Cold start is the product, not a temporary scaffold

Every adaptive system has a moment when it has nothing to go on. The account is new and the model meant to tailor the experience has no evidence. The trade calls this the cold start and treats it as latency. This unit argues the opposite. The cold start is the one moment when a system cannot claim to know you. What it does then is not a scaffold on the way to the product. It is the product itself, because an empty model still shows something, and that something is a decision about a population.

A blank profile forces a population assumption. The fourth of ten units, this one asks what a system does with no evidence about the person, and whether asking beats imputing. The fifth unit asks what an offline score can show about the benefit anyone receives.

The blank profile is the system's first honest statement

Schein and colleagues (2002) named the problem in the paper that made cold start a research object. Their method combines content and collaborative data in one probabilistic framework. They benchmarked it against a naive Bayes classifier on items nobody in the community has rated. They add the CROC curve and three testing methodologies for a setting with almost no evidence. A system with no evidence still answers, so its answer is a design choice rather than a data accident. It is the default that personalisation was meant to remove, and the only part of the interface that every user sees.

The field's instinct is to impute a population prior

Lam and colleagues (2008) build a hybrid model from two probabilistic aspect models combined with user information. On MovieLens they report substantial and consistent improvement over pure collaborative filtering for new users. They note that researchers have studied the item side far more than the user side.

Lika, Kolomvatsos and Hadjiefthymiades (2014) show the imputing instinct in concrete form. Their cold-start model keeps collaborative filtering and adds classification, using demographic data to place a new user with others who behave similarly, then predicting from that group. They tested it on the GroupLens ratings data across several scenarios and reported satisfactory results. The remedy is an estimate of the person assembled from strangers who match on demographics, which is the population assumption the unit questions. And the study measures it offline against a benchmark, so it reports fit to the log rather than the value of anyone's first session.

Imputation reaches the raw signal and the catalogue

Hu, Koren and Volinsky (2008) take imputation down to the raw signal. They characterise implicit feedback, where the system has no direct statement of preference. They treat the data as positive and negative preference with varying confidence. Because absence of an event counts as weak evidence of disinterest, the model reads silence as a signal.

Cremonesi, Tripodi and Turrin (2011) look elsewhere. In cross-domain recommendation a class of neighbourhood algorithms follows paths connecting users or items across domains, tested on scenarios built by partitioning Netflix. The cold start shrinks, because the method imports a profile built elsewhere. That profile is still somebody's judgement of the person, assembled for another purpose and never shown to them.

Saveski and Mantrach (2014) make the same move on the catalogue. They learn a local model for each new item alongside a global model, so the item's own properties stand in for the interactions it lacks.

Asking beats assuming, and the questions are the design

Against that current runs a smaller literature that asks. Rashid and colleagues (2002) compared six techniques for learning a new user's preferences. Each technique selected a sequence of items to rate, from information-theoretic selection to aggregate statistics. The comparison covered offline experiments and a live experiment with over 300 users. The choice of technique significantly affected both the effort the user spent and the accuracy of the predictions that followed, which makes elicitation a design variable.

Christakopoulou, Radlinski and Hofmann (2016) turn the idea into an interaction. They frame restaurant recommendation as a cold-start problem in online learning, and build a framework that decides which questions to ask. They report a 25 per cent improvement over a static model after only two questions, with bandit exploration helping. It is the clearest figure behind the claim that a few questions substitute for imputed history, though at prototype scale.

The default the person meets is a statement about a majority

The first screen of an adaptive system states which majority the design assumes. A default drawn from aggregate behaviour, or an implicit model in which silence reads as disinterest (Hu et al., 2008), describes the median user. A person who is not that median, whether through age or because the household shares one phone, spends their first sessions correcting an error baked in before they arrived. The elicitation evidence says a few questions avoid the correction cheaply (Rashid et al., 2002; Christakopoulou et al., 2016), which makes the default a decision, not a necessity.

Nguyen and colleagues (2014) measured the filter bubble at the level of the individual. They used a diversity metric built from user-generated tags to track how a collaborative filtering recommender changed what a person consumed. They found that recommenders expose users to a slightly narrowing set of items over time. Users who consumed the recommended items saw lessened narrowing, however, and rated them more positively. The mechanism is real and not simply self-reinforcing, which is why the claim stops at exposure and does not run on to belief.

