Hi Emily Tran,
Thank you for your questions 🙂
1) The ways a recommender system can destroy an application are endless. From a pure engineering point of view, a RecSys can be very resource demanding and slow down everything, which will hurt the application. From a recommender point of view, there is nothing that scares people away faster than bad recommendations. The contrary can also be the case - if a recommender knows the user too well, the user might find it scary. Just to mention a few.

2) Again, it’s hard to say what the solution is if you don’t know the exact problem. In the case where it’s low-quality recommendations, you should have a look at getting better data and tune your recommender. If it is an engineering problem, it is worth considering adding more resources or look to see if you can do more things offline such that it won’t hurt the performance of the system

3) again depends on the problem. But rule of thumb look for whether you have unhappy users

4) consider what metrics you want to optimise for and be sure you test your system well. Not just offline, but more importantly, online in production. A lot of research indicates that good offline test results do not ensure good online results. Be sure to have monitoring up and running before adding or updating a recommender such that you can measure any changes in performance.

Thank you very much for your answers. Nr 4 is absolutely helpful for every models! I will apply it from now on😤

Anything is possible 🙂 As with any data applications it depends on what data you have.
There are different ideas about how to do it. One way that seems to be discussed a lot recently is to do federated learning. Where basically you start out with a model trained on the server, this will be sent to each client, where the model will be updated according to the clients actions. This model will then be sent back to the server where the client models will be assimilated, and then it starts over.
Have a look at this article.

As with every data driven product, the life cycle is something the lines of:

1. Understand data/business
   
2. Collect data
   
3. Prepare data
   
4. RecSys Modelling
   
5. Offline evaluation
   
6. Recsys deployment
   
7. A/B testing
   
8. Repeat

I would like to answer this question but it would be many hundreds of pages. In short Content-based recommenders mostly use NLP to create embeddings of descriptions of the content, and then use those embeddings either to train recommenders or simply recommend by finding nearest neighbours to items consumed by the user.

Session-based recommendations on the other hand are done a bit like the language models that predicts next words in something you are writing. If you watched movie A, B, C then the next best movie to suggest would be D.
Have a look at this survey for more information.

I am not sure what to answer here. It depends on the product/content organisation wants the user to consume.
If there are too many products for users to overview themselves, then a recommender system is recommended 😉

It depends on whether the organisation has data enough to implement a recommender system, either usage or content data. And if they monitor their system. If they don’t monitor the system, then they won’t be able to measure if the recommender improves anything, then they might as well not do it.

Got it! Sorry for the vague question. But I got a good answer that I was looking for.

A) There are two aspects for a data engineer, the offline part one where you should have a system which collects the data needed to train the model, and then actually have the model trained, ensuring that it doesn’t blow up or worse create bad recs because of some new trends in user behavior. Secondly is the online part where you have to create a pipeline which provide personalised recs quick enough for the user not to find another page that loads quicker.

B) The data scientist needs to understand the data, the business domain and come up with good models (and tricks) to make the recommendations good.

In a way:) graph-based recommenders are based on the same concepts described in my book. But actual graph-based algorithms are not described in detail. For that, I would refer to Graph-Powered Machine Learning by Alessandro Negro.

Popularity bias is discussed, as well as item and user bias in collaborative filtering.

Usually, you would use a content-based recommender algorithm if you have little data. But since you say long tail, it has to do with something that also has popularity (because if there were no popular items, there would be no long tail either). It will always depend on which algorithm you use, but a way to do it is to penalise recommendations by popularity. Be careful because the idea is to recommend stuff that is not too popular but still something the user would like to watch.

Have a look at this article

yes I have seen some papers on this but what is your take on actually getting this stuff working in production?

Doink Given the chance, I would like to try it out because it sounds very interesting. The thing is that most apps/user only interacts with few items in a catalogue, so I am not sure how applicable it will be in a production system. Because of the amount of data, but also the resources used to train the model. Again I’d rather not say it wouldn’t work, I just think that many other things also need to be working to make it work.

The list is long. The biggest mistake is probably to think it is easy 🙂. As with any data-driven product, it is very hard to evaluate the quality of what you are implementing. It’s only ever in production your recommender gets an accurate estimate of performance.

