PERSONALIZATION · 2026-11-04
Personalization — Paper Daily, 2026-11-04
Today's recommendation and ads track has two threads. The first is where training signals come from: Meta uses auxiliary signals such as pre-click viewing time and add-to-cart events to fill in rare conversion labels, and another paper uses behavior patterns learned from popular items to correct the content embeddings of long-tail items. The second is the details around launch. How to encode numerical features has to be re-tested on your own data, and whether a click opens a website or an app is worth deciding click by click. Headline rewrites by LLMs that push up rankings now also have a defense with no added latency.
4.35
There are a dozen or so ways for a CTR model to turn numerical features such as prices and counts into vectors, and in one fair comparison no single method wins everywhere. This used to be hard to compare, because papers differ in preprocessing and training pipelines. This paper puts 18 of them into one codebase with the same tuning budget and compares them across several datasets and backbones. DEER has the best overall rank, but it does not win everywhere: change the dataset or the backbone and the leader can change. When the inputs switch from standardized values back to their original scale, the winner changes in 11 of 27 settings. Accuracy and inference latency also do not rise and fall together. So how to encode numerical features has to be re-tested on your own data and backbone.
4.25
Ad CVR models are trained on conversion events such as purchases and installs, but conversions are rare: in this Meta model, about half of the training samples have no conversion label. This paper from Meta explains how to fill the gap with other signals. How long a user looks at an ad before clicking, add-to-cart events, the semantics of the ad content and more are trained as auxiliary tasks alongside the main task. At serving time only the main task is used, so there is no extra cost. Auxiliary signals are often noisy. The most effective fix is called LaCo. It combines several auxiliary signals with logic such as AND, OR and NOT, picks the combination that best matches real conversions, especially with the fewest false positives, and trains on it as one new label. It lowers offline NE by 0.25%. After these changes launched together with other improvements, A/B tests showed a significant rise in advertiser conversions. For teams doing conversion prediction, getting the auxiliary labels right matters more than adding more auxiliary tasks.
4.25
News recommendation has a loophole: keep the facts, let an LLM make the headline more sensational, for example by adding words like Thrilling or Ultimate, and the article ranks higher in the list. Existing defenses usually target only one or two rewriting methods. This paper first uses several LLMs and several rewriting strategies to generate, offline, a large pool of fact-checked headline rewrites. During training it samples a few of them each time and asks the model to give the original and the rewritten headline similar scores, while still learning from clicks as usual. At serving time the model architecture is unchanged and there is no extra latency. On three recommendation models on the MIND dataset, facing rewriting models never seen in training, it has the lowest rate of articles pushed up the ranking, up to 3.0 percentage points lower than the strongest defense baseline. The cost is a small drop in ranking metrics for some models.
4.20
Papers on detecting fake shilling accounts in recommender systems usually report a single AUROC. This paper finds that detectors built on model embeddings fail completely under a more realistic attack: the attacker copies a real user's entire item history and slips in only one to three items to promote. On MovieLens-1M, every embedding-based detection score drops to chance. The simplest method works instead: use Jaccard similarity to measure how close each account's item set is to its most similar other account, and AUROC reaches above 0.92. The cause lies in the representation itself. LightGCN's embeddings place a fake account close to the real user it copied, which erases the signal that separates them. So any learned detector should be paired with a similarity check that looks directly at the raw records.
3.95
A user taps an offsite ad: should it open the advertiser's website, or jump straight into the advertiser's app already installed on the phone? Switching the default from web to app for everyone actually lowers overall advertiser value (value-weighted conversions) by 1.05%, with a small gain only on ads optimizing for purchases. In other words, which destination is better depends on each click. This paper from Meta picks the destination for each click separately, and the model went through three generations. The last one directly predicts which of the two destinations is better, with training labels taken from the difference between a larger model's conversion scores for the two destinations. In online A/B tests, this generation raises advertiser value by 0.54% over always opening the website. Where a click sends the user is an often overlooked lever for optimization.
3.75
A post-processing rule on a trained recommender, such as swapping a few items or reordering, may improve the average result but hurt some users. The usual safeguard is conformal risk control: use held-out data to set a threshold that keeps the expected loss within a budget. But it is often too conservative and blocks useful changes too. This paper finds the problem is wasted information during calibration. One existing method only records whether each user got worse, so a slight drop counts the same as a severe one. The paper's calibration method (ACRC) charges the budget by the actual size of each loss, so a slight drop uses only a little budget, and more changes get through under the same risk guarantee. On 6 datasets, changing only this calibration step gives higher net utility than the existing calibration methods.
3.75
Risk screening and recommendation systems often decide whether to pass or block an account, and an account brings both value and loss. If the value lost by blocking has a cap, the best ranking orders accounts by how much value each unit of expected loss carries, blocking the lowest ratio first until the value budget is used up. Another common criterion blocks by expected loss (or loss minus value), from largest to smallest. Under the same budget, the ratio rule may block an account this criterion would pass, and the paper counts this as a contestable block. When training the ranking score, the paper adds a penalty for accounts that would trigger such a block, and the heavier the penalty, the more the account is pushed toward the pass side. Using different penalty multipliers for different groups traces a trade-off curve between how much loss is caught and the gap in complaints between groups. On public recommendation, insurance and credit data, group-specific penalties narrowed the gap in some groups, but on average they were not significantly better than group-blind penalties.
3.65
Sequential recommenders often use LLM item embeddings to help long-tail items, but these embeddings only capture whether item contents are similar, not what users tend to view right after what. Long-tail items have too little behavior data to correct their embeddings on their own. PACERec first computes a behavioral target for each item: it takes the items users viewed just before choosing it and computes a weighted average of their content embeddings, with more recent items weighted higher. The model then learns one correction pattern shared by all items: given where a vector is now, it computes which direction to move next and by how much, moving the content embedding step by step toward the behavioral target. This pattern is learned mainly from popular items, whose behavioral targets are reliable. A long-tail item only needs its content embedding to use it directly. On three datasets and three sequential models, tail-item NDCG@10 is on average 33.31% higher than the strongest baseline in each setting.
3.55
When an LLM agent tries to improve recommendation algorithms automatically, the hard part is that one experiment's result rarely tells it what to change next. RecRSI attaches a mechanism hypothesis to every change: which component to modify, which existing design it draws on, why it should work, and what result to expect. After the experiment, the result and its design context are passed on to later research and used to decide which mechanisms to develop, revise or combine next. Across several initial recommenders and datasets, it raises NDCG@5 by 8.48% on average relative to the initial algorithms, and by 2.60% over the strongest competing method in each setting. This shows that automated algorithm work keeps improving only when each experiment leaves a reusable conclusion.
Meta