is no AC power supply, the icon goes out, the brightness is not detected at this time, and the system enters the standby mode.

Photocell Timer, Photocell Timer Socket, Photocell Sensor, Photocell Sensor Socket, Sensor Plug, Sensor Switch Socket, Digital Photocell Timer, Digital Sensor Timer

NINGBO COWELL ELECTRONICS & TECHNOLOGY CO., LTD , https://www.cowellsockets.com