How to Overcome Recommendation Cold Start: Methods and Huajiao Live’s Real‑World Practices

Cold‑Start Problem Overview

In recommendation systems a cold‑start situation occurs when a user or an item has insufficient interaction history, making personalized ranking unreliable. Three sub‑categories are commonly distinguished:

System cold start : the entire platform is newly launched and no aggregate data exist. Designers must rely on product positioning, seed‑user acquisition, and early‑stage heuristics.

Item cold start : newly added items (e.g., a fresh e‑commerce product or a newly‑joined streamer) have no views, clicks or ratings.

User cold start : a newly registered user has no past behavior; the initial experience strongly influences retention.

Common User Cold‑Start Approaches

Popular‑Content Recommendation Rule‑based selection of currently trending items (e.g., based on recent click‑through rate or retention rank). Simple to deploy but requires frequent updates to avoid staleness.

Group‑Representative Recommendation Present a short list of interest tags or representative items for the user to choose. The chosen tag/item seeds the user’s preference vector, allowing downstream matrix‑factorization or similarity‑based models to generate personalized results. This method is highly interpretable but adds a front‑end interaction step.

Auxiliary‑Information Utilization If the user logs in via a third‑party platform, existing profiles (social‑graph, interests, demographics) can be imported to enrich the cold‑start vector. Privacy compliance and cross‑platform data pipelines increase implementation effort.

Bandit Algorithms (Exploration + Exploitation) The multi‑armed bandit (MAB) framework treats each recommendation option as an arm with an unknown reward distribution. Algorithms continuously balance exploring less‑tried arms and exploiting arms with high estimated reward.

Thompson Sampling

Thompson sampling maintains a Beta( win , lose ) posterior for each arm’s success probability. At each decision step:

for each arm a:
    sample p_a ~ Beta(win_a, lose_a)
select arm a* = argmax_a p_a
observe reward r (1 = click/view, 0 = no click)
if r == 1:
    win_a* += 1
else:
    lose_a* += 1

This stochastic selection naturally balances exploration and exploitation and is well‑suited for user cold‑start where prior knowledge is scarce.

Deep Learning Models Neural networks can automatically learn high‑order feature interactions and generalize from sparse data. Typical architectures include:

Wide&Deep : a linear “wide” component for manual feature crossing (e.g., age × device) combined with a deep feed‑forward network that captures non‑linear patterns.

DeepFM : replaces the wide part with a factorization‑machine (FM) layer that learns pairwise feature embeddings, enabling automatic crossing even for feature pairs unseen in training.

Sequential Models : RNN, LSTM, GRU or the Deep Interest Network (DIN) use attention mechanisms to encode a user’s short‑term behavior sequence (e.g., recent watch or click events) into a dynamic interest vector.

Huajiao Live Practical Solutions for User Cold‑Start

Popular Recommendation Generate a hot‑streamer list using metrics such as new‑user retention rank or overall popularity, then serve this list to newcomers.

Popular + Real‑Time Behavior Feedback When a new user has any recent watch history, retrieve similar streamers via collaborative filtering and rank them together with the hot list. If no history exists, fall back to the hot list.

Bandit‑Based Recommendation Implementation steps:

Train streamer embeddings using matrix factorization or item2vec.

Cluster embeddings with K‑means; each cluster becomes a bandit arm.

Initialize a Beta posterior for each arm.

During serving, sample arm scores with Thompson sampling, retrieve streamers from the selected cluster, and compute a final score = bandit score × streamer intrinsic score.

After each user interaction (click/view), update the corresponding arm’s Beta parameters.

This approach diversifies recommendations while gradually adapting to individual preferences.

Deep Learning Pipeline Features used:

Basic profile: gender, age.

Contextual attributes: region, device model, acquisition channel, time of day.

Short‑term watch sequence (e.g., last 5 streamed videos).

Model architecture:

Wide&Deep component for manual cross features (e.g., age × region).

DeepFM layer for automatic pairwise crossing.

DIN attention layer to aggregate the watch sequence into a real‑time interest vector.

The combined output yields a probability score for each candidate streamer.

Auxiliary‑Information Integration When available, external platform attributes (e.g., social‑network friends, interests) are merged into the user vector to further alleviate sparsity.

Observed Impact

Adding real‑time behavior feedback to the popular‑streamer baseline increased average watch time per new user by >50 %. Incorporating the Wide&Deep + DeepFM + DIN model contributed an additional ~20 % lift on the same metric.

Conclusion

Cold‑start mitigation is essential for sustaining growth in live‑streaming and e‑commerce platforms. Practitioners should start with low‑cost heuristics (popular items, group‑representative tags) and progressively adopt more sophisticated techniques—bandit algorithms for exploration and deep neural models for high‑dimensional feature learning—once a stable data pipeline is in place.