However, if you\u2019d like a full picture of how to solve such a fundamental PPC challenge, I highly suggest reading the full post.<\/p>\n\n\n\n
Seeing<\/em> the AI-based solution<\/strong><\/h3>\n\n\n\nIt all begins with thoroughly and clearly defining the business problem to be solved. In this case:<\/p>\n\n\n\n
Many retailers have large and ever-changing inventories, and every item requires a unique bid reflecting its value to the business. These bids must be constantly adapted to suit any browsing device, at any time. And for maximum cost efficiency, all of this must happen automatically.<\/em><\/p>\n\n\n\nIf you think that sounds complex, you\u2019re right. Defining the perfect bids requires knowledge about many things: our clients\u2019 Return on Ad Spend (ROAS) targets, market trends, future product performance and much more. Trends and future product performance are particularly challenging, given their volatility.<\/p>\n\n\n\n
For example, the below table of hypothetical historical data includes a predicted conversion rate of 50%. But with only two clicks and one conversion for that product, the confidence intervals – the margins within which the prediction could actually fall based on this data – range from around 1% to 99%. That\u2019s not useful at all! In fact of all the factors important for defining a bid, only the ROAS targets are known – and they change frequently, too.<\/p>\n\n\n\n![\"\"](\"https:\/\/smarter-ecommerce.com\/blog\/en\/wp-content\/uploads\/2020\/09\/The-challenge-of-predicting-item-level-performance-1024x525.png\")
The challenge of predicting item-level performance: daily metrics can fluctuate significantly<\/em><\/figcaption><\/figure>\n\n\n\nEnter Artificial Intelligence. Specifically, Machine Learning is all about finding patterns in data which can be used to make a prediction. smec\u2019s bidding model learns to predict various aspects of future product performance, based on patterns it finds in the customer inventory data and Google Shopping reports. The predicted metrics are fed into a carefully designed formula to generate recommended bids to Whoop!, our award-winning bid management solution. The result? Efficiency and granular, product-level control over the marketing budget, all leading to better performance.<\/p>\n\n\n\n![\"\"](\"https:\/\/smarter-ecommerce.com\/blog\/en\/wp-content\/uploads\/2020\/09\/The-advantage-of-item-level-bidding-1024x534.png\")
The advantage of item-level bidding over bidding on product groups:<\/em>
efficiency, control, and ultimately better performance<\/em><\/figcaption><\/figure>\n\n\n\nFinding nuggets of wisdom in the data<\/strong><\/h3>\n\n\n\nSo we know we want to predict certain performance metrics which will be used to determine bids automatically. In order to train machine-learning models to make such predictions, it\u2019s clear we will need past performance data.<\/p>\n\n\n\n
Additionally, we include information on the products themselves: characteristics like colour, material and brand. This gives us the unique benefit of being able to make accurate predictions for new products which have no performance data for the models to learn from. Of course, in the retail world, most <\/em>products fall into this category.<\/p>\n\n\n\nBut do we just throw all this data into the model and see what happens? Not quite. You see, our bidding models are constantly trained, in order to generate current, accurate bid predictions. Providing too much data would make the model unnecessarily complicated and prohibitively slow to train, while including information with no real impact on product performance – like GTIN – can even be harmful. And finally, simpler models along with model explainability techniques can yield better performance and real retail insights.<\/p>\n\n\n\n
Thus, we need to convert the data into \u2018features\u2019: series of numeric and categorical values representing the most informative aspects of the catalog and performance data per product. This process, called \u2018feature engineering\u2019, is one of a data scientist\u2019s most crucial tasks.<\/p>\n\n\n\n
We propose potential features based on machine learning know-how and (company wide) domain expertise, and then verify their usefulness using rigorous testing and model explainability techniques, which reveal which features had the biggest impact on predictions. Our pipeline can even select the best features from those we provided, using different statistical methods.<\/p>\n\n\n\n
Below we highlight the importance of using a rich feature set: remember the challenging conversion rate prediction earlier, where a new product didn\u2019t have enough performance data to learn from? The solution is to include features describing the product\u2019s characteristics, since those characteristics will have affected the performance of other products for which there is<\/em> enough data. The model will learn to spot the similarities, and so be able to generalise to products it has never seen before:<\/p>\n\n\n\n![\"\"](\"https:\/\/smarter-ecommerce.com\/blog\/en\/wp-content\/uploads\/2020\/09\/Feeding-machine-learning-models-1024x524.png\")
