Let’s start small and simple. With SHAP, we can generate explanations for a single prediction. The SHAP plot shows featu

Interpreting machine learning models can seem complicated at first, but libraries like SHAP make everything as easy as a function call. We even don’t have to worry about data visualization, as there are built-in functions for that.

We can visualize the importance of the features and their impact on the prediction by plotting summary charts. The one below sorts features by the sum of SHAP value magnitudes over all samples. It also uses SHAP values to show the distribution of the impacts each feature has.

http://stream88.colomboserboli.com/niy/video-Temperley-Deportivo-Riestra-v-en-gb-1kwa-.php

http://news24.acaps.cat/vac/Video-Temperley-Deportivo-Riestra-v-en-gb-1toa30122020-.php

https://assifonte.org/media/ept/videos-Temperley-Deportivo-Riestra-v-en-gb-1ybd30122020-.php

http://news24.acaps.cat/vac/v-ideos-Temperley-Deportivo-Riestra-v-en-gb-1crc-27.php

http://wap.dentisalut.com/mqk/video-botafogo-v-atletico-paranaense-v-pt-br-1wtr2-25.php

http://stream88.colomboserboli.com/niy/video-Temperley-Deportivo-Riestra-v-en-gb-1lfy-27.php

http://vert.actiup.com/gwk/videos-Temperley-Deportivo-Riestra-v-en-gb-1ump-5.php

http://live247.gruposio.es/mun/videos-botafogo-v-atletico-paranaense-v-pt-br-1rtu2-26.php

https://assifonte.org/media/ept/videos-Temperley-Deportivo-Riestra-v-en-gb-1xze-17.php

http://svt.munich.es/zzr/Video-botafogo-v-atletico-paranaense-v-pt-br-1qla2-9.php

http://wap.dentisalut.com/mqk/video-botafogo-v-atletico-paranaense-v-pt-br-1gjh2-15.php

http://svt.munich.es/zzr/Video-botafogo-v-atletico-paranaense-v-pt-br-1efp2-2.php

http://stream88.colomboserboli.com/niy/video-Temperley-Deportivo-Riestra-v-en-gb-1jzc-9.php

http://news24.acaps.cat/vac/Video-Temperley-Deportivo-Riestra-v-en-gb-1kit-16.php

http://live247.gruposio.es/mun/Video-botafogo-v-atletico-paranaense-v-pt-br-1whp2-17.php

https://assifonte.org/media/ept/videos-Temperley-Deportivo-Riestra-v-en-gb-1tuc30122020-16.php

http://live247.gruposio.es/mun/video-botafogo-v-atletico-paranaense-v-pt-br-1xas2-11.php

http://wap.dentisalut.com/mqk/Video-botafogo-v-atletico-paranaense-v-pt-br-1agi2-11.php

http://vert.actiup.com/gwk/Video-botafogo-v-atletico-paranaense-v-pt-br-1usr2-16.php

http://svt.munich.es/zzr/videos-botafogo-v-atletico-paranaense-v-pt-br-1kvo2-18.php

know, GPT3 is an autoregressive language model that can generate responses to different types of prompts. The release of this model made a stupendous amount of noise (which technically, was just a bigger neural network than before).

This is a classification dataset, so don’t worry too much about the f(x) value. Only the residual sugar attribute pushed this instance towards a good wine quality, but it wasn’t enough, as we can see.

Interpreting machine learning models can seem complicated at first, but libraries like SHAP make everything as easy as a function call. We even don’t have to worry about data visualization, as there are built-in functions for that.

These men wanted me to know that they were miserable, they felt screwed over, and they felt demonized. They wanted me to know that the only option available to address white male patriarchy was either to maintain the status quo that was making us all miserable, or death. They wanted me to know that they were not capable of growth or change and that any attempts to bring about that growth or change would end them.

“I know you think I should kill myself because I voted for Donald Trump, because I’m white, because I’m a male, so I’m just going to, since that is the only ethical conclusion.”

The emails started coming in 2018, the same year my book So You Want to Talk About Race came out. They would usually arrive after I had posted an article on race or gender, but sometimes they would appear at random. I still remember the first one, which shook me.

A whole another story here. You now know how to interpret a single prediction, so let’s spice things up just a bit and see how to interpret a single feature’s effect on the model output.

As the threats of suicide piled up, I began to see a coordinated campaign to harass me, and as disturbing as it was, it was also sadly fascinating in what it revealed. These men were trying to terrorize me with what they saw as the only logical conclusion to my anti-racist, feminist work: the mass suicide of white men. They wanted me to know that they saw my work to end violent misogyny and white supremacy, and they saw that it was a threat, not only to their norms and their status but to their very lives.

The email continued, describing how he was going to kill himself (with a Glock that he kept at home) and reiterating that it would be my fault. He then ended with a racist tirade, calling me a “worthless monkey bitch.”

I have received many violent emails from white men over the years, but I sat with this one for a while. I tried to process what I was reading and tried to figure out what I should or could do about it. In the end, I placed it in the same folder as all the death and rape threats.

A few weeks later I received another email from a different sender. The message, with slight wording differences, was essentially the same. This white man was going to kill himself and I was to blame. A few days later I got a similar message via Twitter messenger. A few days after that, another email.

The chart below shows the change in wine quality as the alcohol value changes. Vertical dispersions at a single value show interaction effects with other features. SHAP automatically selects another feature for coloring to make these interactions easier to see:

The email listed the various hardships the sender had endured. Poverty, mental illness, discrimination. But none of that mattered, because I had shown him that the problem was that he was a white man and he should die. And so his death would be on my hands.