Livestreaming（直播）through channels such as Amazon Live and QVC is an increasingly popular way to sell goods online. It usually lasts between 5 and 10 minutes, and someone promotes a product. Viewers can then readily buy it by clicking on a link.
We analyzed 99, 451 sales cases on a livestream selling platform and matched them with actual sales cases. In terms of time, that is equal to over 2 million 30-second television advertisements.
To determine the emotional（情绪的）expression of the salesperson， we used two deep learning models： a face model and an emotion model. The face model discovers the presence or absence of a face in a frame （镜头） of a video stream. The emotion model then determines the probability that a face is exhibiting any of the six basic human emotions: happiness, sadness, surprise, anger, fear or disgust. For example, smiling signals a high probability of happiness, while an off-putting expression usually points toward anger.
We wanted to see the effect of emotions expressed at different times in the sales cases so we counted probabilities for each emotion for all 62 million frames in our database. We then combined these probabilities with other possible aspects that might drive sales - such as price and product characteristics -to judge the effect of emotion.
We found that, perhaps unsurprisingly, when salespeople show more negative emotions-such as anger and disgust-the volume of sales went down. But we also found that a similar thing happened when the salespeople show high levels of positive emotions, such as happiness or surprise.
A likely explanation, based on our research, is that smiling can be unpleasant because it lacks true feelings and can reduce trust in the seller. A seller's happiness may be taken as a sign that the seller is gaining interests at the customer's expense.