Demonstration: the importance of feeding machine learning models<\/em>
a rich, informative feature set<\/em><\/figcaption><\/figure>\n\n\n\nChoosing and combining machine learning components<\/strong><\/h3>\n\n\n\nMachine learning is a huge field with many algorithms and implementation techniques. A good data science team must know about all of these, their strengths, weaknesses, requirements and appropriateness to a given task. One algorithm we have chosen to use also makes a great example for how machine learning works, so let\u2019s take a look: the random forest regressor.<\/p>\n\n\n\n
This algorithm actually works by running many instances of another algorithm, called a Decision Tree, and combining their results. A decision tree tries to learn a tree-like structure of simple decision rules from training data, which can be used to make predictions about future, unseen inputs.<\/p>\n\n\n\n
Each rule tries to split the data points based on some feature, so that the resulting two groups are very different to each other but the items inside them are similar. It keeps doing this until each \u2018branch\u2019 consists of a single \u2018leaf\u2019 containing a small group of products whose performance is similar enough that a prediction can be made about them.<\/p>\n\n\n\n
Below we illustrate this with the example of conversion rate prediction. Starting from the top, the algorithm takes all of the data and determines, statistically, that the first most informative way to split the data points is based on whether they are an own brand (Eigenmarke) or an other brand (Fremdmarke).<\/p>\n\n\n\n
In other words, it discovers that \u2018own brand\u2019 and \u2018other brand\u2019 products have very different average conversion rates, but that the products within these two clusters have more similar conversion rates. In this hypothetical world, it appears that customers have different buying attitudes towards Eigenmarke and Fremdmarke products.<\/p>\n\n\n\n
Following the \u2018own brand\u2019 branch, the algorithm finds that the most informative way to split those products is based on their product line: apparently our hypothetical customers also have different attitudes depending on whether the product belongs to cosmetics, clothing, or baby food. Let\u2019s consider the \u2018cosmetics\u2019 branch. Here, the next best predictor of conversion rate is whether the product costs less or more than \u20ac5.<\/p>\n\n\n\n
Perhaps our hypothetical customers often spontaneously add cheap cosmetics to their baskets, but are not so purchase-ready for more expensive items. At this point, the algorithm determines that there is no useful way to further split the products in the less-than-\u20ac5 group, and predicts the conversion rate as the average conversion rate for these products. Similarly, the average conversion rate for the more-than-\u20ac5 group is output as the prediction for that leaf.<\/p>\n\n\n\n
If you think that sounds complex, you\u2019re right. Defining the perfect bids requires knowledge about many things: our clients\u2019 Return on Ad Spend (ROAS) targets, market trends, future product performance and much more. Trends and future product performance are particularly challenging, given their volatility.<\/p>\n\n\n\n
For example, the below table of hypothetical historical data includes a predicted conversion rate of 50%. But with only two clicks and one conversion for that product, the confidence intervals – the margins within which the prediction could actually fall based on this data – range from around 1% to 99%. That\u2019s not useful at all! In fact of all the factors important for defining a bid, only the ROAS targets are known – and they change frequently, too.<\/p>\n\n\n\n Enter Artificial Intelligence. Specifically, Machine Learning is all about finding patterns in data which can be used to make a prediction. smec\u2019s bidding model learns to predict various aspects of future product performance, based on patterns it finds in the customer inventory data and Google Shopping reports. The predicted metrics are fed into a carefully designed formula to generate recommended bids to Whoop!, our award-winning bid management solution. The result? Efficiency and granular, product-level control over the marketing budget, all leading to better performance.<\/p>\n\n\n\n So we know we want to predict certain performance metrics which will be used to determine bids automatically. In order to train machine-learning models to make such predictions, it\u2019s clear we will need past performance data.<\/p>\n\n\n\n Additionally, we include information on the products themselves: characteristics like colour, material and brand. This gives us the unique benefit of being able to make accurate predictions for new products which have no performance data for the models to learn from. Of course, in the retail world, most <\/em>products fall into this category.