Why some recommendations fail: recommendation engines and their challenges

Using algorithms to generate buying suggestions is big business. Netflix reported that its recommendation engine contributes $1 billion to the bottom line every year. However, sometimes the suggestions are far away.

Take, for example, an ad I received applying to be a van driver. I’ve never been a professional driver, I don’t even like driving and I’ve never owned a van. It’s clear that this recommendation engine doesn’t know anything about me.

There are several ways that recommendation algorithms can reach incorrect conclusions. Here are just a few examples of each type of recommendation engine.

1. Collaborative Filtering

This filtering method is based on collecting and analyzing information about user preferences. The assumption is that if two users have a common interest, they will also have other common interests, so product recommendations will fit both. The advantage of this type of analysis is that the algorithm does not need to use deep learning inferences to understand the recommended article, it only needs to identify users with similar interests.

However, a disadvantage of collaborative filtering is that it requires a large dataset of active users who have rated or purchased a product in order to make accurate predictions. When you have little user activity, it’s much harder to generate quality recommendations. The number of items sold on major ecommerce sites is extremely large. Therefore, even the most popular items can have very few reviews. This is known as the long tail or data scarcity problem.

There is also no way to deal with new items that have not yet been rated.

Additionally, there are millions of users and products in many environments where these systems make recommendations. As a result, a large amount of computing power is often required to perform the necessary calculations, meaning many companies are forced to limit the amount of data their models ingest, which can negatively impact accuracy.

2. Content-Based Filtering

Content-based filtering methods use keywords that describe an item to match recommendations and people. For example, when recommending jobs, keywords in the job description can be matched with keywords in the user’s resume.

The main disadvantage of this model is that it can only give recommendations based on the user’s existing characteristics. It also requires text analysis, which can lead to errors when the algorithm needs to identify keywords that are spelled differently; for example: trainer, trainer, teacher or moderator.

This type of recommendation engine is also challenged when the solution is multilingual and requires translating and comparing words and phrases in different languages.

3. Hybrid Recommendation Engines

Hybrid recommender systems use collaborative filtering and product-based filtering together to recommend customers a broader range of products with more precision.

Hybrid recommender systems can generate predictions separately and then combine them, or the capabilities of collaborative methods can be added to a content-based approach (and vice versa). Additionally, many hybrid recommendation engines incorporate demographic analysis and knowledge-based algorithms that draw inferences about user needs and preferences based on deep learning.

However, even if hybrid recommendation engines can improve accuracy, they can suffer from longer processing times. The importance of speed varies depending on the application. For example, movie and e-commerce recommendation systems can learn more slowly, while an application that recommends who to follow on Twitter changes frequently, forcing a recommendation engine to make near real-time predictions based on up-to-date data.

In addition, personal interests are time-sensitive in different ways. For example, individual sports like running or swimming are long-term, while tracking sporting events like championships for popular professional teams can change constantly. Recommendations based on real-time matches need to be updated more frequently.

Improving accuracy for all types of recommendation engines

In any case, to be more reliable, recommendations should be diverse, able to quickly adapt to new trends, and quickly scale to handle more data. One way for developers to improve the accuracy of their recommendation engines is to use pre-trained models off the shelf and invest in MLops tools that can help speed up the process of getting models into production and have models up and running on a regular basis to monitor deviations.

Personally, I’m always happy to receive recommendations for restaurants, bars, books and musical performances. Even if the predictions are far-fetched, I can be persuaded to try new things. However, using more complex models, pre-trained with more data, makes it less likely that I’ll be asked to apply to be a van driver.

Michael Galarnyk is an AI evangelist at cnvrg.io.