Deep Learning is a subset of Machine Learning, which is also known as artificial neural networks(ANNs), it is an algorithm inspired by the structure of human’s brain. To explain "deep learning" in a simple way, it is to find out a function to solve a very specific problem. Due to the complexity of the function, we can’t find out on our own, that’s the reason we need Deep Learning to help humans solve the problem.

Deep Learning is widely utilized in businesses, solving the problem of human inability, which can analyze large amounts of data in a short period of time, and even extend to more applications, such as automatic copywriting, optimizing user experience, providing visual data analysis, and making consumer behavior prediction.

The main value of deep learning in advertising is making consumer behavior predictions. We expect the "Deep Learning Model" can be trained to identify the outline of the buyer and the result may be verified in the future.

In this article, we are going to show you how to scientifically measure the performance of deep learning models, even explain how we train deep learning models and how to predict consumers' purchase intent.

What’s the importance of data? 

Today, understanding the needs of the consumers is very important for marketers. How much are consumers willing to pay for your products? How much time are they willing to spend on your store? Will they repurchase on your store? The more you know about your consumers, the more you can interact with them and fulfill them in a more effective way. How do you bring your brand closer to consumers? It all depends on the collected data.

Flexible data for best decision

The deep learning model makes the buying behavior prediction based on the past 7 days’ data, and predicts the audience’s action in the next 14 days. The prediction result will classify the audience as "will buy" and "won't buy", in order to evaluate classification models, we need to verify it with the real data.

Evaluating Deep Learning Models: Precision, Recall and F1-score

Precision

Precision is a measure of models’ positive predictions made are correct. It wants to find out the percentage of people who are actually buying. For example, if the model predicts that 100 people will purchase, but only 20 people have actually purchased, then the model's precision is 20%.

Recall

Recall is a measure of how many of the positive cases the model correctly predicted, over all the positive cases in the data. It gives a clear picture on how many people actually did purchase. According to the last example, if the total number of people who actually purchase is only 40, and the model has predicted 20 people of them, then the recall is 50%.

F1-score

F1-score is a measure combining both precision and recall. It helps balance the two metrics(precision and recall), requiring both to have a higher value for the F1-score value to rise. We can more intuitively evaluate the deep learning models through F1-score.

📓Learning more about Precision, Recall & F1-score? Read an article by Teemu Kanstrén.

From past experiences, increasing in the precision rate is usually lower the recall rate, which means that we might miss out on a portion of potential consumers. That's why we bring in F1-score to balance it out.

Tagtoo AI prediction results vs. Remarketing audience results

In digital marketing, "remarketing" is the most common strategy out there. For e-commerce, "Products Page View" and "Add To Cart" are by far the easiest way in getting potential audiences for "remarketing".

The results above shown that the precision rate of "Add to Cart" is higher than "Products Page View", but the recall rate is much lower, which means the number of audience is too few and the number of people who actually purchase is also relatively less. On the contrary, the recall rate of "Products Page View" is much higher than "Add to cart", but the precision rate is very low, which means that even though the number of audience is large, most of them don’t purchase.

Q: Why is there a very large gap in both precision rate and recall rate between "Products Page View" and "Add to Cart" audience? What is the relationship between the two?

First, we must understand the definition of the two types of audience mentioned above. It is the classification based on the users’ action on your website.

There may be hundreds of thousands or even millions of people browsing your website every day, but how many of them will add items to their cart? Based on the survey by databox, the average add-to-cart rate is 3-4%.

You can imagine that the "Products Page View" and "Add to Cart" audience are like a funnel that filters out users who are more likely to purchase, with fewer and fewer audience in the lower stage. The results of "Tagtoo Deep Learning Model" shown a much higher precision and recall rate compare to "remarketing audience" with a F1-score of 29%. This indicates that the predicted audience has doubled the conversion rate of "remarketing".

Creating a new type of audience

We hope to offer our clients more choices on advertising, so we applied the data from our clients, OB Design, with deep learning models to make the buying behavior prediction. The predicting audience has a wide usage, not only in advertising, but also in membership marketing or email marketing.

Training for Tagtoo Deep Learning Models

Step1: Website Data Processing

E-commerce websites have an average of 5 million data per day. In order to improve the efficiency of the deep learning model, we must compress the data and decide which features to train the model.

Step2: Deep learning model optimization

The characteristic of e-commerce data is chronological, each consumer has a series of purchase signals that must be linked sequentially. Thus, we use Recurrent Neural Networks (RNN) and Long Short-term Memory (LSTM) to train the model, it can be interpreted as a "memory neural network" that allows the model to distinguish the pattern in which each purchase occurs.

Step3: Verify with real data

The results of the deep learning model need to be verified with the real data. As mentioned earlier, the deep learning model will use the past 7 days’ data to predict the buying behavior in the next 14 days, so we need to prove the result and find out the error of the deep learning model. We always want to lower our error when training a model, optimizing a model until it reaches a lowest error rate.

Customized models for different industries

We do not use a single model in every situation or industry. Instead, we train different models accordingly. By extracting the right data, we'll able to predict the buying behavior accurately.

What are the benefits of using deep learning models?

1. Extensive data insight

AI can provide an extremely broad view of massive amounts of data exactly and quickly.

2. Truly understand your customer

Marketers have more beneficial clues, such as customer profiles and product preferences, to adjust marketing strategies.

3. Targeting more precise audiences

AI can identify audiences with higher conversion rather than searching blindly.

4. Customized audience

Flexible adjustment according to particular industries and customer needs.

The Verdict

Big data analytics can bring infinite possibilities for enterprises, nevertheless, only the proper analysis can deduce beneficial information from the data. We believe marketers who know how to use data will be able to stand out in the future.

If you are interested in learning more about how we use deep learning in digital marketing, please contact us for consultation.