Jannach, Manzoor, Cai and Chen (2021) note that recommender research long assumed a one-shot interaction, in which preferences are estimated from past behaviour. They find that researchers evaluate the newer conversational systems inconsistently, and rarely on the conversation itself.

The contested question: does asking beat imputing, and does either escape the default?

Two positions exist, and a third refuses both sets of evidence. The first holds that the cold start is an engineering problem whose answer is to impute. Lam and colleagues (2008) report substantial and consistent improvement on the user side. Cremonesi and colleagues (2011) shrink the cold start by importing a profile from another domain. Saveski and Mantrach (2014) reconstruct the item side from an item's own properties. On this reading the population prior does real work and elicitation is an optional refinement.

The second holds that the imputed prior is the thing under dispute and that a few questions do better. Rashid and colleagues (2002) found the explaining variable to be the technique for choosing items to rate, in a live study of over 300 users. Christakopoulou and colleagues (2016) report a 25 per cent improvement over a static model after two questions. On this reading the first screen is a design surface, and a population average wastes the one moment when the person will state a preference.

The third questions the evidence on both sides. Jannach and colleagues (2021) find conversational recommenders evaluated inconsistently and rarely on the conversation itself, so researchers assert the benefit of asking more often than they measure it. No study in this list holds the recommender fixed while varying only the cold-start policy in a field setting. The headline figure comes from one framework on selected data (Christakopoulou et al., 2016).

What would settle it is a randomised comparison on a real population. Hold the engine fixed and assign new users at random to an imputed prior, a population default, or a short elicitation. Then measure the first session's value, the effort spent, and the diversity of what they consumed over the following weeks, using the individual-level measure Nguyen and colleagues (2014) built. Until then, asking is a viable policy with direct support at small scale. Choosing to impute is a decision about whose preferences the system assumes, and the team takes it by default rather than by argument.

The two positions rest on different kinds of evidence. The imputing position rests on offline benchmarking against a log. It rests on MovieLens for new users (Lam et al., 2008) and on the item side (Saveski and Mantrach, 2014). The asking position rests on a live experiment with over three hundred users (Rashid et al., 2002). It rests on one prototype that reported a gain after two questions (Christakopoulou, Radlinski and Hofmann, 2016). The live study is the stronger evidence. It measures the effort and the prediction rather than the value of a first session. While the dispute is open, ask a few questions on the first screen. The strongest objection is that the unit's confidence in asking outruns its evidence.

Where this is heading

The substitute for a missing profile is becoming a document rather than an average. Di Palma (2023) proposes a retrieval-augmented recommender in which a large language model grounds suggestions in retrieved item descriptions. It targets the sparse conditions where classical remedies struggle, and the paper is candid that it is a proposal rather than a controlled comparison. Saveski and Mantrach (2014) had already rebuilt the item side from properties learned alongside interactions. The open question is the one the thesis raises. If the first recommendation now comes from a description of an item, the system has changed the source of its majority assumption. It has not given the person more chance to see it.

