predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pat Ferrel <>
Subject Re: Tuning of Recommendation Engine
Date Wed, 07 Dec 2016 22:09:18 GMT
No, we find the value of quantile LLR thresholds and use those thresholds to calculate MAP.
Then we look at MAP&* number-of-people-that-get-recs to see if there is a max. This is
basically and analysis of precision vs recall. MAP will often increase monotonically with
higher thresholds until you get no recommendations at all. Hyper-parameter search via trial
and error.

On Dec 4, 2016, at 9:15 PM, Gustavo Frederico <> wrote:

Pat, is this tool that finds the optimal LLR thresholds using MAP@K? Did you model it as a
regression problem? 



On Thu, Dec 1, 2016 at 2:48 PM, Pat Ferrel < <>>
This is a very odd statement. How many tuning knobs do you have with MLlib’s ALS, 1, 2?
There are a large number of tuning knobs for the UR to fit different situations. What other
recommender allows multiple events as input? The UR also has business rules in the form of
filters and boosts on item properties. I think you may have missed a lot in the docs, check
some of the most important tuning here: <>
and the config params for business rule here: <>

But changes must be based on either A/B tests or cross-validation. Guessing at tuning is dangerous,
intuition about how a big-data algorithm works takes a long time to develop and the trade-offs
may do your business harm.

We have a tool that find optimal LLR thresholds based on predictive strength and sets the
threshold per event pair. While you can set these by hand the pattern we follow is called
hyper-parameter search, which finds optimal tuning for you. 

On Dec 1, 2016, at 11:17 AM, Harsh Mathur < <>>

Hi Pat,
I really appreciate the product, but our team was discussing about how little control we have
As in, say some recommendations got delivered to the user and we are tracking conversions
of course, so we can know if it's working or not. Now, say if we see that conversions are
low, as a developer I have very little to experiment with here. I don't mean any disrespect.
I have gone through the code and have put in efforts to understand it too,  the UR is still
better than the explicit or implicit templates as it has filtration for properties, only thing
lacking in my opinion is the weightages.

I read your ppt 
Recommendations = PtP +PtV+...
We were wondering if it could be
Recommendations = a * PtP + b * PtV+ ...

Where a and b are constants for tuning. In my understanding PtP is a matrix so scalar multiplication
should have be possible. Please correct me if I am wrong.

Also I was reading about log likelihood method, but I couldn't find a proper explanation.
I would be happy if anyone here can explain it in more detail. Thanks in advance.

Here is what I understood.
For every item-item pair per expression (PtP, PtV), to calculate a score, it will find 4 things:
1. No of users who posted both events for the pair,
2. No of users who posted event for one but not the other and vice versa,
3. No of users who posted for neither

Then a formula is applied taking the 4 params as input and a score is returned.

For each item and event pair you are storing top 20 items according to score in elastic search.
I didn't understand why the 2nd and third parameters are taken, also if anyone can explain
the correctness of the method, That is why does it work rather how it works?

Harsh Mathur

On Dec 1, 2016 11:01 PM, "Pat Ferrel" < <>>
Exactly so. The weighting of events is done by the algorithm. To add biases would very likely
be wrong and result in worse results. It is therefore not supported in the current code. There
may be a place for this type of bias but it would have to be done in conjunction with a cross-validation
tests we have in our MAP test suite and it is not yet supported. Best to leave them with the
default weighting in the CCO algorithm, which is based on the strength of correlation with
the conversion event, which I guess is purchase in your case.

On Nov 28, 2016, at 2:19 PM, Magnus Kragelund < <>>

It's my understanding that you cannot apply a bias to the event, such as "view" or "purchase"
at query time. How the engine is using your different events to calculate score, is something
that is in part defined by you and in part defined during training.
In the engine.json config file you set an array of event names. The first event in the array
is considered a primary event, and will be the event that the engine is trying to predict.
The other events that you might specify is secondary events, that the engine is allowed to
take in to consideration, when finding correlations to the primary event in your data set.
If no correlation is found for a given event, the event data is not taken into account when
predicting results. 

Your array might look like this, when predicting purchases: ["purchase",  "initiated_payment",
"view", "preview"]

If you use the special $set event to add metadata to your items, you can apply a bias or filter
on those metadata properties at query time.


From: Harsh Mathur < <>>
Sent: Monday, November 28, 2016 3:46:46 PM
To: <>
Subject: Tuning of Recommendation Engine
I have successfully deployed the UR template.

Now I wanted to tune it a little bit, As of now I am sending 4 events, purchase, view, initiated_payment
and preview. Also our products have categories, I am setting that as item properties.

Now, as I query say:
"item": "{item_id}",
"fields": [
"name": "view",
"bias": 0.5
"name": "preview",
"bias": 5
"name": "purchase",
"bias": 20

and query 
        "item": "{item_id}"

For both queries, I get the same number of recommendations just the score varies. The boosting
isn't changing any recommendations, just changing the scores. Is there any way in UR that
we can give more preference to some events, it will help give us more room to try and see
and make the recommendations more relevant to us.

Harsh Mathur <>

“Perseverance is the hard work you do after you get tired of doing the hard work you already

View raw message