Recommender Systems Tutorial (Part 4) -- Challenges
Download as PPT, PDF2 likes568 views
1. Evaluation methods for recommendation systems include experimental design with A/B testing, as well as offline evaluation on logged user data using metrics like predictive accuracy, recall, and replay-match methods. 2. Many recommendation system evaluations focus only on predictive accuracy rather than overall user utility or business objectives. Multi-objective optimization aims to maximize multiple goals like clicks, revenue, and engagement. 3. Challenges in recommendation system evaluation and optimization include multi-position optimization, exploration-exploitation tradeoffs for large item pools, constructing user profiles with sparse data, incorporating social information and whole-page effects, and multi-context learning.
![Details on Replay-Match method (Li, Langford, et al)
? x: feature vector for a visit
? r = [r1,r2,¡,rK]: reward vector for the K items in inventory
? h(x): recommendation algorithm to be evaluated
? Goal: Estimate expected reward for h(x)
? s(x): recommendation scheme that generated logged-data
? x1,..,xT: visits in the logged data
? rti: reward for visit t, where i = s(xt)
Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 5](https://image.slidesharecdn.com/recommendersystems-tutorial-part-4-challenges-121219120044-phpapp01/85/Recommender-Systems-Tutorial-Part-4-Challenges-5-320.jpg)
More Related Content
Similar to Recommender Systems Tutorial (Part 4) -- Challenges (20)
Recommender Systems Tutorial (Part 4) -- Challenges
- 1. Evaluation Methods and Challenges
- 2. Evaluation Methods ? Ideal method ¨C Experimental Design: Run side-by-side experiments on a small fraction of randomly selected traffic with new method (treatment) and status quo (control) ¨C Limitation ? Often expensive and difficult to test large number of methods ? Problem: How do we evaluate methods offline on logged data? ¨C Goal: To maximize clicks/revenue and not prediction accuracy on the entire system. Cost of predictive inaccuracy for different instances vary. ? E.g. 100% error on a low CTR article may not matter much because it always co-occurs with a high CTR article that is predicted accurately Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 2
- 3. Usual Metrics ? Predictive accuracy ¨C Root Mean Squared Error (RMSE) ¨C Mean Absolute Error (MAE) ¨C Area under the Curve, ROC ? Other rank based measures based on retrieval accuracy for top-k ¨C Recall in test data ? What Fraction of items that user actually liked in the test data were among the top-k recommended by the algorithm (fraction of hits, e.g. Karypsis, CIKM 2001) ? One flaw in several papers ¨C Training and test split are not based on time. ? Information leakage ? Even in Netflix, this is the case to some extent ¨C Time split per user, not per event. For instance, information may leak if models are based on user-user similarity. Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 3
- 4. Metrics continued.. ? Recall per event based on Replay-Match method ¨C Fraction of clicked events where the top recommended item matches the clicked one. ? This is good if logged data collected from a randomized serving scheme, with biased data this could be a problem ¨C We will be inventing algorithms that provide recommendations that are similar to the current one ? No reward for novel recommendations Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 4
- 5. Details on Replay-Match method (Li, Langford, et al) ? x: feature vector for a visit ? r = [r1,r2,¡,rK]: reward vector for the K items in inventory ? h(x): recommendation algorithm to be evaluated ? Goal: Estimate expected reward for h(x) ? s(x): recommendation scheme that generated logged-data ? x1,..,xT: visits in the logged data ? rti: reward for visit t, where i = s(xt) Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 5
- 6. Replay-Match continued ? Estimator ? If importance weights and ¨C It can be shown estimator is unbiased ? E.g. if s(x) is random serving scheme, importance weights are uniform over the item set ? If s(x) is not random, importance weights have to be estimated through a model Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 6
- 7. Back to Multi-Objective Optimization EDITORIAL Recommender AD SERVER ?Clicks on FP links influence PREMIUM display downstream supply distribution (GUARANTEED) content Spot Market (Cheaper) Downstream engagement (Time spent) Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 7
- 8. Serving Content on Front Page: Click Shaping ? What do we want to optimize? ? Current: Maximize clicks (maximize downstream supply from FP) ? But consider the following ¨C Article 1: CTR=5%, utility per click = 5 ¨C Article 2: CTR=4.9%, utility per click=10 ? By promoting 2, we lose 1 click/100 visits, gain 5 utils ? If we do this for a large number of visits --- lose some clicks but obtain significant gains in utility? ¨C E.g. lose 5% relative CTR, gain 40% in utility (revenue, engagement, etc) Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 8
- 9. Why call it Click Shaping? other buzz other video autos finance buzz finance videogames gmy.news video tv health autos videogames gmy.news health travel hotjobs tv hotjobs tech movies travel new.music movies tech sports new.music ?AFTER ?BEFORE shopping sports news shine rivals shopping omg news realestate shine 10.00% rivals omg 8.00% realestate 6.00% 4.00% 2.00% ?Supply distribution 0.00% ts ?Changes ing bs s omg s te ch -2.00% e tv ie s s s r o buzz l s tat e ic ce th es tra ve othe aut o shin rival new .ne w vide spor .mus hotjo heal finan gam mov -4.00% p shop rea le gmy -6.00% new o vide -8.00% -10.00% ?SHAPING can happen with respect to any downstream metrics (like engagement) Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 9
- 10. Multi-Objective Optimization K properties m user segments n articles ?A1 ?S1 ?news ?A2 ?S2 ?finance ?¡ ?¡ ?¡ ?omg ?An ?Sm ? CTR of user segment i on article j: pij ? Time duration of i on j: dij ?10 Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 10
- 11. Multi-Objective Program ? Scalarization Goal Programming Simplex constraints on xiJ is always applied Constraints are linear Every 10 mins, solve x Use this x as the serving scheme in the next 10 mins 11 Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 11
- 12. Pareto-optimal solution (more in KDD 2011) ?12 Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 12
- 13. Summary ? Modern recommendation systems on the web crucially depend on extracting intelligence from massive amounts of data collected on a routine basis ? Lots of data and processing power not enough, the number of things we need to learn grows with data size ? Extracting grouping structures at coarser resolutions based on similarity (correlations) is important ¨C ML has a big role to play here ? Continuous and adaptive experimentation in a judicious manner crucial to maximize performance ¨C Again, ML has a big role to play ? Multi-objective optimization is often required, the objectives are application dependent. ¨C ML has to work in close collaboration with engineering, product & business execs Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 13
- 14. Challenges
- 15. Recall: Some examples ? Simple version ¨C I have an important module on my page, content inventory is obtained from a third party source which is further refined through editorial oversight. Can I algorithmically recommend content on this module? I want to drive up total CTR on this module ? More advanced ¨C I got X% lift in CTR. But I have additional information on other downstream utilities (e.g. dwell time). Can I increase downstream utility without losing too many clicks? ? Highly advanced ¨C There are multiple modules running on my website. How do I take a holistic approach and perform a simultaneous optimization? Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 15
- 16. For the simple version ? Multi-position optimization ¨C Explore/exploit, optimal subset selection ? Explore/Exploit strategies for large content pool and high dimensional problems ¨C Some work on hierarchical bandits but more needs to be done ? Constructing user profiles from multiple sources with less than full coverage ¨C Couple of papers at KDD 2011 ? Content understanding ? Metrics to measure user engagement (other than CTR) Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 16
- 17. Other problems ? Whole page optimization ¨C Incorporating correlations ? Incentivizing User generated content ? Incorporating Social information for better recommendation ? Multi-context Learning Deepak Agarwal & Bee-Chung Chen @ ICML¡¯11 17