Sources

• Schein, A. I., Popescul, A., Ungar, L. H. and Pennock, D. M. (2002). Methods and metrics for cold-start recommendations. Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval. doi:10.1145/564376.564421 Combined content and collaborative data in one probabilistic framework and benchmarked it against a naive Bayes classifier on the cold-start problem, adding the CROC curve and three testing methodologies for a setting with almost no evidence. It is the point at which the blank profile became a research object with its own metrics.
• Lam, X. N., Vu, T., Le, T. D. and Duong, A. D. (2008). Addressing cold-start problem in recommendation systems. Proceedings of the 2nd international conference on Ubiquitous information management and communication. doi:10.1145/1352793.1352837 Built a hybrid model from two probabilistic aspect models combined with user information, and reported substantial and consistent improvement over pure collaborative filtering on MovieLens for new users, while noting that the item side has been studied far more than the user side. The remedy is imputation: a population-level estimate placed where a preference should be.
• Lika, B., Kolomvatsos, K. and Hadjiefthymiades, S. (2014). Facing the cold start problem in recommender systems. Expert Systems with Applications. doi:10.1016/j.eswa.2013.09.005 Built a cold-start model that keeps collaborative filtering and adds classification, grouping a new user with demographically similar others, and tested it on the GroupLens ratings data. The remedy is a population-level estimate measured offline.
• Rashid, A. M., Albert, I., Cosley, D., Lam, S. K., McNee, S. M., Konstan, J. A. et al. (2002). Getting to know you: learning new user preferences in recommender systems. Proceedings of the 7th international conference on Intelligent user interfaces. doi:10.1145/502716.502737 Compared six techniques that select a sequence of items for a new user to rate, from information-theoretic selection to aggregate statistics, in offline experiments and a live experiment with over 300 users. The choice of technique significantly affected both the effort the user spent and the accuracy of the resulting predictions.
• Christakopoulou, K., Radlinski, F. and Hofmann, K. (2016). Towards Conversational Recommender Systems. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. doi:10.1145/2939672.2939746 Framed restaurant recommendation as a cold-start problem in online learning and built a framework that selects which questions to ask, comparing feedback types and question strategies on synthetic and real data. It reports a 25 per cent improvement over a static model after asking only two questions.
• Jannach, D., Manzoor, A., Cai, W. and Chen, L. (2021). A Survey on Conversational Recommender Systems. ACM Computing Surveys. doi:10.1145/3453154 Surveys conversational recommenders, notes that research has long assumed a one-shot interaction in which preferences are estimated from past behaviour and a ranked list is pushed out, and categorises the richer interaction patterns that aim to improve elicitation. It reviews how such systems are evaluated and identifies the gaps.
• Cremonesi, P., Tripodi, A. and Turrin, R. (2011). Cross-Domain Recommender Systems. 2011 IEEE 11th International Conference on Data Mining Workshops. doi:10.1109/ICDMW.2011.57 Formalised cross-domain recommendation and defined a class of neighbourhood algorithms that follow paths connecting users or items across domains, tested on three cross-domain scenarios built by partitioning the Netflix dataset. The approach shrinks the cold start by importing a profile built elsewhere.
• Hu, Y., Koren, Y. and Volinsky, C. (2008). Collaborative Filtering for Implicit Feedback Datasets. 2008 Eighth IEEE International Conference on Data Mining. doi:10.1109/ICDM.2008.22 Characterised implicit feedback, where the system has no direct statement of preference, and treated the data as positive and negative preference with vastly varying confidence, with an optimisation routine that scales linearly with data size and was used in a television recommender. Because absence of an event counts as weak evidence of disinterest, the model reads silence as a signal.
• Nguyen, T. T., Hui, P. M., Harper, F. M., Terveen, L. and Konstan, J. A. (2014). Exploring the filter bubble. Proceedings of the 23rd international conference on World wide web. doi:10.1145/2566486.2568012 Measured the filter bubble at the level of the individual with a diversity metric built from user-generated tags, tracking the longitudinal effect of a collaborative filtering recommender. It found a slightly narrowing set of items over time, and that users who consumed the recommended items experienced lessened narrowing and rated them more positively.
• Saveski, M. and Mantrach, A. (2014). Item cold-start recommendations: learning local and global models. Proceedings of the 8th ACM Conference on Recommender systems. doi:10.1145/2645710.2645751 Learns a local model for each new item alongside a global model, exploiting item properties and past user preferences, and reports that the method outperformed six state-of-the-art item cold-start methods on news recommendation and email recipient recommendation. The item's own properties stand in for the interactions it lacks.
• Di Palma, D. (2023). Retrieval-augmented Recommender System: Enhancing Recommender Systems with Large Language Models. Proceedings of the 17th ACM Conference on Recommender Systems. doi:10.1145/3604915.3608889 Proposes a retrieval-augmented recommender in which a large language model grounds suggestions in retrieved item descriptions, aimed at the sparse and cold conditions where classical remedies struggle. The paper is a short proposal rather than a controlled comparison, and says so.