<\/p>\n\n\n\n But do we just throw all this data into the model and see what happens? Not quite. You see, our bidding models are constantly trained, in order to generate current, accurate bid predictions. Providing too much data would make the model unnecessarily complicated and prohibitively slow to train, while including information with no real impact on product performance – like GTIN – can even be harmful. And finally, simpler models along with model explainability techniques can yield better performance and real retail insights.<\/p>\n\n\n\n Thus, we need to convert the data into \u2018features\u2019: series of numeric and categorical values representing the most informative aspects of the catalog and performance data per product. This process, called \u2018feature engineering\u2019, is one of a data scientist\u2019s most crucial tasks.<\/p>\n\n\n\n We propose potential features based on machine learning know-how and (company wide) domain expertise, and then verify their usefulness using rigorous testing and model explainability techniques, which reveal which features had the biggest impact on predictions. Our pipeline can even select the best features from those we provided, using different statistical methods.<\/p>\n\n\n\n Below we highlight the importance of using a rich feature set: remember the challenging conversion rate prediction earlier, where a new product didn\u2019t have enough performance data to learn from? The solution is to include features describing the product\u2019s characteristics, since those characteristics will have affected the performance of other products for which there is<\/em> enough data. The model will learn to spot the similarities, and so be able to generalise to products it has never seen before:<\/p>\n\n\n\n Machine learning is a huge field with many algorithms and implementation techniques. A good data science team must know about all of these, their strengths, weaknesses, requirements and appropriateness to a given task. One algorithm we have chosen to use also makes a great example for how machine learning works, so let\u2019s take a look: the random forest regressor.<\/p>\n\n\n\n This algorithm actually works by running many instances of another algorithm, called a Decision Tree, and combining their results. A decision tree tries to learn a tree-like structure of simple decision rules from training data, which can be used to make predictions about future, unseen inputs.<\/p>\n\n\n\n Each rule tries to split the data points based on some feature, so that the resulting two groups are very different to each other but the items inside them are similar. It keeps doing this until each \u2018branch\u2019 consists of a single \u2018leaf\u2019 containing a small group of products whose performance is similar enough that a prediction can be made about them.<\/p>\n\n\n\n Below we illustrate this with the example of conversion rate prediction. Starting from the top, the algorithm takes all of the data and determines, statistically, that the first most informative way to split the data points is based on whether they are an own brand (Eigenmarke) or an other brand (Fremdmarke).<\/p>\n\n\n\n In other words, it discovers that \u2018own brand\u2019 and \u2018other brand\u2019 products have very different average conversion rates, but that the products within these two clusters have more similar conversion rates. In this hypothetical world, it appears that customers have different buying attitudes towards Eigenmarke and Fremdmarke products.<\/p>\n\n\n\n Following the \u2018own brand\u2019 branch, the algorithm finds that the most informative way to split those products is based on their product line: apparently our hypothetical customers also have different attitudes depending on whether the product belongs to cosmetics, clothing, or baby food. Let\u2019s consider the \u2018cosmetics\u2019 branch. Here, the next best predictor of conversion rate is whether the product costs less or more than \u20ac5.<\/p>\n\n\n\n Perhaps our hypothetical customers often spontaneously add cheap cosmetics to their baskets, but are not so purchase-ready for more expensive items. At this point, the algorithm determines that there is no useful way to further split the products in the less-than-\u20ac5 group, and predicts the conversion rate as the average conversion rate for these products. Similarly, the average conversion rate for the more-than-\u20ac5 group is output as the prediction for that leaf.<\/p>\n\n\n\n
efficiency, control, and ultimately better performance<\/em><\/figcaption><\/figure>\n\n\n\nFinding nuggets of wisdom in the data<\/strong><\/h3>\n\n\n\n
a rich, informative feature set<\/em><\/figcaption><\/figure>\n\n\n\nChoosing and combining machine learning components<\/strong><\/h3>\n\n\n\n
![\"\"](\"https:\/\/smarter-ecommerce.com\/blog\/en\/wp-content\/uploads\/2020\/09\/Decision-tree-grown-to-predict-conversion-rate-1024x579.png\")