Hi WingCode,
Yes, I am planning a 2end edition. What exactly will be the content is still up for debate as I’m no further than working on the table of content. But on my wish list are chapters on deep learning algorithms, multi-armed bandits and reinforcement learning, and Sequence-based. Besides generally updating the rest of the material. Do you have any other good suggestions?
It took my four years to write the first edition, I will be faster the second time around, but I wouldn’t expect it before a couple of years.

I better note that I don’t think my book is outdated, I think the content is still very relevant for somebody who wants to build recommender systems.

Kim Falk Probably more methods and techniques in extracting features from non-text media ( video, audio, picture) which can be used for recommendation systems?

This will be a longer answer 🙂

A search engine takes a seed, and the search engine returns a list of content using that seed. In most cases, the seed is a sequence of words. Nowadays, Google also allows for searching for images and returns images. A recommender system takes a user profile or an item as a seed and returns a list of content items.
Both search engines and recommender systems can personalise the result. So I would say a search engine and a recommender system differs only in what the seed is.
A recommender system can be viewed as a pipeline that does the following three steps:

• candidate selection
  
• apply filtering rules
  
• rerank result.
  Where candidate selection is a rough cut of items that might be interesting, it filters them based on some internal set of business rules and then reranks the remaining items. If you view it this way, then the candidate selection could be made by a search engine. In this view, a search engine could be a component of a recommender system. I’m sure others would look at it as the other way around 🙂.

Great, thank you!

Oh I am glad Alexey Grigorev thought to do this as I was on vacation last week 😉

I am not familiar with any work applying recommender systems on that problem. I heard about patient-medicin matching, but I would have to go and search for it again.

Oh, interesting. Thanks!

Hi Mansi Parikh
Thank you for the question. It depends on what type of data you have. Here are two quick ideas on how to start.
If you have usage data, then have a look at the framework called Implicit or LightFM. Both are pretty straightforward to use.
On the other hand, if you have content descriptions. Then you can use Gensim package to create word2vec embeddings and then find similar items based on cosine similarity.
I think the more advanced methods depends on the problem domain.
Hope that helps.

I love this answer. All of this is fairly new to me, and now I have a real plan to learn it or at least get some exposure. Thank you, Kim!!! Much appreciated.

Hi Kim Falk, I couldn’t understand the difference between usage data and content description with respect to recommendation.
Like I’m chalking a plan for building a Hotel Recommendation System for which I have data around 1gb with 0.9 million reviews of 4000 hotels. Im still a beginner and I cannot decide if I should use LightFM or TensorFlow Recommender or other ML algo.

Now that you have given 2 approaches to build on, i.e. usage data & content description. In which domain the Hotel Recommendation falls in?

The sky is the limit. Audio could be something like is it music or talk, is it one person or many, are they angry or happy. The video could be cut into scenes, and count number of explosions, kisses, cars, houses ect. generally do object detection, if there are humans what are they doing ect. It is also relevant how much the camera moves, how dark the images are.

With the images you should do object detection, atmosphere, style.

This is just a few ideas.

Extracting features from Audio, Video and Images are not something I have tried, I have used data which was extracted as embeddings, but never done it myself. Im not sure which framework to recommend you.

Thank you Kim for all the answers

What are the ways by which recommender systems can react in real time to topical/dynamic trends? As with most of the answers I have given, this also depends on what is the specific domain and scenario. A way could be to to rerank your recommendations based on which items have changed most in frequency the last hour compared to the last day (last day compared to the last week/month). I think a better answer would need more context.

How does one balance the weightages to recommendations derived from historical data and those pertaining to recent/near term trending data? Depends on your domain, but you can create a hybrid model as a linear function of the output of the two models and then run tests to see which weights will provide you with the better recommendations.

To be data driven you will have to create a platform where you allow for data to be collected about different settings and then let the data decide what the right values should be. Bayesian Multiarmed bandits would be away to leave the weights completely up to the machine.

Any framework to do this in a data driven way (with or w//o manual tinkering)? Im not sure what this framework should do exactly?

Kim Falk Thanks for the answers/pointers. Domain is shopping/retail primarily. I’ll check your answers to other questions also to pick up some useful info.

The simple answer is that you just take 10% of the traffic for the test group. The somewhat harder answer is that it needs to be a random sample which is representative of the whole population. If the users has a randomly created identifier you can create a hash function which can split the users into 10 buckets and then simply take all users from that bucket, There are several books written about how to do it best, Fx Tuning Up by David Sweet.